The Intelligence Factory Open the visualizer

Evidence

Where the numbers come from

Every figure shown with a label says what kind of statement it is. Spec: the maker or a standards body publishes it. Vendor: a vendor’s own comparison, attributed to it, with its baseline. Reported: a named third party states it. Calc.: this site’s model calculates it. Assumed: the model picks a value where no single published figure applies. Below is each labeled figure with what backs it, and then every source, with when it was published and when it was checked.

Figures below are fixed to one reference scenario, not whichever campus you set on the visualizer: 100 MW at the meter, GB200 NVL72 racks, 415 V AC to the rack, warm-water cooling. Change the campus on the main page and its own numbers follow; the claims and sources on this page always describe this one scenario.

Claims
911
Spec
239
Vendor
32
Reported
245
Calc.
172
Assumed
223
Sources
395

1Scale across2,000 km across

power

Each campus, its own grid

See it in 3D ↗
Interconnection
345–500 kV
Assumed

Assumed: Typical campus interconnection voltage, 345–500 kV. Why

Example
Amazon Project Rainier: existing Olive 345 kV station, new interconnection
Reported

Sources: Measured AI (Substack) body text: "the campus already adjoined extra-high-voltage (EHV) transmission and a very substantial existing substation — the Olive 345 kV station"

Map symbols
substations, towers and supply lines are representative
Assumed

Assumed: Map symbols on the regional map, representative, heights and sizes exaggerated. Why

power

Generation

See it in 3D ↗
US grid losses
≈5% (EIA)
Reported

Sources: U.S. Energy Information Administration FAQ answer: "annual electricity transmission and distribution (T&D) losses averaged about 5% of the electricity transmitted and distributed in the United States in 2018 through 2022"

Plant symbols
representative types, heights exaggerated
Assumed

Assumed: Map symbols on the regional map, representative, heights and sizes exaggerated. Why

power

This campus

See it in 3D ↗
Meter
100 MW
Assumed

Assumed: Modeled campus size, set by the reader. Why

IT load
88 MW at PUE 1.14
Calc.

Calculated: IT load = meter power ÷ this design’s modeled power usage effectiveness (PUE); the gap is cooling and electrical overhead. Method

power

Grid carbon by state

See it in 3D ↗
Ohio, 2024
1,005 lb/MWh, 456 g/kWh
Reported

Sources: U.S. Energy Information Administration Table 1. 2024 Summary statistics (Ohio), row "Carbon Dioxide (lbs/MWh)"

US average, eGRID 2022
373 g/kWh
Reported

Sources: U.S. Environmental Protection Agency Table 3, State Output Emission Rates (eGRID2022), "U.S." row, CO2 lb/MWh (823.1, matching Table 1’s subregion total)

Lowest shown, Washington
113 g/kWh
Reported

Sources: U.S. Energy Information Administration Table 1. 2024 Summary statistics (Washington), row "Carbon Dioxide (lbs/MWh)": 249

Highest shown, Wisconsin
494 g/kWh
Reported

Sources: U.S. Energy Information Administration Table 1. 2024 Summary statistics (Wisconsin), row "Carbon Dioxide (lbs/MWh)"

power

Stargate Abilene

See it in 3D ↗
Status, 09/24/2026
Partly live · ≈421 MW IT
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" figure

IT power operating, buildings 1–4 of 8
≈421 MW
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" 421MW

Planned
≈1.2 GW (Epoch, 04/2026 estimate)
Reported

Sources: Epoch AI site profile: "Projected capacity: 1.2 GW | 1.0 million H100-equivalents"

Grid
double 345 kV corridor (Midland–Graham), Mulberry Creek substation
Reported

Sources: Yes Energy body text: "The facility is fed by a double 345 kV corridor that goes from Midland to Graham, Texas," and "a substation next to the facility called Abilene Mulberry Creek"

On-site generation
natural gas reported by Epoch (04/2026); not itemized in the current tracker
Reported

Sources: Epoch AI body text: "Power is currently supplied by a mix of on-site natural gas and grid power, which includes local wind power"

power

SpaceXAI Colossus

See it in 3D ↗
Colossus 1: status, 09/23/2026
Operating · 150 MW from the grid, plus turbines
Reported

Sources: Compute Atlas facility profile: operating-status line and grid-power figure

Colossus 2: status, 09/24/2026
Partly live · ≈946 MW IT
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" figure

Colossus 2: ranking, epoch ai
Most powerful operating today
Reported

Sources: Epoch AI ranking page lead text: "Colossus 2 is the largest tracked AI data center at about 946 MW of current IT power"

GPUs, per Elon Musk, 09/25/2026
230k: 150k H100, 50k H200, 30k GB200
Reported

Sources: Elon Musk, on X post on X, 09/25/2026: "Colossus 1 is 150k H100, 50k H200 and 30k GB200. Colossus 2 is 110k GB200 and 440k GB300."

GPUs, reported earlier
≈200,000 (H100, H200, some GB200)
Reported

Sources: Tom's Hardware headline: "fully operational with 200,000 GPUs"

Phase 2 power
≈300 MW
Reported

Sources: Tom's Hardware headline: "Phase 2 to consume 300 MW"

Grid supply
MLGW/TVA, ≈150 MW
Reported

Sources: Interesting Engineering body text: "receiving 150MW of grid power from MLGW and TVA"

On-site generation
≈35 gas turbines, ≈422 MW
Reported

Sources: Compute Atlas facility profile: on-site generation "422 MW" via "approximately 35 natural-gas turbines"

Batteries
Tesla Megapacks, ≈150 MW
Reported

Sources: Interesting Engineering body text: "150MW of Megapack Batteries for stored energy backup have been integrated"

Backup, per SpaceXAI
more than 240 batteries, enough for the site to come fully off the grid
Spec

Sources: SpaceXAI (SpaceX) Power tab, "Installing resilient, sustainable backup power": "more than 240 batteries so Colossus I can come completely offline during emergencies or peak demand"

Cooling water, per SpaceXAI
≈820,000 gal a day, hybrid system
Spec

Sources: SpaceXAI (SpaceX) Water tab, "What Colossus uses today": "Colossus I has a hybrid system that uses about 820,000 gallons a day"

power

Fairwater Atlanta

See it in 3D ↗
Status, 09/24/2026
Partly live · ≈636 MW IT
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" figure

IT power running, satellite estimate
≈636 MW, 4 of 9 Main Campus buildings
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" 636MW; body text on buildings 1–4 liveMeasured AI (Substack) body text: "The QTS Fayetteville Main Campus comprises nine data center buildings — four operational, five under construction"

Planned
≈1.5 GW across all 13 buildings (Fairwater + Main + East Campus)
Reported

Sources: Measured AI (Substack) body text: the full QTS Fayetteville campus "designed to eventually reach approximately 1,500 MW across all 13 buildings"

Accelerators
NVIDIA Blackwell NVL72 (GB200/GB300 family), ≈140 kW per rack
Spec

Sources: Microsoft body text: "~140kW per rack, 1,360 kW per row"Measured AI (Substack) body text: "Each NVL72 rack contains 72 NVIDIA Blackwell GPUs... roughly 140 kW of rack power"

Cooling
closed-loop liquid, no evaporation
Spec

Sources: Microsoft body text: a "closed-loop approach that reuses the liquid continuously after the initial fill with no evaporation"Measured AI (Substack) body text: "The cooling loop is filled once and recirculated for the life of the building"

Backup power
no on-site generation, UPS or dual-corded distribution (grid-only design)
Spec

Sources: Microsoft body text: Microsoft can "forgo traditional resiliency approaches for the GPU fleet (such as on-site generation, UPS systems and dual-corded distribution)"Measured AI (Substack) body text: "The Fairwater template strips on-site power generation, UPS systems, and dual-corded distribution out of the buildings’ electrical topology entirely"

power

Fairwater Wisconsin

See it in 3D ↗
Status, 09/24/2026
Partly live · ≈369 MW IT
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" figure

Power
Building 1 ≈369 MW live; $3.3B initial investment
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" 369MWMicrosoft (On the Issues) body text: "$3.3 billion investment pledge"

Accelerators
GB200 NVL72 racks (NVLink 5.0 inside), racks linked by 800G Ethernet
Reported

Sources: TechTimes body text: "NVIDIA’s GB200 NVL72 systems connect 72 Blackwell GPUs through NVLink 5.0"; separately, "connecting the racks is a two-level networking tree built on 800G Ethernet"

Cooling
closed-loop liquid, filled once
Spec

Sources: Microsoft (On the Issues) body text: "a state-of-the-art closed-loop liquid cooling system, filled during construction and recirculated continuously"

Opened
06/23/2026
Reported

Sources: Microsoft Source headline/dateline: construction completed, announced Tuesday, June 23, 2026

power

Meta Hyperion

See it in 3D ↗
Status, 09/27/2026
Under construction · nothing live yet
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" figure

Phase 1
≈2 GW by 2030
Reported

Sources: IEEE Spectrum body text: "the first phase—a 2-GW version—will be completed by 2030"

Planned
5 GW at full build (~2032), more than $50B
Spec

Sources: Meta body text: "delivering 5 gigawatts of compute capacity to house Hyperion" and "$50B+ Data center investment"Wikipedia body text: "full build-out around 2032"

GPUs, planned
more than 3 million at full build (reporter estimate)
Reported

Sources: IEEE Spectrum body text: "Hyperion could include over 41,000 rack-scale systems, for a total of more than 3 million GPUs"

New gas generation
2.26 GW combined, three turbine plants
Reported

Sources: IEEE Spectrum body text: Entergy’s "three new gas-turbine power plants... will generate a combined 2.26 GW"Wikipedia body text: "Entergy has begun building three gas-powered combined-cycle power plants"

power

Project Rainier

See it in 3D ↗
Status, 09/24/2026
Partly live · ≈910 MW IT
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" figure

IT power operating, ~16 of ~30 buildings live
≈910 MW
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" 910MW

Planned full campus, due ~Q1 2028
≈1,925 MW (1.9 GW)
Reported

Sources: Epoch AI directory entry: "PROJECTED IT POWER" 1,925MW

Chips, Oct. 2025 launch vs. 09/2026
≈500,000 Trainium2 at launch (64-chip UltraServers); ≈1,045,000 by Epoch's 09/24/2026 count
Reported

Sources: Compute Atlas facility profile: "Phase 1 online with ~500,000 Trainium2 chips," operational October 2025Epoch AI directory entry: current Trainium2 count "1,045k"

Grid interconnection
Indiana Michigan Power (AEP), existing Olive 345 kV station
Reported

Sources: Measured AI (Substack) body text: "the campus already adjoined... the Olive 345 kV station" on "a purpose-built 345 kV transmission delivery scheme"Compute Atlas facility profile: utility "Indiana Michigan Power (AEP)"

On-site generation
none for operating power; diesel gensets for backup only
Reported

Sources: Measured AI (Substack) body text: "All operating power comes from the utility grid... The only on-site generation is a diesel fleet held in reserve for emergency backup"

power

Meta Prometheus

See it in 3D ↗
Status, 09/24/2026
Partly live · ≈471 MW IT
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" figure

IT power operating, Epoch satellite estimate
≈471 MW
Reported

Sources: Epoch AI directory entry: "471 MW of IT power," as of 08/31/2026

Compute, satellite estimate
≈536k H100-eq compute; ≈166k Nvidia B300 chips
Reported

Sources: Epoch AI directory entry: current chips "166k" B300 delivering "536k H100-eq" (08/31/2026)

Planned, by Q3 2028
≈1,022 MW IT power, ≈$38.7B total capital
Reported

Sources: Epoch AI directory entry: "PROJECTED IT POWER" 1,022MW by Q3 2028, with the tracker’s total-capital estimate

On-site generation
Socrates North + South gas plants, ≈400 MW, fully off-grid
Reported

Sources: MLQ.ai body text: "400MW of dedicated, off-grid natural gas generation," split between "the 200MW Socrates South plant" and a "twin 200MW Socrates North facility"

Construction
5 tent structures (~125,000 sq ft each) built Apr–Jun 2026, plus traditional colocation buildings
Reported

Sources: MLQ.ai body text: "Five ~125,000-sq-ft tent structures at its New Albany, Ohio campus between April and June 2026," built in about 3 months versus 18–36 months for traditional halls

data

Other campuses

See it in 3D ↗
Microsoft AI WAN fiber added
120,000 miles
Spec

Sources: Microsoft body text: "The company has deployed 120,000 miles of dedicated fiber for the network"

NVIDIA Spectrum-XGS
nearly 2× NCCL across sites
Vendor

Sources: NVIDIA announcement text: "Spectrum-XGS Ethernet nearly doubles the performance of the NVIDIA Collective Communications Library"

Compared with: NVIDIA’s own NCCL (Collective Communications Library) baseline, in geographically distributed multi-site clusters

DeepMind Decoupled DiLoCo
4 US regions over 2–5 Gb/s
Reported

Sources: Google DeepMind blog post: "We successfully trained a 12 billion parameter model across four separate U.S. regions using 2-5 Gbps of wide-area networking"

data

Fiber route

See it in 3D ↗
Speed in fiber
≈4.9 µs per km, one way
Calc.

Calculated: Speed = speed of light ÷ silica fiber’s group refractive index (≈1.47) ≈ 2.04×10⁸ m/s, which is ≈4.9 µs per kilometer, one way. Method

1,000 km, propagation only
≈5 ms one way, ≈10 ms round trip
Calc.

Calculated: One-way delay = 1,000 km × the fiber-speed figure above (≈4.9 µs/km) ≈ 5 ms; round trip doubles it to ≈10 ms. Switching, routing and queueing are not included. Method

Microsoft hollow-core fiber
≈33% lower latency
Reported

Sources: Fiber Broadband Association body text: hollow-core fiber "cutting latency by 33%" versus solid-glass fiberNetwork World article text: Microsoft’s pilot "involved over 1,200 km of fibre, now installed underground and actively carrying live traffic"

Per fiber pair, C-band 800G
32 × 800G = 25.6 Tb/s
Reported

Sources: Converge Digest article text: "the enhanced POLS can support 32 DWDM wavelengths over a fiber pair," an "aggregate capacity of 25.6 Tbps"

data

Amplifier huts

See it in 3D ↗
Spacing
≈80–100 km, rule of thumb
Assumed

Assumed: Optical amplifier hut spacing, ≈80–100 km. Why

data

Line terminals

See it in 3D ↗
Per wavelength, WaveLogic 6
up to 1.6 Tb/s
Spec

Sources: Ciena product announcement: WaveLogic 6 family, up to 1.6 Tb/s per wavelength

800G pluggable, e.g. Marvell COLORZ 800
800 Gb/s to ≈500 km
Spec

Sources: Marvell, via PR Newswire press release: "up to 800 Gbps of bandwidth for DCI links up to 500km"

Field trial
1.6 Tb/s over 1,100 km (Telstra)
Reported

Sources: Converge Digest lead sentence: "Telstra has transmitted four 400GbE client services over a single 1.6 Tbps wavelength across approximately 1,100 km between Melbourne and Sydney"

C+L band
up to 2× capacity per fiber
Reported

Sources: Converge Digest body text: expanding from C-band into L-band "effectively opens a second optical transmission band, increasing the usable spectrum"

Inside a line terminal, e.g. Ciena 6500 RLS
mux/demux, ROADM and amplifier modules in compact shelves
Spec

Sources: Ciena module list: "ROADM with Line Amplifier (RLA) 12x1 C-band Module ... integrating twin 1x12 flexible grid WSS, EDFAs"; "Dual Line Amplifier (DLA) C-band module ... integrating bi-directional C-band EDFA line amplifiers"; "Raman Amplifier (SRA) C-band Module"; "Colorless Channel Mux/Demux (CCMD)"

800G ZR+ reach, Juniper example
800G to 450 km · 600G to 900 km · 400G to 1,800 km
Spec

Sources: Juniper Networks parameters table, "Distance (amplified)": 450km (800G ZR+), 900km (600G 8QAM), 1800km (400G QPSK), all at 118.2GBd; 120km for OIF 800ZR

Slot per 800G wavelength
≈150 GHz at ≈118 GBd, so 32 fill the C-band
Reported

Sources: MapYourTech section 2: "An 800ZR carrier at 118.2 GBd and minus 7 dBm sits in a 150 GHz slot"; Table 1, channel spacing 150 GHzJuniper Networks parameters table: signaling rate "(118.2GBd)" for the 800G modes

data

This campus

See it in 3D ↗
GPUs
≈45,288
Calc.

Calculated: GPU count = this campus’s racks × the GPUs per rack for the chosen accelerator, where rack count follows from the chosen meter power and rack power. Method

heat

Climate picks sites

See it in 3D ↗
Free cooling
most hours in cool climates
Assumed

Assumed: Free cooling depends on climate, more hours in cool, dry climates. Why

Heat drawn
log scale, each 10× in power 5× the motion; none under 0.5 W
Assumed

Assumed: How much heat each part is drawn with, log scale: each 10× in power is 5× the motion; nothing under 0.5 W. Why

heat

This campus

See it in 3D ↗
Heat out
100 MW
Calc.

Calculated: By energy conservation, essentially all electrical power a campus draws is eventually rejected as heat, so the heat-out figure equals the meter figure. Method

2Grid & campus1.6 km across

power

Transmission line

See it in 3D ↗
Voltage
345 kV line-to-line
Assumed

Assumed: Campus interconnection voltage, 345 kV. Why

Current, 100 MW
≈90 A per phase, on each circuit
Calc.

Calculated: Per-phase current on each of two equally loaded circuits = campus MW ÷ (2 × √3 × 345 kV × power factor). Method

US grid losses, 2018–2022
≈5% (EIA)
Reported

Sources: U.S. Energy Information Administration FAQ answer: "annual electricity transmission and distribution (T&D) losses averaged about 5% of the electricity transmitted and distributed in the United States in 2018 through 2022"

Example
Stargate Abilene: double 345 kV corridor
Reported

Sources: Yes Energy blog post: "The facility is fed by a double 345 kV corridor that goes from Midland to Graham, Texas."

power

Campus substation

See it in 3D ↗
Breakers
6 dead-tank SF₆, ring bus
Assumed

Assumed: Substation breaker count and yard size, 6 dead-tank SF₆ breakers, ring bus, ≈200×150 m pad. Why

Equipment drawn
typical forms, representative
Assumed

Assumed: Substation equipment as drawn, representative. Why

Yard
≈200 × 150 m gravel pad
Assumed

Assumed: Substation breaker count and yard size, 6 dead-tank SF₆ breakers, ring bus, ≈200×150 m pad. Why

Interconnection study
1–3 years alone
Reported

Sources: ATK Energy Group blog: "System impact studies, facilities studies, and any required network upgrades can run twelve to thirty-six months depending on the region"

power

Main power transformers

See it in 3D ↗
Rating
3 × 75 MVA, N+1
Calc.

Calculated: Count = ceil(campus MW ÷ (transformer unit MVA × assumed 0.95 power factor)) + 1 spare (N+1); the model buys 300 MVA units above 400 MW and 75 MVA units below. Method

Efficiency, 345 kV class
>99.6% at all loading levels
Spec

Sources: PA Transformer product page: "High efficiency: exceeding 99.6% at all loading levels"

Loss at 100 MW
≈0.4 MW
Calc.

Calculated: Loss = campus MW × (1 − main-transformer efficiency). Method

Fittings drawn
radiator banks, fans, bushings: a typical layout
Assumed

Assumed: Main transformer fittings as drawn, representative. Why

Lead time, 2026
128–144 weeks
Reported

Sources: IndustrialSage quoting a Wood Mackenzie Q2 2025 survey: "standard power transformers average 128 weeks for delivery"; generator step-up units "average 144 weeks"

power

34.5 kV switchgear

See it in 3D ↗
Feeders
≈10, each ≈10 MW
Calc.

Calculated: Feeders = max(2, ceil(campus MW ÷ 10 MW per feeder)). Method

Voltage
34.5 kV (some campuses use 13.8 kV)
Reported

Sources: ATK Energy Group blog: "On a large campus, 34.5 kV has become the standard distribution voltage because it carries more power with fewer and smaller feeders than 13.8 kV"

Loss, cables + gear
≈0.3 MW
Calc.

Calculated: Loss = campus MW × main-transformer efficiency × (1 − cable-and-switchgear efficiency). Method

Buildings drawn
a typical prefab kit, representative
Assumed

Assumed: Prefab e-house, control house and shelter drawings, representative. Why

power

Battery energy storage

See it in 3D ↗
Size here
≈20 MW / 40 MWh
Calc.

Calculated: Power = 0.2 × campus MW; energy = 0.4 × campus MW (a 2-hour duration at that power) — a round sizing this model picks to show scale, not a vendor design. Method

Training load swings
tens to hundreds of MW, seconds
Reported

Sources: Microsoft, OpenAI and NVIDIA researchers, via arXiv Section I (Introduction): "these swings can amount to tens or hundreds of megawatts" (Microsoft, OpenAI, NVIDIA)

Example
SpaceXAI Colossus 1: ≈150 MW of Megapacks
Reported

Sources: Interesting Engineering "150 megawatts of Tesla Megapack batteries have been installed to serve as a stored energy backup"

power

Standby generator yard

See it in 3D ↗
Unit size
2.5–3 MW class
Spec

Sources: Cummins data sheet NAD-5919-EN: DQKAN rated 2500 kW standbyCaterpillar product page: C175-16 rated 2500–3100 kW standby

Units here
≈40, N+20%
Calc.

Calculated: Units = ceil(campus MW ÷ 3 MW × 1.2) for N+20% redundancy on 3 MW-class sets. Method

Fuel, 2.5 MW at full load
173 US gal/h (≈0.26 L/kWh)
Spec

Sources: Cummins data sheet fuel-consumption table: 173.1 gal/hr at full (2500 kW) load

Start to load
≈10 s
Spec

Sources: Cummins reprint of NFPA 110-2016 Table 4.1(b), "Types of EPSSs": Type 10 = "10 sec"Nixon Power Services comparison table: NFPA 110 Type 10 = 10 seconds; "data centers generally use Level 1, Type 10 systems"

power

Bulk fuel storage

See it in 3D ↗
Volume, 48 h at 100 MW
≈1.2 million L
Calc.

Calculated: Volume = campus MW × 48 h × 0.26 L/kWh (the gensets’ published full-load fuel rate), shown in million liters. Method

Tanker deliveries to refill
≈42
Calc.

Calculated: Tanker count = 48-hour fuel volume ÷ 30,000 L per truckload. Method

Tanks drawn
double-wall horizontal, representative fittings
Assumed

Assumed: Bulk fuel tanks as drawn, representative. Why

power

Data halls

See it in 3D ↗
IT load
88 MW at PUE 1.14
Calc.

Calculated: IT MW solves meter MW × upstream efficiency = IT MW ÷ IT-path efficiency + (cooling fraction + misc fraction) × IT MW ÷ side-path efficiency, chaining the model’s own conversion efficiencies; PUE = meter MW ÷ IT MW. Method

Racks
≈629
Calc.

Calculated: Racks = floor(IT MW × 1000 ÷ (one rack’s kW + its GPUs’ share of the fabric’s kW)). Method

Halls
2
Calc.

Calculated: Halls = max(1, ceil(IT MW ÷ 45 MW per hall)). Method

Floor load, one rack
≈1.4 t on 0.6 × 1.1 m
Reported

Sources: Sunbird DCIM blog: "The GB200 NVL72 weighs 1.36 metric tons, or 3,000 pounds"; rack "600mm wide by 1,068mm deep"

power

Dry coolers

See it in 3D ↗
Heat rejected
≈92 MW
Calc.

Calculated: Heat rejected ≈ IT MW × 1.05, adding back the plant’s own electrical overhead as heat. Method

Units, ≈0.8 MW each
≈121
Calc.

Calculated: Units = ceil(IT MW × 1.1 ÷ 0.8 MW per dry-cooler unit). Method

Water classes
ASHRAE W32–W45: 32–45 °C max supply
Spec

Sources: Upsite Technologies blog: classes "W17, W27, W32, NEW class W40, W45" named for their maximum supply temperature in °CASHRAE TC 9.9, hosted via Vertiv p.4, "Change to ASHRAE Water Classifications": "the W classes are being renamed with the upper temperature limits incorporated in the name... W17 (previously W1), W27 (W2), W32 (W3), W40 (new), W45 (W4)"

Water use, dry + adiabatic
≈0.15–0.17 L/kWh total, with wind/solar power
Reported

Sources: arXiv Table 5: WUE for "IT Liquid cooling: dry cooler with adiabatic assist (air-cooled chiller)" = 0.15–0.17 L/kWh, adapted from Lei et al. 2025

power

Cooling towers & tanks

See it in 3D ↗
Towers
≈6
Calc.

Calculated: Towers = warm-water design: ceil(IT MW ÷ 16 MW per tower); air/liquid design: ceil(IT MW × 1.3 ÷ 6 MW per tower). Method

WUE, on site
≈0.16 L/kWh IT
Assumed

Assumed: Water use by cooling design, air 1.0, liquid 0.5, warm water 0.16 L/kWh IT. Why

Use
peak days only
Calc.

Calculated: Follows the design: air/liquid designs run cooling towers year-round; warm-water designs default to dry coolers and only add tower water once outdoor air crosses the adiabatic threshold. Method

Cell drawn
counterflow, representative
Assumed

Assumed: Cooling tower cell as drawn, counterflow cell, representative. Why

power

Fiber entrances

See it in 3D ↗
Routes
2 or more, physically separate
Reported

Sources: TRG Datacenters "A properly designed facility has dual fiber entrances. Fiber enters from two separate locations, following different physical paths into the building."

power

Perimeter and gate

See it in 3D ↗
Microsoft’s layers
facility perimeter, building perimeter, inside the building, the hall floor
Reported

Sources: Microsoft Learn "access approval at the facility’s perimeter, at the building’s perimeter, inside the building, and on the datacenter floor"

Google’s data centers
six layers of physical security
Reported

Sources: Google "Each data center is protected with six layers of physical security designed to thwart unauthorized access."

power

Operations center

See it in 3D ↗
Power monitoring (EPMS)
readings from key distribution points
Reported

Sources: Vertiv "gathers power and energy information from key distribution points across a building"

Infrastructure management (DCIM)
assets, environment, security
Reported

Sources: Schneider Electric "Oversee and control assets, environmental conditions, and security"

data

Fiber entrances

See it in 3D ↗
Routes
2 or more, physically separate
Reported

Sources: TRG Datacenters "A properly designed facility has dual fiber entrances. Fiber enters from two separate locations, following different physical paths into the building."

data

Line terminal hut

See it in 3D ↗
Per wavelength, Ciena WaveLogic 6
up to 1.6 Tb/s
Spec

Sources: Ciena press release: "transmission at rates of up to 1.6 Tbps via a single carrier"Lightwave Online coverage of the same announcement, same figure

400ZR reach, amplified
up to ≈120 km
Reported

Sources: Smartoptics knowledge-bank post: the 400ZR project focused on "400G Ethernet with amplified point-to-point DWDM links over DCI up to 120 KM"

Module power, 400ZR / 800ZR
≈18–20 W / ≈23–25 W
Reported

Sources: AscentOptics guide: 400ZR modules "typically consume 18 to 20 W"; "one 800G link delivers the same capacity at 23 to 25 watts"

data

Inside the line terminal

See it in 3D ↗
Coherent pluggables
one wavelength each, straight from a router port
Spec

Sources: Cisco overview: "Cisco 800G-capable coherent optics can transmit a single wavelength of 800G traffic directly from a router or switch port up to 120 km for 800ZR in a Data Center Interconnect (DCI) application"

Mux/demux drawn, e.g. Ciena CMD64
2RU, up to 64 channels on a 75 GHz C-band grid
Spec

Sources: Ciena module list: "64-channel mux/demux (CMD64): 2RU 75GHz, C-band channel mux/demux for add/drop of up to 64 channels; including support for coherent pluggables"

Slot per 800G wavelength
≈150 GHz, so 32 per fiber, not 64
Reported

Sources: MapYourTech section 2: "An 800ZR carrier at 118.2 GBd and minus 7 dBm sits in a 150 GHz slot"

A 64 × 75 GHz grid, in frequency
191.3625–196.0875 THz, ≈1,529–1,567 nm
Reported

Sources: FS.com product title: "64 Channels DWDM Mux Demux, Super C-band 75GHz 191.3625-196.0875THz"; wavelength = speed of light ÷ frequency

ROADM shelf drawn, Ciena 6500 RLS R8-300
8 slots; 330 × 440 × 281 mm, 7.5U
Spec

Sources: Ciena Physical dimensions: "R8-300 8-slot shelf: 7.5U, 330 mm (H) x 440 mm (W) x 281 mm (D)"

Modules in that family
ROADM with line amplifier (WSS and EDFAs), dual line amplifier, Raman amplifier, colorless mux/demux
Spec

Sources: Ciena module list: "ROADM with Line Amplifier (RLA) 12x1 C-band Module ... integrating twin 1x12 flexible grid WSS, EDFAs"; "Dual Line Amplifier (DLA) C-band module ... integrating bi-directional C-band EDFA line amplifiers"; "Raman Amplifier (SRA) C-band Module"; "Colorless Channel Mux/Demux (CCMD)"

What a ROADM does
switches single wavelengths without converting the rest to electricity
Reported

Sources: Wikipedia lead: "adds the ability to remotely switch traffic from a wavelength-division multiplexing (WDM) system at the wavelength layer ... without the need to convert the signals on all of the WDM channels to electronic signals and back again"

Racks, router, slot layout, port counts and fiber thickness as drawn
representative
Assumed

Assumed: Line-terminal cutaway as drawn, two open racks: an R8-300-size ROADM/amplifier shelf, a 2RU 64-port mux/demux and a router holding eight coherent pluggables. Why

data

Meet-me room and border switches

See it in 3D ↗
NVIDIA’s name for it
the edge network
Spec

Sources: NVIDIA "Edge Network is the network that connects the SuperPOD to the customer’s network."

Meet-me room
where carriers physically interconnect
Reported

Sources: Wikipedia "a place within a colocation center (or carrier hotel) where telecommunications companies can physically connect to one another"

data

Hall-to-hall fiber

See it in 3D ↗
Strands
about 181k strands crossing between halls
Calc.

Calculated: Cross-hall strands = half the campus’s physical scale-out endpoints × fibers per link, summed over each fleet member (two endpoints per Rubin GPU, one otherwise). Method

Drawn spine-to-spine route
≈140 m between the halls, inside DR reach
Assumed

Assumed: Hall-to-hall route as drawn, ≈140 m of duct bank between the two detailed halls’ spine ends, under DR’s 500 m. Why

DR, parallel fiber
500 m, eight fibers on MPO-12/APC
Spec

Sources: NVIDIA DR4 vs FR4 section: "DR4 (data center reach) supports up to 500 meters using four parallel channels over eight fibers with MPO-12/APC connectors"

FR4, e.g. NVIDIA MMS4X50-NM (2 × FR4)
2 km, four wavelengths per 400G port, duplex LC
Spec

Sources: NVIDIA specifications: product name "800Gbps Twin-port OSFP 2xFR4, 2x400Gb/s Single Mode, 2km"; connector "Single-mode Duplex LC PC"; four wavelengths per transmitter, 1264.5–1337.5 nmNVIDIA "FR4 (far reach) supports up to two kilometers by multiplexing four channels of different laser wavelengths into a single fiber pair"; "DR4 is parallel and FR4 is multiplexed, so they can't be directly linked to each other"TIA Fiber Optics Technology Consortium 400GBASE-FR4, IEEE 802.3 clause 151 (first published as 802.3cu): "400 Gb/s wavelength-division multiplexed (WDM) PAM4 serial transmission over 2 single-mode optical fibers, with reach up to at least 2 km"

LR4, 400GBASE-LR4
10 km, four wavelengths
Spec

Sources: IEEE 802.3cu Task Force adopted objective: "Define a four-wavelength 400 Gb/s PHY for operation over SMF with lengths up to at least 10 km"

data

Duct bank

See it in 3D ↗
Strands crossing
≈181k
Calc.

Calculated: Cross-hall strands = half the campus’s physical scale-out endpoints × fibers per link, summed over each fleet member (two endpoints per Rubin GPU, one otherwise). Method

Cable
6,912-fiber ribbon fits a 2-inch duct
Spec

Sources: Prysmian FlexRibbon whitepaper: "6,912 bend-insensitive fibres small enough to fit into a 50.8 mm/2-inch duct"

Duct-bank layout
general telecom practice
Assumed

Assumed: Duct-bank layout, 4-inch conduits, one cable per conduit, a spare row. Why

data

Long-haul route

See it in 3D ↗
Light in fiber
≈4.9 µs per km, one way
Calc.

Calculated: One-way delay per km = 1 ÷ (speed of light ÷ assumed fiber refractive index), ≈ 4.9 µs/km. Method

heat

Dry coolers

See it in 3D ↗
Heat rejected
≈92 MW
Calc.

Calculated: Heat rejected ≈ IT MW × 1.05, adding back the plant’s own electrical overhead as heat. Method

NVIDIA warm-water spec
45 °C in, ≈55 °C out
Spec

Sources: NVIDIA NVIDIA blog: "the coolant entering a fully liquid-cooled chip at 45 degrees Celsius exits at roughly 55 degrees"

Water classes
ASHRAE W32–W45
Spec

Sources: Upsite Technologies blog: classes "W17, W27, W32, NEW class W40, W45" named for their maximum supply temperature in °CASHRAE TC 9.9, hosted via Vertiv p.4, "Change to ASHRAE Water Classifications": "the W classes are being renamed with the upper temperature limits incorporated in the name... W17 (previously W1), W27 (W2), W32 (W3), W40 (new), W45 (W4)"

heat

Cooling towers

See it in 3D ↗
WUE, this design
≈0.16 L/kWh IT
Assumed

Assumed: Water use by cooling design, air 1.0, liquid 0.5, warm water 0.16 L/kWh IT. Why

Water treatment
anti-scale, anti-corrosion and disinfectant dosing, plus blowdown
Reported

Sources: CDC "Automate anti-corrosion, anti-scale, and disinfectant addition and monitoring."

Water per day
≈336 m³
Calc.

Calculated: Water = WUE (L/kWh IT) × IT MW × 24 h, converted to cubic meters. Method

At the power plant, typical thermal
≈1.8 L/kWh (NREL)
Spec

Sources: NREL review’s synthesis figure: 0.47 gal (1.8 L) evaporated per kWh of end-use electricity for typical thermoelectric generation

heat

Where 100 MW goes

See it in 3D ↗
Heat out
100 MW
Calc.

Calculated: Heat out ≈ meter MW: energy conservation means virtually every watt drawn eventually becomes heat. Method

heat

Heat reuse

See it in 3D ↗
Meta Odense, Denmark
≈165,000 MWh a year, ≈11,000 homes
Reported

Sources: Munters case study: Meta’s Odense heat reuse provides "up to 165,000 MWh of energy a year" to warm "11,000 homes and businesses"

Microsoft + Fortum, Finland
up to 180 MW of district heat
Reported

Sources: Fortum press release: "72 units, producing up to 180 megawatts of district heating"

Stockholm Data Parks
30+ data centers selling heat
Reported

Sources: Stockholm Exergi / City of Stockholm program siteEurelectric "The platform now connects 30+ DCs across 16 providers."

Tie-in drawn here
an illustration, not built
Assumed

Assumed: Heat-export tie-in drawn in the heat layer, illustration, not built here. Why

3Power room & data hall70 m across

power

Unit substation

See it in 3D ↗
Rating
2.5 MVA
Assumed

Assumed: Unit substation size, 2.5 MVA. Why

Secondary current
≈3,000 A at full load
Calc.

Calculated: Current = 2.5 MVA ÷ (√3 × 480 V); the rating is already apparent power, so no separate power-factor term applies. Method

Efficiency
≈99%
Assumed

Sources: U.S. Department of Energy DOE's 04/04/2024 final rule updating distribution-transformer efficiency standards, which supersedes the 2013 rule

Assumed: Unit substation efficiency (cooling/building side), ≈99%. Why

Enclosure as drawn
representative
Assumed

Assumed: Unit substation as drawn, representative. Why

power

480 V switchgear

See it in 3D ↗
Transfer
automatic, utility ↔ generator
Assumed

Assumed: Standard switchgear, battery and busway design practice, automatic utility/generator transfer; Li-ion (LFP or NMC) UPS batteries; tap-off busway. Why

Markings
nameplate and arc-flash signs, representative
Assumed

Assumed: Nameplates, voltage stencils and hazard signs, model values only; designations and sign artwork representative. Why

power

UPS modules

See it in 3D ↗
Module
1.25–1.5 MW
Assumed

Assumed: UPS module size, 1.25 MW modeled. Why

Modules
≈75 across the campus, ≈38 per hall (2 halls)
Calc.

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

Assumed: UPS module size, 1.25 MW modeled. Why

Efficiency, Eaton 9395XR
up to 97.5% online, 99% eco
Reported

Sources: CEIE (distributor reprint of Eaton figures) "tested efficiency of Eaton 9395XR UPS dual conversion in online mode is as high as 97.5%"; eco/AC-direct mode "can be improved to 99%" (Eaton’s own product page could not be opened this pass -- see its unchecked note)

Loss at 100 MW
≈3.2 MW
Calc.

Calculated: Loss = the downstream load this stage feeds ÷ its assumed efficiency, minus that load itself. Method

Assumed: UPS double-conversion efficiency, 96.5%. Why

power

Battery cabinets

See it in 3D ↗
Runtime
set by string count, often ≈5 min
Assumed

Assumed: Hall battery runtime, ≈5 minutes. Why

Chemistry
Li-ion (LFP or NMC)
Assumed

Assumed: Standard switchgear, battery and busway design practice, automatic utility/generator transfer; Li-ion (LFP or NMC) UPS batteries; tap-off busway. Why

power

Overhead busway

See it in 3D ↗
Rack voltage
415 V three-phase
Reported

Sources: The Definitive Guide to AI Data Centers "415V (and its 400V European twin) is the de facto rack standard for liquid-density AI rows", vs. 208V three-phase in legacy air-cooled halls (the page does not separately discuss 480 V upstream distribution)

Per rack
≈190 A per phase at 131 kW
Calc.

Calculated: AC case: current = rack kW ÷ (√3 × 415 V × assumed power factor). DC case: current = rack kW ÷ 800 V. Method

Assumed: Power factor, 0.95. Why

Why busway
tap-offs move without rewiring
Assumed

Assumed: Standard switchgear, battery and busway design practice, automatic utility/generator transfer; Li-ion (LFP or NMC) UPS batteries; tap-off busway. Why

Hardware as drawn
representative
Assumed

Assumed: Busway hardware as drawn, representative. Why

power

GB200 NVL72 racks

See it in 3D ↗
Power, this model
≈131 kW
Calc.

Calculated: kW = GPU package power + CPU power + scale-up switching + NICs/DPUs + drives/fans/management, run up through the assumed 12 V and core-voltage conversion efficiencies, then divided by the rack’s AC/DC or DC/DC input-stage efficiency. Method

Published range
120–132 kW
Reported

Sources: ServeTheHome confirms the 120 kW figure ("the 120kW flagship system stacked in a single rack"); this site’s 132 kW upper figure is a commonly repeated peak/TDP figure not independently re-confirmed in this pass

GPUs
72 Blackwell
Spec

Sources: NVIDIA product page: the NVL72 platform name and specifications describe a 72-GPU rack

Liquid / air
114 kW / 17 kW
Calc.

Calculated: Split the rack total by the accelerator’s liquid-cooled heat share, this site’s estimate since vendors publish a rack-level figure, not a fixed engineering constant. Method

power

Hot aisle containment

See it in 3D ↗
Air share of heat
≈13%
Calc.

Calculated: Air share = 1 − the liquid-cooled share used for the rack’s liquid/air split above. Method

Doors and roof as drawn
representative
Assumed

Assumed: Containment doors as drawn, representative. Why

power

Coolant distribution unit

See it in 3D ↗
Capacity range
70 kW – 2.3 MW
Spec

Sources: Vertiv CoolChip CDU family: models from CDU 70 (70 kW) to CDU 2300 (2300 kW)Motivair (Schneider Electric) "COOLING UP TO 2.3MW", MCDU-4U (102 kW) through MCDU-60 (2.3 MW) rated-capacity table

Units, ≈1.25 MW each
≈61 across the campus, ≈31 per hall (2 halls)
Calc.

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

Assumed: Coolant distribution unit size, ≈1.25 MW modeled. Why

Rule
rack loop stays above dew point
Spec

Sources: Motivair (Schneider Electric) "The CDU maintains a secondary loop water temperature above the dew point in the data center to eliminate the possibility of condensation"

Cabinet drawn
representative
Assumed

Assumed: CDU / in-row cabinet as drawn, 0.8 m row slot; doors, louvres, touchscreen and roof valve ports representative. Why

power

Facility water loop

See it in 3D ↗
Supply → return
≈42 → 52 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Temperature rise
≈10 °C across the racks
Assumed

Assumed: Facility/rack water temperature rise, ≈10 °C. Why

Pipework as drawn
representative
Assumed

Assumed: Facility pipework as drawn, representative. Why

Pipe markers
green with white letters and a flow arrow, ASME A13.1 style
Assumed

Sources: Project Materials color code chart: "Green water, potable water, cooling water, condensate", green background with white text; "Flow direction arrows are a mandatory element of every pipe label"

Assumed: Pipe markers on the cooling water, ASME A13.1 style; wording and spacing representative. Why

power

Fan wall

See it in 3D ↗
Share of rack heat
≈13%
Calc.

Calculated: Air share = 1 − the liquid-cooled share used for the rack’s liquid/air split above. Method

Cells as drawn
representative
Assumed

Assumed: Fan wall as drawn, representative. Why

power

Optical modules

See it in 3D ↗
Supply
3.3 V on 4 power contacts, up to 2.5 A each (3.25 A on OSFP1600)
Spec

Sources: OSFP MSA Rev 5.22 §15.6: "each power pin allows up to 2.5 Amps for a total of 10.0 Amps... For OSFP1600, each power pin allows up to 3.25 Amps"

Edge connector
60 contacts: 32 high-speed (8 TX + 8 RX differential pairs), 4 control, 4 power, 20 ground
Spec

Sources: OSFP MSA Module Electrical Connector section, checked in the MSA text mirrored by fluxlight.com (the rev 5.22 PDF would not extract): "16 contacts for 8 differential pairs of high-speed transmit signals, 16 contacts for 8 differential pairs of high-speed receive signals, 4 contacts for low-speed control signals, 4 contacts for power and 20 contacts for ground"

power

Co-packaged optics

See it in 3D ↗
Per port, NVIDIA figure
as low as 9 W
Vendor

Sources: NVIDIA body text: 30 W per port with pluggable transceivers, "as low as 9 W" per port with co-packaged optics

Compared with: 30 W per port with pluggable transceivers

Q3450 switch, NVIDIA figure
3.95 kW
Vendor

Sources: Lambda power comparison: "NVIDIA Photonics CPO switch: 3.95 kW. Standard switch: 7.0 kW."

Compared with: a pluggable-optics equivalent switch at 7.0 kW

power

Network spine

See it in 3D ↗
Per GPU
400 Gb/s scale-out
Spec

Sources: NVIDIA "NVIDIA Quantum-2 InfiniBand networking that delivers 400Gb/s bandwidth per GPU"

data

Fiber distribution frames

See it in 3D ↗
Fabric strands, whole campus
≈1.1 million
Calc.

Calculated: Physical endpoints = GPUs × scale-out NICs per GPU (two for Rubin, one for the other modeled generations). The selected fabric uses each physical port’s rate, not aggregate per-GPU bandwidth. Switch, module and fiber counts follow that endpoint inventory and the modeled tiers. Method

Housing density, Corning EDGE8
144 fibers per 1U
Spec

Sources: Corning EDGE8-01U-SP product page: "Number of Modules: 18", "Fiber Capacity: 144" (18 modules × 8 fibers)

4U housings for this campus
≈1,887
Calc.

Calculated: Housings = total fabric strands ÷ fibers per housing (four EDGE8 1U housings’ worth, 4 × 144 fibers, per 4U enclosure). Method

Bay layout as drawn
representative
Assumed

Assumed: Fiber frame bays as drawn, representative. Why

data

To the other halls

See it in 3D ↗
Strands between halls, campus
≈181k
Calc.

Calculated: Half of this campus’s spine-to-core links are modeled as crossing between halls; strands = (physical scale-out endpoints ÷ 2) × fibers per link, summed over each fleet member. Method

data

One model, four racks

See it in 3D ↗
Traffic
point to point, per micro-batch
Reported

Sources: Meta AI, arXiv §3.3.2, Parallelism for Model Scaling: activations pass point-to-point between pipeline stages

Llama 3 405B
pipeline parallel 16
Spec

Sources: Meta AI, arXiv §3.3.2, Table 4: 4D parallelism for 405B pretraining, TP=8, CP=16, PP=16, DP=8 on up to 16K H100 GPUs

Layout drawn here
illustrative
Assumed

Assumed: Pipeline-stage layout drawn on the hall floor, illustrative. Why

data

Copies of the model

See it in 3D ↗
Traffic
large gradient sync, once per step
Reported

Sources: Meta AI, arXiv §3.3.2, Parallelism for Model Scaling: data-parallel replicas synchronize after each training step

Llama 3 405B
TP 8 × CP 16 × PP 16 × DP 8 = 16,384 GPUs
Spec

Sources: Meta AI, arXiv §3.3.2, Table 4: "Llama 3 405B is trained on up to 16K H100 GPUs" with TP=8, CP=16, PP=16, DP=8

DeepSeek-V3
no tensor parallel; EP 64, PP 16, ZeRO-1 DP
Spec

Sources: DeepSeek-AI, arXiv §3.2, Training Framework: "train DeepSeek-V3 without using costly Tensor Parallelism (TP)"; "applies 16-way Pipeline Parallelism (PP), 64-way Expert Parallelism (EP) spanning 8 nodes, and ZeRO-1 Data Parallelism (DP)"

data

Rack uplinks

See it in 3D ↗
Per GPU
400 Gb/s
Spec

Sources: NVIDIA "NVIDIA Quantum-2 InfiniBand networking that delivers 400Gb/s bandwidth per GPU"

Ports per rack
72
Spec

Sources: NVIDIA NVL72 platform page: 72 GPUs per rack

data

Leaf switches

See it in 3D ↗
Leaf switches, campus
≈1,416
Calc.

Calculated: Physical endpoints = GPUs × scale-out NICs per GPU (two for Rubin, one for the other modeled generations). The selected fabric uses each physical port’s rate, not aggregate per-GPU bandwidth. Switch, module and fiber counts follow that endpoint inventory and the modeled tiers. Method

Hops, same rail
1
Assumed

Assumed: Rail-optimized fabric hop count, 1 hop, same rail. Why

Switch hop, InfiniBand
under ≈100 ns; NVIDIA publishes none
Assumed

Assumed: InfiniBand switch hop latency, under ≈100 ns. Why

Switch drawn, QM9700
1U, 32 OSFP cages
Spec

Sources: NVIDIA QM97xx specifications: 32 OSFP cages; 438 mm × 43.6 mm (1U) × 660 mm

Face arrangement
representative
Assumed

Assumed: Switch faces as drawn, representative arrangement, true-size parts. Why

data

Spine switches

See it in 3D ↗
Switch
Quantum-2, 64 × 400G
Spec

Sources: NVIDIA QM97xx specifications: 32 OSFP cages, 25.6 Tbps total (64 logical 400G NDR ports)

Merchant switch chips, same role
Broadcom Tomahawk 6 (102.4 Tb/s), Marvell Teralynx 10 (51.2 Tb/s)
Spec

Sources: Broadcom page title: "Broadcom Now Shipping World’s First 102.4 Tbps Switch in Production Volume"Marvell, via PR Newswire press release: "a low power, programmable 51.2 Tbps Ethernet device"

Spine + core switches, campus
≈2,124
Calc.

Calculated: Physical endpoints = GPUs × scale-out NICs per GPU (two for Rubin, one for the other modeled generations). The selected fabric uses each physical port’s rate, not aggregate per-GPU bandwidth. Switch, module and fiber counts follow that endpoint inventory and the modeled tiers. Method

Switch drawn, QM9700
1U, 32 OSFP cages
Spec

Sources: NVIDIA QM97xx specifications: 32 OSFP cages; 438 mm × 43.6 mm (1U) × 660 mm

Face arrangement
representative
Assumed

Assumed: Switch faces as drawn, representative arrangement, true-size parts. Why

data

Storage and its network

See it in 3D ↗
Separate networks, NVIDIA’s design
compute, storage, in-band and out-of-band management, besides NVLink
Spec

Sources: NVIDIA the Network Fabrics page describes the NVLink fabric, the compute fabric, the storage fabric and network, the in-band and out-of-band management networks; storage: "The storage network embodies the performance for the high-speed storage while keeping the support for high availability."

Storage attach
RDMA over Ethernet (RoCEv2), a partner’s shared file system
Spec

Sources: NVIDIA "High-Performance Storage is provided via RDMA over Converged Ethernet v2 (RoCEv2) connected storage from a DGX SuperPOD certified storage partner"

data

Management and login servers

See it in 3D ↗
In NVIDIA’s B300 design
2 for Base Command Manager, 3 for Kubernetes, 2 Slurm login nodes
Spec

Sources: NVIDIA Management Server Quantities and Connectivity: "Base Command Manager in High Availability (HA): 2 Nodes"; "K8s Management Server: 3 Nodes"; "SLURM Login Nodes, 2 Nodes"

data

Fiber runways

See it in 3D ↗
Fibers per link
8
Assumed

Assumed: Fibers per fabric link, 8 (400G/800G tier) or 16 (1.6T tier). Why

Multimode fiber jacket
aqua
Spec

Sources: NVIDIA Introduction: "Multimode optics is denoted by a tan-colored pull tab and aqua-colored optical fiber."NVIDIA §2 Application: "Multimode fibers use an industry standard Aqua fiber jacket color"

Single-mode fiber jacket
yellow
Spec

Sources: NVIDIA §2 Application: "Single-mode fibers use an industry standard yellow fiber jacket color"

Splitter cable reach
up to 50 m, through two patch panels
Spec

Sources: NVIDIA Key Features: "50m max reach"; §2: "50-meter specification assumes two optical patch panels in the link with total of 4 optical connector junctions"NVIDIA Key Features: "Up to 50 m reach"

Aqua cables in this runway
representative
Assumed

Assumed: Aqua multimode links at one leaf switch, as drawn, representative. Why

data

Optical modules

See it in 3D ↗
At the switch: twin-port 800G OSFP, 2 × 400G
17 W max
Spec

Sources: NVIDIA §4.2, Recommended Operating Conditions and Power Supply Requirements: Maximum Power Dissipation, Max 17 WNVIDIA "The 400G IB/EN switches require finned-top 2x400G transceivers for additional cooling"

Short runs, multimode option
2 × SR4, OM4 up to 50 m (OM3 30 m)
Spec

Sources: NVIDIA MMA4Z00-NS specifications: "Operating Distance (OM3) 2–30 m", "Operating Distance (OM4) 2–50 m"; two MPO-12/APC connectors

Multimode light source
850 nm VCSEL
Spec

Sources: NVIDIA Key Features, MMA4Z00-NS400 (the single-port 400G SR4 adapter end): "850nm VCSEL"NVIDIA MMA4Z00-NS transmitter and receiver tables: "Wavelength λC 844 850 863 nm"

How multimode is marked
tan pull tab, aqua fiber
Spec

Sources: NVIDIA "Tan pull-tab denotes multimode optics"NVIDIA Introduction: "Multimode optics is denoted by a tan-colored pull tab and aqua-colored optical fiber."

1:2 splitter cables
one 4-channel MPO-12 port to two 2-channel ends, 200G each
Spec

Sources: NVIDIA Introduction: "a multimode, 4-channel-to-two 2-channel splitter fiber cable"; "The 2-channel ends are inserted into two, single-port 400Gb/s OSFP and/or QSFP112 transceivers which with only 2 fibers can output 200G rates"NVIDIA Introduction: the single-mode version, the same 4-channel-to-two 2-channel split

Twin-port module with two splitters
one switch cage to four 200G adapters
Spec

Sources: NVIDIA Introduction: "Two splitter fiber cables are used in the twin-port OSFP transceiver enabling four, 2-channel ends to four transceivers"; the typical use is "linking OSFP switches to" ConnectX-7 adapters and/or BlueField-3 DPUsNVIDIA §2.1 Connectivity Scenarios, 200Gb/s mode: "This case creates links to four 200Gb/s ConnectX-7/OSFP adapter cards"

Adapter module on a splitter end
2 lanes lit, 200G; ≈5.5 W typical instead of 8 W
Spec

Sources: NVIDIA §2.1.1 Use cases: two-fiber ends "only activate two of the lanes in the 400G transceiver creating a 200G device and automatically reduces the power consumption of only the 400G transceivers from 8 Watts typical to 5.5 Watts typical"

Straight and split on one module
not mixed: both ports straight, or both split
Spec

Sources: NVIDIA §2.1: both fibers in the twin-port transceiver "must be the same type - straight or splitter and cannot be mixed"NVIDIA §2 Application: twin-port transceivers "must use the same fiber type in both MPO-12/APC ports (straight or 1:2 splitter) and cannot be mixed"

Aqua straight links and splitters at one leaf, as drawn
representative
Assumed

Assumed: Aqua multimode links at one leaf switch, as drawn, representative. Why

The DSP inside each module
e.g. Marvell Ara, Broadcom Sian, Credo Bluebird
Spec

Sources: Marvell Marvell’s own 1.6T optical DSP portfolio announcementBroadcom, via GlobeNewswire Broadcom’s own Sian3 200G-lane DSP announcementCredo Credo’s own Bluebird 1.6T optical DSP product page

Linear-drive (LPO), Semtech target
about 10 W versus 23–25 W retimed
Vendor

Sources: Semtech "200G LPO Power, Reach and Loss: Real Numbers" -- Semtech’s own published LPO-vs-DSP power comparison

Compared with: Semtech’s 200G-lane discussion: target linear-module power versus current retimed modules, not a universal measured saving

data

Co-packaged optics

See it in 3D ↗
NVIDIA Quantum-X / Spectrum-X Photonics
5× power efficiency, 4× fewer lasers (Aug. 2026 reporting; NVIDIA’s own March 2025 launch claimed 3.5×)
Vendor

Sources: StorageReview "5x lower power consumption" and "4x fewer lasers" vs. conventional pluggable-optics switches, 08/15/2026NVIDIA launch announcement: "4x fewer lasers to deliver 3.5x more power efficiency... compared with traditional methods", 03/18/2025NVIDIA 08/26/2025 post restates "reducing the total number of lasers in the data center by a factor of four compared to legacy designs" but not a power-efficiency multiplier

Compared with: conventional switches using pluggable optical transceivers

Broadcom Davisson
102.4 Tb/s, ≈3.5 W per 800G port
Vendor

Sources: Broadcom page title: "Broadcom Now Shipping World’s First 102.4 Tbps Switch in Production Volume" -- Davisson is Broadcom’s CPO variant built on this same Tomahawk 6 ASIC (per nextplatform-broadcom-cpo)The Next Platform "An 800 Gb/sec port will burn about 3.5 watts, says Broadcom, which is 36.4 percent lower than with the Tomahawk 5 CPO port at the same bandwidth and more than 70 percent lower than pluggable optics at the same bandwidth"

Compared with: Broadcom’s prior-generation Tomahawk 5 CPO port and pluggable optics, both at 800 Gb/s

Stand-in faceplate, Q3450 counts
144 MPO, 18 laser modules, 4 capped UDQ4 ports
Reported

Sources: Lambda front panel: 144 MPO optical connectors; "The switch supports 18 removable external light-source modules"; "four UDQ4 liquid cooling connections"

Faceplate layout
representative
Assumed

Assumed: CPO stand-in faceplate arrangement, representative. Why

Where CPO ships today
standalone switches like this stand-in, from NVIDIA and Broadcom
Reported

Sources: TrendForce first paragraph: "NVIDIA has begun shipping its next-generation Spectrum-X CPO switch to select partners. Meanwhile, Broadcom is continuing limited shipments of its 51.2T Bailly CPO switch"NVIDIA body: Spectrum-X Ethernet Photonics, "the world’s first co-packaged-optics (CPO)-based switches with 200Gb/s SerDes — now in production"

data

NVL72 racks

See it in 3D ↗
NVLink per GPU
1.8 TB/s
Spec

Sources: NVIDIA Blackwell platform launch release: "the latest iteration of NVIDIA NVLink delivers groundbreaking 1.8TB/s bidirectional throughput per GPU" (the NVL72 product page itself states only the 130 TB/s system aggregate, not a per-GPU figure)

Domain
72 GPUs
Spec

Sources: NVIDIA NVL72 platform page: all 72 GPUs in one NVLink domain

heat

Coolant distribution unit

See it in 3D ↗
Capacity range
70 kW – 2.3 MW
Spec

Sources: Vertiv CoolChip CDU family: models from CDU 70 (70 kW) to CDU 2300 (2300 kW)Motivair (Schneider Electric) "COOLING UP TO 2.3MW", MCDU-4U (102 kW) through MCDU-60 (2.3 MW) rated-capacity table

Approach, facility to rack loop
a few °C (≈3 °C here)
Assumed

Assumed: CDU approach temperature, facility to rack loop, a few °C (≈3 °C in this model). Why

Rack loop, supply → return
≈45 → 55 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

heat

Hot aisle

See it in 3D ↗
Air share of rack heat
≈13%
Calc.

Calculated: Air share = 1 − the liquid-cooled share used for the rack’s liquid/air split above. Method

Doors and roof as drawn
representative
Assumed

Assumed: Containment doors as drawn, representative. Why

heat

Fan wall

See it in 3D ↗
Moves
the ≈13% air share
Calc.

Calculated: Air share = 1 − the liquid-cooled share used for the rack’s liquid/air split above. Method

Cells as drawn
representative
Assumed

Assumed: Fan wall as drawn, representative. Why

heat

Facility water loop

See it in 3D ↗
Rise
≈10 °C
Assumed

Assumed: Facility/rack water temperature rise, ≈10 °C. Why

Supply → return
≈42 → 52 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Pipework as drawn
representative
Assumed

Assumed: Facility pipework as drawn, representative. Why

Pipe markers
green with white letters and a flow arrow, ASME A13.1 style
Assumed

Sources: Project Materials color code chart: "Green water, potable water, cooling water, condensate", green background with white text; "Flow direction arrows are a mandatory element of every pipe label"

Assumed: Pipe markers on the cooling water, ASME A13.1 style; wording and spacing representative. Why

heat

Co-packaged optics

See it in 3D ↗
Cooling, Q3450 switch
liquid; 4 UDQ4 connections, dual internal loops
Reported

Sources: Lambda cooling: "Cooling runs through four UDQ4 liquid cooling connections with dual internal loops."

heat

Fire detection

See it in 3D ↗
Air-sampling detection
continuous, the earliest warning
Reported

Sources: Xtralis (Honeywell) "continuous air sampling provide the earliest possible warning of an impending fire hazard"

Devices as drawn
representative
Assumed

Assumed: Detector and sprinkler heads as drawn, representative. Why

NFPA 75 covers
fire, smoke, corrosion, heat and water damage
Reported

Sources: Consulting-Specifying Engineer NFPA 75’s purpose: "minimum requirements for the protection of IT equipment and IT equipment areas from damage by fire or its associated effects"

heat

Risers to the roof

See it in 3D ↗
Carries
nearly all of the hall’s heat
Assumed

Assumed: Standard switchgear, battery and busway design practice, automatic utility/generator transfer; Li-ion (LFP or NMC) UPS batteries; tap-off busway. Why

Pipework as drawn
representative
Assumed

Assumed: Facility pipework as drawn, representative. Why

Pipe markers
green with white letters and a flow arrow, ASME A13.1 style
Assumed

Sources: Project Materials color code chart: "Green water, potable water, cooling water, condensate", green background with white text; "Flow direction arrows are a mandatory element of every pipe label"

Assumed: Pipe markers on the cooling water, ASME A13.1 style; wording and spacing representative. Why

4The rack2.2 m tall

power

Rack feed

See it in 3D ↗
Feeds
A + B
Assumed

Assumed: Two power feeds into a rack (A/B), A + B tap-offs. Why

Tap-off hardware as drawn
representative
Assumed

Assumed: Busway and tap-off units as drawn, representative. Why

power

Power shelves

See it in 3D ↗
Shelf
≈33 kW, 6 × 5.5 kW
Spec

Sources: Flex product page: "consist of 6 PSUs with a max output power of 33kW"; component diagram separately labels one PSU "5500W PSU" (33 kW ÷ 6 = 5.5 kW matches)

Shelves per rack
8
Spec

Sources: NVIDIA Hardware: "The power shelf uses six air-cooled 5.5kW PSUs in eight power shelves that provide N+N redundancy and the required input power of 33kW per power shelf"

Efficiency
≈97.5% peak
Spec

Sources: Flex product page: "High efficiency up to 97.5% (peak)"

power

DC busbar

See it in 3D ↗
Voltage
≈50 V DC (OCP ORv3)
Spec

Sources: NVIDIA Hardware: shelves "convert AC power into nominal 50V-51V DC output and distribute it through the bus bar"

Busbar rating
1,400 A
Spec

Sources: NVIDIA "The new design supports a substantially higher 1,400 amp current flow"

Next: NVIDIA Kyber, 2027
800 V DC, 45% less copper
Vendor

Sources: NVIDIA "reducing copper requirements by 45%"; "Full-scale production... will coincide with NVIDIA Kyber rack-scale systems in 2027"

Compared with: NVIDIA’s comparison is with traditional 415 V AC distribution, not with today’s 54 V/50 V busbar

power

Compute trays

See it in 3D ↗
Trays
18
Spec

Sources: NVIDIA Hardware: "18x 1RU compute trays, each with 2 Grace CPUs and 4 Blackwell GPUs"; "9x 1RU NVLink switch trays" of "2x NVLink NVSwitches" eachNVIDIA System Hardware & Components: 18 compute trays, 9 NVLink switch trays of 2 NVSwitch ASICs each

GPUs per tray
4
Spec

Sources: NVIDIA Hardware: "18x 1RU compute trays, each with 2 Grace CPUs and 4 Blackwell GPUs"; "9x 1RU NVLink switch trays" of "2x NVLink NVSwitches" eachNVIDIA System Hardware & Components: 18 compute trays, 9 NVLink switch trays of 2 NVSwitch ASICs each

CPUs per tray
2
Spec

Sources: NVIDIA Hardware: "18x 1RU compute trays, each with 2 Grace CPUs and 4 Blackwell GPUs"; "9x 1RU NVLink switch trays" of "2x NVLink NVSwitches" eachNVIDIA System Hardware & Components: 18 compute trays, 9 NVLink switch trays of 2 NVSwitch ASICs each

Tray power
≈6.5 kW
Calc.

Calculated: NVL72: tray power = (rack DC-bus power − NVLink scale-up power − busbar loss) ÷ 18 trays. DGX H100: tray power = rack power ÷ 4 servers. Method

Front panels as drawn
representative
Assumed

Assumed: NVL72 tray and shelf front panels as drawn, representative. Why

power

NVLink switch trays

See it in 3D ↗
Trays
9
Spec

Sources: NVIDIA Hardware: "18x 1RU compute trays, each with 2 Grace CPUs and 4 Blackwell GPUs"; "9x 1RU NVLink switch trays" of "2x NVLink NVSwitches" eachNVIDIA System Hardware & Components: 18 compute trays, 9 NVLink switch trays of 2 NVSwitch ASICs each

Switch chips per tray
2
Spec

Sources: NVIDIA Hardware: "18x 1RU compute trays, each with 2 Grace CPUs and 4 Blackwell GPUs"; "9x 1RU NVLink switch trays" of "2x NVLink NVSwitches" eachNVIDIA System Hardware & Components: 18 compute trays, 9 NVLink switch trays of 2 NVSwitch ASICs each

Bandwidth per GPU
1.8 TB/s
Spec

Sources: NVIDIA "The revolutionary 1.8 TB/s of bidirectional throughput per GPU"

Domain total
130 TB/s
Spec

Sources: NVIDIA "130 terabytes per second (TB/s) of low-latency GPU communications"

Front handles
gold, for removing the tray
Reported

Sources: ServeTheHome NVSwitch shelves: "these gold features are handles to remove the shelves"

power

NVLink spine

See it in 3D ↗
Links
more than 5,000 copper
Spec

Sources: NVIDIA "These cartridges accommodate over 5,000 active copper cables"NVIDIA Hardware: opening paragraph calls the same backplane the "NVLink passive copper cable cartridge backplane"

Total length
≈2 miles
Reported

Sources: The Register rack tour: "Both the NVLink switch and compute sleds slot into a blind mate backplane with more than 2 miles (3.2 km) of copper cabling"

Signaling
224G PAM4
Reported

Sources: NADDOD interconnect analysis: NVLink 5 cartridge links run 224 Gb/s PAM4; not stated in these terms by NVIDIA itself

Tray connection
blind-mate backplane, no loose cables
Reported

Sources: The Register rack tour: "Both the NVLink switch and compute sleds slot into a blind mate backplane"

Cartridge hardware as drawn
representative
Assumed

Assumed: NVLink cartridge and busbar hardware as drawn, representative. Why

power

Coolant manifolds

See it in 3D ↗
Liquid-cooled parts
GPUs and CPUs
Spec

Sources: NVIDIA Hardware: compute trays "cooled by liquid that runs up and down the rack through manifolds, then through the cold plates that are attached to the CPUs and the GPUs"

Tray connection
blind-mate nozzles
Reported

Sources: ServeTheHome "The rack is designed to blind-mate the power via the bus bar, as well as the liquid cooling nozzles and the NVLink connections for each component."

Manifold hardware as drawn
representative
Assumed

Assumed: Coolant manifold hardware as drawn, representative. Why

Pipe markers
TCS supply and return, ASME A13.1 style
Assumed

Assumed: Pipe markers on the cooling water, ASME A13.1 style; wording and spacing representative. Why

data

NVLink: one layer, 72 GPUs

See it in 3D ↗
Traffic
every layer, many times per token
Reported

Sources: Meta AI, arXiv describes tensor-parallel GPUs exchanging activations inside every transformer layer, many times per forward/backward pass

Llama 3 405B, H100
tensor parallel 8, inside each server
Spec

Sources: Meta AI, arXiv training section: 405B model trained with tensor parallelism 8

NVL72 wide expert parallel
up to 72 GPUs, the NVLink domain
Reported

Sources: NVIDIA blog: "The NVLink domain can now support up to 72 NVIDIA Blackwell GPUs"; its own DeepSeek R1 example spreads 256 routed experts across 64 of those GPUs, not all 72

data

NVLink switch trays

See it in 3D ↗
Trays
9, 2 switch chips each (18)
Spec

Sources: NVIDIA Hardware: "18x 1RU compute trays, each with 2 Grace CPUs and 4 Blackwell GPUs"; "9x 1RU NVLink switch trays" of "2x NVLink NVSwitches" eachNVIDIA System Hardware & Components: 18 compute trays, 9 NVLink switch trays of 2 NVSwitch ASICs each

Per GPU
NVLink 5, 1.8 TB/s
Spec

Sources: NVIDIA "The revolutionary 1.8 TB/s of bidirectional throughput per GPU"

data

NVLink spine

See it in 3D ↗
Links
more than 5,000
Spec

Sources: NVIDIA "These cartridges accommodate over 5,000 active copper cables"

Power saved vs optics, NVIDIA
≈20 kW per rack
Vendor

Sources: SemiAnalysis quotes Jensen Huang: copper NVLink saves roughly 20 kW per rack vs an equivalent optical scale-up fabric

Compared with: an equivalent NVL72-scale optical scale-up fabric

Passive copper reach at 224G
≈1 m
Reported

Sources: IEEE 802.3 IEEE 802.3dj electrical ad hoc contribution on copper reach objectives at 224 Gb/s PAM4

data

Copper and optical fabrics

See it in 3D ↗
Ironwood superpod
9,216 chips, 144 cubes of 64
Spec

Sources: Google Cloud "a small ‘pod’ (e.g., a 256-chip Ironwood pod with four cubes) to a massive ‘superpod’ (e.g., a 9,216-chip system with 144 cubes)"; each cube is "64 Ironwood chips"

ICI per chip, inside a cube
1.2 TB/s; mostly copper
Reported

Sources: Google Cloud "linked via a breakthrough Inter-Chip Interconnect (ICI) network operating at 9.6 Tb/s" (= 1.2 TB/s); Google’s own pages do not state the link mediumSemiAnalysis reports that ICI links interior to the 4×4×4 cube run over direct-attached copper (DAC) cables, while chips at the cube’s face/edge/corner use optical transceivers instead

TPU v4: 4,096 chips in 64 racks
48 OCSes join them
Spec

Sources: Google (Jouppi et al., ISCA 2023) Section 2.2, Construction of the TPU v4 Supercomputer: "48 OCSes connect the 48 pairs of cables from 64 4^3 blocks (each 64 chips), yielding the desired total of 4096 TPU v4 chips"; "The 48 OCSes join eight rows together to form the complete 64-rack system."

Packet switches for the same job, Google’s estimate
568 InfiniBand switches
Vendor

Sources: Google (Jouppi et al., ISCA 2023) Section 7.3, What if TPU v4 used IB versus OCS?: "To replace the 48 128-port OCSes, 4096 TPU v4s need 568 IB switches." (a full 3-level fat tree, following NVIDIA’s guidance)

Compared with: the 48 OCSes that join the 4,096-chip TPU v4 system

Palomar OCS ports
136 × 136: 128 in use + 8 spares
Spec

Sources: Google (Jouppi et al., ISCA 2023) Section 2.2: "The Palomar OCS is 136×136 (128 ports plus 8 spares for link testing and repairs)"

How it switches
3D MEMS mirrors, in milliseconds
Spec

Sources: Google (Jouppi et al., ISCA 2023) Section 2.1, Optical Circuit Switching: "based on 3D Micro-Electro-Mechanical Systems (MEMS) mirrors that switch in milliseconds. They employ circulators to send light both ways in a fiber, halving the number of required ports and cables."

OCS power, reported
≈108 W vs ≈3,000 W for a 136-port packet switch
Reported

Sources: SemiAnalysis "The Apollo switch uses only 108 watts of power consumption compared to a standard 136 port electrical packet switch (EPS), which would be in the 3,000 watts range."

Jupiter data center network
an OCS layer replaces the spine
Spec

Sources: Google (Poutievski et al., SIGCOMM 2022) Section 1, Introduction: OCSes "move Jupiter from a Clos to a block-level direct-connect topology that eliminates the spine switching layer and its associated challenges altogether"

data

Scale-out ports

See it in 3D ↗
Ports
72
Reported

Sources: SemiAnalysis describes NVL72’s scale-out network as 72 OSFP ports, one per GPU

Aggregate per GPU
400 Gb/s
Spec

Sources: NVIDIA Hardware: each compute tray has "4x NVIDIA ConnectX-7 single port 400G OSFP NIC"

DPU cages on the tray front, as drawn
QSFP112 form, two per DPU; positions representative
Assumed

Sources: NVIDIA Network ports: every BlueField-3 DPU and SuperNIC model lists "QSFP112 (copper and optical)", dual-port on the DPU models; the guide does not place the ports on a DGX GB tray front

Assumed: BlueField-3 network cages on the tray front, as drawn, QSFP112 form, two per DPU; positions representative. Why

data

Compute trays

See it in 3D ↗
GPUs
4
Spec

Sources: NVIDIA Hardware: "18x 1RU compute trays, each with 2 Grace CPUs and 4 Blackwell GPUs"; "9x 1RU NVLink switch trays" of "2x NVLink NVSwitches" eachNVIDIA System Hardware & Components: 18 compute trays, 9 NVLink switch trays of 2 NVSwitch ASICs each

data

Management switch

See it in 3D ↗
Rate
1–10 GbE class
Assumed

Assumed: Management/BMC network speed, 1–10 GbE class. Why

TOR switches
2
Spec

Sources: NVIDIA Hardware: "2x TOR Switches for management"

Out-of-band network, GB200 reference
every compute and switch tray’s management port, on SN2201 switches
Spec

Sources: NVIDIA Out-of-Band Management Network: "The OOB management network use SN2201 switches"; "It connects the management ports of all devices including DGX GB200 compute trays, switch trays, and management servers, storage, networking gear, rack PDUs, and all other devices"; "The OOB network carries all IPMI related control traffic"

Leads as drawn
one RJ45 copper lead per tray, to the 1 GbE ports
Assumed

Sources: NVIDIA Connector/Port Specifications: "48 RJ45 ports of 1GbE and 4 QSFP28 ports of 100GbE"

Assumed: Management leads to the trays as drawn, one RJ45 copper lead per tray; routing representative. Why

Ports as drawn
48 × 1 GbE + 4 × 100 GbE
Assumed

Sources: NVIDIA Connector/Port Specifications: "48 RJ45 ports of 1GbE and 4 QSFP28 ports of 100GbE"

Assumed: Management switch front panel as drawn, SN2201-class: 48 × 1 GbE RJ45, 4 × 100 GbE QSFP28. Why

Free units above
cable manager + blanking panels
Assumed

Assumed: Unused rack units above the switches, 1U brush-strip cable manager and 1U blanking panels. Why

heat

Coolant manifolds

See it in 3D ↗
To liquid, this model
≈114 kW
Calc.

Calculated: Split the rack total by the liquid-cooled share this site assumes for this rack (the cited vendor spec sheet could not be re-read in the 10/08/2026 audit). Method

Rise across the rack
≈10 °C
Reported

Sources: Alliance Chemical "roughly 170–195 liters per minute of coolant at a 10°C inlet-to-outlet rise (about 1.5 LPM per kW)" for a GB200 NVL72 rack

Supply → return, this design
≈45 → 55 °C
Spec

Sources: NVIDIA NVIDIA blog: "the coolant entering a fully liquid-cooled chip at 45 degrees Celsius exits at roughly 55 degrees"

Flow rate
sources disagree ≈5×
Assumed

Sources: Alliance Chemical "roughly 170–195 liters per minute"; notes "the OEM/CDU specification governs" the built system

Assumed: Coolant rise and flow rate across an NVL72 rack, ≈10 °C rise; ≈170–195 L/min. Why

heat

Rear exhaust

See it in 3D ↗
To air, this model
≈17 kW
Calc.

Calculated: Split the rack total by the liquid-cooled share this site assumes for this rack (the cited vendor spec sheet could not be re-read in the 10/08/2026 audit). Method

heat

Compute trays

See it in 3D ↗
Per tray
≈6.5 kW
Calc.

Calculated: NVL72: tray power = (rack DC-bus power − NVLink scale-up power − busbar loss) ÷ 18 trays. DGX H100: tray power = rack power ÷ 4 servers. Method

5GB200 compute tray44 cm wide, schematic

power

Busbar clip

See it in 3D ↗
Tray power
≈6.5 kW
Calc.

Calculated: NVL72: tray power = (rack DC-bus power − NVLink scale-up power − busbar loss) ÷ 18 trays. DGX H100: tray power = rack power ÷ 4 servers. Method

Current at 50 V
≈130 A
Calc.

Calculated: Current = tray power ÷ 50 V busbar voltage. Method

Finger count and housing
representative
Assumed

Assumed: Compute tray mechanical detail, representative. Why

power

Bus converters

See it in 3D ↗
Efficiency
≈97–98%
Assumed

Assumed: Bus-converter (IBC) efficiency, 97–98%. Why

Loss, campus-wide
≈1.4 MW
Calc.

Calculated: Per-rack loss = power passing through the stage × (1 ÷ assumed stage efficiency − 1); the campus figure sums this across every rack the scenario builds. Method

power

Voltage regulators

See it in 3D ↗
Phases per GPU
≈20–30
Assumed

Assumed: VRM phases per GPU, ≈20–30. Why

Efficiency
≈91%
Assumed

Assumed: Voltage-regulator (VRM) efficiency, 90–92%. Why

Loss, campus-wide
≈5.4 MW
Calc.

Calculated: Per-rack loss = power passing through the stage × (1 ÷ assumed stage efficiency − 1); the campus figure sums this across every rack the scenario builds. Method

Core current
≈1,300 A
Calc.

Calculated: Current = GPU power less its HBM share (the HBM stacks have their own supply) ÷ an assumed 0.8 V core voltage. Method

Board layout, traces and passives
representative
Assumed

Assumed: Compute tray mechanical detail, representative. Why

power

Blackwell GPUs

See it in 3D ↗
Power
≈1,200 W
Reported

Sources: SemiAnalysis / InferenceX chip spec table: "TDP per chip: 1,200 W"

Transistors
208 billion
Spec

Sources: NVIDIA "NVIDIA Blackwell-architecture GPUs pack 208 billion transistors"

Memory, from rounded rack total
≈186 GB HBM3e
Calc.

Inputs: NVIDIA "13.4 TB HBM3E" total ÷ 72 GPUs

Calculated: Per-GPU capacity = NVIDIA’s published rack-wide HBM total ÷ 72 GPUs. This is usable capacity as NVIDIA states it for the rack; the 8-stack, 24 GB/stack nominal figure vendors also cite (192 GB for GB200, 288 GB for GB300) is a few percent higher, likely raw die capacity before ECC/redundancy reservation — the two are not reconciled here. Method

power

Grace CPUs

See it in 3D ↗
Cores
72 Arm Neoverse V2 cores
Spec

Sources: NVIDIA "144 Arm Neoverse V2 cores into a single module" (two Grace CPUs); 72 cores per CPU

CPU–GPU link
900 GB/s NVLink-C2C
Assumed

Sources: NVIDIA "The Grace CPU Superchip is composed of two Grace CPUs connected coherently over NVIDIA NVLink-C2C at 900 GB/s" — Grace-to-Grace, not Grace-to-GPUNVIDIA "900 gigabytes per second (GB/s) of coherent interface" for the Grace-to-Hopper CPU-to-GPU link, one generation earlier

Assumed: Grace-to-GPU NVLink-C2C bandwidth, GB200/GB300, 900 GB/s. Why

power

LPDDR5X memory

See it in 3D ↗
Capacity
≈470 GB per CPU (rounded 17 TB ÷ 36)
Calc.

Inputs: NVIDIA "17 TB LPDDR5X" total ÷ 36 Grace CPUs

Calculated: The rounded 17 TB rack memory total divided among 36 CPUs gives about 470 GB per CPU. This quotient is approximate because the rack total is rounded. Method

power

Cold plates

See it in 3D ↗
Heat per GPU
≈1.2 kW
Reported

Sources: SemiAnalysis / InferenceX chip spec table: "TDP per chip: 1,200 W"

Plate height, as drawn
lifted off the chips
Assumed

Assumed: Tray cold plates, as drawn, lifted off the chips. Why

Plate, fitting and screw shapes
representative
Assumed

Assumed: Compute tray mechanical detail, representative. Why

power

NICs, DPU and SSDs

See it in 3D ↗
Scale-out
400 Gb/s per GPU
Spec

Sources: NVIDIA Hardware: each compute tray has "4x NVIDIA ConnectX-7 single port 400G OSFP NIC"

NIC mounting
mezzanine boards on Mirror Mezz connectors
Reported

Sources: SemiAnalysis "the ConnectX-7/8 ICs now sit directly on top of the Bianca board using a mezzanine board via Mirror Mezz connectors"

To the front cages
DensiLink flyover cables
Reported

Sources: SemiAnalysis "The electrical lanes are routed to the OSFP cages at the front of the chassis with DensiLink connectors from the mezzanine board."

Cable path and cage board
representative
Assumed

Assumed: NIC-to-cage cables as drawn, representative. Why

power

Optical modules

See it in 3D ↗
NIC module, NVIDIA single-port 400G DR4
9 W max
Spec

Sources: NVIDIA specifications table: "Maximum Power consumption (400G)", 9 W

power

NVLink connectors

See it in 3D ↗
Per GPU
NVLink 5, 1.8 TB/s
Spec

Sources: NVIDIA "The revolutionary 1.8 TB/s of bidirectional throughput per GPU"

data

NVLink connectors

See it in 3D ↗
Per GPU
1.8 TB/s
Spec

Sources: NVIDIA "The revolutionary 1.8 TB/s of bidirectional throughput per GPU"

data

NVLink-C2C

See it in 3D ↗
Bandwidth
900 GB/s
Assumed

Sources: NVIDIA "The Grace CPU Superchip is composed of two Grace CPUs connected coherently over NVIDIA NVLink-C2C at 900 GB/s" — Grace-to-Grace, not Grace-to-GPUNVIDIA "900 gigabytes per second (GB/s) of coherent interface" for the Grace-to-Hopper CPU-to-GPU link, one generation earlier

Assumed: Grace-to-GPU NVLink-C2C bandwidth, GB200/GB300, 900 GB/s. Why

data

ConnectX-7 NICs

See it in 3D ↗
NIC
ConnectX-7
Spec

Sources: NVIDIA Hardware: each compute tray has "4x NVIDIA ConnectX-7 single port 400G OSFP NIC"

Per GPU
400 Gb/s per GPU, NVIDIA’s reference design; ConnectX-8 upgrades to 800 Gb/s
Spec

Sources: NVIDIA Hardware: each compute tray has "4x NVIDIA ConnectX-7 single port 400G OSFP NIC"

PCIe retimers, CPU to GPU
none on the reference board
Reported

Sources: SemiAnalysis Compute Tray Diagrams & Cabling: "there is no longer any need for switches or retimers between the CPU and GPU on the reference design." Frontend Networking: with hyperscalers’ own backend NICs, "a dedicated PCIe switch from Broadcom / Astera Labs will be required"

data

Optical modules

See it in 3D ↗
NIC to cage
DensiLink flyover cables
Reported

Sources: SemiAnalysis "The electrical lanes are routed to the OSFP cages at the front of the chassis with DensiLink connectors from the mezzanine board."

Cable path
representative
Assumed

Assumed: NIC-to-cage cables as drawn, representative. Why

A 400G module, e.g.
NVIDIA single-port 400G OSFP, DR4
Spec

Sources: NVIDIA specifications: single-port 400G OSFP DR4 transceiver

Its power
9 W max
Spec

Sources: NVIDIA specifications table: "Maximum Power consumption (400G)", 9 W

data

BlueField DPU

See it in 3D ↗
DPUs per tray
2
Spec

Sources: NVIDIA Compute tray includes two BlueField-3 DPUs.

BlueField-3
up to 400 Gb/s
Spec

Sources: NVIDIA "BlueField-3 offers speeds up to 400 gigabits per second (Gb/s)"

Role
storage, security, tenant networking
Spec

Sources: NVIDIA describes BlueField-3 offloading "software-defined networking, storage, security, and management functions"

data

Blackwell GPUs

See it in 3D ↗
Links per GPU
NVLink, C2C, PCIe to the NIC
Reported

Sources: NVIDIA describes each compute tray’s GPU connecting out over NVLink to the spine, NVLink-C2C to its CPU, and PCIe to its NIC

heat

Cold plates

See it in 3D ↗
Heat per GPU
≈1.2 kW
Reported

Sources: SemiAnalysis / InferenceX chip spec table: "TDP per chip: 1,200 W"

Plate height, as drawn
lifted off the chips
Assumed

Assumed: Tray cold plates, as drawn, lifted off the chips. Why

heat

The heat source

See it in 3D ↗
Tray heat
≈6.5 kW
Calc.

Calculated: NVL72: tray power = (rack DC-bus power − NVLink scale-up power − busbar loss) ÷ 18 trays. DGX H100: tray power = rack power ÷ 4 servers. Method

heat

Fans

See it in 3D ↗
Air-cooled parts
peripheral electronics
Reported

Sources: NVIDIA Hardware: CPU/GPU liquid cooling and air-cooled peripheral components.

heat

Optical modules

See it in 3D ↗
NIC module, NVIDIA single-port 400G DR4
9 W max
Spec

Sources: NVIDIA specifications table: "Maximum Power consumption (400G)", 9 W

OSFP body
107.8 × 22.58 × 13.0 mm
Reported

Sources: AscentOptics form factor dimensions: the integrated-heat-sink body as 22.58 mm wide × 107.8 mm long × 13.0 mm tall, restating the OSFP MSA

heat

Quick disconnects

See it in 3D ↗
Per tray
one supply, one return per board
Reported

Sources: NVIDIA Hardware: the rear provides "inlets and outlets to the liquid cooling manifolds"

6GPU package & tokens11 cm view

power

Solder balls & substrate

See it in 3D ↗
Core voltage
≈0.7–0.9 V
Assumed

Assumed: GPU core voltage, ≈0.8 V. Why

Core current, die watts ÷ 0.8 V
≈1,300 A over several rails
Calc.

Calculated: Current = GPU power less its HBM share (the HBM stacks have their own supply) ÷ an assumed 0.8 V core voltage. Method

Stiffener ring and capacitors, as drawn
representative
Assumed

Assumed: GPU package stiffener ring and substrate surface, one flat ring and clustered capacitors; sizes and placement representative. Why

Substrate size, as drawn
≈8.4 cm square, representative
Assumed

Assumed: GPU package footprints and placement, as drawn, one package per accelerator at the tray and package levels; sizes and stack positions representative. Why

Balls and bumps, as drawn
fewer, larger and coarser than real
Assumed

Assumed: Solder balls and bumps, as drawn, fewer, larger and farther apart than real; representative. Why

power

Interposer

See it in 3D ↗
Packaging
TSMC CoWoS-L
Reported

Sources: Tom's Hardware headline and body: NVIDIA "shifts to CoWoS-L packaging for Blackwell GPU production"

Structure
bridge-based, not a monolithic silicon interposer
Reported

Sources: TechInsights body text: GB100 "utilizes the local area silicon interconnect (-L) variant of CoWoS instead of a monolithic silicon interposer (-S)", "NVIDIA’s first use of a bridge-based 2.5D integration technology"

Bridges and microbumps, as drawn
representative
Assumed

Assumed: Interposer bridges and microbumps, as drawn, representative count, size and placement. Why

power

Two GPU dies

See it in 3D ↗
Transistors
208 billion
Spec

Sources: NVIDIA Newsroom body text: "Packed with 208 billion transistors"

Die-to-die link
10 TB/s NV-HBI
Spec

Sources: NVIDIA Newsroom body text: "a 10 TB/second chip-to-chip link" joining the two reticle-limit dies

Process
TSMC 4NP
Spec

Sources: NVIDIA Newsroom body text: "manufactured using a custom-built 4NP TSMC process"

Die and stack sizes, as drawn
dies 26 × 33 mm, the reticle limit; stacks ≈11 × 10 mm
Assumed

Assumed: GPU package footprints and placement, as drawn, one package per accelerator at the tray and package levels; sizes and stack positions representative. Why

Floorplan shown
illustrative x-ray
Assumed

Assumed: GPU die surface and floorplan, as drawn, illustrative x-ray; not a real floorplan. Why

power

HBM3e stacks

See it in 3D ↗
Capacity
192 GB nominal; rack total implies ≈186 GB
Spec

Sources: NVIDIA specifications: "372 GB" HBM3e and "16 TB/s" per Grace Blackwell Superchip (2 GPUs)Micron specifications table: 24 GB at 8-Hi, >1.2 TB/s per stack

Bandwidth
8 TB/s
Spec

Sources: NVIDIA specifications: "372 GB" HBM3e and "16 TB/s" per Grace Blackwell Superchip (2 GPUs)Micron specifications table: 24 GB at 8-Hi, >1.2 TB/s per stack

Layers per stack
8
Spec

Sources: Micron specifications table: "8H, 12H HBM3E" at "24GB, 36GB"

Share of GPU power
≈13.2%
Assumed

Assumed: HBM memory, share of GPU package power, h100 10%, gb200 13.2%, gb300 15%, rubin (pre-launch) 16%. Why

Stack height, as drawn
about 3× real
Assumed

Assumed: HBM stacks, as drawn, height about 3x real; layer bands, top print and spacer representative. Why

Not drawn: Google Ironwood TPU
192 GB HBM at ≈7.37 TB/s per chip; Google publishes no chip power
Spec

Sources: Google Cloud documentation "Each chip is equipped with 192 GB of HBM" with "bandwidth of approximately 7.37 TB/s"; no power figure

Not drawn: AWS Trainium2
1.5 TB HBM3 at 46 TB/s per 16-chip instance; AWS publishes no chip power
Spec

Sources: AWS "16 Trainium2 chips"; "1.5 TB HBM3 with 46 terabytes per second (TBps) of memory bandwidth"

Not drawn: AMD Instinct MI355X
288 GB HBM3E at 8 TB/s; 1,400 W, liquid-cooled
Spec

Sources: AMD ROCm documentation "288GB of total capacity (36GB per stack)"; "up to 8 TB/s of peak theoretical memory bandwidth"; MI355X "1400W, direct liquid-cooled"

power

Tokens

See it in 3D ↗
Google, median Gemini text prompt
0.24 Wh, all-in
Spec

Sources: Google Cloud body text: "the median Gemini Apps text prompt uses 0.24 watt-hours (Wh) of energy"

LLaMA-65B on V100, 2023
≈3–4 J per token
Reported

Sources: Samsi et al., arXiv Section IV.C: "with length 512, we see that it takes about 3-4 Joules for a output token" — Figs. 6-7's 8/16/32-shard x-axis is the paper's V100 config (Table II: 65B needs 8 V100s but only 4 A100s)

GB200 vs H200
≈8× tokens per MW
Reported

Sources: SemiAnalysis / InferenceX body text: "single node H200 FP8 vs a GB200 NVL72 FP4 (without Multi Token Prediction)... ~8x improvement in token/s processed per all-in provisioned MW" (DeepSeek R1)

Model in the math
Llama 3.1 70B at FP8
Assumed

Sources: Meta Model Information: “a collection of pretrained and instruction tuned generative models in 8B, 70B and 405B sizes”NVIDIA, on Hugging Face Model Overview: quantizes “weights and activations of Meta-Llama-3.1-70B-Instruct to FP8 data type”, “reducing the number of bits per parameter from 16 to 8”

Assumed: The model in the token math, Llama 3.1 70B (70 billion parameters, dense) with FP8 weights, 1 byte per parameter. Why

Arithmetic per token
≈140 GFLOP (2 × 70B weights)
Calc.

Calculated: FLOPs per token ≈ 2 × parameter count: in one decode step every weight multiplies one number of the token’s vector and adds it to a running sum, a multiply and an add. For a dense model every weight takes part in every token. Attention over the context adds more that grows with its length; it is not counted. Method

Read from HBM per token
≈70 GB (70B × 1 byte)
Calc.

Calculated: Bytes per token ≈ parameter count × bytes per parameter at the model’s precision (1 byte at FP8): one stream reads every weight once per token. Weights only; the KV cache reads, which grow with the context, are not counted. Method

Arithmetic per byte read
≈2 FLOP; GB200 can do ≈625 (FP8)
Calc.

Calculated: For one stream: FLOPs per token ÷ bytes per token ≈ 2 FLOP per byte. The GPU’s balance: its dense peak at the same precision as the model ÷ its HBM bandwidth, in FLOPs per byte. When the first is far below the second, the step waits on memory, not math: decode is memory-bound. Batching many streams onto each read of the weights raises the first; that is why servers batch. Method

One stream on one GPU
≈8.8 ms reading weights vs ≈0.028 ms of math
Calc.

Calculated: Reading: bytes per token ÷ HBM bandwidth. Math: FLOPs per token ÷ dense peak at the model precision; omitted when that peak is unavailable. Both are ideal floors for a GPU that holds every weight and runs one stream; real steps take longer, and a model split across GPUs divides the reads between them. Method

data

HBM3e

See it in 3D ↗
Bandwidth
8 TB/s
Spec

Sources: NVIDIA specifications: "372 GB" HBM3e and "16 TB/s" per Grace Blackwell Superchip (2 GPUs)Micron specifications table: 24 GB at 8-Hi, >1.2 TB/s per stack

Drawn flow
lifted over the parts; the wires run beneath
Assumed

Assumed: HBM traffic, as drawn, schematic arcs over the parts; the real route is buried. Why

Activity on the dies
schematic, not a floorplan or a trace
Assumed

Assumed: Activity on the GPU dies, as drawn, schematic: not a floorplan, not a scheduling trace. Why

data

Die-to-die link

See it in 3D ↗
Bandwidth
10 TB/s
Spec

Sources: NVIDIA Newsroom body text: "a 10 TB/second chip-to-chip link" joining the two reticle-limit dies

data

NVLink SerDes

See it in 3D ↗
Per GPU
1.8 TB/s
Spec

Sources: NVIDIA product page: "1.8 TB/s of GPU-to-GPU interconnect", fifth-generation NVLinkNVIDIA Newsroom body text: "1.8TB/s bidirectional throughput per GPU"

data

Light on the package

See it in 3D ↗
GB200
electrical I/O only
Reported

Sources: NVIDIA hardware overview: "connected by NVLink through the NVLink passive copper cable cartridge backplane"

data

Tokens

See it in 3D ↗
Per token of text
a few bytes
Assumed

Assumed: Bytes per output token, text, a few bytes. Why

Model in the math
Llama 3.1 70B at FP8
Assumed

Sources: Meta Model Information: “a collection of pretrained and instruction tuned generative models in 8B, 70B and 405B sizes”NVIDIA, on Hugging Face Model Overview: quantizes “weights and activations of Meta-Llama-3.1-70B-Instruct to FP8 data type”, “reducing the number of bits per parameter from 16 to 8”

Assumed: The model in the token math, Llama 3.1 70B (70 billion parameters, dense) with FP8 weights, 1 byte per parameter. Why

Arithmetic per token
≈140 GFLOP (2 × 70B weights)
Calc.

Calculated: FLOPs per token ≈ 2 × parameter count: in one decode step every weight multiplies one number of the token’s vector and adds it to a running sum, a multiply and an add. For a dense model every weight takes part in every token. Attention over the context adds more that grows with its length; it is not counted. Method

Read from HBM per token
≈70 GB (70B × 1 byte)
Calc.

Calculated: Bytes per token ≈ parameter count × bytes per parameter at the model’s precision (1 byte at FP8): one stream reads every weight once per token. Weights only; the KV cache reads, which grow with the context, are not counted. Method

Arithmetic per byte read
≈2 FLOP; GB200 can do ≈625 (FP8)
Calc.

Calculated: For one stream: FLOPs per token ÷ bytes per token ≈ 2 FLOP per byte. The GPU’s balance: its dense peak at the same precision as the model ÷ its HBM bandwidth, in FLOPs per byte. When the first is far below the second, the step waits on memory, not math: decode is memory-bound. Batching many streams onto each read of the weights raises the first; that is why servers batch. Method

One stream on one GPU
≈8.8 ms reading weights vs ≈0.028 ms of math
Calc.

Calculated: Reading: bytes per token ÷ HBM bandwidth. Math: FLOPs per token ÷ dense peak at the model precision; omitted when that peak is unavailable. Both are ideal floors for a GPU that holds every weight and runs one stream; real steps take longer, and a model split across GPUs divides the reads between them. Method

heat

The dies

See it in 3D ↗
Package power
≈1,200 W
Reported

Sources: SemiAnalysis / InferenceX the chip’s specifications table: "1,200 W" per GPU

Compute dies, this model
≈1,042 W; the rest is HBM
Calc.

Calculated: Per GPU, the package power less its HBM share; for the campus, that times the GPU count, and as a share, that over the meter MW. The ledger’s rows, this, and the unallocated remainder add up to the meter power. It covers everything on the compute dies (computation, on-chip memory, communication, control and leakage); the model does not estimate the part that is useful arithmetic. Method

At full load, this operating point
≈75 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Throttle point
near ≈85 °C; NVIDIA publishes none
Assumed

Assumed: GPU throttle temperature, ≈85 °C. Why

heat

Heat flux

See it in 3D ↗
Die area, two dies
≈16 cm²
Reported

Sources: WCCFTech body text: "AI Superchip - 208 Billion Transistors (TSMC 4NP, >1600mm2)"

Average flux
≈65 W/cm²
Calc.

Calculated: Flux = GPU power less its HBM share ÷ die area. Hot spots run well above this average. Method

Hot spots, cooling trade press
500+ W/cm²
Reported

Sources: Alliance Chemical body text: "At 1,000 W TDP with an active die area of approximately 1.5–2 cm², the resulting heat flux at the cold-plate interface reaches 500–600 W/cm²" (B200)

heat

Thermal interface and lid

See it in 3D ↗
Lid, as drawn
for clarity; may ship bare-die
Assumed

Assumed: Layers between die and coolant, die → thermal interface → lid → thermal interface → cold plate or heat sink; H100: no lid. Why

Layers to coolant
die, interface, lid, interface, cold plate
Assumed

Assumed: Layers between die and coolant, die → thermal interface → lid → thermal interface → cold plate or heat sink; H100: no lid. Why

heat

HBM stacks

See it in 3D ↗
Share of package power
≈10–16%
Assumed

Assumed: HBM’s share of GPU package power, 10–16% assumed across the modeled generations. Why

HBM3e limit, Micron
105 °C
Spec

Sources: Micron specifications table: "Operating Temperature 0°C ≤ TOPER ≤ +105°C"

+Pluggable optics10.8 cm long

power

Edge connector

See it in 3D ↗
This scenario’s switch module
NVIDIA MMS4X00, 800G twin-port OSFP, 2 × DR4, finned top
Spec

Sources: NVIDIA overview: "two transceiver engines in a single OSFP form-factor plug creating 800Gb/s electrical to the switch and 2x400G optics"NVIDIA overview: "one, 17-Watt twin-port, finned-top 2xDR4 transceivers in each switch"

Edge connector
60 contacts: 32 high-speed (8 TX + 8 RX differential pairs), 4 control, 4 power, 20 ground
Spec

Sources: OSFP MSA Module Electrical Connector section, checked in the MSA text mirrored by fluxlight.com (the rev 5.22 PDF would not extract): "16 contacts for 8 differential pairs of high-speed transmit signals, 16 contacts for 8 differential pairs of high-speed receive signals, 4 contacts for low-speed control signals, 4 contacts for power and 20 contacts for ground"

Supply
3.3 V on 4 power contacts, up to 2.5 A each (3.25 A on OSFP1600)
Spec

Sources: OSFP MSA Rev 5.22 §15.6: "each power pin allows up to 2.5 Amps for a total of 10.0 Amps... For OSFP1600, each power pin allows up to 3.25 Amps"

Max power, NVIDIA MMS4X00
17 W
Spec

Sources: NVIDIA specifications table: "Maximum Power Dissipation", 17 WNVIDIA overview: "Twin-port single mode OSFP transceivers remain at 17 Watts for all configurations"

OSFP body
107.8 × 22.58 × 13.0 mm
Reported

Sources: AscentOptics form factor dimensions: the integrated-heat-sink body as 22.58 mm wide × 107.8 mm long × 13.0 mm tall, restating the OSFP MSA

Internals drawn
one layout for the 800G and 1.6T twin-ports
Assumed

Assumed: Which module the module level opens, the module at the end of the link you came from; the NIC side is one engine of the same drawing. Why

power

DSP

See it in 3D ↗
One DSP for both ports, e.g.
Credo Robin 800: 800G DR8 or 2 × 400G DR4, 8:8 at 100G/lane
Spec

Sources: Credo Technology Group product table, 100G/Lane DSPs: Robin 800, 8:8, 100G/lane, "800G SR8/DR8, 2x400G SR4/DR4/FR4"

DSPs in NVIDIA’s module
not published: one for both ports or one per port
Assumed

Assumed: One DSP or two in the twin-port module, one, as drawn; not published. Why

Max power, NVIDIA MMS4X00
17 W
Spec

Sources: NVIDIA specifications table: "Maximum Power Dissipation", 17 WNVIDIA overview: "Twin-port single mode OSFP transceivers remain at 17 Watts for all configurations"

Inside a twin-port module
two independent 400G ports
Spec

Sources: NVIDIA overview: "two transceiver engines in a single OSFP form-factor plug creating 800Gb/s electrical to the switch and 2x400G optics"

Layout
representative, not one product
Assumed

Assumed: Layout inside the pluggable module, representative. Why

Module power at 200G/lane, Semtech figures
LRO ≈16 W · full DSP 23–25 W · LPO target ≈10 W
Vendor

Sources: Semtech power section: "currently 23 to 25 watts per module... LRO/RTLR (Half-Retimed) Modules: currently approximately 16 watts... LPO Modules: targeting approximately 10 watts"

Compared with: Semtech’s comparison of current full-DSP and LRO modules with its LPO target; not a measured saving for one module

Credo Dove 850 (800G LRO DSP), Credo figure
up to 50% less DSP power
Vendor

Sources: Credo Technology Group release body: "In an LRO transceiver or Active Optical Cable (AOC), only the transmit path from the electrical input to the optical line side output includes a DSP for signal retiming and equalization"; "Dove 850 reduces DSP power by up to 50%"

Compared with: a baseline Credo’s release does not name

Transmit-only DSP, Marvell
Ara T: 8 × 200G TRO; sampling from Q1 2026
Spec

Sources: Marvell body text: "Ara T, the first 8x200G transmit-retimed optics (TRO) DSP"; "Ara X, Ara T, Petra and Aquila M DSPs are sampling to customers beginning in Q1 2026."

LRO as drawn
same package, receive half hatched; representative
Assumed

Assumed: The half-retimed (LRO) module, as drawn, representative. Why

power

Power conversion

See it in 3D ↗
Layout
representative, not one product
Assumed

Assumed: Layout inside the pluggable module, representative. Why

power

Driver

See it in 3D ↗
LPO keeps
a linear driver (TX), the TIA (RX)
Reported

Sources: Flexoptix "LPO technology removes the DSP with complex CDR functionality and keeps only high-linearity analog components such as drivers, lasers, photodiodes, and TIAs"

Layout
representative, not one product
Assumed

Assumed: Layout inside the pluggable module, representative. Why

power

Receive amplifier

See it in 3D ↗
Layout
representative, not one product
Assumed

Assumed: Layout inside the pluggable module, representative. Why

power

Lasers

See it in 3D ↗
Design drawn
silicon photonics, separate lasers
Assumed

Assumed: A silicon photonics module, one design of several. Why

Short-reach DR8
directly modulated lasers are also used
Spec

Sources: Juniper Networks "DMLs are used for single-mode optics such as DR8"

NVIDIA’s 800G twin-port (MPO-16)
1310 nm EML lasers
Spec

Sources: NVIDIA product page: "1310nm EML laser"

data

Edge connector

See it in 3D ↗
Host lanes
8 × 100G electrical, each way
Spec

Sources: NVIDIA overview: "two transceiver engines in a single OSFP form-factor plug creating 800Gb/s electrical to the switch and 2x400G optics"OSFP MSA Module Electrical Connector section: "16 contacts for 8 differential pairs of high-speed transmit signals, 16 contacts for 8 differential pairs of high-speed receive signals"

Line rate per lane
53.125 GBd, 106.25 Gb/s (PAM4, 100G payload), on the electrical and the optical side
Calc.

Calculated: Eight lanes each way. A 100G PAM4 lane runs 53.125 GBd × 2 bits = 106.25 Gb/s on the wire (100 Gb/s of payload plus FEC and framing); a 200G PAM4 lane runs 106.25 GBd × 2 bits = 212.5 Gb/s. Eight lanes × 100G = 800G (2 × 400G DR4); eight × 200G = 1.6T (2 × 800G DR4). The host electrical lanes run at the same rate as the optical lanes. Method

Edge connector
60 contacts: 32 high-speed (8 TX + 8 RX differential pairs), 4 control, 4 power, 20 ground
Spec

Sources: OSFP MSA Module Electrical Connector section, checked in the MSA text mirrored by fluxlight.com (the rev 5.22 PDF would not extract): "16 contacts for 8 differential pairs of high-speed transmit signals, 16 contacts for 8 differential pairs of high-speed receive signals, 4 contacts for low-speed control signals, 4 contacts for power and 20 contacts for ground"

This scenario’s switch module
NVIDIA MMS4X00, 800G twin-port OSFP, 2 × DR4, finned top
Spec

Sources: NVIDIA overview: "two transceiver engines in a single OSFP form-factor plug creating 800Gb/s electrical to the switch and 2x400G optics"NVIDIA overview: "one, 17-Watt twin-port, finned-top 2xDR4 transceivers in each switch"

Internals drawn
one layout for the 800G and 1.6T twin-ports
Assumed

Assumed: Which module the module level opens, the module at the end of the link you came from; the NIC side is one engine of the same drawing. Why

data

DSP

See it in 3D ↗
One DSP for both ports, e.g.
Credo Robin 800: 800G DR8 or 2 × 400G DR4, 8:8 at 100G/lane
Spec

Sources: Credo Technology Group product table, 100G/Lane DSPs: Robin 800, 8:8, 100G/lane, "800G SR8/DR8, 2x400G SR4/DR4/FR4"

DSPs in NVIDIA’s module
not published: one for both ports or one per port
Assumed

Assumed: One DSP or two in the twin-port module, one, as drawn; not published. Why

Optical lanes, MMS4X00
8 × 100G PAM4 each way, 2 × DR4; 53.125 GBd
Spec

Sources: NVIDIA specifications table: "Signaling Rate per Lane", 53.125 GBdNVIDIA overview: "The line rate is 400Gb/s for both 400GbE Ethernet and NDR InfiniBand based on the 100G-PAM4 modulation"; "two 4-channel MPO-12/APC optical connectors"

What it does
retiming, equalization, error correction
Spec

Sources: Juniper Networks "The CDR is responsible for re-timing incoming data to reduce jitter. The DSP handles functions like equalization, error correction, and other signal processing tasks"

Example 800G DSPs, 8 × 100G
Credo Robin and Dove families, 8:8
Spec

Sources: Credo Technology Group product table, 100G/Lane DSPs: Robin 800/802/850/852 and Dove 800/850, each 8:8 at 100G/lane for 800G DR8 and 2x400G DR4

Half-retimed (LRO, RTLR)
DSP on transmit only; the host’s processing replaces a receive DSP
Spec

Sources: OIF announcement body: RTLR "eliminates the need for a receive DSP in the optical module and leveraging signal processing already available in the host device"LPO MSA FAQ answer: "Links that use a linear receiver and a retimed transmitter (i.e., half-linear or half-retimed modules) are referred to as LRO modules and will be designated with “-LRO” on any optical specifications."

OIF RTLR interface
OIF-EEI-112G-RTLR 1.0, 112 Gb/s lanes, published 11/18/2025
Spec

Sources: OIF title and body: "112 Gb/s Retimed Transmitter Linear Receiver (RTLR)"; "published on November 18, 2025, with the document designation OIF-EEI-112G-RTLR version 01.0"

Transmit-only DSP, Marvell
Ara T: 8 × 200G TRO; sampling from Q1 2026
Spec

Sources: Marvell body text: "Ara T, the first 8x200G transmit-retimed optics (TRO) DSP"; "Ara X, Ara T, Petra and Aquila M DSPs are sampling to customers beginning in Q1 2026."

LRO as drawn
same package, receive half hatched; representative
Assumed

Assumed: The half-retimed (LRO) module, as drawn, representative. Why

data

Driver

See it in 3D ↗
Driver channels as drawn
8 each way: 4 per DR4 engine, one per 100G lane
Assumed

Assumed: Layout inside the pluggable module, representative. Why

LPO keeps
a linear driver (TX), the TIA (RX)
Reported

Sources: Flexoptix "LPO technology removes the DSP with complex CDR functionality and keeps only high-linearity analog components such as drivers, lasers, photodiodes, and TIAs"

Layout
representative, not one product
Assumed

Assumed: Layout inside the pluggable module, representative. Why

data

Lasers

See it in 3D ↗
Lasers as drawn
4 continuous-wave lasers, each split to 2 modulators (8 lanes)
Assumed

Assumed: A silicon photonics module, one design of several. Why

Design drawn
silicon photonics, separate lasers
Assumed

Assumed: A silicon photonics module, one design of several. Why

data

Modulators

See it in 3D ↗
Modulators as drawn
8 each way: 4 per DR4 engine, one per 100G lane
Assumed

Assumed: Layout inside the pluggable module, representative. Why

Kind
Mach-Zehnder, a common silicon-photonics modulator design
Reported

Sources: Optica (OFC 2019, Th4A.6) abstract: single-lane 200G optical interconnects demonstrated with a silicon photonic Mach-Zehnder modulator (the earlier "commercial mainstream" wording came from a trade newsletter whose link is dead and is not repeated here)

Design drawn
silicon photonics, separate lasers
Assumed

Assumed: A silicon photonics module, one design of several. Why

data

Fiber connectors

See it in 3D ↗
Optical lanes, MMS4X00
8 × 100G PAM4 each way, 2 × DR4; 53.125 GBd
Spec

Sources: NVIDIA specifications table: "Signaling Rate per Lane", 53.125 GBdNVIDIA overview: "The line rate is 400Gb/s for both 400GbE Ethernet and NDR InfiniBand based on the 100G-PAM4 modulation"; "two 4-channel MPO-12/APC optical connectors"

Reach, MMS4X00-NS
up to 100 m; a 500 m version (-NM) exists
Spec

Sources: NVIDIA overview: "two straight fibers cables (MFP7E30-Nxxx) up to 100-meters"; "A 500-meter version (-NM) is also available for longer switch-to-switch reaches"

Connectors
dual MPO-12
Spec

Sources: NVIDIA overview: the 800G twin-port uses "two 4-channel MPO-12/APC optical connectors"Juniper Networks optical interface: DR8 / 2×DR4 modules use dual MPO-12/APC connectorsNVIDIA optical interface: two MPO-12/APC connectors on the twin-port DR8 module

Fibers lit
16: 4 out and 4 in per connector
Spec

Sources: NVIDIA optical interface: a twin-port DR8 module on two MPO-12/APC connectors, eight fibers active on each (four transmit, four receive)

Receptacles
two, ferrules vertical, 10.0 mm apart
Spec

Sources: OSFP MSA Rev 5.22 Fig 14-48, "Optical receptacle and channel orientation for Dual MPO connector": two vertical fiber columns, "10.0±0.6mm" apart (PDF p.154)NVIDIA Option 2 drawing, "TRANCEIVEIR PORT VIEW": two vertical ferrules side by side (PDF p.16)

Splitter cables, 800G twin-port generation
one MPO-12 port split into two 2-channel ends
Spec

Sources: NVIDIA Introduction: "a single mode, 4-channel-to-two 2-channel splitter fiber" cable; the 4-channel end goes into a twin-port 800G OSFP, the 2-channel ends into two single-port 400G transceivers that run at 200GNVIDIA Introduction: the multimode version, a "4-channel-to-two 2-channel splitter fiber cable"

Splitter fibers
8 lit at the 4-channel end, shared between the two ends
Spec

Sources: NVIDIA Introduction: the MPO-12/APC "uses 8 active fibers to transmit light and 4 inactive fibers as strength members"; §2.1: "These cables have 8 individual fibers, 4 in each direction"NVIDIA Introduction: the 2-channel ends run "with only 2 fibers" at 200G

Nose, saddle and receptacle bodies
shape representative
Assumed

Assumed: Exterior details of the pluggable module, representative. Why

data

Photodiodes

See it in 3D ↗
Line rate per lane
53.125 GBd, 106.25 Gb/s (PAM4, 100G payload), on the electrical and the optical side
Calc.

Calculated: Eight lanes each way. A 100G PAM4 lane runs 53.125 GBd × 2 bits = 106.25 Gb/s on the wire (100 Gb/s of payload plus FEC and framing); a 200G PAM4 lane runs 106.25 GBd × 2 bits = 212.5 Gb/s. Eight lanes × 100G = 800G (2 × 400G DR4); eight × 200G = 1.6T (2 × 800G DR4). The host electrical lanes run at the same rate as the optical lanes. Method

Receive, per module
8 photodiodes, one per lane
Spec

Sources: Juniper Networks receive section: eight photodetectors and eight TIAs, one pair per optical lane

Layout
representative, not one product
Assumed

Assumed: Layout inside the pluggable module, representative. Why

data

Receive amplifier

See it in 3D ↗
TIA channels as drawn
8 each way: 4 per DR4 engine, one per 100G lane
Assumed

Assumed: Layout inside the pluggable module, representative. Why

What it does
photodiode current in, voltage out
Spec

Sources: Juniper Networks receive section: the TIA "converts and amplifies the electrical current from the photodiode into an electrical voltage level"

LPO keeps
a linear driver (TX), the TIA (RX)
Reported

Sources: Flexoptix "LPO technology removes the DSP with complex CDR functionality and keeps only high-linearity analog components such as drivers, lasers, photodiodes, and TIAs"

In an LRO module
linear receive: the TIA output goes to the host
Spec

Sources: OIF announcement body: RTLR "eliminates the need for a receive DSP in the optical module and leveraging signal processing already available in the host device"LPO MSA FAQ answer: "Links that use a linear receiver and a retimed transmitter (i.e., half-linear or half-retimed modules) are referred to as LRO modules and will be designated with “-LRO” on any optical specifications."

Linear receive, Semtech on LPO
the host SerDes is calibrated to supply the equalization
Reported

Sources: Semtech host requirements: "Not necessarily new ASICs, but the host SerDes must be correctly calibrated to deliver the equalization LPO modules expect."

heat

DSP

See it in 3D ↗
Max power, NVIDIA MMS4X00
17 W
Spec

Sources: NVIDIA specifications table: "Maximum Power Dissipation", 17 WNVIDIA overview: "Twin-port single mode OSFP transceivers remain at 17 Watts for all configurations"

Module power at 200G/lane, Semtech figures
LRO ≈16 W · full DSP 23–25 W · LPO target ≈10 W
Vendor

Sources: Semtech power section: "currently 23 to 25 watts per module... LRO/RTLR (Half-Retimed) Modules: currently approximately 16 watts... LPO Modules: targeting approximately 10 watts"

Compared with: Semtech’s comparison of current full-DSP and LRO modules with its LPO target; not a measured saving for one module

Credo Dove 850 (800G LRO DSP), Credo figure
up to 50% less DSP power
Vendor

Sources: Credo Technology Group release body: "In an LRO transceiver or Active Optical Cable (AOC), only the transmit path from the electrical input to the optical line side output includes a DSP for signal retiming and equalization"; "Dove 850 reduces DSP power by up to 50%"

Compared with: a baseline Credo’s release does not name

Heat drawn
log scale, each 10× in power 5× the motion; none under 0.5 W
Assumed

Assumed: How much heat each part is drawn with, log scale: each 10× in power is 5× the motion; nothing under 0.5 W. Why

heat

Shell and fins

See it in 3D ↗
OSFP body
107.8 × 22.58 × 13.0 mm
Reported

Sources: AscentOptics form factor dimensions: the integrated-heat-sink body as 22.58 mm wide × 107.8 mm long × 13.0 mm tall, restating the OSFP MSA

Lid print
OSFP 800G 2xDR4, as on this scenario’s switches; the LPO view adds LPO
Assumed

Assumed: Printed lid labels on pluggable modules, module class, reach and no part numbers. Why

Max power, NVIDIA MMS4X00
17 W
Spec

Sources: NVIDIA specifications table: "Maximum Power Dissipation", 17 WNVIDIA overview: "Twin-port single mode OSFP transceivers remain at 17 Watts for all configurations"

Layout
representative, not one product
Assumed

Assumed: Layout inside the pluggable module, representative. Why

+Inside the CPO packagepackage, representative

power

Switch ASIC

See it in 3D ↗
Q3450 switch, NVIDIA figure
3.95 kW
Vendor

Sources: Lambda power comparison: "NVIDIA Photonics CPO switch: 3.95 kW. Standard switch: 7.0 kW."

Compared with: a pluggable-optics equivalent switch at 7.0 kW

Layout
representative; counts are NVIDIA’s
Assumed

Assumed: Layout of the CPO switch package, representative. Why

power

Optical engines

See it in 3D ↗
Engine stack, as reported
electronic die above photonic die, TSMC SoIC-X
Reported

Sources: TSMC PDF page 2, Silicon Photonics Integration: COUPE uses SoIC-X to stack an electrical die on top of a photonic die; no process-node sizes asserted here.NVIDIA / Hot Chips PDF page 14: Electronic IC / Photonic IC stack and COUPE microlens surface coupling; does not disclose the illustrated driver/TIA floorplan.

Per switch chip
6 subassemblies × 3 engines = 18
Spec

Sources: NVIDIA Quantum-X section: "Each Quantum-X switch ASIC delivers 28.8Tbps full duplex bandwidth, harnessed through six high-capacity optical subassemblies", each with three COUPE-based optical engines

Per engine
1.6 Tb/s each way, 8 × 200G
Spec

Sources: NVIDIA Quantum-X section: engines "each supporting 1.6Tbps transmit and 1.6Tbps receive", with "eight 200Gbps PAM4 lanes for both transmit and receive"

Engine and package rate
1.6 Tb/s per engine = 8 lanes × 200 Gb/s each way; 18 engines × 1.6 Tb/s = 28.8 Tb/s
Calc.

Calculated: 8 lanes × 200 Gb/s = 1.6 Tb/s per engine each way; 18 engines × 1.6 Tb/s = 28.8 Tb/s per switch chip. Method

Optics power per 800G, SemiAnalysis estimate
≈4–5 W for the Q3450’s engines and lasers, against ≈16–17 W for an 800G DR4 transceiver
Reported

Sources: SemiAnalysis (Dylan Patel, Daniel Nishball, Myron Xie and others) power section: "While an 800G DR4 optical transceiver consumes about 16-17W, we estimate that the optical engine together with external laser sources used in Nvidia’s Q3450 CPO switch consume about 4-5W per 800G of bandwidth, a 73% reduction in power."

power

External laser sources

See it in 3D ↗
Laser modules
18 per Q3450 switch, 8 lasers each
Reported

Sources: NVIDIA external laser source section: "Each ELS contains eight high-quality lasers"Lambda "The switch supports 18 removable external light-source modules, each feeding eight MPO ports."

Transmit lanes per module
32
Spec

Sources: NVIDIA external laser source section: "An individual ELS is capable of powering 32 of the Quantum-X switch’s 576 transmit lanes"

power

Switch ASIC

See it in 3D ↗
Package, Broadcom
Tomahawk 5 switch chip and eight 6.4 Tb/s optical engines: 51.2 Tb/s
Spec

Sources: Broadcom opening: "The product integrates eight silicon photonics based 6.4-Tbps optical engines with Broadcom’s best-in-class StrataXGS® Tomahawk®5 switch chip."SemiAnalysis Broadcom 6.4T Optical Engine - Paper 23.4: "One CPO package consists of eight 6.4T OEs, each with a PIC and an EIC, on TSMC's N7 process."

Optics power, Broadcom figure
70% lower than pluggable transceivers
Vendor

Sources: Broadcom opening: "Bailly enables the optical interconnect to operate at 70% lower power consumption and delivers an 8x improvement in silicon area efficiency as compared to pluggable transceiver solutions."

Compared with: pluggable transceiver solutions

Layout
representative; counts are Broadcom’s
Assumed

Assumed: The Mach-Zehnder package, as drawn, representative; counts Broadcom’s, proportions from its package images. Why

power

Optical engines

See it in 3D ↗
Engine, Broadcom
6.4T-FR4 engines with a Broadcom Fiber Connector (BFC)
Spec

Sources: Broadcom product highlights: "Broadcom 6.4T-FR4 Bailly SCIP optical engines with Broadcom Fiber Connector (BFC) for CPO systems"

Broadcom’s engine, as reported
one 7 nm electronic die per engine: all 64 transmit and receive channels, MZM drivers and TIAs
Reported

Sources: 光芯 (光芯前沿 column), Tencent Cloud developer community write-up of Broadcom’s ISSCC 2026 paper 23.4, chip section: "该CPO ASIC采用7nm FinFET工艺制造(图23.4.7),单片裸片集成64条发射/接收通道、TXPLL、辅助公共电路及ADC/DAC组件;6.4Tb/s光引擎通过7nm ASIC与PIC硅片的3D封装实现" (the CPO ASIC is made in a 7 nm FinFET process; a single die integrates the 64 transmit/receive channels, the transmit PLL, shared support circuits and ADC/DAC blocks; the 6.4 Tb/s optical engine is this 7 nm ASIC 3D-packaged with the photonic chip). Its transmit path ends in "MZM驱动器" (MZM drivers); its receive path is a "直接驱动TIA" (direct-drive TIA). This is the 51.2T (Tomahawk 5) generation at 106.25 Gb/s per lane; the paper itself was not opened.SemiAnalysis Broadcom 6.4T Optical Engine - Paper 23.4: "One CPO package consists of eight 6.4T OEs, each with a PIC and an EIC, on TSMC's N7 process."

Broadcom’s engine, lanes
64 × 106.25 Gb/s PAM4, 6.4 Tb/s, for 51.2T (Tomahawk 5) switching
Reported

Sources: 光芯 (光芯前沿 column), Tencent Cloud developer community opening: "发布了一款用于51.2T交换的6.4Tb/s CPO专用ASIC,该7nm ASIC与硅光芯片PIC进行3D封装,能量效率达到4.2pJ/b。" (it presented a 6.4 Tb/s CPO ASIC for 51.2T switching, the 7 nm ASIC 3D-packaged with the silicon-photonics PIC); conclusion: "该6.4Tb/s CPO ASIC通过64条106.25Gb/s PAM-4通道实现最高集成度和吞吐量" (the 6.4 Tb/s CPO ASIC reaches its integration and throughput through 64 lanes of 106.25 Gb/s PAM-4). The 4.2 pJ/b figure is not used here: the same post’s title says 4.5 pJ/b.SemiAnalysis Broadcom 6.4T Optical Engine - Paper 23.4: "The optical engines were tested in a Tomahawk 5 51.2T CPO system."

Engine and package rate
6.4 Tb/s per engine = 16 × 400G FR4 ports = 64 lanes × 100 Gb/s each way (106.25 Gb/s line rate); 8 engines × 6.4 Tb/s = 51.2 Tb/s
Calc.

Calculated: 16 × 400G FR4 ports = 6.4 Tb/s per engine; each FR4 port is four 100 Gb/s wavelength lanes, so 16 × 4 = 64 lanes × 100 Gb/s each way (reported as 64 × 106.25 Gb/s PAM4 line rate); 8 engines × 6.4 Tb/s = 51.2 Tb/s. Method

Layout
representative; counts are Broadcom’s
Assumed

Assumed: The Mach-Zehnder package, as drawn, representative; counts Broadcom’s, proportions from its package images. Why

power

Remote laser modules

See it in 3D ↗
Lasers, Broadcom reference system
remote laser modules, field-replaceable
Spec

Sources: Broadcom product highlights: "System design compatible to support multiple remote laser modules (RLM) for field replaceability"

Layout
representative; counts are Broadcom’s
Assumed

Assumed: The Mach-Zehnder package, as drawn, representative; counts Broadcom’s, proportions from its package images. Why

power

A CPO switch, shipping now

See it in 3D ↗
NVIDIA Spectrum-X Ethernet Photonics
in production (NVIDIA, 05/31/2026); shipping to select partners (TrendForce, 07/27/2026)
Reported

Sources: NVIDIA summary line: "Vera Rubin introduces NVIDIA Spectrum-X Ethernet Photonics — now in production — combining co-packaged optics with Spectrum-X switching to enable million-GPU AI factories."TrendForce first paragraph: "NVIDIA has begun shipping its next-generation Spectrum-X CPO switch to select partners."StorageReview opening: "NVIDIA has moved Spectrum-X Ethernet Photonics, its co-packaged optics (CPO) Ethernet switch platform, into full production."

NVIDIA Quantum-X InfiniBand Photonics
Q3450-LD, 144 × 800G ports; engineering samples at Lambda, 06/2026
Reported

Sources: NVIDIA Quantum-X section: "The NVIDIA Quantum-X800 InfiniBand platform includes CPO-based switches, including the Q3450-LD with 144 ports of 800 gigabits-per-second (Gb/s) InfiniBand."Lambda opening: "Lambda is taking an early look at co-packaged optics, starting with the NVIDIA Quantum-X InfiniBand Photonics Q3450-LD switch"; timing: "Engineering samples are a chance to work through rack design, cooling, power delivery, fiber routing, and the installation process with the vendor while the product is still becoming real."

Broadcom TH6-Davisson
102.4 Tb/s, 16 × 6.4T engines; announced 10/08/2025 as shipping, sampled to early-access customers
Spec

Sources: Broadcom opening: "now shipping Tomahawk® 6 Davisson (TH6-Davisson), the company's third-generation Co-Packaged Optics (CPO) Ethernet switch"; product highlights: "16 x 6.4 Tbps Davisson DR Optical Engines"; availability: "Broadcom is currently sampling the TH6-Davisson BCM78919 device to its early access customers and partners."

Broadcom TH5-Bailly, 51.2 Tb/s
in volume manufacturing with Delta and Micas Networks (TrendForce, 07/27/2026)
Reported

Sources: TrendForce Broadcom paragraph: "Broadcom's 51.2T Bailly CPO switch has entered volume manufacturing with support from Delta Electronics and Micas Networks."

Bailly in Meta’s lab, Broadcom figure
one million flap-free 400G port-hours
Vendor

Sources: Broadcom opening: "one million cumulative 400G equivalent port device hours of flap-free CPO operation at Meta"; "The absence of link flaps in Meta’s high-temperature lab characterization environment"; footnote: ECOC 2025 talk by Siamak Amiralizadeh, 09/30/2025

Compared with: pluggable modules: "Compared to pluggable module solutions, the data highlights that CPO reduces optics power by 65 percent and also demonstrates higher link reliability."

power

Next: CPO on the GPU

See it in 3D ↗
Marvell compute tray, a concept
4 XPUs, each with four 6.4T light engines: 102.4 Tb/s on 1,024 fibers, in 1U
Vendor

Sources: Marvell (Chris McCormick) The CPO Server Tray: "a conceptualized AI compute tray with CPO developed with products from SENKO Advanced Components and Marvell. The design contains room for four XPUs and up to 102.4 Tbps of bandwidth delivered through 1024 optical fibers, all in a 1U tray."; "Each XPU is connected to four Marvell 6.4T light engines for opto-electric conversions." The post later puts the total, with laser fibers, at 1,152.

Compared with: copper scale-up, in Marvell’s words: "The density and reach enabled by CPO opens the door to scale-up domains far beyond what is possible with copper alone."

Marvell CPO switch, a reference design
switch chip + 16 light engines; 16 laser modules at the faceplate; 1OU with a cold plate
Vendor

Sources: Marvell (Chris McCormick) Heat and Space: "Marvell, SENKO, Jabil, and Mikros Technologies also recently unveiled a reference design of a data center CPO switch."; "surrounded by 16 light engine tiles and 1,152 fibers (128 laser fibers and 1,024 data fibers). The light engines are driven by 16 laser modules, which are connected to the faceplate for better serviceability."; "maintains a system height of 1OU"

Compared with: air cooling: "conventional air cooling would require a chassis that is two to three times thicker"

OCI MSA, founded 03/12/2026
an open optical scale-up spec: AMD, Broadcom, Meta, Microsoft, NVIDIA, OpenAI
Reported

Sources: Broadcom (for the OCI MSA) opening: "The Optical Compute Interconnect (OCI) Multi-Source Agreement (MSA) group today announced its formation, led by founding members AMD, Broadcom, Meta, Microsoft, NVIDIA and OpenAI."; "OCI will enable migration from copper-based to optical-based scale-up architectures"; "Support for pluggable, on-board, and co-packaged optics (CPO)."

Ayar Labs, with GUC
optical engines on an XPU package: over 100 Tb/s, a package design
Vendor

Sources: Ayar Labs "The new XPU multi-chip package (MCP) design replaces traditional electrical interconnects with Ayar Labs’ optical engines attached directly to the MCP organic substrate. This architecture enables more than 100 Tbps full-duplex optical interface from the XPU package, more than an order of magnitude improvement over current XPUs."

Compared with: current XPUs, which Ayar Labs puts more than ten times lower

Lightmatter Passage L200
32 and 64 Tb/s 3D CPO engines for XPUs and switches; announced 03/31/2025 for 2026
Vendor

Sources: Lightmatter "The L200 3D CPO family includes both 32 Tbps and 64 Tbps versions, representing a 5 to 10x improvement over existing solutions."; "Available in 2026, Lightmatter’s L200 and L200X 3D CPO chips are designed to accelerate time to market and performance of next generation XPUs and switches"

Compared with: existing interconnect solutions, which Lightmatter puts 5 to 10 times lower

Celestial AI
bought by Marvell, 02/02/2026, for optical scale-up
Reported

Sources: Marvell (SEC Form 8-K exhibit 99.1) opening: "completed its previously announced acquisition of Celestial AI, a pioneer in optical interconnect technology for scale-up connectivity"; outlook: initial revenue "in the second half of fiscal 2028"

Ranovus Odin
one monolithic die: micro-ring modulators, photodetectors, RF drivers, TIAs, control logic
Spec

Sources: Ranovus release body: "a groundbreaking monolithic Electronic and Photonic Integrated Circuit (EPIC) platform"; "Featuring monolithic integration of silicon photonics, RF drivers, transimpedance amplifiers (TIA), and control logic"; "Key innovations include patented silicon photonics-based Micro Ring Modulators (MRM), internal isolator-free laser sources, and high-performance photodetectors."

Ayar Labs TeraPHY
micro-ring based, electronics and photonics on one die; later designs may split them
Reported

Sources: Ayar Labs, release reprinted by HPCwire opening: "Ayar Labs has successfully demonstrated its patented monolithic electronic/photonic solution on GLOBALFOUNDRIES (GF) next generation photonics solution based on its 45nm platform."; CEO quote: "Ayar Labs has been perfecting our micro-ring based monolithic electronic/photonic solution for nearly a decade."SemiAnalysis (Dylan Patel, Daniel Nishball, Myron Xie and others) TSMC COUPE section: "Ayar Labs, who has previously relied on Global Foundries’ Fotonix platform for monolithic optical engines, now also has COUPE on their roadmap."Gazettabyte From monolithic optics to modular chiplets: "This [TeraPHY optical I/O chiplet] architecture lets us move seamlessly through different foundry processes"; "The 8Tbps TeraPHY device is built using GlobalFoundries’ 45SPCLO 45nm silicon-photonics process"; "But the design can also be migrated to TSMC’s more advanced CMOS nodes for the electrical IC while benefiting from TSMC’s silicon photonics and packaging flows."

data

Switch ASIC

See it in 3D ↗
Per switch chip
28.8 Tb/s each way
Spec

Sources: NVIDIA Quantum-X section: 28.8 Tb/s full-duplex per switch ASIC

Per port, NVIDIA figure
as low as 9 W
Vendor

Sources: NVIDIA body text: 30 W per port with pluggable transceivers, "as low as 9 W" per port with co-packaged optics

Compared with: 30 W per port with pluggable transceivers

data

Package traces

See it in 3D ↗
Electrical loss, NVIDIA figures
≈4 dB, from 20–22 dB
Vendor

Sources: NVIDIA body text: 22 dB for the pluggable path against approximately 4 dB with co-packaged opticsLambda "Signal loss drops from roughly 20dB to 4dB"

Compared with: the path to a pluggable module, 20–22 dB

Layout
representative; counts are NVIDIA’s
Assumed

Assumed: Layout of the CPO switch package, representative. Why

data

Electronic chip

See it in 3D ↗
Engine stack, as reported
electronic die above photonic die, TSMC SoIC-X
Reported

Sources: TSMC PDF page 2, Silicon Photonics Integration: COUPE uses SoIC-X to stack an electrical die on top of a photonic die; no process-node sizes asserted here.NVIDIA / Hot Chips PDF page 14: Electronic IC / Photonic IC stack and COUPE microlens surface coupling; does not disclose the illustrated driver/TIA floorplan.

Engine partition, as reported
drivers, TIAs and control logic on the electronic die; modulators and detectors on the photonic die (COUPE)
Reported

Sources: SemiAnalysis (Dylan Patel, Daniel Nishball, Myron Xie and others) Quantum-X Photonics section: "Each optical engine integrates a Photonic Integrated Circuit (PIC) built on a mature N65 process node, and an Electronic Integrated Circuit (EIC) fabricated on an advanced N6 node. The PIC leverages the older node because it contains optical components such as modulators, waveguides, and detectors—devices that do not benefit from scaling, and often perform better at larger geometries. In contrast, the EIC includes drivers, TIAs, and control logic, which benefit significantly from higher transistor density and improved power efficiency enabled by advanced nodes." TSMC COUPE section, Die fabrication: "The EIC is manufactured on the N7 node, integrating high-speed optical modulator drivers and TIAs." A report on the COUPE engines, not NVIDIA’s published floorplan.

Broadcom’s engine, as reported
one 7 nm electronic die per engine: all 64 transmit and receive channels, MZM drivers and TIAs
Reported

Sources: 光芯 (光芯前沿 column), Tencent Cloud developer community write-up of Broadcom’s ISSCC 2026 paper 23.4, chip section: "该CPO ASIC采用7nm FinFET工艺制造(图23.4.7),单片裸片集成64条发射/接收通道、TXPLL、辅助公共电路及ADC/DAC组件;6.4Tb/s光引擎通过7nm ASIC与PIC硅片的3D封装实现" (the CPO ASIC is made in a 7 nm FinFET process; a single die integrates the 64 transmit/receive channels, the transmit PLL, shared support circuits and ADC/DAC blocks; the 6.4 Tb/s optical engine is this 7 nm ASIC 3D-packaged with the photonic chip). Its transmit path ends in "MZM驱动器" (MZM drivers); its receive path is a "直接驱动TIA" (direct-drive TIA). This is the 51.2T (Tomahawk 5) generation at 106.25 Gb/s per lane; the paper itself was not opened.SemiAnalysis Broadcom 6.4T Optical Engine - Paper 23.4: "One CPO package consists of eight 6.4T OEs, each with a PIC and an EIC, on TSMC's N7 process."

Electronic die floorplan drawn
eight driver and eight TIA blocks; placement representative
Assumed

Assumed: CPO electronic die floorplan, as drawn, eight driver and eight TIA blocks on the EIC; modulators and photodiodes on the PIC; placement representative. Why

Marvell’s 6.4T engine
32 × 200G; modulators, drivers and TIAs in one 3D-packaged device
Spec

Sources: Marvell opening: "the industry’s first highly integrated SiPho engine featuring 32 channels of 200G electrical and optical interfaces"; body: "The Marvell 3D SiPho engine combines hundreds of components such as waveguides and modulators, photodetectors, modulator drivers, trans-impedance amplifiers, microcontrollers, and a host of other passive components into a single, unified device"; "The SiPho Engine leverages advanced 3D packaging and other Marvell technologies to integrate hundreds of components into a single device."Marvell body: "The Marvell 6.4T 3D SiPho Engine is a highly integrated optical engine with 32 channels of 200G electrical and optical interfaces"

data

Ring modulators

See it in 3D ↗
Modulators, NVIDIA figure
micro-rings, 200G PAM4 per wavelength
Spec

Sources: NVIDIA How TSMC helped solve Micro Ring Modulator problems: NVIDIA identifies microring modulation and a direct 200 Gb/s PAM4 rate per wavelength.

Broadcom, for comparison
Mach-Zehnder modulators (MZM), not rings
Reported

Sources: SemiAnalysis Broadcom 6.4T Optical Engine - Paper 23.4: "Broadcom showcased progress on their 6.4T MZM optical engine (OE) consisting of 64 lanes of ~100G using PAM4 modulation."光芯 (光芯前沿 column), Tencent Cloud developer community title: "Broadcom基于7nm ASIC+硅光MZM的3D集成6.4Tb/s" (Broadcom: a 7 nm ASIC plus silicon-photonics MZM, 3D-integrated, 6.4 Tb/s)

Per engine
1.6 Tb/s each way, 8 × 200G
Spec

Sources: NVIDIA Quantum-X section: engines "each supporting 1.6Tbps transmit and 1.6Tbps receive", with "eight 200Gbps PAM4 lanes for both transmit and receive"

Engine and package rate
1.6 Tb/s per engine = 8 lanes × 200 Gb/s each way; 18 engines × 1.6 Tb/s = 28.8 Tb/s
Calc.

Calculated: 8 lanes × 200 Gb/s = 1.6 Tb/s per engine each way; 18 engines × 1.6 Tb/s = 28.8 Tb/s per switch chip. Method

Modulator size, SemiAnalysis
Mach-Zehnder: millimeters long; rings: tens of microns
Reported

Sources: SemiAnalysis (Dylan Patel, Daniel Nishball, Myron Xie and others) modulator section, Mach-Zehnder drawbacks: "Large form factor with dimensions measured in millimeter scale for length (compared to MRM in micron scale), since they require two waveguide arms and a combining region"; micro-ring advantages: "The are extremely compact (scale in the tens of microns), allowing far higher modulator density than MZMs."

data

Photodiodes

See it in 3D ↗
Per engine
1.6 Tb/s each way, 8 × 200G
Spec

Sources: NVIDIA Quantum-X section: engines "each supporting 1.6Tbps transmit and 1.6Tbps receive", with "eight 200Gbps PAM4 lanes for both transmit and receive"

Layout
representative; counts are NVIDIA’s
Assumed

Assumed: Layout of the CPO switch package, representative. Why

Electronic die floorplan drawn
eight driver and eight TIA blocks; placement representative
Assumed

Assumed: CPO electronic die floorplan, as drawn, eight driver and eight TIA blocks on the EIC; modulators and photodiodes on the PIC; placement representative. Why

Engine partition, as reported
drivers, TIAs and control logic on the electronic die; modulators and detectors on the photonic die (COUPE)
Reported

Sources: SemiAnalysis (Dylan Patel, Daniel Nishball, Myron Xie and others) Quantum-X Photonics section: "Each optical engine integrates a Photonic Integrated Circuit (PIC) built on a mature N65 process node, and an Electronic Integrated Circuit (EIC) fabricated on an advanced N6 node. The PIC leverages the older node because it contains optical components such as modulators, waveguides, and detectors—devices that do not benefit from scaling, and often perform better at larger geometries. In contrast, the EIC includes drivers, TIAs, and control logic, which benefit significantly from higher transistor density and improved power efficiency enabled by advanced nodes." TSMC COUPE section, Die fabrication: "The EIC is manufactured on the N7 node, integrating high-speed optical modulator drivers and TIAs." A report on the COUPE engines, not NVIDIA’s published floorplan.

data

External laser sources

See it in 3D ↗
Laser modules
18 per Q3450 switch, 8 lasers each
Reported

Sources: NVIDIA external laser source section: "Each ELS contains eight high-quality lasers"Lambda "The switch supports 18 removable external light-source modules, each feeding eight MPO ports."

Transmit lanes per module
32
Spec

Sources: NVIDIA external laser source section: "An individual ELS is capable of powering 32 of the Quantum-X switch’s 576 transmit lanes"

Fibers per engine
8 transmit, 8 receive, 2 laser in
Spec

Sources: NVIDIA Quantum-X section: "eight transmit, eight receive, and two laser input fibers per engine"

data

Fiber out of the package

See it in 3D ↗
Interface, Quantum-X
socketed subassemblies, sealed fiber interface
Reported

Sources: NVIDIA How modular optical subassemblies ensure rapid deployment: the subassembly has a socket-based electrical interface to the main switch package and a hermetically sealed fiber interface to the engines.

Spectrum-X, for comparison
a detachable optical connector
Spec

Sources: NVIDIA body text: "Spectrum-X employs a detachable optical connector"

Fibers per engine
8 transmit, 8 receive, 2 laser in
Spec

Sources: NVIDIA Quantum-X section: "eight transmit, eight receive, and two laser input fibers per engine"

Front panel, Q3450
144 MPO connectors
Reported

Sources: Lambda front panel: 144 MPO optical connectors for 144 × 800G ports

data

Switch ASIC

See it in 3D ↗
Package, Broadcom
Tomahawk 5 switch chip and eight 6.4 Tb/s optical engines: 51.2 Tb/s
Spec

Sources: Broadcom opening: "The product integrates eight silicon photonics based 6.4-Tbps optical engines with Broadcom’s best-in-class StrataXGS® Tomahawk®5 switch chip."SemiAnalysis Broadcom 6.4T Optical Engine - Paper 23.4: "One CPO package consists of eight 6.4T OEs, each with a PIC and an EIC, on TSMC's N7 process."

Broadcom’s engine, lanes
64 × 106.25 Gb/s PAM4, 6.4 Tb/s, for 51.2T (Tomahawk 5) switching
Reported

Sources: 光芯 (光芯前沿 column), Tencent Cloud developer community opening: "发布了一款用于51.2T交换的6.4Tb/s CPO专用ASIC,该7nm ASIC与硅光芯片PIC进行3D封装,能量效率达到4.2pJ/b。" (it presented a 6.4 Tb/s CPO ASIC for 51.2T switching, the 7 nm ASIC 3D-packaged with the silicon-photonics PIC); conclusion: "该6.4Tb/s CPO ASIC通过64条106.25Gb/s PAM-4通道实现最高集成度和吞吐量" (the 6.4 Tb/s CPO ASIC reaches its integration and throughput through 64 lanes of 106.25 Gb/s PAM-4). The 4.2 pJ/b figure is not used here: the same post’s title says 4.5 pJ/b.SemiAnalysis Broadcom 6.4T Optical Engine - Paper 23.4: "The optical engines were tested in a Tomahawk 5 51.2T CPO system."

Engine and package rate
6.4 Tb/s per engine = 16 × 400G FR4 ports = 64 lanes × 100 Gb/s each way (106.25 Gb/s line rate); 8 engines × 6.4 Tb/s = 51.2 Tb/s
Calc.

Calculated: 16 × 400G FR4 ports = 6.4 Tb/s per engine; each FR4 port is four 100 Gb/s wavelength lanes, so 16 × 4 = 64 lanes × 100 Gb/s each way (reported as 64 × 106.25 Gb/s PAM4 line rate); 8 engines × 6.4 Tb/s = 51.2 Tb/s. Method

data

Package traces

See it in 3D ↗
Signal conditioning, Broadcom
less needed with the engines on the switch chip’s substrate
Spec

Sources: Broadcom body: "The high degree of integration enables the placement of the optical engines on a common substrate with complex logic ASICs minimizing the need for signal conditioning circuitry."

Layout
representative; counts are Broadcom’s
Assumed

Assumed: The Mach-Zehnder package, as drawn, representative; counts Broadcom’s, proportions from its package images. Why

data

Electronic chip

See it in 3D ↗
Broadcom’s engine, as reported
one 7 nm electronic die per engine: all 64 transmit and receive channels, MZM drivers and TIAs
Reported

Sources: 光芯 (光芯前沿 column), Tencent Cloud developer community write-up of Broadcom’s ISSCC 2026 paper 23.4, chip section: "该CPO ASIC采用7nm FinFET工艺制造(图23.4.7),单片裸片集成64条发射/接收通道、TXPLL、辅助公共电路及ADC/DAC组件;6.4Tb/s光引擎通过7nm ASIC与PIC硅片的3D封装实现" (the CPO ASIC is made in a 7 nm FinFET process; a single die integrates the 64 transmit/receive channels, the transmit PLL, shared support circuits and ADC/DAC blocks; the 6.4 Tb/s optical engine is this 7 nm ASIC 3D-packaged with the photonic chip). Its transmit path ends in "MZM驱动器" (MZM drivers); its receive path is a "直接驱动TIA" (direct-drive TIA). This is the 51.2T (Tomahawk 5) generation at 106.25 Gb/s per lane; the paper itself was not opened.SemiAnalysis Broadcom 6.4T Optical Engine - Paper 23.4: "One CPO package consists of eight 6.4T OEs, each with a PIC and an EIC, on TSMC's N7 process."

Broadcom’s transmit drive, as reported
retimed, then three driver segments per MZM: one for the low bit, two for the high bit
Reported

Sources: 光芯 (光芯前沿 column), Tencent Cloud developer community system architecture: "发射路径中,模拟前端(AFE)后接时钟数据恢复(CDR)模块,对来自主机的输入数据进行重定时和串并转换" (in the transmit path the analog front end is followed by a clock-and-data-recovery block that retimes and deserializes the host’s data); "串并转换后的数据流通过发射数字信号处理器(TX DSP)送入MZM驱动器,其中PAM-4信号的温度计编码支持利用最高有效位(MSB)数据同时驱动两个调制器分段。" (the data passes through a transmit DSP into the MZM driver, whose thermometer coding drives two modulator segments with the PAM-4 high bit); driver section: "这些位于8倍时钟(8t)域的数据流被分配至3个驱动器分段——第一个分段分配给LSB,后两个分配给MSB。" (the streams go to three driver segments: the first for the low bit, the other two for the high bit)

Broadcom’s engine, lanes
64 × 106.25 Gb/s PAM4, 6.4 Tb/s, for 51.2T (Tomahawk 5) switching
Reported

Sources: 光芯 (光芯前沿 column), Tencent Cloud developer community opening: "发布了一款用于51.2T交换的6.4Tb/s CPO专用ASIC,该7nm ASIC与硅光芯片PIC进行3D封装,能量效率达到4.2pJ/b。" (it presented a 6.4 Tb/s CPO ASIC for 51.2T switching, the 7 nm ASIC 3D-packaged with the silicon-photonics PIC); conclusion: "该6.4Tb/s CPO ASIC通过64条106.25Gb/s PAM-4通道实现最高集成度和吞吐量" (the 6.4 Tb/s CPO ASIC reaches its integration and throughput through 64 lanes of 106.25 Gb/s PAM-4). The 4.2 pJ/b figure is not used here: the same post’s title says 4.5 pJ/b.SemiAnalysis Broadcom 6.4T Optical Engine - Paper 23.4: "The optical engines were tested in a Tomahawk 5 51.2T CPO system."

Layout
representative; counts are Broadcom’s
Assumed

Assumed: The Mach-Zehnder package, as drawn, representative; counts Broadcom’s, proportions from its package images. Why

data

Mach-Zehnder modulators

See it in 3D ↗
Broadcom, for comparison
Mach-Zehnder modulators (MZM), not rings
Reported

Sources: SemiAnalysis Broadcom 6.4T Optical Engine - Paper 23.4: "Broadcom showcased progress on their 6.4T MZM optical engine (OE) consisting of 64 lanes of ~100G using PAM4 modulation."光芯 (光芯前沿 column), Tencent Cloud developer community title: "Broadcom基于7nm ASIC+硅光MZM的3D集成6.4Tb/s" (Broadcom: a 7 nm ASIC plus silicon-photonics MZM, 3D-integrated, 6.4 Tb/s)

Broadcom’s transmit drive, as reported
retimed, then three driver segments per MZM: one for the low bit, two for the high bit
Reported

Sources: 光芯 (光芯前沿 column), Tencent Cloud developer community system architecture: "发射路径中,模拟前端(AFE)后接时钟数据恢复(CDR)模块,对来自主机的输入数据进行重定时和串并转换" (in the transmit path the analog front end is followed by a clock-and-data-recovery block that retimes and deserializes the host’s data); "串并转换后的数据流通过发射数字信号处理器(TX DSP)送入MZM驱动器,其中PAM-4信号的温度计编码支持利用最高有效位(MSB)数据同时驱动两个调制器分段。" (the data passes through a transmit DSP into the MZM driver, whose thermometer coding drives two modulator segments with the PAM-4 high bit); driver section: "这些位于8倍时钟(8t)域的数据流被分配至3个驱动器分段——第一个分段分配给LSB,后两个分配给MSB。" (the streams go to three driver segments: the first for the low bit, the other two for the high bit)

Modulator size, SemiAnalysis
Mach-Zehnder: millimeters long; rings: tens of microns
Reported

Sources: SemiAnalysis (Dylan Patel, Daniel Nishball, Myron Xie and others) modulator section, Mach-Zehnder drawbacks: "Large form factor with dimensions measured in millimeter scale for length (compared to MRM in micron scale), since they require two waveguide arms and a combining region"; micro-ring advantages: "The are extremely compact (scale in the tens of microns), allowing far higher modulator density than MZMs."

400G FR4
four wavelengths share one fiber each way
Spec

Sources: NVIDIA "FR4 (far reach) supports up to two kilometers by multiplexing four channels of different laser wavelengths into a single fiber pair"TIA Fiber Optics Technology Consortium 400GBASE-FR4, IEEE 802.3 clause 151: "400 Gb/s wavelength-division multiplexed (WDM) PAM4 serial transmission over 2 single-mode optical fibers, with reach up to at least 2 km"

Layout
representative; counts are Broadcom’s
Assumed

Assumed: The Mach-Zehnder package, as drawn, representative; counts Broadcom’s, proportions from its package images. Why

Engine and package rate
6.4 Tb/s per engine = 16 × 400G FR4 ports = 64 lanes × 100 Gb/s each way (106.25 Gb/s line rate); 8 engines × 6.4 Tb/s = 51.2 Tb/s
Calc.

Calculated: 16 × 400G FR4 ports = 6.4 Tb/s per engine; each FR4 port is four 100 Gb/s wavelength lanes, so 16 × 4 = 64 lanes × 100 Gb/s each way (reported as 64 × 106.25 Gb/s PAM4 line rate); 8 engines × 6.4 Tb/s = 51.2 Tb/s. Method

data

Photodiodes

See it in 3D ↗
Broadcom’s receive, as reported
each TIA drives the switch chip directly through the package substrate
Reported

Sources: 光芯 (光芯前沿 column), Tencent Cloud developer community receive section: "TIA差分输出通过有机基板直接驱动主机交换ASIC" (the TIA’s differential output drives the host switch ASIC directly through the organic substrate)

400G FR4
four wavelengths share one fiber each way
Spec

Sources: NVIDIA "FR4 (far reach) supports up to two kilometers by multiplexing four channels of different laser wavelengths into a single fiber pair"TIA Fiber Optics Technology Consortium 400GBASE-FR4, IEEE 802.3 clause 151: "400 Gb/s wavelength-division multiplexed (WDM) PAM4 serial transmission over 2 single-mode optical fibers, with reach up to at least 2 km"

Layout
representative; counts are Broadcom’s
Assumed

Assumed: The Mach-Zehnder package, as drawn, representative; counts Broadcom’s, proportions from its package images. Why

data

Remote laser modules

See it in 3D ↗
Lasers, Broadcom reference system
remote laser modules, field-replaceable
Spec

Sources: Broadcom product highlights: "System design compatible to support multiple remote laser modules (RLM) for field replaceability"

Layout
representative; counts are Broadcom’s
Assumed

Assumed: The Mach-Zehnder package, as drawn, representative; counts Broadcom’s, proportions from its package images. Why

data

Fiber out of the package

See it in 3D ↗
Engine, Broadcom
6.4T-FR4 engines with a Broadcom Fiber Connector (BFC)
Spec

Sources: Broadcom product highlights: "Broadcom 6.4T-FR4 Bailly SCIP optical engines with Broadcom Fiber Connector (BFC) for CPO systems"

Fiber attach, Broadcom
edge-coupled, high-density, automated
Spec

Sources: Broadcom body: "Broadcom’s innovative manufacturing approach that utilizes proven CMOS foundry processes, advanced packaging technologies and a highly automated high-density, edge-coupled fiber attach capability"

Front panel, Broadcom reference system
128 × 400G FR4 ports on 128 duplex LC connectors
Spec

Sources: Broadcom product highlights: "4RU system design with high-efficiency air cooling to deliver 128 ports of 400G FR4 connectivity externally fiber-coupled with 128 duplex LC optical connectors"

Data fibers per engine
16 transmit and 16 receive
Calc.

Calculated: 128 duplex 400G FR4 ports ÷ 8 engines = 16 ports per engine, each with one transmit and one receive fiber. Method

data

A CPO switch, shipping now

See it in 3D ↗
NVIDIA Spectrum-X Ethernet Photonics
in production (NVIDIA, 05/31/2026); shipping to select partners (TrendForce, 07/27/2026)
Reported

Sources: NVIDIA summary line: "Vera Rubin introduces NVIDIA Spectrum-X Ethernet Photonics — now in production — combining co-packaged optics with Spectrum-X switching to enable million-GPU AI factories."TrendForce first paragraph: "NVIDIA has begun shipping its next-generation Spectrum-X CPO switch to select partners."StorageReview opening: "NVIDIA has moved Spectrum-X Ethernet Photonics, its co-packaged optics (CPO) Ethernet switch platform, into full production."

NVIDIA Quantum-X InfiniBand Photonics
Q3450-LD, 144 × 800G ports; engineering samples at Lambda, 06/2026
Reported

Sources: NVIDIA Quantum-X section: "The NVIDIA Quantum-X800 InfiniBand platform includes CPO-based switches, including the Q3450-LD with 144 ports of 800 gigabits-per-second (Gb/s) InfiniBand."Lambda opening: "Lambda is taking an early look at co-packaged optics, starting with the NVIDIA Quantum-X InfiniBand Photonics Q3450-LD switch"; timing: "Engineering samples are a chance to work through rack design, cooling, power delivery, fiber routing, and the installation process with the vendor while the product is still becoming real."

Broadcom TH6-Davisson
102.4 Tb/s, 16 × 6.4T engines; announced 10/08/2025 as shipping, sampled to early-access customers
Spec

Sources: Broadcom opening: "now shipping Tomahawk® 6 Davisson (TH6-Davisson), the company's third-generation Co-Packaged Optics (CPO) Ethernet switch"; product highlights: "16 x 6.4 Tbps Davisson DR Optical Engines"; availability: "Broadcom is currently sampling the TH6-Davisson BCM78919 device to its early access customers and partners."

Broadcom TH5-Bailly, 51.2 Tb/s
in volume manufacturing with Delta and Micas Networks (TrendForce, 07/27/2026)
Reported

Sources: TrendForce Broadcom paragraph: "Broadcom's 51.2T Bailly CPO switch has entered volume manufacturing with support from Delta Electronics and Micas Networks."

Bailly in Meta’s lab, Broadcom figure
one million flap-free 400G port-hours
Vendor

Sources: Broadcom opening: "one million cumulative 400G equivalent port device hours of flap-free CPO operation at Meta"; "The absence of link flaps in Meta’s high-temperature lab characterization environment"; footnote: ECOC 2025 talk by Siamak Amiralizadeh, 09/30/2025

Compared with: pluggable modules: "Compared to pluggable module solutions, the data highlights that CPO reduces optics power by 65 percent and also demonstrates higher link reliability."

data

Next: CPO on the GPU

See it in 3D ↗
Marvell compute tray, a concept
4 XPUs, each with four 6.4T light engines: 102.4 Tb/s on 1,024 fibers, in 1U
Vendor

Sources: Marvell (Chris McCormick) The CPO Server Tray: "a conceptualized AI compute tray with CPO developed with products from SENKO Advanced Components and Marvell. The design contains room for four XPUs and up to 102.4 Tbps of bandwidth delivered through 1024 optical fibers, all in a 1U tray."; "Each XPU is connected to four Marvell 6.4T light engines for opto-electric conversions." The post later puts the total, with laser fibers, at 1,152.

Compared with: copper scale-up, in Marvell’s words: "The density and reach enabled by CPO opens the door to scale-up domains far beyond what is possible with copper alone."

Marvell CPO switch, a reference design
switch chip + 16 light engines; 16 laser modules at the faceplate; 1OU with a cold plate
Vendor

Sources: Marvell (Chris McCormick) Heat and Space: "Marvell, SENKO, Jabil, and Mikros Technologies also recently unveiled a reference design of a data center CPO switch."; "surrounded by 16 light engine tiles and 1,152 fibers (128 laser fibers and 1,024 data fibers). The light engines are driven by 16 laser modules, which are connected to the faceplate for better serviceability."; "maintains a system height of 1OU"

Compared with: air cooling: "conventional air cooling would require a chassis that is two to three times thicker"

OCI MSA, founded 03/12/2026
an open optical scale-up spec: AMD, Broadcom, Meta, Microsoft, NVIDIA, OpenAI
Reported

Sources: Broadcom (for the OCI MSA) opening: "The Optical Compute Interconnect (OCI) Multi-Source Agreement (MSA) group today announced its formation, led by founding members AMD, Broadcom, Meta, Microsoft, NVIDIA and OpenAI."; "OCI will enable migration from copper-based to optical-based scale-up architectures"; "Support for pluggable, on-board, and co-packaged optics (CPO)."

Ayar Labs, with GUC
optical engines on an XPU package: over 100 Tb/s, a package design
Vendor

Sources: Ayar Labs "The new XPU multi-chip package (MCP) design replaces traditional electrical interconnects with Ayar Labs’ optical engines attached directly to the MCP organic substrate. This architecture enables more than 100 Tbps full-duplex optical interface from the XPU package, more than an order of magnitude improvement over current XPUs."

Compared with: current XPUs, which Ayar Labs puts more than ten times lower

Lightmatter Passage L200
32 and 64 Tb/s 3D CPO engines for XPUs and switches; announced 03/31/2025 for 2026
Vendor

Sources: Lightmatter "The L200 3D CPO family includes both 32 Tbps and 64 Tbps versions, representing a 5 to 10x improvement over existing solutions."; "Available in 2026, Lightmatter’s L200 and L200X 3D CPO chips are designed to accelerate time to market and performance of next generation XPUs and switches"

Compared with: existing interconnect solutions, which Lightmatter puts 5 to 10 times lower

Celestial AI
bought by Marvell, 02/02/2026, for optical scale-up
Reported

Sources: Marvell (SEC Form 8-K exhibit 99.1) opening: "completed its previously announced acquisition of Celestial AI, a pioneer in optical interconnect technology for scale-up connectivity"; outlook: initial revenue "in the second half of fiscal 2028"

Ranovus Odin
one monolithic die: micro-ring modulators, photodetectors, RF drivers, TIAs, control logic
Spec

Sources: Ranovus release body: "a groundbreaking monolithic Electronic and Photonic Integrated Circuit (EPIC) platform"; "Featuring monolithic integration of silicon photonics, RF drivers, transimpedance amplifiers (TIA), and control logic"; "Key innovations include patented silicon photonics-based Micro Ring Modulators (MRM), internal isolator-free laser sources, and high-performance photodetectors."

Ayar Labs TeraPHY
micro-ring based, electronics and photonics on one die; later designs may split them
Reported

Sources: Ayar Labs, release reprinted by HPCwire opening: "Ayar Labs has successfully demonstrated its patented monolithic electronic/photonic solution on GLOBALFOUNDRIES (GF) next generation photonics solution based on its 45nm platform."; CEO quote: "Ayar Labs has been perfecting our micro-ring based monolithic electronic/photonic solution for nearly a decade."SemiAnalysis (Dylan Patel, Daniel Nishball, Myron Xie and others) TSMC COUPE section: "Ayar Labs, who has previously relied on Global Foundries’ Fotonix platform for monolithic optical engines, now also has COUPE on their roadmap."Gazettabyte From monolithic optics to modular chiplets: "This [TeraPHY optical I/O chiplet] architecture lets us move seamlessly through different foundry processes"; "The 8Tbps TeraPHY device is built using GlobalFoundries’ 45SPCLO 45nm silicon-photonics process"; "But the design can also be migrated to TSMC’s more advanced CMOS nodes for the electrical IC while benefiting from TSMC’s silicon photonics and packaging flows."

heat

Switch ASIC and engines

See it in 3D ↗
Cooling, Q3450 switch
liquid; 4 UDQ4 connections, dual internal loops
Reported

Sources: Lambda cooling: "Cooling runs through four UDQ4 liquid cooling connections with dual internal loops."

Cold plate drawn
one per package, representative
Assumed

Assumed: Layout of the CPO switch package, representative. Why

Q3450 switch, NVIDIA figure
3.95 kW
Vendor

Sources: Lambda power comparison: "NVIDIA Photonics CPO switch: 3.95 kW. Standard switch: 7.0 kW."

Compared with: a pluggable-optics equivalent switch at 7.0 kW

Per port, NVIDIA figure
as low as 9 W
Vendor

Sources: NVIDIA body text: 30 W per port with pluggable transceivers, "as low as 9 W" per port with co-packaged optics

Compared with: 30 W per port with pluggable transceivers

Heat drawn
log scale, each 10× in power 5× the motion; none under 0.5 W
Assumed

Assumed: How much heat each part is drawn with, log scale: each 10× in power is 5× the motion; nothing under 0.5 W. Why

heat

Cold plate

See it in 3D ↗
Cooling, Q3450 switch
liquid; 4 UDQ4 connections, dual internal loops
Reported

Sources: Lambda cooling: "Cooling runs through four UDQ4 liquid cooling connections with dual internal loops."

Cold plate drawn
one per package, representative
Assumed

Assumed: Layout of the CPO switch package, representative. Why

heat

Switch ASIC and engines

See it in 3D ↗
Cooling, Broadcom reference system
4RU, high-efficiency air cooling
Spec

Sources: Broadcom product highlights: "4RU system design with high-efficiency air cooling to deliver 128 ports of 400G FR4 connectivity"

Heat sink drawn
representative shape
Assumed

Assumed: The Mach-Zehnder package, as drawn, representative; counts Broadcom’s, proportions from its package images. Why

heat

Heat sink

See it in 3D ↗
Cooling, Broadcom reference system
4RU, high-efficiency air cooling
Spec

Sources: Broadcom product highlights: "4RU system design with high-efficiency air cooling to deliver 128 ports of 400G FR4 connectivity"

Heat sink drawn
representative shape
Assumed

Assumed: The Mach-Zehnder package, as drawn, representative; counts Broadcom’s, proportions from its package images. Why

+Coherent optics10.8 cm long

power

Coherent DSP

See it in 3D ↗
800ZR module, FiberMall ZR page
24–25 W
Reported

Sources: FiberMall power section: "800G ZR draws 24 to 25 watts"

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

power

Tunable laser

See it in 3D ↗
Shared laser, same-wavelength link
CW light splits to the modulator and receiver local oscillator
Reported

Sources: Acacia Communications Page 10, Design optimization for Coherent BiDi inset: a same-wavelength transmitter and receiver can share a CW laser; its conventional coherent diagram shows the splitter branches to modulator and receiver LO. Different-wavelength BiDi needs a separate LO laser.

Nano-ITLA, JLT 2023 research example
25.0 × 15.6 × 6.5 mm, under 3 W
Reported

Sources: Optica / IEEE abstract: "The Nano-ITLA is 25.0 mm (L) × 15.6 mm (W) × 6.5 mm (H)" and "less than 3 W total power consumption"

EFFECT Photonics nITLA17
2.9 W
Spec

Sources: EFFECT Photonics specifications: 2.9 W power consumption

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

power

Modulator driver IC

See it in 3D ↗
800G-class research example
separate 128 GBd driver IC
Reported

Sources: Sumitomo Electric Section 2 and Photo 1: separately developed driver IC; Section 4.1 later assembles it with an InP modulator in a CDM. Not a teardown of this OSFP.

Drawn as
discrete driver and TIA packages; separate optical assemblies, representative
Assumed

Assumed: Discrete coherent electronics and optical assemblies, representative separate board packages. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

power

Dual-polarization IQ modulator

See it in 3D ↗
Power and heat shown
optics’ bias well under half a watt; heat not drawn
Assumed

Assumed: Coherent optical-component power and heat, well under half a watt each; no heat drawn. Why

Drawn as
discrete driver and TIA packages; separate optical assemblies, representative
Assumed

Assumed: Discrete coherent electronics and optical assemblies, representative separate board packages. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

power

Receiver optics (hybrids + photodiodes)

See it in 3D ↗
Power and heat shown
optics’ bias well under half a watt; heat not drawn
Assumed

Assumed: Coherent optical-component power and heat, well under half a watt each; no heat drawn. Why

Drawn as
discrete driver and TIA packages; separate optical assemblies, representative
Assumed

Assumed: Discrete coherent electronics and optical assemblies, representative separate board packages. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

power

Transimpedance amplifier IC

See it in 3D ↗
800G-class research example
separate 128 GBd TIA IC
Reported

Sources: Sumitomo Electric Section 3 and Photo 2: separately developed TIA; Section 4.1 later assembles it with detector optics in an ICR. Not a teardown of this OSFP.

Drawn as
discrete driver and TIA packages; separate optical assemblies, representative
Assumed

Assumed: Discrete coherent electronics and optical assemblies, representative separate board packages. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

data

Coherent DSP

See it in 3D ↗
Skew, OIF IC-TROSA
high-speed contacts grouped to suit common coherent DSP ball maps
Spec

Sources: OIF sec. 4.3: "To reduce high speed channel skew between IC-TROSA and most common coherent DSP ball maps the TX and RX contacts are grouped in a single outer row."

Makers, per one 2026 survey
Cisco (Acacia), Marvell, Ciena, Nokia (Infinera)
Reported

Sources: Cignal AI coherent DSP supply chain survey: the companies with their own coherent DSP

Electrical routes drawn
four host-path groups; schematic, not a lane count or pinout
Assumed

Assumed: Coherent DSP interface drawing, schematic routing groups. Why

800ZR module, FiberMall ZR page
24–25 W
Reported

Sources: FiberMall power section: "800G ZR draws 24 to 25 watts"

800G ZR tier, FiberMall guide
20–30 W
Reported

Sources: FiberMall power table: 800G ZR modules "20-30W"

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

data

Modulator driver IC

See it in 3D ↗
800G-class research example
separate 128 GBd driver IC
Reported

Sources: Sumitomo Electric Section 2 and Photo 1: separately developed driver IC; Section 4.1 later assembles it with an InP modulator in a CDM. Not a teardown of this OSFP.

Drawn as
discrete driver and TIA packages; separate optical assemblies, representative
Assumed

Assumed: Discrete coherent electronics and optical assemblies, representative separate board packages. Why

Optical packages, OIF agreements
RF pads at one end, fibers at the opposite end
Spec

Sources: OIF sec. 7.1: "For both form factors the fiber inputs and the RF electrical outputs are located on opposite ends of the package."OIF Fig. 4-2: RF landing pads RF1–RF13 on the end opposite the input and output fiber bootsOIF sec. 9.1: "For both form factors the optical fiber interfaces and the RF electrical interfaces are located on opposite ends of the package."

Order as drawn
DSP → driver/TIA → optics → laser, short direct RF paths
Assumed

Assumed: Inside the 800ZR module, representative. Why

Common alternative, NeoPhotonics COSA
modulator co-packaged with drivers, receiver with TIAs
Vendor

Sources: NeoPhotonics (PR Newswire) release text: "The COSA contains a coherent I/Q modulator co-packaged with drivers, as well as a coherent receiver co-packaged with trans-impedance amplifiers."

Compared with: the discrete packages drawn here; a 64 GBd product announcement, not a teardown of this OSFP.

Commercial example, 400G ZR/ZR+
CHR2094 driver paired with CHR1094 TIA
Spec

Sources: Coherent Product paragraph: CHR1094/CHR2094 chipset is for 400G ZR/ZR+ coherent links. The separately listed CHR2075 targets 800G/1.6T pluggables; that is not an 800ZR qualification for this coherent chipset.

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

data

Dual-polarization IQ modulator

See it in 3D ↗
Optical function
nested modulators for two polarizations
Vendor

Sources: Lumentum Product description identifies two nested modulators and a four-channel driver as distinct components within its co-packaged CDM; this drawing separates the components for inspection.

Compared with: Lumentum HB-CDM product architecture; not a performance comparison or a teardown of this OSFP.

Modulation
DP-16QAM at ≈118 GBd
Reported

Sources: MapYourTech body text: DP-16QAM at roughly 118 Gbaud for 800ZRFiberMall body text: DP-16QAM at roughly 118 Gbaud

Drawn as
discrete driver and TIA packages; separate optical assemblies, representative
Assumed

Assumed: Discrete coherent electronics and optical assemblies, representative separate board packages. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

data

Tunable laser

See it in 3D ↗
Shared laser, same-wavelength link
CW light splits to the modulator and receiver local oscillator
Reported

Sources: Acacia Communications Page 10, Design optimization for Coherent BiDi inset: a same-wavelength transmitter and receiver can share a CW laser; its conventional coherent diagram shows the splitter branches to modulator and receiver LO. Different-wavelength BiDi needs a separate LO laser.

Nano-ITLA, JLT 2023 research example
25.0 × 15.6 × 6.5 mm, under 3 W
Reported

Sources: Optica / IEEE abstract: "The Nano-ITLA is 25.0 mm (L) × 15.6 mm (W) × 6.5 mm (H)" and "less than 3 W total power consumption"

EFFECT Photonics nITLA17
2.9 W
Spec

Sources: EFFECT Photonics specifications: 2.9 W power consumption

Pigtail, vendor nano-ITLA example
PANDA fiber, bend radius 5 mm or more; Molex board connector
Spec

Sources: LD-PD specifications table: "Fiber Type: PANDA Fiber, or equivalents", "Minimum Fiber Bend Radius: Equal or better than 5mm", "Electrical Interface: Molex 5054761010". The drawn flex tail and routing are representative.

Tunes across
C-band, 1528.58–1567.34 nm
Spec

Sources: Cisco optical specifications table: C-band tuning range 1528.58–1567.34 nm

Integrated example, Furukawa IC-TROSA
laser at the LC end; driver and TIA at the RF flex end
Reported

Sources: Furukawa Electric Fig. 1, top view of the IC-TROSA Type-2: RF flexible printed circuit with the 4-ch TIA and 4-ch DRV at one end, the laser chip and LC receptacle at the other. A single-package design, not this discrete board.

Order as drawn
DSP → driver/TIA → optics → laser, short direct RF paths
Assumed

Assumed: Inside the 800ZR module, representative. Why

data

Receiver optics (hybrids + photodiodes)

See it in 3D ↗
Optical packages, OIF agreements
RF pads at one end, fibers at the opposite end
Spec

Sources: OIF sec. 7.1: "For both form factors the fiber inputs and the RF electrical outputs are located on opposite ends of the package."OIF Fig. 4-2: RF landing pads RF1–RF13 on the end opposite the input and output fiber bootsOIF sec. 9.1: "For both form factors the optical fiber interfaces and the RF electrical interfaces are located on opposite ends of the package."

Optical receiver function
two 90° hybrids, four balanced photodiode pairs
Spec

Sources: OIF Section 6, page 8: two 90-degree hybrid mixers, four balanced detector pairs and four linear amplifiers. Supports functional architecture, not this physical layout.

Drawn as
discrete driver and TIA packages; separate optical assemblies, representative
Assumed

Assumed: Discrete coherent electronics and optical assemblies, representative separate board packages. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

data

Transimpedance amplifier IC

See it in 3D ↗
800G-class research example
separate 128 GBd TIA IC
Reported

Sources: Sumitomo Electric Section 3 and Photo 2: separately developed TIA; Section 4.1 later assembles it with detector optics in an ICR. Not a teardown of this OSFP.

Optical packages, OIF agreements
RF pads at one end, fibers at the opposite end
Spec

Sources: OIF sec. 7.1: "For both form factors the fiber inputs and the RF electrical outputs are located on opposite ends of the package."OIF Fig. 4-2: RF landing pads RF1–RF13 on the end opposite the input and output fiber bootsOIF sec. 9.1: "For both form factors the optical fiber interfaces and the RF electrical interfaces are located on opposite ends of the package."

Order as drawn
DSP → driver/TIA → optics → laser, short direct RF paths
Assumed

Assumed: Inside the 800ZR module, representative. Why

Common alternative, NeoPhotonics COSA
modulator co-packaged with drivers, receiver with TIAs
Vendor

Sources: NeoPhotonics (PR Newswire) release text: "The COSA contains a coherent I/Q modulator co-packaged with drivers, as well as a coherent receiver co-packaged with trans-impedance amplifiers."

Compared with: the discrete packages drawn here; a 64 GBd product announcement, not a teardown of this OSFP.

Commercial example, 400G ZR/ZR+
CHR1094: four channels, 64 GBd, wire-bonded die
Spec

Sources: Coherent CHR1094 product-table row: coherent, 4 channels, 64 GBaud, die/wire bonded, 400G ZR/ZR+. This example does not specify the TIA used in an 800ZR module.

Drawn as
discrete driver and TIA packages; separate optical assemblies, representative
Assumed

Assumed: Discrete coherent electronics and optical assemblies, representative separate board packages. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

data

One fiber pair

See it in 3D ↗
Connector
LC duplex
Spec

Sources: Cisco specifications: "Connector: LC duplex"

Receptacle as drawn
body, sleeves and port pitch representative
Assumed

Assumed: Duplex LC receptacle as drawn, representative. Why

Reach, Cisco modules
120 km amplified (800ZR) · over 1,000 km amplified (ZR+) · 75–80 km unamplified
Spec

Sources: Cisco tables 1 and 2: "Up to 120 km amplified DWDM" (800ZR), "Over 1000 km amplified DWDM" (ZR+), and "Up to 75 km with 800ZR and 80 km with 800G ZR+" unamplified

Standard
OIF 800ZR, published 10/30/2024
Spec

Sources: OIF release: OIF-800ZR-01.0 implementation agreement, 10/30/2024

data

Pull tab and label

See it in 3D ↗
Form factors, Cisco 800G ZR/ZR+
QSFP-DD and OSFP
Spec

Sources: Cisco ordering table and compliance: QSFP-DD and OSFP versions; "the Cisco 800G OSFP optical modules are mechanically compliant with the OSFP MSA"

OSFP module, MSA
22.58 mm wide, 13.0 mm tall
Spec

Sources: OSFP MSA Fig. 3-2, OSFP overall dimensions: width 22.58 ± 0.10 mm, height 13.00 mm inside the cage

Label location, MSA
top face at the fiber end, about 15 × 20 mm
Spec

Sources: OSFP MSA sec. 3.2: "Figure 3-4 shows the recommended reference locations for the label"; Fig. 3-4 draws a (15.0) × (20.0) mm label at the front of the top face

Length with pull tab, Cisco OSFP 800G
116 mm max
Spec

Sources: Cisco Table 6, "Module dimension with pull tab": "(H x W x D) 13 x 22.58 x 116 mm (0.51 x 0.89 x 4.57 in.) max"

Compared with: Cisco’s non-coherent OSFP 800G modules; used here only as a length bound for the drawn tab.

Pull tab color as drawn
white, the MSA color for 1550 nm modules; no coherent row
Assumed

Assumed: OSFP shell, label and pull tab as drawn, representative. Why

Label and tab as drawn
wording, stock and tab shape representative
Assumed

Assumed: OSFP shell, label and pull tab as drawn, representative. Why

heat

DSP heat

See it in 3D ↗
Heat drawn
log scale, each 10× in power 5× the motion; none under 0.5 W
Assumed

Assumed: How much heat each part is drawn with, log scale: each 10× in power is 5× the motion; nothing under 0.5 W. Why

800ZR module, FiberMall ZR page
24–25 W
Reported

Sources: FiberMall power section: "800G ZR draws 24 to 25 watts"

Drawn above the DSP
finned top housing, pedestal and gap pad; representative
Assumed

Assumed: OSFP housing, pedestal and gap pad, representative. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

heat

Laser heat

See it in 3D ↗
Nano-ITLA, JLT 2023 research example
25.0 × 15.6 × 6.5 mm, under 3 W
Reported

Sources: Optica / IEEE abstract: "The Nano-ITLA is 25.0 mm (L) × 15.6 mm (W) × 6.5 mm (H)" and "less than 3 W total power consumption"

EFFECT Photonics nITLA17
2.9 W
Spec

Sources: EFFECT Photonics specifications: 2.9 W power consumption

Heat drawn
log scale, each 10× in power 5× the motion; none under 0.5 W
Assumed

Assumed: How much heat each part is drawn with, log scale: each 10× in power is 5× the motion; nothing under 0.5 W. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

heat

Driver IC heat

See it in 3D ↗
800G-class research example
separate 128 GBd driver IC
Reported

Sources: Sumitomo Electric Section 2 and Photo 1: separately developed driver IC; Section 4.1 later assembles it with an InP modulator in a CDM. Not a teardown of this OSFP.

Heat drawn
log scale, each 10× in power 5× the motion; none under 0.5 W
Assumed

Assumed: How much heat each part is drawn with, log scale: each 10× in power is 5× the motion; nothing under 0.5 W. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

heat

Modulator losses

See it in 3D ↗
Power and heat shown
optics’ bias well under half a watt; heat not drawn
Assumed

Assumed: Coherent optical-component power and heat, well under half a watt each; no heat drawn. Why

Heat drawn
log scale, each 10× in power 5× the motion; none under 0.5 W
Assumed

Assumed: How much heat each part is drawn with, log scale: each 10× in power is 5× the motion; nothing under 0.5 W. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

heat

Receiver-optics losses

See it in 3D ↗
Power and heat shown
optics’ bias well under half a watt; heat not drawn
Assumed

Assumed: Coherent optical-component power and heat, well under half a watt each; no heat drawn. Why

Heat drawn
log scale, each 10× in power 5× the motion; none under 0.5 W
Assumed

Assumed: How much heat each part is drawn with, log scale: each 10× in power is 5× the motion; nothing under 0.5 W. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

heat

TIA IC heat

See it in 3D ↗
800G-class research example
separate 128 GBd TIA IC
Reported

Sources: Sumitomo Electric Section 3 and Photo 2: separately developed TIA; Section 4.1 later assembles it with detector optics in an ICR. Not a teardown of this OSFP.

Heat drawn
log scale, each 10× in power 5× the motion; none under 0.5 W
Assumed

Assumed: How much heat each part is drawn with, log scale: each 10× in power is 5× the motion; nothing under 0.5 W. Why

Layout
OSFP envelope and cited nano-ITLA case to scale; other parts representative
Assumed

Assumed: Inside the 800ZR module, representative. Why

+Inside the copper cablesone plug end

power

Direct Attach Copper (DAC)

See it in 3D ↗
Power, NVIDIA DAC
≈0.1 W per end
Spec

Sources: NVIDIA DAC section: "Power consumption is 0.1 Watts per end"

In the signal path
no redriver, retimer or DSP
Reported

Sources: Flexoptix "pure twinax copper with no active signal processing at all"

GB200 NVL72’s NVLink spine
passive copper
Spec

Sources: NVIDIA hardware overview: "connected by NVLink through the NVLink passive copper cable cartridge backplane"

Glow drawn
log scale, each 10× in power 5× the glow; none under 0.5 W
Assumed

Assumed: How much each part glows in the Power layer, log scale, per view: each 10× in power is 5× the glow; nothing under 0.5 W. Why

power

Active Copper Cable (ACC)

See it in 3D ↗
Power and reach
a couple of watts per end, ≈3 m at 200G/lane
Reported

Sources: Vik's Newsletter "pushes copper out to 3 meters at 200G/lane for just a couple of watts per end"

Inside
a redriver on the receive direction; NVIDIA’s LACC has one in each end
Reported

Sources: Marvell Inside the Cable: ACC equalizers amplify signals received from the opposite cable end; the equalizer is analogNVIDIA "LACCs are essentially DAC cables with an additional IC in each end to boost the signal power and noise reduction"

Glow drawn
log scale, each 10× in power 5× the glow; none under 0.5 W
Assumed

Assumed: How much each part glows in the Power layer, log scale, per view: each 10× in power is 5× the glow; nothing under 0.5 W. Why

power

Active Electrical Cable (AEC)

See it in 3D ↗
Power
≈20 W per end at 200G/lane; 2.5–3.5 W in one guide, lane rate unstated
Reported

Sources: Vik's Newsletter "[AEC] burns around 20 watts per end"AscentOptics AEC section: "higher power consumption (2.5–3.5W)", no lane rate given

Inside
a DSP retimer in each end: CTLE, DFE, CDR, FIR
Reported

Sources: Marvell Inside the Cable: AEC DSPs at each cable end maintain signal integrity during both transmission and receptionAscentOptics AEC section: "The Retimer chip in AEC is more advanced than the Redriver used in ACC. It integrates CTLE (Continuous-Time Linear Equalizer), DFE (Decision Feedback Equalization), CDR (Clock Data Recovery), and FIR drivers."

Glow drawn
log scale, each 10× in power 5× the glow; none under 0.5 W
Assumed

Assumed: How much each part glows in the Power layer, log scale, per view: each 10× in power is 5× the glow; nothing under 0.5 W. Why

data

Direct Attach Copper (DAC)

See it in 3D ↗
In the signal path
no redriver, retimer or DSP
Reported

Sources: Flexoptix "pure twinax copper with no active signal processing at all"

Power, NVIDIA DAC
≈0.1 W per end
Spec

Sources: NVIDIA DAC section: "Power consumption is 0.1 Watts per end"

Reach, 800G-class
0.5–2 m (1–2 m typical)
Reported

Sources: Flexoptix reach section: up to 5 m at lower data rates, 1–2 m at 800G-class signalingNVIDIA "0.5m to 2-meters for straight cables"

GB200 NVL72’s NVLink spine
passive copper
Spec

Sources: NVIDIA hardware overview: "connected by NVLink through the NVLink passive copper cable cartridge backplane"

Layout
representative, not one product
Assumed

Assumed: Inside a copper cable head, representative. Why

data

Active Copper Cable (ACC)

See it in 3D ↗
Inside
a redriver on the receive direction; NVIDIA’s LACC has one in each end
Reported

Sources: Marvell Inside the Cable: ACC equalizers amplify signals received from the opposite cable end; the equalizer is analogNVIDIA "LACCs are essentially DAC cables with an additional IC in each end to boost the signal power and noise reduction"

Power and reach
a couple of watts per end, ≈3 m at 200G/lane
Reported

Sources: Vik's Newsletter "pushes copper out to 3 meters at 200G/lane for just a couple of watts per end"

Layout
representative, not one product
Assumed

Assumed: Inside a copper cable head, representative. Why

data

Active Electrical Cable (AEC)

See it in 3D ↗
Inside
a DSP retimer in each end: CTLE, DFE, CDR, FIR
Reported

Sources: Marvell Inside the Cable: AEC DSPs at each cable end maintain signal integrity during both transmission and receptionAscentOptics AEC section: "The Retimer chip in AEC is more advanced than the Redriver used in ACC. It integrates CTLE (Continuous-Time Linear Equalizer), DFE (Decision Feedback Equalization), CDR (Clock Data Recovery), and FIR drivers."

Power
≈20 W per end at 200G/lane; 2.5–3.5 W in one guide, lane rate unstated
Reported

Sources: Vik's Newsletter "[AEC] burns around 20 watts per end"AscentOptics AEC section: "higher power consumption (2.5–3.5W)", no lane rate given

Layout
representative, not one product
Assumed

Assumed: Inside a copper cable head, representative. Why

heat

Redriver heat

See it in 3D ↗
Power and reach
a couple of watts per end, ≈3 m at 200G/lane
Reported

Sources: Vik's Newsletter "pushes copper out to 3 meters at 200G/lane for just a couple of watts per end"

Inside
a redriver on the receive direction; NVIDIA’s LACC has one in each end
Reported

Sources: Marvell Inside the Cable: ACC equalizers amplify signals received from the opposite cable end; the equalizer is analogNVIDIA "LACCs are essentially DAC cables with an additional IC in each end to boost the signal power and noise reduction"

Heat drawn
log scale, each 10× in power 5× the motion; none under 0.5 W
Assumed

Assumed: How much heat each part is drawn with, log scale: each 10× in power is 5× the motion; nothing under 0.5 W. Why

Layout
representative, not one product
Assumed

Assumed: Inside a copper cable head, representative. Why

heat

Retimer heat

See it in 3D ↗
Power
≈20 W per end at 200G/lane; 2.5–3.5 W in one guide, lane rate unstated
Reported

Sources: Vik's Newsletter "[AEC] burns around 20 watts per end"AscentOptics AEC section: "higher power consumption (2.5–3.5W)", no lane rate given

Inside
a DSP retimer in each end: CTLE, DFE, CDR, FIR
Reported

Sources: Marvell Inside the Cable: AEC DSPs at each cable end maintain signal integrity during both transmission and receptionAscentOptics AEC section: "The Retimer chip in AEC is more advanced than the Redriver used in ACC. It integrates CTLE (Continuous-Time Linear Equalizer), DFE (Decision Feedback Equalization), CDR (Clock Data Recovery), and FIR drivers."

Heat drawn
log scale, each 10× in power 5× the motion; none under 0.5 W
Assumed

Assumed: How much heat each part is drawn with, log scale: each 10× in power is 5× the motion; nothing under 0.5 W. Why

Layout
representative, not one product
Assumed

Assumed: Inside a copper cable head, representative. Why

2.1The ledgerwhere 100 MW goes

Main power transformers
0.4 MW
Calc.

Calculated: Loss = the power flowing into this stage × (1 − its assumed or published efficiency). Method

Assumed: Main power transformer efficiency, 99.6%. Why

Campus cables & switchgear
0.3 MW
Calc.

Calculated: Loss = the power flowing into this stage × (1 − its assumed or published efficiency). Method

Assumed: Campus cable & switchgear efficiency, 99.7%. Why

Unit substations
1.0 MW
Calc.

Inputs: U.S. Department of Energy DOE's own summary of the 04/04/2024 final rule updating distribution-transformer efficiency standards -- supersedes the 2013 rule this site cited before

Calculated: Loss = the power flowing into this stage × (1 − its assumed or published efficiency). Method

Assumed: Unit substation efficiency (cooling/building side), ≈99%. Why

UPS, double conversion
3.2 MW
Calc.

Inputs: CEIE (distributor reprint of Eaton figures) distributor summary of Eaton's tested figures: "dual conversion in online mode is as high as 97.5%", eco/AC-direct mode "can be improved to 99%"

Calculated: Loss = the downstream load this stage feeds ÷ its assumed efficiency, minus that load itself. Method

Assumed: UPS double-conversion efficiency, 96.5%. Why

Busway & whips, with 480 → 415 V transformers
1.6 MW
Calc.

Calculated: Loss = the power flowing into this stage × (1 − its assumed or published efficiency). Method

Assumed: Busway, whips and 480 → 415 V transformers, busway and whips 99.5%; transformers 98.7%. Why

Cooling: warm water, dry coolers
4.4 MW
Calc.

Inputs: Google Efficiency page: Google’s own fleet-wide PUE, anchoring the low end of the warm-water/liquid band

Calculated: MW = the IT load × an assumed share of it, picked so the design lands in the cooling class’s published PUE band. Method

Assumed: Cooling and lighting overhead, share of IT load, air 42%, liquid 10%, warm-water 5%; lighting/controls/offices 1.7%. Why

Lighting, controls, offices
1.5 MW
Calc.

Calculated: MW = the IT load × an assumed share of it, picked so the design lands in the cooling class’s published PUE band. Method

Assumed: Cooling and lighting overhead, share of IT load, air 42%, liquid 10%, warm-water 5%; lighting/controls/offices 1.7%. Why

Scale-out switches, 3 tiers
2.7 MW
Calc.

Calculated: MW = a published or estimated per-unit power (per switch or per optical module) × how many units this campus’s fabric needs for its GPU count. Method

Optical transceivers
2.3 MW
Calc.

Inputs: NVIDIA section 4.2, Recommended Operating Conditions and Power Supply Requirements: Maximum Power Dissipation, Max 17 W (the 400G-class per-module figure the fabric model scales from)

Calculated: MW = a published or estimated per-unit power (per switch or per optical module) × how many units this campus’s fabric needs for its GPU count. Method

Rack power shelves, AC → DC
2.1 MW
Calc.

Inputs: NVIDIA system power figures behind the ≈96% AC→54 V PSU efficiency assumed for the DGX H100 caseNVIDIA OCP contribution post describing the GB200 ORv3 power-shelf conversion this figure covers

Calculated: Loss = the downstream load this stage feeds ÷ its assumed efficiency, minus that load itself. Method

Busbar
0.2 MW
Calc.

Calculated: MW = a published or estimated power budget for this component in one rack × the number of racks the site’s power buys. Method

NVLink switch trays (scale-up)
6.7 MW
Calc.

Calculated: MW = a published or estimated power budget for this component in one rack × the number of racks the site’s power buys. Method

Assumed: Per-rack component power budgets (scale-up switching, NICs/DPUs, CPUs, drives/fans/management), from each accelerator’s own table. Why

Grace CPUs + LPDDR5X
8.2 MW
Calc.

Calculated: MW = a published or estimated power budget for this component in one rack × the number of racks the site’s power buys. Method

Assumed: Per-rack component power budgets (scale-up switching, NICs/DPUs, CPUs, drives/fans/management), from each accelerator’s own table. Why

NICs & DPUs
2.7 MW
Calc.

Inputs: NVIDIA "NVIDIA Quantum-2 InfiniBand networking that delivers 400Gb/s bandwidth per GPU" -- the per-GPU NIC rate this row's per-rack figure is built from

Calculated: MW = a published or estimated power budget for this component in one rack × the number of racks the site’s power buys. Method

Assumed: Per-rack component power budgets (scale-up switching, NICs/DPUs, CPUs, drives/fans/management), from each accelerator’s own table. Why

SSDs, fans, management
1.5 MW
Calc.

Calculated: MW = a published or estimated power budget for this component in one rack × the number of racks the site’s power buys. Method

Assumed: Per-rack component power budgets (scale-up switching, NICs/DPUs, CPUs, drives/fans/management), from each accelerator’s own table. Why

Bus converters, 50 → 12 V
1.4 MW
Calc.

Inputs: SemiAnalysis board-level power-delivery survey behind the assumed ≈97–98% intermediate-bus-converter efficiency (no vendor publishes an absolute IBC efficiency at a stated voltage/current point)

Calculated: Loss = the downstream load this stage feeds ÷ its assumed efficiency, minus that load itself. Method

Voltage regulators, 12 → 0.8 V
5.4 MW
Calc.

Inputs: SemiAnalysis board-level power-delivery survey behind the assumed VRM efficiency at ≈0.8 V core voltage (no vendor publishes an absolute figure at a stated current)

Calculated: Loss = the downstream load this stage feeds ÷ its assumed efficiency, minus that load itself. Method

HBM3e memory
7.2 MW
Calc.

Calculated: MW = a published or estimated power budget for this component in one rack × the number of racks the site’s power buys. Method

Assumed: HBM memory, share of GPU package power, h100 10%, gb200 13.2%, gb300 15%, rubin (pre-launch) 16%. Why

Unallocated: rounds down to a whole rack
0.1 MW
Calc.

Calculated: MW = the meter figure, minus every other ledger row, minus the GPU-silicon figure. Building a whole number of racks always leaves a sliver of the meter’s IT budget short of one more rack; this row is that rounding remainder, not a separate piece of equipment. Method

2.2The inventorycounts sized from the scenario

Main power transformers, 75 MVA
3
Calc.

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

34.5 kV feeders
≈10
Calc.

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

Diesel generators, 3 MW
≈40
Calc.

Inputs: Cummins DQKAN generator-set data sheet: 2500 kW (2.5 MW) standby rating for one commercial unit in this class -- this site’s own model rounds to an illustrative 3 MW genset unit for its count, not this specific product’s rating

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

Diesel on site, 48 h
≈1.2 million L
Calc.

Calculated: Volume = campus MW × 48 h × 0.26 L/kWh (the gensets’ published full-load fuel rate), shown in million liters. Method

Battery storage
≈20 MW / 40 MWh
Calc.

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

Rooftop dry coolers
≈121
Calc.

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

Cooling towers
≈6
Calc.

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

Data halls
2
Calc.

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

Unit substations, 2.5 MVA
≈45
Calc.

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

Assumed: Unit substation size, 2.5 MVA. Why

UPS modules, 1.25 MW
≈75
Calc.

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

Assumed: UPS module size, 1.25 MW modeled. Why

Coolant distribution units
≈61
Calc.

Inputs: Vertiv CoolChip CDU family page: models from CDU 70 (70 kW) to CDU 2300 (2300 kW)Motivair (Schneider Electric) brochure: "COOLING UP TO 2.3MW", MCDU-4U (102 kW) through MCDU-60 (2.3 MW) rated-capacity table

Calculated: Count = the meter MW (and, for cooling equipment, the IT load and cooling choice) ÷ the catalog unit size this site’s model assumes for that piece of equipment, with N+1 or N+20% margin where the card notes it. Method

Busway runs
≈63
Calc.

Calculated: Count = the racks (or servers, or per-rack part multiples) the model’s power-and-fabric solve builds for this scenario. Method

GB200 NVL72 racks
≈629
Calc.

Calculated: Racks = the IT-power budget ÷ (one rack’s kW + its share of the scale-out fabric’s per-GPU power), rounded down to a whole rack. Method

Power shelves
≈5,032
Calc.

Calculated: Count = the racks (or servers, or per-rack part multiples) the model’s power-and-fabric solve builds for this scenario. Method

Rectifiers
≈30,192
Calc.

Calculated: Count = the racks (or servers, or per-rack part multiples) the model’s power-and-fabric solve builds for this scenario. Method

NVLink copper connections
≈3.3 million
Calc.

Calculated: Count = the racks (or servers, or per-rack part multiples) the model’s power-and-fabric solve builds for this scenario. Method

Blackwell GPUs
≈45,288
Calc.

Calculated: Count = GPUs (or CPUs, HBM stacks, VRM phases) × the fixed number of that part per GPU or per rack that this accelerator’s bill of materials uses. Method

Grace CPUs
≈22,644
Calc.

Calculated: Count = GPUs (or CPUs, HBM stacks, VRM phases) × the fixed number of that part per GPU or per rack that this accelerator’s bill of materials uses. Method

HBM3e stacks
≈362k
Calc.

Calculated: Count = GPUs (or CPUs, HBM stacks, VRM phases) × the fixed number of that part per GPU or per rack that this accelerator’s bill of materials uses. Method

VRM phases
≈1.1 million
Calc.

Calculated: Count = GPUs (or CPUs, HBM stacks, VRM phases) × the fixed number of that part per GPU or per rack that this accelerator’s bill of materials uses. Method

Transistors in GPUs
≈9.4 quadrillion
Calc.

Calculated: Count = GPUs (or CPUs, HBM stacks, VRM phases) × the fixed number of that part per GPU or per rack that this accelerator’s bill of materials uses. Method

NVLink switch chips
≈11,322
Calc.

Calculated: Count = the racks (or servers, or per-rack part multiples) the model’s power-and-fabric solve builds for this scenario. Method

ConnectX-7 NICs
≈45,288
Calc.

Calculated: Physical endpoints = GPUs × scale-out NICs per GPU (two for Rubin, one for the other modeled generations). The selected fabric uses each physical port’s rate, not aggregate per-GPU bandwidth. Switch, module and fiber counts follow that endpoint inventory and the modeled tiers. Method

Leaf / spine / core switches, 64-port
≈3,540
Calc.

Inputs: NVIDIA QM97xx specifications: 32 OSFP cages, 25.6 Tbps (64 logical 400G ports) -- the 400G-tier radix the fabric model encodesNVIDIA Quantum-X800 XDR specifications -- the 800G-tier radix the fabric model encodes

Calculated: Physical endpoints = GPUs × scale-out NICs per GPU (two for Rubin, one for the other modeled generations). The selected fabric uses each physical port’s rate, not aggregate per-GPU bandwidth. Switch, module and fiber counts follow that endpoint inventory and the modeled tiers. Method

Optical modules
≈159k
Calc.

Inputs: NVIDIA section 4.2: Maximum Power Dissipation, Max 17 W -- the per-module figure the fabric’s optics-power model scales from

Calculated: Physical endpoints = GPUs × scale-out NICs per GPU (two for Rubin, one for the other modeled generations). The selected fabric uses each physical port’s rate, not aggregate per-GPU bandwidth. Switch, module and fiber counts follow that endpoint inventory and the modeled tiers. Method

Fiber strands in the fabric
≈1.1 million
Calc.

Calculated: Physical endpoints = GPUs × scale-out NICs per GPU (two for Rubin, one for the other modeled generations). The selected fabric uses each physical port’s rate, not aggregate per-GPU bandwidth. Switch, module and fiber counts follow that endpoint inventory and the modeled tiers. Method

3.2Linkscopper and optics

DAC — direct-attach copper
reach 2–5 m at 25–100 Gb/s per lane (ratified clauses); at 200 Gb/s per lane, IEEE’s own P802.3dj objective is a floor of at least 1 m — trade estimates for what a design actually achieves spread from under 1 m to about 3 m; power 0 W added — passive
Reported

Sources: Wikipedia "25GBASE-CR (Direct Attach): 5 meters maximum reach... 25GBASE-CR-S: 3 meters"Wikipedia 100GBASE-CR2 (2 lanes, 100G port) PHY row: 3 meters, 802.3cd-2018IEEE 802.3 objectives table: "Reach of up to at least 1.0 meter," stated for 200/400/800/1600 Gb/s passive twin-ax alikeEthernet Alliance the project’s own scope table lists 400 Gb/s-per-lane copper cable reach as not yet defined

ACC — active copper cable
reach ≈3 m at 200 Gb/s per lane, commonly reaching one adjacent rack; 3–5 m at 100 Gb/s per lane (NVIDIA LACC); power ≈2–3 W per end at 200 Gb/s per lane; 1.5 W max per end at 100 Gb/s per lane (NVIDIA LACC, 800G port)
Reported

Sources: NVIDIA "LACCs are essentially DAC cables with an additional IC in each end to boost the signal power and noise reduction. This cable uses a pre-emphasis type of circuit..." (an amplifying IC, not a DSP); the companion MCA4J80-Nxxx datasheet below calls the same part a "pre-emphasis equalizer IC"NVIDIA MCA4J80-Nxxx datasheet: 3/4/5 m reach by wire gauge, 1.5 W max per end, at 100 Gb/s per lane on an 800G portVik's Newsletter "pushes copper out to 3 meters at 200G/lane for just a couple of watts per end"

AEC — active electrical cable
reach Commonly ≈7 m, demonstrated to 9 m; power ≈20 W per end at 200 Gb/s per lane
Reported

Sources: NVIDIA "AECs use digital signal processors (DSPs) at each end to restore and retime signals"Credo product page: a 7 m 800G AEC, built on ahead of the 9 m part belowInfraeo, via PRWeb 9 m 800G OSFP AEC, pre-FEC BER < 1E-8, built on Infraeo’s own prior 7 m 800G AEC (its release names no other maker; Credo’s separate 7 m part is the row above)Vik's Newsletter "[AEC] burns around 20 watts per end"Marvell, via PR Newswire Marvell’s own framing: AEC DSPs for server/NIC-to-ToR and switch-to-switch links

LPO — linear pluggable optics
reach 500 m standard; ≈2 km in a DR variant; power ≈10 W target per port at 200 Gb/s per lane, vs. ≈23–25 W for a fully-retimed DSP module at the same lane rate — port/module capacity is not stated for either figure (Semtech)
Spec

Sources: Semtech "200G LPO Power, Reach and Loss": Semtech’s own 200 Gb/s-per-lane ladder ("currently 23 to 25 watts per module" retimed DSP, "targeting approximately 10 watts" LPO) — the article never states either figure’s port/module capacityLightwave Online LPO MSA’s 100 Gbps/lane spec completed March 31, 2025, a 200 Gbps/lane follow-on in development

DSP pluggable optics
reach 500 m (DR8) to 2 km (FR4/FR8); power 12 W max (400G QSFP-DD, today’s 100 Gb/s/lane generation); ≈15 W (800G 2×FR4, same generation); on the newest 3 nm, 200 Gb/s/lane DSPs: sub-13 W (800G) and sub-23 W (1.6T) — a different chip generation, not a lower price for the same part
Spec

Sources: Cisco datasheet: 400G QSFP-DD (DR4/FR4), 12 W maxSemiAnalysis (Dylan Patel, Daniel Nishball, Myron Xie and others) newsletter figure: 800G 2×FR4 ≈ 15 WBroadcom, via GlobeNewswire Broadcom’s own DSP spec: "sub-13 W" (800G modules) / "sub-23 W" (1.6T modules) on its newest 200 Gb/s-per-lane DSPMarvell, via PRNewswire Marvell’s own release: >20% lower 1.6T module power on its 3 nm, 200 Gbps/lane Ara platform vs. its prior Nova 2 generation

LR4 — direct-detect campus optics
reach 10 km, single-mode — sized to absorb a campus’s routing and conduit slack, no optical amplification needed; power Same DSP-pluggable class as the backbone above (roughly 12–15 W in today’s generation) — a longer-reach optic, not a separate power class
Reported

Sources: Cisco the same datasheet’s LR4 variant sits alongside DR4/FR4 in the DSP-pluggable family cited for the rung aboveL-Com 10 km, single-mode LR4 reach, sized to absorb a campus’s routing and conduit slack

Coherent (400ZR/800ZR-class)
reach ≈40 km unamplified at 400ZR’s standard 11 dB loss budget, extending toward ≈75 km in longer-reach variants (both per OIF); 80–120 km per amplified span for longer routes — repeating amplifiers every ≈80 km extends reach, but noise accumulates with every span, so it is not unlimited; power ≈15–20 W at 400ZR; ≈23–25 W at 800ZR (≈26–30 W in longer-reach 800ZR+ variants)
Reported

Sources: EDGE Optical Solutions "≈40 km unamplified" / "≈120 km amplified" for 400ZR, and 50–75 km unamplified for the longer-reach MZR-class variantAscentOptics "400ZR (Standard)... Typical Power 18 to 20 W... Maximum Power 22 W"; 800ZR "typically... 23 to 25 watts"; 800ZR+ "approximately 20–30W" with real modules rated to 30 W max

CPO — co-packaged optics
reach Same reach class as the pluggables it replaces; this is a switch-side packaging change, without its own separate reach spec; power ≈5.5 W per 800G port, optics plus external laser (Broadcom Bailly, 2025); ≈3.5 W per 800G port (Broadcom Tomahawk 6 “Davisson,” newest) — both exclude the switch ASIC’s own host-side SerDes power
Spec

Sources: The Next Platform body text: "an 800 Gb/sec port on the CPO consumed about 6.4 watts..." vs. Bailly (Tomahawk 5) "took 5.5 watts, which was a 14.1 percent reduction"; and, on Davisson (Tomahawk 6), "An 800 Gb/sec port will burn about 3.5 watts, says Broadcom, which is 36.4 percent lower than... the Tomahawk 5 CPO port" — the 3.5 W figure is this outlet’s direct quote of Broadcom, not a number Broadcom’s own release states in watts (see next ref)Broadcom Broadcom’s own release states "102.4 Tbps" and a ratio, "a 70% reduction in optical interconnect power consumption—more than 3.5x lower than traditional pluggable solutions," not an absolute per-port wattageServeTheHome independent write-up of the same Davisson launch: 102.4 Tbps, sixteen 6.4 Tbps Davisson DR optical engines, "on the order of 70% lower power consumption" — this write-up does not itself give a watts-per-port figure

Where copper runs out, by lane rate
Calc.

Inputs: Wikipedia 25GBASE-CR/CR-S reach, the chart’s 25 Gb/s/lane barWikipedia 100GBASE-CR2/CR1 reach, the chart’s 50 and 100 Gb/s/lane barsIEEE 802.3 the draft P802.3dj objective (≥ 1 m), the chart’s 200 Gb/s/lane bar and its floor tickEthernet Alliance the project’s own scope table lists 400 Gb/s/lane copper reach as not yet defined, the chart’s "not yet defined" bar

Calculated: Each bar plots the reach figure cited for that PAM4 lane rate in the media ladder above (a ratified IEEE clause, a draft objective, or "not yet defined"); the dashed line marks the ≈7 m band where trade sources put the general copper/optics crossover, any lane rate. Method

Inside a DSP pluggable module
Spec

Sources: Cisco a full DSP retimer regenerates the signal at each end of a QSFP-DD/OSFP module, alongside a laser and photodetectorsBroadcom, via GlobeNewswire Broadcom’s own description of its Sian3 DSP’s place inside a retimed pluggable module

Inside a linear pluggable module
Spec

Sources: Semtech the module keeps the laser and photodetectors but drops the DSP; the host chip’s own SerDes drives and reads the line directlyLightwave Online the LPO MSA’s own scope: a linear pluggable module specification with no retiming DSP

Co-packaged optics
Spec

Sources: NVIDIA NVIDIA’s own description of the co-packaged optical engines and "4x fewer lasers" vs. pluggablesNVIDIA NVIDIA’s own description: "Lasers reside on front-panel external laser source pluggable OSFP modules, enabling quick diagnosis and replacement while the switch core remains sealed" — the ELS is a separate side feed, not inline with the traffic fibersBroadcom Broadcom’s own description of Tomahawk 6 "Davisson": optical engines built onto the switch package with field-replaceable laser modules

4The clocknotes under the simulations

Production clusters swing in the 0.2–3 Hz band (Microsoft, OpenAI and NVIDIA, 2025). The 2-second step drawn here sits inside it.
Spec

Sources: Microsoft, OpenAI and NVIDIA researchers, via arXiv Section III-A, "Frequency-domain spec": "AI workload power traces... show FFT energy concentrated between 0.2–3 Hz" (Microsoft, OpenAI and NVIDIA researchers, submitted Aug. 20, 2025)

This generation has no on-rack storage, so every swing reaches the site batteries directly.
Reported

Sources: NVIDIA developer blog: the rack-level capacitor smoothing it describes is a GB300 feature, not present on earlier NVLink generations

GPUs are 61% of the meter here, so one step moves the campus by about 40 MW. Site batteries are rated 20 MW; a full step swing is more than the site batteries can absorb, so the grid still sees part of it.
Calc.

Calculated: GPU share of the meter = GPU MW ÷ meter MW; swing size = GPU MW × (1 − the model’s low-power floor between steps). Site- and rack-battery ratings are read from this model’s own BESS and rack-storage sizing. Method

NFPA 110 Type 10 requires standby power to assume its full rated load within 10 seconds of a utility failure; data centers commonly specify this class for their generators, the timing this scenario assumes.
Spec

Sources: Nixon Power Services comparison table: Type 10 = 10 seconds; "data centers generally use Level 1, Type 10 systems"Cummins Cummins "Ask the Experts" sheet on NFPA 110 Type/Level classes

UPS battery runtime commonly ranges from 1–2 minutes at hyperscale facilities up to 10–15 minutes in the financial sector, far longer than the 10 s they need here.
Reported

Sources: Data Centre Review "hyperscale data centres are being designed with 1-2 minutes of battery runtime... in the financial industry, you will typically see 10-15 minutes of battery runtime"

Chiller restart (≈2 min), outage length (15 min) and the 5 minutes of grid stability before transfer back are illustrative.
Assumed

Assumed: Outage scenario timing, chiller restart ≈2 min, outage 15 min, grid-stable wait 5 min. Why

A 24–40 °C day. Dry coolers alone cannot hold the loop above about 35 °C, so adiabatic sprays switch on and the campus starts using water.
Calc.

Calculated: Chiller-based cooling power scales as clamp(1 + 0.025 × (T − 25 °C), 0.7, 1.6), ≈ 2.5%/°C above 25 °C; warm-water designs hold the loop with dry coolers alone until the adiabatic threshold (35 °C), then draw water. Method

No data-center dry-cooler derating curve was found, so the slopes here are estimates.
Assumed

Assumed: Hot-day cooling response slope, chiller ≈2.5%/°C above 25 °C; dry-cooler/adiabatic curve fit to the adiabatic threshold. Why

The shape (overnight trough, morning rise, midday dip, afternoon peak) comes from measurement studies of real LLM services.
Reported

Sources: Microsoft Azure Research – Systems and UIUC, via arXiv paper: "LLM inference workloads, as user-facing applications, exhibit a typical diurnal pattern with peaks during working hours and valleys at night and weekends," from production Azure LLM traces

No provider publishes a peak-to-trough ratio; 2.5× is illustrative. Idle GPUs are assumed to draw 30% of full power, and the busiest hour runs at 90%.
Assumed

Assumed: Inference day peak-to-trough ratio and idle/peak power, 2.5× peak:trough; idle GPUs at 30% of full power; busiest hour at 90%. Why

4.1Hot to coldone operating point per cooling design

GPU die
≈75 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Rack loop return
≈55 °C
Spec

Sources: NVIDIA NVIDIA blog: "the coolant entering a fully liquid-cooled chip at 45 degrees Celsius exits at roughly 55 degrees"

Facility return
≈52 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Rack loop supply
≈45 °C
Spec

Sources: NVIDIA NVIDIA blog: "the coolant entering a fully liquid-cooled chip at 45 degrees Celsius"

Facility supply
≈42 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Outdoor air, hot day
≈35 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

1Real campusesfacts behind the presets

IT power operating, buildings 1–4 of 8
≈421 MW
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" 421MW

Planned
≈1.2 GW (Epoch, 04/2026 estimate)
Reported

Sources: Epoch AI site profile: "Projected capacity: 1.2 GW | 1.0 million H100-equivalents"

Grid
double 345 kV corridor (Midland–Graham), Mulberry Creek substation
Reported

Sources: Yes Energy body text: "The facility is fed by a double 345 kV corridor that goes from Midland to Graham, Texas," and "a substation next to the facility called Abilene Mulberry Creek"

On-site generation
natural gas reported by Epoch (04/2026); not itemized in the current tracker
Reported

Sources: Epoch AI body text: "Power is currently supplied by a mix of on-site natural gas and grid power, which includes local wind power"

GPUs, per Elon Musk, 09/25/2026
230k: 150k H100, 50k H200, 30k GB200
Reported

Sources: Elon Musk, on X post on X, 09/25/2026: "Colossus 1 is 150k H100, 50k H200 and 30k GB200. Colossus 2 is 110k GB200 and 440k GB300."

GPUs, reported earlier
≈200,000 (H100, H200, some GB200)
Reported

Sources: Tom's Hardware headline: "fully operational with 200,000 GPUs"

Phase 2 power
≈300 MW
Reported

Sources: Tom's Hardware headline: "Phase 2 to consume 300 MW"

Grid supply
MLGW/TVA, ≈150 MW
Reported

Sources: Interesting Engineering body text: "receiving 150MW of grid power from MLGW and TVA"

On-site generation
≈35 gas turbines, ≈422 MW
Reported

Sources: Compute Atlas facility profile: on-site generation "422 MW" via "approximately 35 natural-gas turbines"

Batteries
Tesla Megapacks, ≈150 MW
Reported

Sources: Interesting Engineering body text: "150MW of Megapack Batteries for stored energy backup have been integrated"

Backup, per SpaceXAI
more than 240 batteries, enough for the site to come fully off the grid
Spec

Sources: SpaceXAI (SpaceX) Power tab, "Installing resilient, sustainable backup power": "more than 240 batteries so Colossus I can come completely offline during emergencies or peak demand"

Cooling water, per SpaceXAI
≈820,000 gal a day, hybrid system
Spec

Sources: SpaceXAI (SpaceX) Water tab, "What Colossus uses today": "Colossus I has a hybrid system that uses about 820,000 gallons a day"

GPUs, per Elon Musk, 09/25/2026
550k: 110k GB200 + 440k GB300
Reported

Sources: Elon Musk, on X post on X, 09/25/2026: "Colossus 1 is 150k H100, 50k H200 and 30k GB200. Colossus 2 is 110k GB200 and 440k GB300."Tom’s Hardware article, 09/25/2026, quoting the postGround News (aggregating wire reporting) aggregated wire coverage (~09/25/2026): "110k GB200 and 440k GB300"

Coming, per Elon Musk
+220k GB300 “next week”, +220k in November, +220k by late December “if we get lucky”
Reported

Sources: Elon Musk, on X same post: "Another 220k GB300 will be fully operational next week and another 220k in November. If we get lucky, yet another 220k GB300 by late December."

Why 110k, per Elon Musk
the number of fiber optic cables that can plug into a central switch
Reported

Sources: Elon Musk, on X follow-up post on X, 09/25/2026: "The strange multiple of 110k is due to the number of fiber optic cables that can be plugged into a central switch"

IT power running, satellite estimate, 09/24/2026
≈946 MW
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" 946MW

Chips, satellite count, 09/24/2026
≈440,000: 110k GB200, 330k GB300
Reported

Sources: Epoch AI directory entry, hardware table: GB200 "110k", GB300 "330k"

GPUs planned, per SpaceXAI
1M+
Spec

Sources: SpaceXAI (SpaceX) Colossus II tab: "GPUs planned 1M+"

Battery, per SpaceXAI
3.3 GWh grid-connected pack, planned (Riley Trettel to the TVA board, 08/20/2026; TVA approved a direct grid hookup that day)
Spec

Sources: Canary Media body text: on Aug. 20 SpaceXAI energy and data center developer Riley Trettel told the TVA board the battery has "3.3 GWh of storage"; "The board approved a direct hookup to the TVA grid for SpaceXAI later that day"SpaceXAI (SpaceX) Colossus II tab, Power: "America’s largest grid-connected battery pack will provide 3.3 gigawatt hours"

Grid, per SpaceXAI
$55M for two MLGW substations in Memphis, one of them 150 MW (which campus they serve is not said)
Spec

Sources: SpaceXAI (SpaceX) Power tab, "SpaceXAI is investing in a stronger electric grid": "$35 million in a 150 MW substation to support MLGW and another $20 million in a second substation"

Gas turbines, per SpaceXAI
temporary units in TN and MS since 08/01/2025; all out by July 2027
Spec

Sources: SpaceXAI (SpaceX) Air tab: "After temporary turbines started on August 1, 2025"; "all remaining temporary turbines must be removed by July 2027" under an agreed order with the Mississippi DEQ

Diesel backup
not mentioned by SpaceXAI
Spec

Sources: SpaceXAI (SpaceX) all five tabs (Colossus I, Colossus II, Water, Power, Air), checked 09/27/2026: no diesel generators are mentioned

Cooling, per SpaceXAI
closed loop, domestic water only
Spec

Sources: SpaceXAI (SpaceX) Water tab, "What Colossus uses today": "Colossus II uses closed-loop cooling and takes only domestic water"

Cooling plant, Aug 2025
119 air-cooled chillers, ≈200 MW
Reported

Sources: SemiAnalysis body text: "119 air-cooled chillers on site, i.e. roughly 200MW of cooling capacity" as of August 22, 2025

SemiAnalysis framing, 09/2025
titled "First Gigawatt Datacenter"; ≈200 MW of cooling then, ≈946 MW IT now
Reported

Sources: SemiAnalysis headline: "xAI’s Colossus 2 — First Gigawatt Datacenter In The World"

IT power running, satellite estimate
≈636 MW, 4 of 9 Main Campus buildings
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" 636MW; body text on buildings 1–4 liveMeasured AI (Substack) body text: "The QTS Fayetteville Main Campus comprises nine data center buildings — four operational, five under construction"

Planned
≈1.5 GW across all 13 buildings (Fairwater + Main + East Campus)
Reported

Sources: Measured AI (Substack) body text: the full QTS Fayetteville campus "designed to eventually reach approximately 1,500 MW across all 13 buildings"

Accelerators
NVIDIA Blackwell NVL72 (GB200/GB300 family), ≈140 kW per rack
Spec

Sources: Microsoft body text: "~140kW per rack, 1,360 kW per row"Measured AI (Substack) body text: "Each NVL72 rack contains 72 NVIDIA Blackwell GPUs... roughly 140 kW of rack power"

Cooling
closed-loop liquid, no evaporation
Spec

Sources: Microsoft body text: a "closed-loop approach that reuses the liquid continuously after the initial fill with no evaporation"Measured AI (Substack) body text: "The cooling loop is filled once and recirculated for the life of the building"

Backup power
no on-site generation, UPS or dual-corded distribution (grid-only design)
Spec

Sources: Microsoft body text: Microsoft can "forgo traditional resiliency approaches for the GPU fleet (such as on-site generation, UPS systems and dual-corded distribution)"Measured AI (Substack) body text: "The Fairwater template strips on-site power generation, UPS systems, and dual-corded distribution out of the buildings’ electrical topology entirely"

Power
Building 1 ≈369 MW live; $3.3B initial investment
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" 369MWMicrosoft (On the Issues) body text: "$3.3 billion investment pledge"

Accelerators
GB200 NVL72 racks (NVLink 5.0 inside), racks linked by 800G Ethernet
Reported

Sources: TechTimes body text: "NVIDIA’s GB200 NVL72 systems connect 72 Blackwell GPUs through NVLink 5.0"; separately, "connecting the racks is a two-level networking tree built on 800G Ethernet"

Cooling
closed-loop liquid, filled once
Spec

Sources: Microsoft (On the Issues) body text: "a state-of-the-art closed-loop liquid cooling system, filled during construction and recirculated continuously"

Opened
06/23/2026
Reported

Sources: Microsoft Source headline/dateline: construction completed, announced Tuesday, June 23, 2026

Phase 1
≈2 GW by 2030
Reported

Sources: IEEE Spectrum body text: "the first phase—a 2-GW version—will be completed by 2030"

Planned
5 GW at full build (~2032), more than $50B
Spec

Sources: Meta body text: "delivering 5 gigawatts of compute capacity to house Hyperion" and "$50B+ Data center investment"Wikipedia body text: "full build-out around 2032"

GPUs, planned
more than 3 million at full build (reporter estimate)
Reported

Sources: IEEE Spectrum body text: "Hyperion could include over 41,000 rack-scale systems, for a total of more than 3 million GPUs"

New gas generation
2.26 GW combined, three turbine plants
Reported

Sources: IEEE Spectrum body text: Entergy’s "three new gas-turbine power plants... will generate a combined 2.26 GW"Wikipedia body text: "Entergy has begun building three gas-powered combined-cycle power plants"

IT power operating, ~16 of ~30 buildings live
≈910 MW
Reported

Sources: Epoch AI directory entry: "CURRENT IT POWER" 910MW

Planned full campus, due ~Q1 2028
≈1,925 MW (1.9 GW)
Reported

Sources: Epoch AI directory entry: "PROJECTED IT POWER" 1,925MW

Chips, Oct. 2025 launch vs. 09/2026
≈500,000 Trainium2 at launch (64-chip UltraServers); ≈1,045,000 by Epoch's 09/24/2026 count
Reported

Sources: Compute Atlas facility profile: "Phase 1 online with ~500,000 Trainium2 chips," operational October 2025Epoch AI directory entry: current Trainium2 count "1,045k"

Grid interconnection
Indiana Michigan Power (AEP), existing Olive 345 kV station
Reported

Sources: Measured AI (Substack) body text: "the campus already adjoined... the Olive 345 kV station" on "a purpose-built 345 kV transmission delivery scheme"Compute Atlas facility profile: utility "Indiana Michigan Power (AEP)"

On-site generation
none for operating power; diesel gensets for backup only
Reported

Sources: Measured AI (Substack) body text: "All operating power comes from the utility grid... The only on-site generation is a diesel fleet held in reserve for emergency backup"

IT power operating, Epoch satellite estimate
≈471 MW
Reported

Sources: Epoch AI directory entry: "471 MW of IT power," as of 08/31/2026

Compute, satellite estimate
≈536k H100-eq compute; ≈166k Nvidia B300 chips
Reported

Sources: Epoch AI directory entry: current chips "166k" B300 delivering "536k H100-eq" (08/31/2026)

Planned, by Q3 2028
≈1,022 MW IT power, ≈$38.7B total capital
Reported

Sources: Epoch AI directory entry: "PROJECTED IT POWER" 1,022MW by Q3 2028, with the tracker’s total-capital estimate

On-site generation
Socrates North + South gas plants, ≈400 MW, fully off-grid
Reported

Sources: MLQ.ai body text: "400MW of dedicated, off-grid natural gas generation," split between "the 200MW Socrates South plant" and a "twin 200MW Socrates North facility"

Construction
5 tent structures (~125,000 sq ft each) built Apr–Jun 2026, plus traditional colocation buildings
Reported

Sources: MLQ.ai body text: "Five ~125,000-sq-ft tent structures at its New Albany, Ohio campus between April and June 2026," built in about 3 months versus 18–36 months for traditional halls

▸The toursthe figures each narrated stop states

Meter power
100 MW
Assumed

Assumed: Modeled campus size, set by the reader. Why

American homes, same power
≈83,000
Calc.

Calculated: Homes = meter kW ÷ 1.2 kW, a round figure for the average continuous electricity draw of an American home (1.2 kW × 8,760 hours ≈ 10,500 kWh a year), rounded to two figures. For scale only. Method

US grid losses, 2018–2022
≈5% (EIA)
Reported

Sources: U.S. Energy Information Administration FAQ answer: "annual electricity transmission and distribution (T&D) losses averaged about 5% of the electricity transmitted and distributed in the United States in 2018 through 2022"

1,000 km, propagation only
≈5 ms one way, ≈10 ms round trip
Calc.

Calculated: One-way delay = 1,000 km × the fiber-speed figure above (≈4.9 µs/km) ≈ 5 ms; round trip doubles it to ≈10 ms. Switching, routing and queueing are not included. Method

DeepMind Decoupled DiLoCo
4 US regions over 2–5 Gb/s
Reported

Sources: Google DeepMind blog post: "We successfully trained a 12 billion parameter model across four separate U.S. regions using 2-5 Gbps of wide-area networking"

Voltage
345 kV line-to-line
Assumed

Assumed: Campus interconnection voltage, 345 kV. Why

Current, 100 MW
≈90 A per phase, on each circuit
Calc.

Calculated: Per-phase current on each of two equally loaded circuits = campus MW ÷ (2 × √3 × 345 kV × power factor). Method

Same power at 415 V
≈146,000 A
Calc.

Calculated: Current = meter W ÷ (√3 × 415 V × 0.95) for three-phase AC, at the same 0.95 power factor the voltage staircase uses, or meter W ÷ 800 V for DC, rounded to the nearest thousand amps. It shows why power travels at high voltage; no conductor carries a whole campus at rack voltage. Method

Rating
3 × 75 MVA, N+1
Calc.

Calculated: Count = ceil(campus MW ÷ (transformer unit MVA × assumed 0.95 power factor)) + 1 spare (N+1); the model buys 300 MVA units above 400 MW and 75 MVA units below. Method

Efficiency, 345 kV class
>99.6% at all loading levels
Spec

Sources: PA Transformer product page: "High efficiency: exceeding 99.6% at all loading levels"

Loss at 100 MW
≈0.4 MW
Calc.

Calculated: Loss = campus MW × (1 − main-transformer efficiency). Method

Units here
≈40, N+20%
Calc.

Calculated: Units = ceil(campus MW ÷ 3 MW × 1.2) for N+20% redundancy on 3 MW-class sets. Method

Size here
≈20 MW / 40 MWh
Calc.

Calculated: Power = 0.2 × campus MW; energy = 0.4 × campus MW (a 2-hour duration at that power) — a round sizing this model picks to show scale, not a vendor design. Method

Training load swings
tens to hundreds of MW, seconds
Reported

Sources: Microsoft, OpenAI and NVIDIA researchers, via arXiv Section I (Introduction): "these swings can amount to tens or hundreds of megawatts" (Microsoft, OpenAI, NVIDIA)

Cooling power
4.4 MW
Calc.

Calculated: Cooling power = coolFrac × IT MW, where coolFrac is calibrated so each cooling design lands in its publicly reported PUE band once the power chain’s own losses are added. Method

PUE, this design
1.14
Calc.

Calculated: IT MW solves meter MW × upstream efficiency = IT MW ÷ IT-path efficiency + (cooling fraction + misc fraction) × IT MW ÷ side-path efficiency, chaining the model’s own conversion efficiencies; PUE = meter MW ÷ IT MW. Method

Water on site, a day
≈336 m³
Calc.

Calculated: Water = WUE (L/kWh IT) × IT MW × 24 h, converted to cubic meters. Method

Assumed: Water use by cooling design, air 1.0, liquid 0.5, warm water 0.16 L/kWh IT. Why

Efficiency, Eaton 9395XR
up to 97.5% online, 99% eco
Reported

Sources: CEIE (distributor reprint of Eaton figures) "tested efficiency of Eaton 9395XR UPS dual conversion in online mode is as high as 97.5%"; eco/AC-direct mode "can be improved to 99%" (Eaton’s own product page could not be opened this pass -- see its unchecked note)

Loss at 100 MW
≈3.2 MW
Calc.

Calculated: Loss = the downstream load this stage feeds ÷ its assumed efficiency, minus that load itself. Method

Assumed: UPS double-conversion efficiency, 96.5%. Why

IT load
88 MW
Calc.

Calculated: IT MW solves meter MW × upstream efficiency = IT MW ÷ IT-path efficiency + (cooling fraction + misc fraction) × IT MW ÷ side-path efficiency, chaining the model’s own conversion efficiencies; PUE = meter MW ÷ IT MW. Method

Compute racks, campus
629 × ≈131 kW = 83 MW
Calc.

Calculated: Racks = floor(IT MW × 1000 ÷ (one rack’s kW + its GPUs’ share of the fabric’s kW)). Method

Network, switches and optics
5.0 MW
Calc.

Calculated: MW = a published or estimated per-unit power (per switch or per optical module) × how many units this campus’s fabric needs for its GPU count. Method

GPUs
45,288
Calc.

Calculated: GPUs = racks × GPUs per rack. Method

Scale-out switches, 3 tiers
2.7 MW
Calc.

Calculated: MW = a published or estimated per-unit power (per switch or per optical module) × how many units this campus’s fabric needs for its GPU count. Method

Optical transceivers
2.3 MW
Calc.

Inputs: NVIDIA section 4.2, Recommended Operating Conditions and Power Supply Requirements: Maximum Power Dissipation, Max 17 W (the 400G-class per-module figure the fabric model scales from)

Calculated: MW = a published or estimated per-unit power (per switch or per optical module) × how many units this campus’s fabric needs for its GPU count. Method

Fabric strands, whole campus
≈1.1 million
Calc.

Calculated: Physical endpoints = GPUs × scale-out NICs per GPU (two for Rubin, one for the other modeled generations). The selected fabric uses each physical port’s rate, not aggregate per-GPU bandwidth. Switch, module and fiber counts follow that endpoint inventory and the modeled tiers. Method

Efficiency
≈97.5% peak
Spec

Sources: Flex product page: "High efficiency up to 97.5% (peak)"

Busbar current
≈2,600 A
Calc.

Calculated: Current = rack DC bus power ÷ busbar voltage. Method

Per GPU
NVLink 5, 1.8 TB/s
Spec

Sources: NVIDIA "The revolutionary 1.8 TB/s of bidirectional throughput per GPU"

Trays
9, 2 switch chips each (18)
Spec

Sources: NVIDIA Hardware: "18x 1RU compute trays, each with 2 Grace CPUs and 4 Blackwell GPUs"; "9x 1RU NVLink switch trays" of "2x NVLink NVSwitches" eachNVIDIA System Hardware & Components: 18 compute trays, 9 NVLink switch trays of 2 NVSwitch ASICs each

A 400G module, e.g.
NVIDIA single-port 400G OSFP, DR4
Spec

Sources: NVIDIA specifications: single-port 400G OSFP DR4 transceiver

Kind
Mach-Zehnder, a common silicon-photonics modulator design
Reported

Sources: Optica (OFC 2019, Th4A.6) abstract: single-lane 200G optical interconnects demonstrated with a silicon photonic Mach-Zehnder modulator (the earlier "commercial mainstream" wording came from a trade newsletter whose link is dead and is not repeated here)

Core current
≈1,300 A
Calc.

Calculated: Current = GPU power less its HBM share (the HBM stacks have their own supply) ÷ an assumed 0.8 V core voltage. Method

Efficiency
≈91%
Assumed

Assumed: Voltage-regulator (VRM) efficiency, 90–92%. Why

Loss, campus-wide
≈5.4 MW
Calc.

Calculated: Per-rack loss = power passing through the stage × (1 ÷ assumed stage efficiency − 1); the campus figure sums this across every rack the scenario builds. Method

GPU dies, whole campus
47 MW
Calc.

Calculated: Per GPU, the package power less its HBM share; for the campus, that times the GPU count, and as a share, that over the meter MW. The ledger’s rows, this, and the unallocated remainder add up to the meter power. It covers everything on the compute dies (computation, on-chip memory, communication, control and leakage); the model does not estimate the part that is useful arithmetic. Method

Share of the meter
47%
Calc.

Calculated: Per GPU, the package power less its HBM share; for the campus, that times the GPU count, and as a share, that over the meter MW. The ledger’s rows, this, and the unallocated remainder add up to the meter power. It covers everything on the compute dies (computation, on-chip memory, communication, control and leakage); the model does not estimate the part that is useful arithmetic. Method

Bandwidth
8 TB/s
Spec

Sources: NVIDIA specifications: "372 GB" HBM3e and "16 TB/s" per Grace Blackwell Superchip (2 GPUs)Micron specifications table: 24 GB at 8-Hi, >1.2 TB/s per stack

Tokens a second, campus
54.3 million
Calc.

Calculated: Tokens a second = GPUs × tokens a second per GPU × utilization, from the calculator’s settings. The per-GPU rate defaults to a figure scaled from each accelerator’s HBM bandwidth until the reader sets one. Method

Assumed: Default tokens a second per GPU, 2,000 for an 8 TB/s GPU, scaled with HBM bandwidth. Why

Tokens per kWh at the meter
1.8 million
Calc.

Calculated: Joules per token = meter W ÷ tokens a second, plus, when the calculator includes it, a share of training: training energy ÷ the tokens the model serves over its life. Tokens per kWh = 3.6 million J ÷ joules per token. Method

Meter power
100 MW
Assumed

Assumed: Modeled campus size, set by the reader. Why

This step, per watt at the meter
0.007 W
Calc.

Calculated: Each step’s slice = the ledger rows for that step ÷ meter MW: the main transformers and campus cables; cooling and lighting; the power room; the network switches and optics; rack power; the NVLink switches; host CPUs, NICs and drives; converters and regulators; memory. What is left reaches the GPU dies. The unallocated remainder, never drawn, gets no step of its own. Method

This step, per watt at the meter
0.059 W
Calc.

Calculated: Each step’s slice = the ledger rows for that step ÷ meter MW: the main transformers and campus cables; cooling and lighting; the power room; the network switches and optics; rack power; the NVLink switches; host CPUs, NICs and drives; converters and regulators; memory. What is left reaches the GPU dies. The unallocated remainder, never drawn, gets no step of its own. Method

PUE, this design
1.14
Calc.

Calculated: IT MW solves meter MW × upstream efficiency = IT MW ÷ IT-path efficiency + (cooling fraction + misc fraction) × IT MW ÷ side-path efficiency, chaining the model’s own conversion efficiencies; PUE = meter MW ÷ IT MW. Method

This step, per watt at the meter
0.058 W
Calc.

Calculated: Each step’s slice = the ledger rows for that step ÷ meter MW: the main transformers and campus cables; cooling and lighting; the power room; the network switches and optics; rack power; the NVLink switches; host CPUs, NICs and drives; converters and regulators; memory. What is left reaches the GPU dies. The unallocated remainder, never drawn, gets no step of its own. Method

This step, per watt at the meter
0.050 W
Calc.

Calculated: Each step’s slice = the ledger rows for that step ÷ meter MW: the main transformers and campus cables; cooling and lighting; the power room; the network switches and optics; rack power; the NVLink switches; host CPUs, NICs and drives; converters and regulators; memory. What is left reaches the GPU dies. The unallocated remainder, never drawn, gets no step of its own. Method

This step, per watt at the meter
0.023 W
Calc.

Calculated: Each step’s slice = the ledger rows for that step ÷ meter MW: the main transformers and campus cables; cooling and lighting; the power room; the network switches and optics; rack power; the NVLink switches; host CPUs, NICs and drives; converters and regulators; memory. What is left reaches the GPU dies. The unallocated remainder, never drawn, gets no step of its own. Method

This step, per watt at the meter
0.067 W
Calc.

Calculated: Each step’s slice = the ledger rows for that step ÷ meter MW: the main transformers and campus cables; cooling and lighting; the power room; the network switches and optics; rack power; the NVLink switches; host CPUs, NICs and drives; converters and regulators; memory. What is left reaches the GPU dies. The unallocated remainder, never drawn, gets no step of its own. Method

This step, per watt at the meter
0.125 W
Calc.

Calculated: Each step’s slice = the ledger rows for that step ÷ meter MW: the main transformers and campus cables; cooling and lighting; the power room; the network switches and optics; rack power; the NVLink switches; host CPUs, NICs and drives; converters and regulators; memory. What is left reaches the GPU dies. The unallocated remainder, never drawn, gets no step of its own. Method

This step, per watt at the meter
0.068 W
Calc.

Calculated: Each step’s slice = the ledger rows for that step ÷ meter MW: the main transformers and campus cables; cooling and lighting; the power room; the network switches and optics; rack power; the NVLink switches; host CPUs, NICs and drives; converters and regulators; memory. What is left reaches the GPU dies. The unallocated remainder, never drawn, gets no step of its own. Method

This step, per watt at the meter
0.072 W
Calc.

Calculated: Each step’s slice = the ledger rows for that step ÷ meter MW: the main transformers and campus cables; cooling and lighting; the power room; the network switches and optics; rack power; the NVLink switches; host CPUs, NICs and drives; converters and regulators; memory. What is left reaches the GPU dies. The unallocated remainder, never drawn, gets no step of its own. Method

Reaches the GPU dies, per watt
0.472 W
Calc.

Calculated: Per GPU, the package power less its HBM share; for the campus, that times the GPU count, and as a share, that over the meter MW. The ledger’s rows, this, and the unallocated remainder add up to the meter power. It covers everything on the compute dies (computation, on-chip memory, communication, control and leakage); the model does not estimate the part that is useful arithmetic. Method

Timing
illustrative, the same for every GPU
Assumed

Assumed: An illustrative request timeline, internet ≈25 ms each way, fiber vault 1 ms, queue 50 ms, front-end network 0.2 ms, prefill 200 ms, 60 tokens a second per stream, a 500-token reply. Why

Speed in fiber
≈4.9 µs per km, one way
Calc.

Calculated: Speed = speed of light ÷ silica fiber’s group refractive index (≈1.47) ≈ 2.04×10⁸ m/s, which is ≈4.9 µs per kilometer, one way. Method

Routes
2 or more, physically separate
Reported

Sources: TRG Datacenters "A properly designed facility has dual fiber entrances. Fiber enters from two separate locations, following different physical paths into the building."

A serving copy, this tour
one rack, 72 GPUs
Assumed

Assumed: One serving copy per NVLink domain, one NVL72 rack (72 GPUs), or one 8-GPU server for H100. Why

Timing
illustrative, the same for every GPU
Assumed

Assumed: An illustrative request timeline, internet ≈25 ms each way, fiber vault 1 ms, queue 50 ms, front-end network 0.2 ms, prefill 200 ms, 60 tokens a second per stream, a 500-token reply. Why

Traffic
every layer, many times per token
Reported

Sources: Meta AI, arXiv describes tensor-parallel GPUs exchanging activations inside every transformer layer, many times per forward/backward pass

NVLink per GPU
1.8 TB/s
Spec

Sources: NVIDIA Blackwell platform launch release: "the latest iteration of NVIDIA NVLink delivers groundbreaking 1.8TB/s bidirectional throughput per GPU" (the NVL72 product page itself states only the 130 TB/s system aggregate, not a per-GPU figure)

DPUs per tray
2
Spec

Sources: NVIDIA Compute tray includes two BlueField-3 DPUs.

Timing
illustrative, the same for every GPU
Assumed

Assumed: An illustrative request timeline, internet ≈25 ms each way, fiber vault 1 ms, queue 50 ms, front-end network 0.2 ms, prefill 200 ms, 60 tokens a second per stream, a 500-token reply. Why

Prefill, this example
200 ms
Assumed

Assumed: An illustrative request timeline, internet ≈25 ms each way, fiber vault 1 ms, queue 50 ms, front-end network 0.2 ms, prefill 200 ms, 60 tokens a second per stream, a 500-token reply. Why

Time to first token, this example
≈301 ms
Calc.

Calculated: The sum of the illustrative steps before the first word is back on the phone: the internet in (25 ms), the fiber vault (1 ms), the queue (50 ms), the front-end network (0.2 ms), prefill (200 ms) and the internet back (25 ms). Method

Stream rate and reply length
60 tokens/s, 500 tokens
Assumed

Assumed: An illustrative request timeline, internet ≈25 ms each way, fiber vault 1 ms, queue 50 ms, front-end network 0.2 ms, prefill 200 ms, 60 tokens a second per stream, a 500-token reply. Why

Energy per reply, at the meter
≈0.28 Wh, training share included
Calc.

Calculated: Wh per reply = joules per token × 500 tokens ÷ 3,600. That is meter energy, cooling included; the IT equipment’s part is that ÷ PUE. Method

Bandwidth
8 TB/s
Spec

Sources: NVIDIA specifications: "372 GB" HBM3e and "16 TB/s" per Grace Blackwell Superchip (2 GPUs)Micron specifications table: 24 GB at 8-Hi, >1.2 TB/s per stack

Energy per reply, at the meter
≈0.28 Wh, training share included
Calc.

Calculated: Wh per reply = joules per token × 500 tokens ÷ 3,600. That is meter energy, cooling included; the IT equipment’s part is that ÷ PUE. Method

IT energy per reply
≈0.25 Wh
Calc.

Calculated: Wh per reply = joules per token × 500 tokens ÷ 3,600. That is meter energy, cooling included; the IT equipment’s part is that ÷ PUE. Method

Water on site, per reply
≈0.039 mL
Calc.

Calculated: mL = the reply’s IT energy in Wh × WUE in liters per kWh of IT energy (Wh × L/kWh = mL). WUE is defined on IT energy, so the meter figure is divided by PUE first. Method

Assumed: Water use by cooling design, air 1.0, liquid 0.5, warm water 0.16 L/kWh IT. Why

Package power
≈1,200 W
Reported

Sources: SemiAnalysis / InferenceX the chip’s specifications table: "1,200 W" per GPU

Compute dies, this model
≈1,042 W
Calc.

Calculated: Per GPU, the package power less its HBM share; for the campus, that times the GPU count, and as a share, that over the meter MW. The ledger’s rows, this, and the unallocated remainder add up to the meter power. It covers everything on the compute dies (computation, on-chip memory, communication, control and leakage); the model does not estimate the part that is useful arithmetic. Method

Die area, two dies
≈16 cm²
Reported

Sources: WCCFTech body text: "AI Superchip - 208 Billion Transistors (TSMC 4NP, >1600mm2)"

Average flux
≈65 W/cm²
Calc.

Calculated: Flux = GPU power less its HBM share ÷ die area. Hot spots run well above this average. Method

Die, this operating point
≈75 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Layers to coolant
die, interface, lid, interface, cold plate
Assumed

Assumed: Layers between die and coolant, die → thermal interface → lid → thermal interface → cold plate or heat sink; H100: no lid. Why

Lid, this operating point
≈67 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Heat per GPU
≈1.2 kW
Reported

Sources: SemiAnalysis / InferenceX chip spec table: "TDP per chip: 1,200 W"

NVIDIA warm-water spec
45 °C in, ≈55 °C out
Spec

Sources: NVIDIA NVIDIA blog: "the coolant entering a fully liquid-cooled chip at 45 degrees Celsius exits at roughly 55 degrees"

To liquid, this model
≈114 kW
Calc.

Calculated: Split the rack total by the liquid-cooled share this site assumes for this rack (the cited vendor spec sheet could not be re-read in the 10/08/2026 audit). Method

Rise across the rack
≈10 °C
Reported

Sources: Alliance Chemical "roughly 170–195 liters per minute of coolant at a 10°C inlet-to-outlet rise (about 1.5 LPM per kW)" for a GB200 NVL72 rack

Supply → return, this design
≈45 → 55 °C
Spec

Sources: NVIDIA NVIDIA blog: "the coolant entering a fully liquid-cooled chip at 45 degrees Celsius exits at roughly 55 degrees"

Approach, facility to rack loop
a few °C (≈3 °C here)
Assumed

Assumed: CDU approach temperature, facility to rack loop, a few °C (≈3 °C in this model). Why

Facility water, supply → return
42 → 52 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Carries
nearly all of the hall’s heat
Assumed

Assumed: Standard switchgear, battery and busway design practice, automatic utility/generator transfer; Li-ion (LFP or NMC) UPS batteries; tap-off busway. Why

Facility water return
≈52 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Heat rejected
≈92 MW
Calc.

Calculated: Heat rejected ≈ IT MW × 1.05, adding back the plant’s own electrical overhead as heat. Method

Facility water supply, back to the hall
≈42 °C
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Sprays needed above
≈35 °C outside
Assumed

Assumed: One operating point for the cooling loops, rack loop +10 °C across a rack; the CDU ≈3 °C apart on each side; die ≈20 °C above the coolant leaving its plate; chiller-made water 12 °C (air) or 20 °C (liquid); supply air 22 °C with an 18 °C rise (air); condenser water 35 °C; a 35 °C design afternoon. Why

Heat out
100 MW
Calc.

Calculated: Heat out ≈ meter MW: energy conservation means virtually every watt drawn eventually becomes heat. Method

Bibliography

Every source

Grouped by who published it, with how many claims on the site rest on each.

Acacia Communications

Alliance Chemical

AMD ROCm documentation

arXiv

AscentOptics

ASHRAE

ASHRAE TC 9.9, hosted via Vertiv

Astera Labs

ATK Energy Group

AWS

AWS Neuron documentation

Ayar Labs

Ayar Labs, release reprinted by HPCwire

Baltimore Aircoil Company

BlackRidge Research

Broadcom

Broadcom, via GlobeNewswire

Business Insider, via Yahoo Finance

Canary Media

Caterpillar

CDC

CEIE (distributor reprint of Eaton figures)

Ciena

Cignal AI

Cisco

CNBC

Coherent

Compute Atlas

Consulting-Specifying Engineer

Converge Digest

Corning

Credo

Credo Technology Group

Cummins

Data Center Dynamics

Data Center Frontier

Data Centre Review

Daudlin et al. / OFC 2021, author-hosted copy

DeepSeek-AI, arXiv

Delta Electronics

Eaton

EDGE Optical Solutions

EFFECT Photonics

Eoptolink

Epoch AI

Ethernet Alliance

Eurelectric

EVAPCO

  • APEX dry coolerprimary · published 2026 · checked 09/27/2026 · background

Fiber Broadband Association

FiberMall

Flexoptix

Fortum

FS.com

Gazettabyte

GIGABYTE

Google

Google (Jouppi et al., ISCA 2023)

Google (Poutievski et al., SIGCOMM 2022)

Google Cloud

Google Cloud documentation

Google DeepMind

  • Decoupled DiLoCoprimary · published 2026 · checked 09/27/2026 · backs 2 figures

Google, arXiv

GridReadiness

Ground News (aggregating wire reporting)

HVAC Tools Kit

IEEE 802.3

IEEE 802.3cu Task Force

IEEE Spectrum

IndustrialSage

Infraeo, via PRWeb

InnoLight

Intel

Interesting Engineering

Introl

Juniper Networks

L-Com

Lambda

LD-PD

Lenovo Press

Lightmatter

Lightwave Online

link-pp.com

Linux Foundation / Ultra Ethernet Consortium

LITE-ON

LITEON

LITEON, via PR Newswire

Locuza (hardware analyst)

Louisiana Economic Development

LPO MSA

  • FAQsprimary · published undated · checked 09/30/2026 · backs 2 figures

Lumentum

MapYourTech

Marvell

Marvell (Chris McCormick)

Marvell (Radha Nagarajan)

Marvell (SEC Form 8-K exhibit 99.1)

Marvell, via PR Newswire

Marvell, via PRNewswire

Measured AI (Substack)

Meta

Meta AI, arXiv

Micas Networks

Micron

  • HBM4primary · checked 09/27/2026 · background
  • HBM3E product briefprimary · published 10/2023 · checked 09/27/2026 · backs 7 figures

Microsoft

Microsoft (On the Issues)

Microsoft Azure Research – Systems and UIUC, via arXiv

Microsoft Learn

Microsoft Source

Microsoft, OpenAI and NVIDIA researchers, via arXiv

MLQ.ai

Motivair (Schneider Electric)

Munters

NADDOD

Navitas Semiconductor

Navitas Semiconductor, via GlobeNewswire

NeoPhotonics (PR Newswire)

Network World

Nixon Power Services

Nokia

NREL

NVIDIA

NVIDIA / Hot Chips

NVIDIA Newsroom

NVIDIA, on Hugging Face

OIF

Open Compute Project (NVIDIA authors)

Open Compute Project marketplace

Optica (OFC 2019, Th4A.6)

Optica / IEEE

Optical Connections News

Optical Internetworking Forum (OIF)

PA Transformer

photoncap.net

Photonics Explained (Substack)

Power Electronics News

Project Materials

Prysmian

Ranovus

Samsi et al., arXiv

Samsung Semiconductor

Schneider Electric

SemiAnalysis

SemiAnalysis (Dylan Patel, Daniel Nishball, Myron Xie and others)

SemiAnalysis / InferenceX

semiconductor-today.com

SemiEngineering

Semtech

ServeTheHome

SK hynix

Smartoptics

SpaceXAI (SpaceX)

Stockholm Exergi / City of Stockholm

StorageReview

STORDIS

Sumitomo Electric

Sunbird DCIM

TechCrunch

TechInsights

TechTimes

TES IT Solutions

The Definitive Guide to AI Data Centers

The Diligence Stack

The Next Platform

The Register

TIA Fiber Optics Technology Consortium

Tom's Hardware

Tom’s Hardware

TRG Datacenters

TSMC

U.S. Department of Energy, via Cornell Legal Information Institute

U.S. Energy Information Administration

U.S. Environmental Protection Agency

UALink Consortium, via Business Wire

Upsite Technologies

Vertiv

Vik's Newsletter

Wikipedia

Wing VC

Xtralis (Honeywell)

Yes Energy

Zayo

光芯 (光芯前沿 column), Tencent Cloud developer community

Personal educational project based on cited public sources. Not an official publication of my employer or the companies discussed. Models are schematic; estimates and assumptions are identified.