The Intelligence Factory Open the visualizer

Glossary

The words, grid to token

Every term you meet on this site, in plain words, with one number from this campus where it helps. Most entries open the part they describe in 3D.

#

34.5 kV

power

The medium-voltage level a campus distributes power at once the main transformers have stepped it down from the incoming transmission line; some campuses use 13.8 kV instead. Prefabricated switchgear splits it into feeders that run underground to each data hall.

  • ATK Energy Group: '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'; describes MV switchgear at the substation feeding unit substations at each data hall

50 V

power

The DC voltage on an NVL72 rack's busbar (the OCP ORv3 standard), stepped down there from the rack's incoming AC or DC feed by power shelves or DC-DC shelves. Bus converters on every tray take it the rest of the way down to 12 V.

400ZR / 800ZR

data

Standardized coherent pluggable formats for metro and data-center-interconnect links, running 400 Gb/s or 800 Gb/s per wavelength. OIF's own implementation agreement targets 80 km or more for amplified 400ZR links (120 km class for the higher-loss-budget variant); an unamplified link's reach instead depends on the specific module's transmit power and receiver sensitivity, commonly quoted around 40 km. Module power runs about 15-20 W for 400ZR; 800ZR draws about 23-25 W, and the longer-reach 800ZR+ variant about 26-30 W.

  • Optical Internetworking Forum (OIF): Section 7.1 'use cases': 'The 400ZR targeted reach for these applications is 80km or more' for amplified, 100/75 GHz DWDM links (Application Codes 0x01/0x03); Section 7.2 states unamplified reach 'is dependent on the transmit output power, input receive sensitivity, and the channel's loss characteristics' rather than giving one fixed number

415 V

power

The AC voltage the overhead busway delivers to each rack, the standard OCP ORv3 rack power shelves are built for. Line to neutral, three-phase 415 V is 240 V, the actual voltage server power supplies take.

800 V DC

power

A higher-voltage rack feed NVIDIA is moving toward for 2027-class racks, made in one step from 34.5 kV AC by solid-state transformers, with batteries sitting right on the DC bus instead of behind a UPS. NVIDIA says it cuts copper in the rack power path by 45% compared with 415 V AC at the same power. Google, Meta and Microsoft's Mt Diablo design feeds racks at ±400 V DC (800 V between its two poles) from a sidecar power rack; on 08/11/2026 NVIDIA said it, Google and Microsoft are working to make 800 V DC the open standard, with Mt Diablo's ±400 V spec serving current racks and a native-800 V version to follow.

  • NVIDIA: Row-level power management section: 'With lower current, thinner conductors can handle the same load, reducing copper requirements by 45%'

A

ACC

also active copper cabledata

A copper cable with a redriver chip at each end that amplifies and cleans the signal without fully regenerating it. At 200 Gb/s per lane it stretches copper to about 3 meters for a couple of watts per end, enough to reach the rack next door.

Accelerator

compute

This page's word for a GPU generation and its rack platform together, since a GPU's power draw, memory, and cooling needs change together from one generation to the next. It tracks four: H100, GB200, GB300, and the pre-launch Vera Rubin.

AEC

also active electrical cabledata

A copper cable with a full DSP retimer at each end that regenerates the signal from scratch, at the cost of far more power than a redriver. It commonly reaches 7-9 meters; AWS uses it to link Trainium racks, and trade photos of xAI's Colossus are reported to show the same cable type.

Amplifier hut

data

A small building along a long-haul fiber route, roughly every 80-100 km, that boosts the light optically without ever converting it back to electricity. Spacing this close is a rule of thumb, not a fixed standard.

ASHRAE water classes (W17-W45)

also ASHRAE liquid cooling classesheat

ASHRAE's naming for facility-water temperature bands, W17 through W45 and W+, where the number is the maximum supply temperature in degrees Celsius. This page's dry-cooler design runs in the W32-W45 band, warm enough to reject heat to outside air most of the year without a chiller.

  • Upsite Technologies: 'the 5th Edition redesignated the classes based on their upper temperature limit — therefore the classes are now W17, W27, W32, [W40,] W45, and the now named W+', redesignated from the prior W1-W5 bands (upper limits 17/27/32/45/over-45°C)

B

Basis (Spec / Vendor / Reported / Derived / Assumed)

also Spec, Vendor, Reported, Derived, Assumedgeneral

The label this page puts on every figure that carries a basis chip, one of five: "Spec" means the maker or a standards body states the figure for the named product or standard, with at least one primary source; "Vendor" means a vendor's own comparison or performance claim, attributed to it and stating what it's compared against, not checked independently here; "Reported" means a named third party — a government agency, researcher, analyst or the trade press — states it; "Derived" means this page's model calculates it from the scenario and its cited inputs, with the formula on the Method page; "Assumed" means the model chose a value where no single published figure applies, with the reason on the Method page. Two earlier labels, "Typical" and "Est.", covered figures not yet traced to one of these five; the site's build checks (`tools/claims.mjs`) fail if either still appears.

BESS

also battery energy storage systempower

Grid-side batteries, sized on this page at roughly a fifth of the campus's meter MW and twice that in MWh (on the Colossus 2 preset, SpaceXAI's planned 3.3 GWh pack instead, assumed to carry the whole campus), that absorb the megawatt swings a synchronized fleet of GPUs can put on the grid in under a second. SpaceXAI's (formerly xAI's) Colossus campuses are reported to use large Tesla Megapack installations for exactly this, though public reporting on their total rated power is inconsistent — from a couple hundred megawatts at the first Memphis site to gigawatt-scale estimates at the newer, much larger Colossus 2.

  • Data Center Dynamics: reports TVA/MLGW granting xAI 150 MW of grid power plus 'discounted Tesla Megapack battery storage to improve stability of the Memphis power grid,' without a Megapack power rating
  • Canary Media: 09/11/2026: counts 720 Tesla Megapack containers at Colossus 2 by satellite imagery and reports estimates 'between 720 megawatts and 1,400 MW' depending on model, and a Memphis utility CEO's figure of '2,000 MW of batteries behind the meter' — the article itself calls the exact size 'unclear'
  • SpaceXAI (SpaceX): Colossus II tab, Power: 'America's largest grid-connected battery pack will provide 3.3 gigawatt hours' (planned; no power rating given)

Busbar

power

A vertical copper bar running the full height of an NVL72 rack that every compute and switch tray clips onto, so there are no power cables to individual trays. NVIDIA's next-generation Kyber design replaces this AC/DC busbar with an 800 V DC version, claiming 45% less copper.

Busway

power

Copper bars in an aluminum housing that run overhead down each row of a data hall, carrying 415 V AC (or 800 V DC, in that design) to every rack through plug-in tap-off boxes. Moving a rack means moving a plug, not rewiring the row.

  • The Definitive Guide to AI Data Centers: 'LV Distribution: Busway, PDUs, RPPs & Rack Power' describes overhead busway with plug-in tap-off boxes as the standard way to run low-voltage power down a data hall row

C

CDU

also coolant distribution unitheat

A cabinet at the end of a row that keeps a rack's own filtered coolant loop separate from the building's facility water, passing heat between the two through a plate heat exchanger without mixing them. In this model the facility water runs about 3 °C below the rack loop on each side. Units on this page range 70 kW to 2.3 MW of capacity.

  • Vertiv: CoolChip CDU family page lists models at 70, 121, 600, 1,350 and 2,300 kW
  • Motivair (Schneider Electric): brochure: 'COOLING UP TO 2.3MW' from '102kW up to 2.3MW, depending on the model'

Chiller

heat

A refrigeration unit that spends electricity to make cold water, about 12 degrees C for air-cooled halls or about 20 degrees C for direct-to-chip liquid racks. Chiller compressors are the single biggest power draw in that kind of cooling plant, which is why it lands at a higher PUE than warm-water cooling.

Coherent optics

data

Optical transmission that encodes data in a light wave's amplitude, phase and polarization together, not just on and off, so one wavelength carries far more bits per symbol than direct-detect optics like LR4. A coherent transceiver — a pluggable (400ZR/800ZR-class) or a transponder in a line-terminal shelf — produces and receives one such wavelength; a separate mux/demux combines many wavelengths onto one fiber pair, and separate optical amplifiers extend the run. Coherent detection isn't what makes DWDM possible — direct-detect wavelengths can be multiplexed too — it's what lets each wavelength on a long run carry the most data.

Cold plate

heat

A copper plate with fine internal fins sitting directly on a hot chip, GPU, CPU, sometimes a switch chip, carrying coolant that enters cool and leaves a few degrees warmer. On this page's fully liquid-cooled NVL72 rack, even the switch trays and power shelves sit on cold plates.

Cooling tower

also cooling towersheat

An evaporative structure that carries away the chillers' rejected heat, or, in a warm-water design, trims the loop only on the hottest days. Evaporating a kilogram of water this way removes about 2.4 megajoules of heat, which is where most of a data center's water use goes.

Copper wall

data

Trade-press shorthand for the point, a few meters of reach at 200-plus Gb/s per lane, past which copper stops working even with an active redriver or retimer chip, and every design switches to optics regardless of the extra power. Every interconnect vendor hits the same physics as lane speeds climb; no single vendor sets this number.

CPO

also co-packaged opticsdata

Optical engines built onto or next to a switch ASIC's own package, shortening the electrical channel between SerDes and modulator, instead of living in separate pluggable modules at the faceplate — a packaging change, not a claim about how many traffic fibers still leave the switch. NVIDIA's Quantum-X and Spectrum-X Photonics claim four times fewer lasers and, as of its August 2026 update, 5x lower power (up from an initial 3.5x); Broadcom's Davisson reaches 102.4 Tb/s at about 3.5 W per 800G port.

  • NVIDIA: 03/18/2025 announcement: 'integrate optics innovations with 4x fewer lasers to deliver 3.5x more power efficiency'
  • StorageReview: 08/2026: production Spectrum-X Photonics delivers '5x lower power consumption'
  • Broadcom: Broadcom's own 10/08/2025 announcement names the 102.4-Tbps figure; the power-per-port figure is reported by NextPlatform (see links:cpo), not stated in this release itself

D

DAC

also direct-attach copperdata

A passive copper cable with no active parts at either end, the cheapest and lowest-power way to move a signal, but its reach collapses as lane speed climbs. At 224 Gb/s per lane it is good for about a meter, enough for NVLink's copper spine inside one rack.

Data parallel

compute

Running many full copies of a model on different data and averaging their gradients across the fabric once per training step. It's the largest-scale, least-frequent form of parallelism on this page short of syncing loosely across whole campuses.

  • Meta AI, arXiv: Table 4: data parallel degree scales with GPU count (DP=64 at 8,192 GPUs, up to DP=128 at 16,384)
  • DeepSeek-AI, arXiv: Section 3: training combines expert and pipeline parallelism with 'ZeRO-1 Data Parallelism (DP)'

DCI

also data center interconnectdata

The equipment and fiber that link one campus to another: coherent transponders in a line-terminal hut turning a handful of wavelengths into a shared long-haul connection. Ciena's WaveLogic 6, cited on this page, reaches up to 1.6 Tb/s per wavelength.

  • Ciena: 02/21/2023 release, headlined (per its own indexed title and a BusinessWire mirror) 'Ciena Unveils WaveLogic 6, Industry's First 1.6Tb/s Coherent Optic Solution'; ciena.com blocked direct and Wayback access on every attempt this pass, so this rests on that title and third-party coverage, not a page this workstream opened itself

Decode

compute

The second phase of answering a prompt: each new reply token is written one at a time. For a small batch, every token re-reads the model's weights, so decode speed usually tracks HBM bandwidth more than raw GPU math; serving many users in one batch shares each read and leans more on compute.

Die

compute

A single piece of silicon, cut at the reticle limit, the largest a chip can be made in one lithography exposure. Blackwell-class GPUs join two such dies with a 10 TB/s link so software sees one GPU; H100 uses a single die.

  • NVIDIA: 'Blackwell Ultra is composed of two reticle-sized dies connected using NVIDIA High-Bandwidth Interface (NV-HBI)...that provides 10 TB/s of bandwidth'

Double conversion

power

A UPS design that always turns utility AC into DC and back into AC, rather than passing it straight through, so the output never carries the grid's own dips or noise. Among every conversion before the rack in an AC power path, it is the single biggest loss.

DPU

also BlueField DPUdata

A separate network processor, NVIDIA's BlueField, that runs storage, security, and the front-end request network, so a user's traffic never shares hardware with the GPU-to-GPU fabric. An NVL72 rack carries two DPUs per compute tray, 36 in all.

  • NVIDIA: Compute Trays table: '2x NVIDIA BlueField-3 DPU, dual port 400G Infiniband or Ethernet' per tray; 18 compute trays per rack give 36 total

Dry cooler

heat

A rooftop coil-and-fan unit that rejects heat from warm facility water, 30-45 degrees C, straight into outside air with fans alone, no water spent and no chiller needed most of the year. This page's warm-water design uses dry coolers, backed by evaporative sprays only on the hottest afternoons.

DSP (optics)

also digital signal processordata

The signal-processing chip inside a fully-retimed pluggable module that regenerates the electrical signal on transmit and recovers it on receive, before it reaches the laser or after it leaves the photodiode. Linear optics (LPO) cut module power by removing this chip and letting the host chip's own SerDes drive and read the line directly. Co-packaged optics (CPO) takes a different path: it moves the optical engine itself onto or next to the switch ASIC's package, shortening the electrical run between SerDes and modulator — it is not simply this same chip relocated, and some CPO designs still retime.

DWDM

also dense wavelength-division multiplexingdata

Putting dozens of separate wavelengths on one fiber pair, each an independent channel of light. This page's model runs 32 wavelengths of 800G each on a route, 25.6 Tb/s total, which is how a campus's few physical fiber routes carry all of its long-haul traffic.

E

eGRID

also Emissions & Generation Resource Integrated Databasegeneral

The EPA database this page cites for the US national average grid carbon figure, 373 g CO2/kWh (823 lb/MWh) on 2022 data, alongside EIA's separate state-by-state electricity profiles for 2024. EPA has since published a newer eGRID2023 release with a lower national figure, so this page's number is a specific, dated historical reading rather than today's current average.

Expert parallel

compute

Spreading a mixture-of-experts model's experts across GPUs so each token only visits the few experts it's routed to. DeepSeek-V3's own training used expert parallel 64 with no tensor parallel at all; NVIDIA describes NVL72's wide NVLink domain as enabling wide expert parallelism for inference too — its own reference deployment of DeepSeek R1 spreads 256 routed experts across 64 of a rack's 72 GPUs.

  • DeepSeek-AI, arXiv: Section 3.2.1: 'DeepSeek-V3 applies 16-way Pipeline Parallelism (PP)...64-way Expert Parallelism (EP)...spanning 8 nodes' and elsewhere: trained 'without using costly Tensor Parallelism (TP)'
  • NVIDIA: 'For the DeepSeek R1 model, this is typically around four experts per GPU, which requires 64 GPUs to accommodate the full 256 routed experts during decoding'

External laser source

also ELSdata

In a co-packaged optics design, a laser supplied from its own field-replaceable module at the front panel instead of one tiny laser built into every optical engine. NVIDIA's Quantum-X Photonics feeds all 144 of a chassis's 800G channels from 18 such modules, each packing eight individual laser diodes and each one lighting eight of those channels — the "one laser for every eight links" figure quoted for this design is at that module level, not a count of individual diodes. Consolidating into fewer, larger, swappable units is most of where NVIDIA's "fewer lasers" claim comes from.

  • NVIDIA: confirms the Q3450 system's '144 ports with 800Gbps apiece' (115.2 Tbps full-duplex) and 'Each ELS contains eight high-quality lasers'; an individual ELS is stated to power 32 of the switch's 576 transmit lanes — 576 ÷ 32 = 18 ELS modules, and 144 ports ÷ 18 modules = 8 ports per module, which is where this page's 'eight channels per module' figure comes from

F

Facility water

heat

The building-wide water loop carrying heat between the CDUs, or in-row cooling units, and the chiller plant or dry coolers outside. This page draws its supply line blue and its warmer return line red.

Fan wall

heat

A wall of fans and water coils that pulls room air through and cools it, handling whatever share of a rack's heat, all of it in an air-cooled design, doesn't leave through liquid cooling.

Fat tree

data

The non-blocking Clos topology a leaf-spine (or leaf-spine-core) fabric uses, where every tier carries as much bandwidth going up as coming down. Scaling it further means adding switches, not just links, at whichever tier runs out of ports.

G

Gensets

also standby generatorspower

Containerized diesel generators, 2.5-3 MW class, that start within about ten seconds of a grid failure and take the load once UPS or DC-bus batteries have carried it that far. They run only a few hours a year, mostly for testing.

  • Cummins: DQKAN generator set data sheet: 2,500 kW (2.5 MW) standby rating
  • Cummins: course listing: 'The 10-second start: NFPA 110 Type 10 starting requirements for generator set applications'

GPU

compute

The chip that does a model's math: one or two reticle-size dies plus stacked HBM memory in a single package. This page follows four generations, from the 700 W H100 (2023) to the pre-launch, 1,800 W Rubin (2026-27), marked as an estimate since it hasn't shipped.

Grid carbon intensity

general

Grams of CO2 emitted per kilowatt-hour of grid electricity, which this page's state map shows varying more than fourfold: 113 g/kWh in hydro-heavy Washington versus 494 g/kWh in coal-and-gas-heavy Wisconsin. The same campus can emit three to four times more in one state than another.

H

H100-equivalent

compute

An industry shorthand, used by trackers like Epoch AI, for normalizing different GPU generations into one common compute unit, since a GB200 or Rubin GPU does many times an H100's math per chip. This page avoids the shorthand and instead reports each real campus's own disclosed chip mix and generation, marking any undisclosed split as an estimate.

HBM

also high-bandwidth memorycompute

Stacked DRAM built beside the GPU die on the same interposer, feeding it at several terabytes per second over millimeters of wiring. This page assumes it runs 10-16% of a GPU's power (10% for H100, 16% for the pre-launch Rubin), an estimate drawn from academic GPU power-characterization studies, which commonly cite about 8-15% for HBM3-class memory, rather than a vendor disclosure; moving weights out of it for every token is a large share of inference energy.

  • Micron: cited for HBM3E's own bandwidth/capacity specs; the WAF on assets.micron.com blocked every attempt to open it this pass (see its 'unchecked' note), so it is not relied on here for the 8-15% power-share figure, which the model treats as its own assumption

Hot aisle

heat

The sealed space between rack backs where every rack's exhaust heat collects before reaching a cooling unit. On an air-cooled rack the whole load leaves this way; on a liquid-cooled one only the fraction water doesn't carry away directly does.

I

IBC

also intermediate bus converter, bus converterpower

A fixed-ratio converter brick on each compute tray that cuts the busbar's roughly 50 V by about four to hand 12 V to the board. Because it doesn't regulate, it runs efficient, around 97-98%, though vendors don't publish the figure so this page estimates it.

ICI (Inter-Chip Interconnect)

data

Google's scale-up fabric for its TPUs: in Ironwood, each chip links straight to six neighbors in a 3D torus across a 64-chip cube (one rack), and optical circuit switches join cubes into pods of up to 9,216 chips. It does the job NVLink does in the NVIDIA racks drawn here.

  • Google Cloud: 'Each chip is connected to 6 other chips in the 3D torus'; 'a 9,216-chip system with 144 cubes'

In-row cooler

also in-row cooling unitheat

A cabinet the size of a rack, standing in the row itself, that pulls hot-aisle air through chilled-water coils and blows it back out cold at the rack fronts. This page's air-cooled hall design uses these instead of a CDU.

Interconnection

power

The process, and the physical tie-in, that connects a new campus's substation to the wider grid, typically at 230-500 kV. It commonly takes two to four years including the required interconnection study, part of why gigawatt campuses cluster where that grid capacity already exists.

Interposer

compute

A silicon layer that wires a GPU's dies and HBM stacks together with connections far finer than any circuit board could carry. NVIDIA calls its packaging CoWoS-L for Blackwell-class GPUs and CoWoS-S for H100.

IT load

power

The power that actually reaches computing equipment, as distinct from the larger figure drawn at the meter, which also covers cooling and building overhead. The ratio of meter power to IT load is PUE.

J

Junction temperature

heat

The temperature at the transistors themselves, the hottest point anywhere in the building. GPUs are thought to throttle near about 85 degrees C, though NVIDIA doesn't publish an exact limit; this page's warm-water design keeps the die around 70 degrees C.

K

KV cache

also key-value cachecompute

A model's running memory of a conversation, the keys and values computed from every earlier token in the prompt and reply so far. Every new token during decode reads both the model's weights and this cache from HBM, which is a second reason decode speed usually follows memory bandwidth.

L

Leaf

data

Switches at the row ends of a data hall that every rack's GPUs connect up into first. In a rail-optimized layout, GPU number n in every rack plugs into the same leaf switch, so most traffic crosses just one hop.

LPO

also linear pluggable opticsdata

A pluggable module with no DSP: the host chip's own SerDes drives the modulator and reads the photodiode directly, doing the equalization work the DSP used to do instead of regenerating the signal first. This page notes it runs at roughly half the power of a fully-retimed module, at some cost to reach and design margin.

M

Main power transformer (MPT)

also MPTpower

An oil-filled transformer, each about the weight of a loaded freight car, that steps the 345 kV transmission line down to the campus's 34.5 kV distribution voltage. Units this size run above 99.6% efficient, and reported 2026 lead times for large HV transformers ran from about a year up to five years depending on the manufacturer and tier.

  • PA Transformer: 345kV Power Transformer Performance Highlights: 'efficiencies exceeding 99.6%' for a 90/120/150 MVA unit; Key Specifications lists total shipping weight 543,350 lb
  • GridReadiness: June 2026 lead-time tracker by manufacturer: Tier 1 OEMs (Hitachi Energy, Siemens, GE Vernova) 48-60+ months, Tier 2 EU/Asian manufacturers 12-36 months

Meta Hyperion

general

Meta's Richland Parish, Louisiana campus: still under construction, with nothing live yet in the latest satellite imagery (04/2026). Phase 1, roughly 1.5-1.6 GW, is projected around late 2027 to early 2028; Meta itself has said the full build will reach 5 GW, though Epoch AI notes it has found no concrete evidence yet that all 5 GW lands at this specific Richland Parish site rather than being spread across Meta's broader build-out.

  • Epoch AI: directory entry: latest satellite imagery '04/17/2026' shows nothing serving yet; Epoch's own projection has 'first phase operational' around 01/01/2028 at about 1,632 MW IT power, and notes 'no concrete evidence' the full 5 GW Meta has announced lands entirely at this one site

Microsoft Fairwater Atlanta

general

Microsoft's Fayetteville, Georgia campus: partly operating, partly under construction. 4 of 9 main-campus buildings were live at about 636 MW of IT power as of 09/24/2026; the rest of the 13 planned buildings, across the main, east and Fairwater campuses, are under construction toward about 1.5 GW.

  • Epoch AI: directory entry, updated 09/24/2026: 636 MW current IT power, 4 buildings operational (Buildings 1-2 by 10/2025, 3-4 by 06/2026)

Microsoft Fairwater Wisconsin

general

Microsoft's Mount Pleasant, Wisconsin campus: partly operating. Building 1 has run since 04/16/2026 at about 369 MW of IT power; Building 2 is under construction, due in 2028.

  • Epoch AI: directory entry, updated 09/24/2026: Building 1 operational, dated to a 04/16/2026 Microsoft post; 369 MW of current IT power. Epoch's own construction-pace estimate for Building 2 (~early 2027) is earlier than the 2028 date reported elsewhere for this campus — see techtimes-fairwater-wisconsin, which this page follows

MPO

also multi-fiber push-on connectordata

A connector that bundles several fibers, commonly 12, behind one push-on plug so a single optical module can light many lanes at once. An 800G DR4 module uses 8 of the 12 positions on one MPO connector; a twin-port DR8 module uses two.

N

NVL72

data

NVIDIA's rack-scale platform that wires 72 GPUs into a single NVLink domain through copper switch trays in the middle of the rack. Both GB200 and GB300 ship as NVL72 racks on this page; DGX H100 does not, since its NVLink domain stops at eight GPUs on one server board.

NVSwitch

data

The switch chip that connects every GPU's NVLink links so any GPU can read another's memory at full speed. An NVL72 rack holds 18 of them, two per switch tray across nine trays; a DGX H100 server puts four on its own baseboard instead.

  • NVIDIA: NVIDIA NVLink Switch Tray section: 'Each GB300 NVL72 rack includes 9 NVLink Fifth-Generation switch trays, with 2 NVSwitch ASICs per tray' (18 total)
  • NVIDIA: DGX H100 platform (nvidia-dgx-h100) pairs eight H100 GPUs with four third-generation NVSwitch chips on one system baseboard

O

OSFP

data

Octal Small Form-factor Pluggable, the module housing scale-out ports use to carry light on and off a switch or NIC. NVIDIA's 800G DR8 module, a twin-port OSFP, draws up to 17 W.

P

PAM4

data

Four-level signaling that packs two bits into every symbol, used on both NVLink's copper spine and scale-out fiber links at speeds like 224 Gb/s per lane. It doubles the data rate of simple two-level signaling at the same symbol rate, at the cost of needing a cleaner signal to read correctly.

PCIe

data

The general-purpose bus connecting a GPU to its CPU and network card inside a DGX H100 server, PCIe Gen5 at about 64 GB/s per x16 direction. NVL72 racks replace it with NVLink-C2C to the CPU and dedicated NVLink for anything chattier.

  • NVIDIA: specifications table: 'Interconnect... PCIe Gen5: 128GB/s' (bidirectional total for an x16 link, i.e. about 64 GB/s per direction)

Pipeline parallel

compute

Splitting a model's layers into consecutive stages, each living on a different rack, and passing activations down the line like an assembly line. Llama 3 405B ran pipeline parallel 16 this way.

  • Meta AI, arXiv: Table 4, 'Scaling configurations and MFU for each stage of Llama 3 405B pre-training': PP=16 at every listed GPU count (8,192 and 16,384)

Pluggable optics

data

Transceiver modules that plug into a switch or NIC port and convert its electrical signal to light for fiber, the default way scale-out and campus links are lit today. Power runs 8-9 W for a 400G module up to 25-30 W for an early 1.6T module.

Power ledger

general

This page's meter-to-silicon accounting: a running list of every conversion and loss between the utility meter and the GPU dies, in megawatts, that sums back to the campus's total draw. It's how the page shows where a watt actually goes, not just the final PUE ratio.

Power shelf

power

A 1U shelf holding six hot-swap rectifiers, wired 3+3 for redundancy, that turns a rack's AC feed into about 50 V DC for the busbar. GB300's shelves add capacitors storing 65 J per GPU, which NVIDIA says cuts peak grid demand from training swings by 30%.

  • NVIDIA: 'Measured benefits and results': LITEON-optimized power electronics 'filled the remaining space with 65 joules/GPU of energy storage'; 'the peak power demand seen by the grid is reduced by 30% when training the Megatron LLM'

Prefill

compute

The first phase of answering a prompt: the model takes in the prompt's tokens in parallel, often in chunks alongside other requests. Each weight it reads serves many prompt tokens, so prefill is usually limited by compute rather than by memory bandwidth. A long prompt can take a few hundred milliseconds; together with the queue and the network, that sets the time to first token.

PUE

also power usage effectivenesspower

The ratio of power drawn at the campus meter to the power that actually reaches IT equipment; the gap is cooling, conversion losses, and building overhead. This page calibrates its own cooling designs to land around 1.5 for air, 1.10-1.20 for chilled liquid, and 1.05-1.15 for warm water — bands chosen to bracket real operating figures rather than measured from a specific campus.

  • Google: Efficiency page: Google's own 2025 fleet-wide PUE was 1.09, against 'the global average PUE of respondents' data centers was 1.54' per the Uptime Institute's 2025 survey it cites — the real-world spread this page's own bands sit inside

R

Rail-optimized

data

A leaf-and-spine wiring pattern where GPU number n in every rack lands on the same leaf switch as GPU number n everywhere else. Most all-reduce and pipeline traffic then crosses only one switch instead of climbing the whole fabric.

Reticle limit

also reticle-limit diecompute

The largest area a chip can be made in a single lithography exposure, a hard ceiling on one die's size no matter the process node. H100's 814 mm² die sits at this ceiling; Blackwell-class GPUs get more silicon per package by joining two reticle-limit dies with a fast die-to-die link instead of trying to build one larger die.

  • NVIDIA: confirms the dual-reticle join at 10 TB/s (NV-HBI); H100's single-die, 814 mm² figure is the well-corroborated industry figure for the GH100 die rather than a number stated on NVIDIA's own two-page H100 datasheet, which gives only 80 billion transistors and TDP

S

Scale-across

data

The long-haul fiber tier linking whole campuses hundreds of kilometers apart over coherent DWDM optics. Light in fiber still needs about 5 milliseconds to cross 1,000 km, so only loosely synced training crosses this tier; interactive traffic mostly does not.

Scale-out

data

The optical fabric outside a rack or server: one port per GPU into a leaf-and-spine network that lets any GPU on the campus reach any other. It carries far less traffic per link than scale-up, which is why it can afford to be fiber instead of copper.

Scale-up

data

The network tier that joins a small group of accelerators tightly enough to split one model layer's own math across them. Here it is NVLink inside an NVL72 rack or a single DGX H100 server, on copper. Other chips use their own: ICI inside a 64-chip Google TPU cube, whose cubes are joined by optical circuit switches; NeuronLink across a 64-chip AWS Trainium2 UltraServer; and UALink, an open standard for the same job.

SerDes

also serializer/deserializerdata

Circuits along a die's edge that turn parallel on-chip data into the fast serial lanes a link like NVLink sends off the package. Every high-speed interconnect on this page, copper or optical, starts and ends at a SerDes.

Single-mode fiber

data

Fiber with a core narrow enough that light travels one path through it, avoiding the mode-mixing that limits shorter-reach multimode fiber. It is what carries scale-out, campus, and long-haul optics on this page, since only single-mode fiber reaches past a few hundred meters.

Solid-state transformer (SST)

also SSTpower

Power electronics switching at high frequency that replace the 60 Hz transformer, the UPS, and the rack rectifiers with a single conversion from 34.5 kV AC straight to 800 V DC. Navitas and NVIDIA both describe this step as eliminating conversion stages to cut losses, but neither publishes a single efficiency figure for it; this page's better-than-98% figure is its own assumption, informed by that direction, for hardware that is not yet shipping.

  • Navitas Semiconductor: 'Navitas Supports 800 VDC Power Architecture...' (10/13/2025): describes the architecture as eliminating 'multiple traditional AC/DC and DC/DC conversion stages, maximizing energy efficiency, reducing losses' without stating a percentage for this specific conversion
  • NVIDIA: Key benefits section: 'Improves end-to-end efficiency by up to 5% compared to current 54 V systems' — a whole-chain figure, not a single-stage SST number

SpaceXAI Colossus 1

also Colossus, xAI Colossus 1general

The first Memphis campus of SpaceXAI (xAI until its July 2026 rebrand, after SpaceX acquired it in February): fully built and operating. About 200,000 GPUs (H100, H200, some GB200), leased in full to Anthropic since 05/06/2026; confirmed grid supply is 150 MW from MLGW/TVA. Trade reports counted about 35 on-site gas turbines, about 422 MW in all; SpaceXAI says its remaining temporary turbines must all be removed by July 2027, and that more than 240 batteries let the site come fully off the grid in emergencies or at peak demand.

  • Compute Atlas: 'About 35 on-site natural-gas turbines (a combined 422 MW per SELC and aerial imagery) supplement a 150 MW grid substation and Tesla Megapack storage'; ~100,000 H100 GPUs Phase 1, later expanded with H200s
  • Data Center Dynamics: reports TVA/MLGW's board approving 'an additional 150MW of power' for Colossus
  • SpaceXAI (SpaceX): Power tab: 'more than 240 batteries so Colossus I can come completely offline during emergencies or peak demand'; Air tab: 'all remaining temporary turbines must be removed by July 2027'
  • Business Insider, via Yahoo Finance: reports xAI's rebrand to SpaceXAI, 07/06/2026

SpaceXAI Colossus 2

also xAI Colossus 2general

SpaceXAI's second Memphis campus (xAI until July 2026), about 3 km from Colossus 1: live and still growing, and called the "First Gigawatt Datacenter" by SemiAnalysis in 09/2025. By satellite estimate it ran about 946 MW of IT power and 440,000 Nvidia chips as of 09/24/2026; a day later Elon Musk put it at 550,000 GPUs, 110k GB200 and 440k GB300, with another 220k GB300 due the following week, 220k in November and, "if we get lucky", 220k more by late December. SpaceXAI says it plans 1M+ GPUs. It lists a planned 3.3 GWh grid-connected battery pack for it (its energy developer Riley Trettel told the TVA board so on 08/20/2026), says its temporary gas turbines must be removed by July 2027, mentions no diesel generators, and says it uses closed-loop cooling that takes only domestic water. By Epoch AI's tracking, it is the most powerful AI data center operating today, by both IT power and compute; Amazon and Anthropic's New Carlisle campus is next, at about 910 MW with more chips but less compute.

  • Epoch AI: directory entry, updated 09/24/2026: '1,112k H100-eq AI compute, supported by 946 MW of IT power,' with '110k' B200 and '330k' B300 chips (≈440k total)
  • Epoch AI: 'Colossus 2 is the largest tracked AI data center at about 946 MW of current IT power, followed by Anthropic-Amazon New Carlisle at about 910 MW'
  • SemiAnalysis: headline: 'xAI's Colossus 2 — First Gigawatt Datacenter In The World'
  • SpaceXAI (SpaceX): Colossus II tab: 'GPUs planned 1M+'; 'America's largest grid-connected battery pack will provide 3.3 gigawatt hours'. Water tab: 'Colossus II uses closed-loop cooling and takes only domestic water'. Air tab: 'all remaining temporary turbines must be removed by July 2027'. No tab mentions diesel generators (checked 09/27/2026)

Spine

data

The switch tier above the leaves that connects every leaf to every other, so any GPU on the campus can reach any other in a couple of hops. Two tiers of 144-port switches reach roughly 10,000 GPUs before a third tier or parallel fabric planes are needed.

Stargate Abilene

also OpenAI Stargate Abilenegeneral

OpenAI, Oracle, Crusoe, and SoftBank's Texas campus: partly operating, partly still being built. Buildings 1-4 of 8 were live at about 421 MW of IT power as of 09/24/2026; buildings 5-8 are roofed and being fitted out, with the full campus, over 1 GW, due between Q4 2026 and Q1 2027.

  • Epoch AI: directory entry, updated 09/24/2026: '421 MW of IT power'; buildings 5-8 'structurally complete and roofed but not yet confirmed operational'
  • Epoch AI: 'Projected capacity: 1.2 GW | 1.0 million H100-equivalents...Projected completion: Q4 2026' for the full site

Substation

also campus substationpower

Where the transmission line dead-ends on steel gantries and lands on a ring of SF6 circuit breakers and disconnect switches, with instrument transformers to measure the power and surge arresters to clip lightning. Building one, with its interconnection study, commonly takes two to four years — see Interconnection.

SuperNIC

also ConnectX SuperNICdata

NVIDIA's per-GPU network card for scale-out traffic, letting a GPU reach other racks without going through its CPU. GB300 pairs each GPU with a ConnectX-8 SuperNIC running 800 Gb/s.

  • NVIDIA: Specifications, C8180 single-port variant: InfiniBand speed 'XDR/NDR/HDR/HDR100/EDR/SDR' — XDR is NVIDIA's 800 Gb/s InfiniBand generation (see nvidia-xdr-switch-specs)

Switchgear

power

A lineup of drawout breakers that protects every outgoing circuit in a room and switches it automatically between utility and generator power when the grid drops. A campus has 34.5 kV switchgear feeding the halls, and each hall has its own 480 V (or medium-voltage, in a DC design) switchgear inside.

T

TDP

also thermal design powerheat

The power a chip is designed to dissipate as heat, in practice the same number as its electrical power draw. This page's GPUs run 700 W (H100) up to 1,800 W (Rubin, pre-launch, an estimate) at the package.

  • NVIDIA: specifications table: 'Max thermal design power (TDP): Up to 700W (configurable)' for the SXM form factor

Tensor parallel

compute

Splitting one model layer's own math across several GPUs, which trade partial results inside every layer, many times per token. It's the chattiest form of parallelism, which is why Llama 3 405B kept it to tensor parallel 8, the size of one NVLink domain on H100.

  • Meta AI, arXiv: Table 4, 'Scaling configurations and MFU for each stage of Llama 3 405B pre-training': all three listed stages run TP=8 (with PP=16, CP=1 or 16, and DP=64/128/8)

TIM

also thermal interface materialheat

A thin layer of thermally conductive paste or pad between a die and whatever sits on it: its lid, where the package has one, or else the cold plate or heat sink directly (a bare-die package such as the H100 SXM5). A lidded package has a second layer between the lid and the cooler. Each layer of TIM the heat crosses costs a few degrees before it reaches coolant.

Token

compute

The unit a language model reads and writes, roughly a few characters of text. What leaves a GPU package after all its power is a few bytes per token, and this page's calculator turns a campus's tokens-per-second into a figure per kilowatt-hour.

Transmission line

also high-voltage linepower

The 345 kV lattice-tower line that carries power from distant plants to a campus substation, two three-phase circuits bundled together with a shield wire on top for lightning. High voltage keeps the current small: a 500 MW campus rides on a few hundred amps per phase instead of the tens of thousands a rack-level voltage would need.

U

Ultra Ethernet

data

An open Ethernet stack for AI and HPC scale-out networks from the Ultra Ethernet Consortium, whose founders include AMD, Arista, Broadcom, Cisco, HPE, Intel, Meta and Microsoft. Its 1.0 specification (06/11/2025) adds RDMA transport and congestion control across NICs, switches, optics and cables: an open alternative to InfiniBand for the compute fabric.

Unit substation

power

A pad-mounted transformer along a data hall that drops 34.5 kV to 480 V, right outside the electrical rooms so the low-voltage, high-current run stays short. In an 800 V DC design it serves only cooling and building loads, since the IT load skips it through solid-state transformers instead.

UPS

also uninterruptible power supplypower

A unit that turns incoming AC into DC and back to clean AC, with batteries on the DC link, so the racks never see a flicker between a grid failure and the generators taking load. Eaton's 9395XR, cited on this page, is reported at up to 97.5% efficiency online — eaton.com itself would not load for this page's own check, so the figure rests on reseller and distributor listings that reproduce Eaton's spec sheet rather than a page this project opened directly.

  • Eaton: eaton.com timed out on every direct attempt (see the 'unchecked' note on this source); the 97.5% online-mode figure is reproduced by multiple third-party resellers (e.g. distributor spec sheets) citing Eaton's own datasheet

V

Voltage staircase

general

A chart following one figure, current or voltage, down through every step of the power chain: 345 kV at the transmission line, 34.5 kV at the campus feeders, down to under a volt at the GPU core. Each rung links to the part of the 3D view where that step happens.

VRM

also voltage regulator modulepower

Dozens of switching phases ringing each GPU, an inductor and a power stage apiece switching around a megahertz, that make the final step from 12 V to under a volt at the die. They sit as close to the chip as possible, because every millimeter of copper at a thousand amps costs power.

W

WUE

also water usage effectivenessheat

Liters of water used on site per kilowatt-hour of IT power. This page's own cooling-design calibration spans about 0.16 L/kWh for warm-water dry coolers up to about 1.0 L/kWh for an air-cooled chiller-and-tower plant — figures this model chose to sit near, not the literal numbers introl-wue reports (which gives a wider industry range: 'ideal' 0.0, 'best-in-class' 0.3-0.7, and an industry average of 1.8-1.9 L/kWh).

  • Introl: gives 'Ideal WUE: 0.0 L/kWh', 'Best-in-class: 0.3-0.7 L/kWh' and 'Industry average: 1.8-1.9 L/kWh' — context this page's own air (≈1.0) and warm-water (≈0.16) design values sit inside and below, respectively

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.