The Intelligence Factory Open the visualizer

Method

How the numbers are made

The site is a scenario model of one campus. Four choices set up a generic campus; named sites can also supply reported fleets and plant overrides. Hardware figures follow these inputs through the code described below. Token estimates add their own workload and environment inputs on top of that hardware model, and the Evidence page always shows one stated reference scenario rather than following the campus you last set. This page names the constants, shows where each comes from, and says what the model leaves out. The names in code are the site's own, in src/model/engine.ts and its neighbors.

1What the model is, and what it is not

This page carries one scenario model of one campus. Its hardware follows from four choices: campus size at the meter, accelerator generation, power design to the rack, and cooling design. For a generic campus, hardware counts and power are derived from those four choices by compute(scenario) in src/model/engine.ts. It models NVIDIA systems only (H100, GB200, GB300, Vera Rubin): Google TPUs, AWS Trainium and AMD Instinct appear as cited comparisons, not as models, and a campus that runs them is drawn with the closest NVIDIA match and says so. Change a choice and that whole model recomputes; reported site fleets and plant overrides provide additional inputs for named campuses.

Tokens and energy per token (Section 8) sit on top of that hardware model rather than following from it alone: they add their own workload and environment inputs, tokens per GPU per second, utilization, a grid-carbon figure, and a training-energy assumption, set in src/model/tokens.js. The Evidence page shows one fixed reference scenario for its claims and sources; it does not follow whatever campus you have set here.

This page is not a bill of materials for a real, named building. No vendor has published a construction drawing for any campus on this page, and this model does not claim to reproduce one. Where a real campus is picked as a preset (Section 9), the model rebuilds the closest scenario it can from disclosed facts, and says plainly which of the four choices had to be assumed rather than confirmed.

The model conserves energy by construction: every megawatt drawn at the meter is accounted for in the ledger (Section 3) plus what reaches the GPU silicon, down to floating-point rounding. src/model/engine.test.ts checks this for every accelerator, power design and cooling choice at once.

2The four choices, and what each drives

The Scenario type in engine.ts has exactly four fields that a reader sets: meterMW, accel, power, cooling; these drive the generic campus. The optional site and stage fields can also select a reported fleet mix and plant overrides, changing hardware counts, backup and cooling. Picking one does change token arithmetic, though: pickSite() in src/app/scenario.js also sets the grid-carbon figure Section 8 uses, since each real campus sits on a different grid.

ChoiceWhat it setsWhat it drives
Campus size at the meter (meterMW)Power drawn at the utility meter, 10 MW to several GWEverything scales from this number: rack count, GPU count, hall count, transformer count, fuel volume
Accelerator (accel: h100, gb200, gb300, rubin)GPU power, GPUs per rack, CPU power, NVLink/NIC generation, board-level efficiencies (vrmEff, ibcEff, psuEff), HBM type and bandwidthRack power bottom-up, the fabric's per-GPU NIC speed, HBM bandwidth and token throughput
Power to the rack (power: ac415, dc800)Which facility conversion chain runs (unit substation/UPS/busway, or solid-state transformer/DC bus/battery) and the final rack-level conversionFacility losses, PUE, and (for accelerators that support it) a small IT-power gain
Cooling (cooling: air, liquid, warm)coolFrac, cooling power as a share of IT load, and wue, water use per unit of energyPUE and on-site water use

The engine also enforces which combinations are physical: an accelerator's coolingOptions array and dc800 flag silently fall back to a supported choice if the reader picks one the hardware cannot take (H100 cannot run on 800 V DC or warm water, for example).

3Power: how rack power and PUE come out of the meter

Rack power is built bottom-up, one rack at a time, in kilowatts. Package power comes first: pkgKW = accel.gpuW * accel.gpusPerRack / 1000, split into an HBM share (accel.hbmShare) and GPU silicon. Voltage regulation loss follows from the regulator's own efficiency: vrmLossKW = pkgKW / accel.vrmEff - pkgKW. GPU package power, its VRM loss, CPU power, NVLink switching (scaleupKW), NICs (nicKW) and the rest (otherKW) sum into a 12 V board load (load12), which the bus converter's efficiency (accel.ibcEff) turns into a small further loss, then a busbar loss (busbarKW), then a final rack-level conversion (accel.psuEff on AC, or EFF.rackDcDc on 800 V DC) to reach the full rack draw, rack.kw.

At the facility level the engine solves for IT power rather than assuming a PUE. It writes one equation, meter times the upstream efficiency equals IT power divided by the IT conversion path, plus cooling and building loads divided by the cooling/building path, and solves it for IT_MW. PUE, meterMW / IT_MW, falls out of that solve; it is never set directly. What is set directly is each conversion step's efficiency and each cooling design's coolFrac, chosen so the whole chain lands the design in its published PUE band (air near 1.5, chilled liquid 1.10 to 1.20, warm water 1.05 to 1.15).

StepEfficiency, as codedBasis
Main power transformersEFF.mpt = 0.996typical
Campus cables & switchgearEFF.campus = 0.997est
Unit substations (AC path)EFF.unitSub = 0.99typical
UPS, double conversion (AC path)EFF.ups = 0.965typical
480 → 415 V transformers (AC path)EFF.pdu = 0.987est
Busway & whips (AC path)EFF.busway = 0.995est
Solid-state transformers, MV → 800 V DC (DC path)EFF.sst = 0.98 (Navitas claim)est
800 V DC bus & batteries (DC path)EFF.dcBus × EFF.dcBattery = 0.997 × 0.998est
In-rack conversion to the busbarEFF.rackDcDc = 0.985 (DC), or per-accelerator psuEff, 0.96 to 0.975 (AC)est / typical
Bus converters, board rail down to 12 Vper-accelerator ibcEff, 0.9828 to 0.983est
Voltage regulators, 12 V to the coreper-accelerator vrmEff, 0.90 to 0.915est
CoolingcoolFrac by cooling choicetypical
Lighting, controls, officesMISC_FRAC = 0.017est

The ledger the page shows (compute().ledger) turns each of these losses into a megawatt figure at campus scale, with gpuSiliconMW calculated from the installed GPU packages minus their assumed HBM share. A generic campus's unused power budget is recorded separately as unallocated capacity, not GPU silicon. engine.test.ts checks that the ledger's rows plus GPU silicon always sum back to the meter, to nine decimal places, for every accelerator, power and cooling combination.

4Racks, GPUs and halls

Rack and GPU counts follow from IT power, not the other way around. A fabric tier's cost per GPU (tierConfig(tiers).kwPerGpu) depends only on the tier count, not on how many GPUs the campus ends up with, so the engine solves for racks directly instead of guessing and re-checking: racks = Math.floor(IT_MW * 1000 / (rackKW + accel.gpusPerRack * fab.kwPerGpu)) with a two-tier fabric's cost first. If the GPU count that buys actually fits a two-tier fabric's own non-blocking port budget, that is the final rack count; if the site's power would buy more GPUs than two tiers can carry, the same formula is solved again with a three-tier fabric's higher (so always more conservative) per-GPU cost, and that is final. GPU count is simply racks * accel.gpusPerRack.

Data halls are sized at roughly 45 MW of IT load each: halls = Math.max(1, Math.ceil(IT_MW / 45)). This is the model's own working assumption, not a vendor figure, and it is the number the campus scene uses to decide how many hall blocks to draw. Everything downstream in the layout, transformer counts, generator counts, chiller or dry-cooler counts, cooling distribution units, is sized the same way: from a power figure (meter, IT, or cooling power) divided by a typical catalog unit size, then rounded up, in layout at the end of compute().

5The network: fat-tree sizing

Fabric sizing keys off NIC speed. FABRICS in engine.ts holds three speed classes, 400G, 800G and 1600G, each with a switch radix, a switch power draw, and per-port and per-GPU optics power. GB200 defaults to the 400G class, NVIDIA's own reference design of one 400G ConnectX-7 port per GPU: a 64-port switch (radix: 64, NVIDIA's Quantum-2 QM9700) drawing 0.75 kW. GB300's documented ConnectX-8 upgrade uses the 800G class instead, a 144-port switch (radix: 144) drawing 2.9 kW, with a 17 W GPU-side module and a 13.5 W port-side module, two ports per module, 8 fibers per link.

tierConfig(tiers) gives each tier count's per-GPU port, link and switch-power cost from that radix; a two-tier, non-blocking fat tree fits up to radix² / 2 GPUs. The engine tries two tiers first, and only falls back to three (up to radix³ / 4 GPUs, and beyond that parallel fabric planes, planes = Math.ceil(G / threeTierMax)) when the GPU count a two-tier fabric would actually buy exceeds its own port budget — which a three-tier fabric's costlier, more conservative per-GPU price can never do in turn, so the choice needs no further checking once made. Two tiers cost each GPU 3 ports and 2 links; three tiers cost 5 ports and 3 links, because traffic now crosses leaf, spine and core.

Switch counts follow directly: leaf = Math.ceil(gpus / (radix / 2)), spine and core by the same division at each tier, and switches = leaf + spine + core. Links are gpus * fab.linksPerGpu, and fibers are links * fibersPerLink. Modules per GPU work out to 1 at the GPU-side NIC plus, in a two-tier build, 3 switch ports carried on twin-port modules, for 1 + 3/2 = 2.5; a three-tier build with 5 ports per GPU comes to 3.5 — the figure engine.test.ts checks directly for the 100 MW reference scenario, which at GB200's 400G fabric is itself a three-tier build. Scale-across fiber (the DCI block) is sized separately at roughly 10 Gb/s per GPU to other sites, converted into a number of lit wavelength pairs at 32 lambdas of 800 Gb/s each.

6Heat and water

Cooling power and water use are both set by the cooling choice, not computed from first principles. Each entry in COOLING carries a coolFrac, cooling power as a fraction of IT load, and a wue, liters of water per kilowatt-hour of energy delivered.

CoolingcoolFracwue (L/kWh)Design PUE band it targets
Air-cooled halls with chillers0.421.0≈1.5
Liquid to the chip, chilled water0.100.51.10–1.20
Warm-water liquid, dry coolers0.050.161.05–1.15

Both columns are the model's own calibration, chosen so that once the power chain's own losses (Section 3) are added on top, each design lands in its published PUE band; the code comment above COOLING says so directly. Water use then follows arithmetically, but against IT power rather than meter power: waterM3h = IT_MW * wue (engine.ts's waterM3h helper), because The Green Grid defines WUE as liters per kilowatt-hour of IT equipment energy, not facility energy — the two differ by a factor of PUE, so using meter power here would overstate water use by however much the design's PUE exceeds 1. On a hot day (clock.ts) it rises further once outdoor air reaches a shared adiabatic-assist threshold, 35°C, for warm-water designs, or as chiller efficiency falls roughly 2.5% per degree above 25°C for chilled designs; that threshold and the design's chilled-water and warm-water supply temperatures live in one place, engine.ts's WATER constant, so the cards and the clock never disagree on what the design assumes.

Every other temperature the page names, from the die to the outside air, comes from one operating point per cooling design built on those constants, loopTemps() in engine.ts (the assumption "One operating point for the cooling loops"). A liquid design has two loops that meet at the CDU: the rack loop warms about 10 °C across a rack, facility water runs about 3 °C below it on each side, and a die sits about 20 °C above the coolant leaving its cold plate. An air design has one chilled-water loop between the in-row coils and the chillers, with 22 °C supply air and an 18 °C rise through the servers. The heat tour, the cards and the Hot to cold chart all read this one set, and each figure says which loop it is on and whether it is the supply or the return. Only NVIDIA's 45 °C warm-water inlet and ≈55 °C outlet, and ASHRAE's 18–27 °C supply-air band, are published; the rest is illustrative.

7Time: the four clocks

src/model/clock.ts samples the same static model over time along four independent clocks, each a pure function of (model, time):

ClockTimescaleWhat it assumes
Training stepseconds (a 60 s window)GPU load swings in a 2-second step with a 0.35 low point and a checkpoint pause at 38–44 s; rests on the 0.2–3 Hz swing band a joint Microsoft/OpenAI/NVIDIA measurement paper found in production clusters (spec). For GB300 and Vera Rubin, a rack-level capacitor buffer (trainingStorage) aims for a low-pass target (τ ≈ 0.6 s, an estimate) but is held to NVIDIA's stated 65 J/GPU capacity and an estimated charge/discharge rate: past that limit it stops tracking the target and the raw swing shows through, exactly what the tiny real capacity does across a multi-second checkpoint. A second buffer, the site batteries, does the same against their own rating (layout.bessMW) and a far larger capacity, so what the grid actually sees can still show part of a swing neither buffer could absorb.
Grid outageseconds, then minutesGrid fails at t=5 s; generators ramp up to carry the full load by t=15 s, matching NFPA 110's Type 10 timing class (Level 1 is a separate classification, for loads whose loss could cost lives — data-center IT is not itself one, but many mission-critical campuses design to Type 10 timing anyway); batteries bridge the gap until then, including the last couple of seconds of the generators' own ramp; chillers (non-warm designs) take about 2 minutes to restart; the grid returns 15 minutes in and the site waits 5 more minutes of stability before transferring back. Sample.meterMW is the utility import alone (zero while islanded); Sample.siteMW is what the site actually draws from every source, the figure the live counters integrate for site-wide energy. The chiller restart time, the outage length and the stability window are this page's own illustrative choices.
Hot day24 hoursOutdoor temperature follows a cosine curve, coolest near 3 a.m., hottest at 3 p.m., with a 16°C swing by default; warm-water designs hold steady until air crosses 35°C, then adiabatic sprays add water; chilled designs lose roughly 2.5% cooling efficiency per degree above 25°C. No sourced dry-cooler derating curve was found, so these slopes are estimates.
Inference day24 hoursDemand follows a qualitative shape drawn from measurement studies of real LLM services, overnight trough, morning rise, midday dip, afternoon peak, evening decline, with a default 2.5× peak-to-trough ratio and GPUs idling at 30% of full power. No hyperscaler has published a peak-to-trough ratio, so that number is explicitly illustrative in the code and on the page.

8Tokens and energy per token

src/model/tokens.js holds the reader's own settings in one object, calc: tokens per GPU per second (defaulting to an accelerator-specific reference, see below), utilization (0.6 default), a grid carbon figure (370 g CO₂/kWh, close to the US average of 373), a training energy allowance (50 GWh) amortized over a model's lifetime output (10¹⁵ tokens by default), and whether training's share is included at all.

tokenFigures(M, c) computes a serving rate, M.gpus * tps * c.util, then joules per token from two sources: jOps = meterMW * 1e6 / rate, the campus's whole draw divided by tokens served, and jTrain = c.trainGWh * 3.6e12 / c.lifeTokens, training energy spread across a model's lifetime tokens. The two sum to total joules per token; a 500-token reply's watt-hours, carbon and water then follow by simple unit conversion. The watt-hours are meter energy, cooling included; the water figure divides them by PUE first, because the wue that governs the cooling chart is defined on IT energy.

The request tour adds its own timeline on top: 25 ms across the internet each way, 50 ms in a queue, 200 ms of prefill and 60 tokens a second for one user's stream. Those are a separate illustrative assumption ("An illustrative request timeline"), the same for every accelerator, not a benchmark of the hardware selected; only the energy and water per reply come from the scenario.

The reference tokens-per-GPU figure, tokPerGpuRef in engine.ts, is Math.round(2000 * accel.hbm.tbs / 8 / 50) * 50: it scales with HBM bandwidth on the reasoning that decode, generating one token at a time, is memory-bound rather than compute-bound, so a GPU with twice the memory bandwidth of another can plausibly serve roughly twice the tokens per second. This is the model's own derived estimate, not a benchmark result for any specific accelerator.

9Real campuses: presets, status and where the numbers come from

src/model/sites.ts holds eight real campuses as presets, each mapped onto the model's four choices plus a status. Colossus 2 is sized the other way round: from the GPU counts Elon Musk posted on 09/25/2026 (110k GB200 + 440k GB300, and his dated plans for 220k more GB300 at a time), each accelerator at its own bottom-up rack power and on its own scale-out fabric, so the racks set the IT load and the IT load sets the meter; the halls, CDUs, cables and token rates that follow are this model's, not SpaceXAI's. Status is checked against Epoch AI's satellite-based Frontier Data Centers tracker (updated 09/24/2026) and current reporting, as of 09/27/2026; the tracker's live figures are IT power, not total facility draw, which is why every site's facts array is careful to say "IT power" rather than "meter power."

CampusStatusLive figure, as of 09/24/2026
SpaceXAI Colossus 1 (Memphis, TN)Operating~200,000 GPUs, fully built; leased in full to Anthropic since 05/06/2026
SpaceXAI Colossus 2 (Memphis, TN)Partly live, still growing~946 MW IT power, ~440,000 chips by satellite count; on 09/25/2026 Elon Musk put it at 550,000 GPUs (110k GB200 + 440k GB300), the counts the preset is sized from
Stargate Abilene (Abilene, TX)Partly liveBuildings 1–4 of 8, ~421 MW IT power; buildings 5–8 roofed and being fitted out, full campus due Q4 2026–Q1 2027
Microsoft Fairwater Atlanta (Fayetteville, GA)Partly live4 of 9 main-campus buildings, ~636 MW IT power; the rest of the 13 planned buildings under construction toward ~1.5 GW
Microsoft Fairwater Wisconsin (Mount Pleasant, WI)Partly liveBuilding 1 running since 04/16/2026, ~369 MW IT power; Building 2 under construction, due 2028
Meta Hyperion (Richland Parish, LA)Under constructionNothing serving yet in the latest satellite imagery (Epoch AI, updated 09/24/2026); a first phase of ~2 GW due by 2030, the full 5 GW around 2032
Amazon Project Rainier (New Carlisle, IN)Partly liveAbout 16 of ~30 planned buildings, ~910 MW IT power and about a million Trainium2 chips; toward ~1.9 GW around Q1 2028
Meta Prometheus (New Albany, OH)Partly live~471 MW IT power across tents, colocation space and traditional buildings, about half the ~1,022 MW Epoch AI expects by Q3 2028; fed by two off-grid gas plants

Colossus 2 is, by Epoch AI's count, the single most powerful AI data center operating anywhere today, by both IT power and by compute, as of 09/24/2026, even though it is itself still under construction and its own status is "partly live" against its own further plans. The next campus by that measure is Amazon and Anthropic's New Carlisle site, at roughly 910 MW, which has more chips installed but less compute than Colossus 2. Colossus 1, by contrast, is the one fully operating, complete campus among the eight: fully built, leased in full to Anthropic, with no further construction pending.

Each preset's unknowns array names exactly what the model had to assume rather than confirm: Abilene's meter power is estimated from its 421 MW IT figure at an assumed PUE near 1.2, and neither its accelerator generation nor its cooling design is disclosed; Colossus 1 is modeled as air-cooled H100 racks on NVIDIA's own reference design, though the sources do not say how Colossus 1 actually cools its racks; Colossus 2 keeps its GB200s and GB300s apart, each at its own rack power, memory and transistor count (its 3D rack, tray and package show a GB300, four of five of its GPUs), and models what its operator publishes: SpaceXAI describes closed-loop cooling that takes only domestic water, so the preset draws air-cooled chillers with no cooling towers and counts no cooling water; and it lists a 3.3 GWh battery pack, planned, and mentions no diesel, so the preset assumes the pack is the backup, draws no generator yard, and assumes the pack's unpublished power rating can carry the whole campus.

10The basis labels, and how a figure earns each

Every figure the page shows beside a basis chip carries one of five labels, defined in src/evidence.js's BASIS object: spec, the maker or a standards body states this exact figure for the named product or standard, backed by at least one primary source; vendor, a vendor's own comparison or performance claim, attributed to it and stating what it's measured against, not checked independently here; reported, a named third party — a government agency, a researcher, an analyst, or the trade press — states it; derived, this page's model calculates it from the scenario and its cited inputs; and assumed, the model chose a value where no single published figure applies. Two earlier labels, "Typical" and "Est.", are on the way out: they mark a figure not yet traced to one of the five, and tools/claims.mjs's strict pass fails on any that remain within the prefixes it's told to check.

Evidence rides with the claim, not just the label. A spec, vendor or reported figure carries a refs list of [sourceId, where] pairs, pointing at a section, table, page or figure in a source listed in src/sources.js — a spec needs at least one of those sources marked primary. A derived figure names a calc id; its formula is written out once in CALCS, in src/evidence.js, and repeated nowhere else. An assumed figure names an assume id; its reasoning lives the same way in ASSUMPTIONS. Both registries are reproduced in full below (Sections 13 and 14), so a reader never has to open the source to see why a number is what it is. Clicking any basis chip on the page opens that one claim's own popover; the Evidence page lists every claim for one reference scenario, and tools/claims.mjs runs the same strict checks on every claim across all 104 scenarios.

11Known gaps and estimates the site carries

The research files in research/ name these gaps directly, and the model's numbers reflect them rather than papering over them:

  • No sourced peak-to-trough ratio exists for inference traffic; the model's 2.5× default is stated as illustrative in both the code comment and the page's own notes.
  • No named vendor dry-cooler derating curve was found; the hot-day clock's cooling-power slopes are estimates, flagged as such in clock.ts's own notes.
  • Bus-converter and voltage-regulator efficiency at the board level (the 50 V/54 V to 12 V step, and 12 V to the GPU core) have no published vendor figure for any of these accelerators; every number in that part of the ledger is labeled Assumed, with its reasoning in Section 14.
  • HBM's share of total GPU power (10–16% across generations, applied per accelerator as hbmShare) is this model's own assumption. No vendor publishes it, and the paper once cited for it measures only relative power and gives no such share.
  • NVIDIA's own claim for 800 V DC is "up to 5% end-to-end efficiency gain"; a separately circulated "83% to 92%" framing is a different, unconfirmed construction and this model uses only NVIDIA's own figure, which engine.test.ts checks the DC-800 scenario against directly.
  • GB200's per-GPU NVLink figure has a direct-fetch anomaly in NVIDIA's own material (1.8 TB/s versus a 3.6 TB/s figure on the same page); the model uses 1.8 TB/s, the figure corroborated everywhere else.
  • No single sourced ratio exists for gensets or unit substations per 100 MW; the counts in layout (for example gensets = Math.ceil(meterMW / 3 * 1.2), a 3 MW class unit at N+20%) are this model's own sizing, not an industry-standard ratio.
  • Microsoft's own post says Fairwater Atlanta forgoes on-site generation, UPS systems and dual-corded distribution for its GPU fleet. The preset still runs the model's generic power chain, UPS and generators included, and says so in its unknowns.

12How the site is tested

The Vitest suite gates model correctness. Four of its core files are described here. engine.test.ts checks energy conservation (ledger plus GPU silicon equals the meter, to nine decimal places, for every accelerator, power and cooling combination), that bottom-up rack power lands within the published range for each accelerator, that cooling choices land in and stay ordered within their published PUE bands, that 800 V DC's IT-power gain over AC stays inside NVIDIA's own "up to 5%" claim, and that the fabric never falls short of the ports it needs. clock.test.ts samples all four clocks across every scenario at 300 points each and checks every value is finite, non-negative and inside a sane range. data.test.ts generates the page's full text content for every scenario, every real-campus preset, and a spread of campus sizes, and checks no string is empty, undefined or NaN, and every number on the page carries a valid basis. journeys.test.ts checks every beat of every guided tour, across the same scenario spread, is complete and lands on a real part in the 3D scene, and that no tour skips a level or returns to one it has left.

The tools/ scripts test the rendered page itself, on the real GPU, rather than the model's arithmetic: cycle.mjs, views.mjs and links.mjs walk every scenario, scene and tour stop and check nothing errors, every view is unobstructed, and every link lands where it claims to; perf.mjs measures frame cost, draw calls and triangle counts; ui.mjs checks on-screen controls never overlap at desktop or phone size; coplanar.mjs and flicker.mjs catch z-fighting; gpushot.mjs, hero.mjs, look.mjs and shot.mjs capture screenshots for review. npm test runs the Vitest suite; the tools run by hand against a live dev server.

13Calculations

What the model calculates, how, and from what. A figure labeled Calc. names one of these.

Mach-Zehnder engine and package rate
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.From: 6.4 Tb/s engines, eight per package (Broadcom), 400G FR4: four wavelengths per fiber each way (IEEE 802.3 clause 151, as summarized by TIA FOTC), 64 × 106.25 Gb/s PAM-4 channels (write-up of ISSCC 2026 paper 23.4), 16 ports per engine (calc bailly-fibers-per-engine)
Pluggable module lane rate, payload and line
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.From: 53.125 GBd per lane, MMS4X00 specifications (NVIDIA), 106.25 GBd per lane, MMS4A00 specifications (NVIDIA), PAM4 carries 2 bits per symbol, eight lanes each way, two DR4 engines of four (module cards)
Micro-ring engine and package rate
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.From: eight 200 Gb/s PAM4 lanes per engine (NVIDIA), 18 engines per package as drawn (NVIDIA’s Quantum-X package image), 28.8 Tb/s per switch ASIC (NVIDIA)
Data fibers per Bailly-class engine
128 duplex 400G FR4 ports ÷ 8 engines = 16 ports per engine, each with one transmit and one receive fiber.From: 128 duplex LC front-panel ports (Broadcom), 8 optical engines (Broadcom)
Transmission line current
Per-phase current on each of two equally loaded circuits = campus MW ÷ (2 × √3 × 345 kV × power factor).From: campus MW at the meter, assumed power factor
Fuel deliveries
Tanker count = 48-hour fuel volume ÷ 30,000 L per truckload.From: generator fuel burn, assumed truckload
Unit substation secondary current
Current = 2.5 MVA ÷ (√3 × 480 V); apparent power already includes the power factor.From: unit substation rating
Per-rack busway current
Current = rack kW ÷ (√3 × 415 V × power factor).From: rack kW, assumed power factor
Liquid and air heat per rack
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).From: rack kW, assumed liquid share
Rack busbar current
Current = rack DC bus power ÷ busbar voltage.From: rack DC bus power, busbar voltage
GPU core current
Current = GPU power less its HBM share (the HBM stacks have their own supply) ÷ an assumed 0.8 V core voltage.From: GPU power, HBM share, assumed core voltage
Average heat flux through the die
Flux = GPU power less its HBM share ÷ die area. Hot spots run well above this average.From: GPU power, HBM share, die area
Fan airflow
Airflow from Q = ṁ · cp · ΔT: rack heat carried by air, air density and specific heat, at the stated temperature rise.From: air-cooled heat, assumed temperature rise
Arithmetic per generated token
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.From: parameter count of the model in the math (assumed model, published size)
Bytes read from HBM per generated token
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.From: parameter count, bytes per parameter at the assumed precision
Arithmetic per byte read, against what the GPU can do
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.From: FLOPs per token, bytes per token, GPU dense peak at the model precision, if published, GPU HBM bandwidth, published
Time for one decode step on one GPU, one stream
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.From: bytes per token, FLOPs per token, GPU HBM bandwidth, published, GPU dense peak at the model precision, if published
Signal speed in optical fiber
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.From: speed of light (c), typical silica-fiber group index ≈1.47
Propagation delay over an illustrative 1,000 km route
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.From: route length (1,000 km, illustrative), fiber signal speed
GPUs at the meter power chosen
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.From: meter MW (scenario choice), rack power and GPUs per rack for the chosen accelerator
IT load from meter power and PUE
IT load = meter power ÷ this design’s modeled power usage effectiveness (PUE); the gap is cooling and electrical overhead.From: meter MW (scenario choice), modeled PUE for the chosen cooling design
Heat out equals power in
By energy conservation, essentially all electrical power a campus draws is eventually rejected as heat, so the heat-out figure equals the meter figure.From: meter MW (scenario choice)
Hours a site battery pack could carry the campus
Hours = the pack’s energy (MWh) ÷ its assumed power (MW). On the Colossus 2 preset the energy is SpaceXAI’s planned 3.3 GWh and the power is assumed equal to the campus’s meter MW, so the hours shrink as the campus grows.From: pack energy, per the operator, assumed pack power (see the assumption “Site batteries carry the whole campus”), meter MW
Main transformer count
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.From: campus MW at the meter, transformer unit rating class, assumed power factor 0.95
Main transformer loss
Loss = campus MW × (1 − main-transformer efficiency).From: campus MW at the meter, the transformers' published efficiency
34.5 kV feeder count
Feeders = max(2, ceil(campus MW ÷ 10 MW per feeder)).From: campus MW at the meter, assumed 10 MW per feeder
Campus cable & switchgear loss
Loss = campus MW × main-transformer efficiency × (1 − cable-and-switchgear efficiency).From: campus MW at the meter, assumed cable & switchgear efficiency
Generator count
Units = ceil(campus MW ÷ 3 MW × 1.2) for N+20% redundancy on 3 MW-class sets.From: campus MW at the meter, genset unit class
Bulk fuel volume
Volume = campus MW × 48 h × 0.26 L/kWh (the gensets’ published full-load fuel rate), shown in million liters.From: campus MW at the meter, the gensets' published fuel rate
Battery energy storage sizing
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.From: campus MW at the meter, assumed BESS sizing fraction
Unit substation count
Count = ceil(campus MW × 0.97 ÷ 2.2 MW per 2.5 MVA unit at typical loading).From: campus MW at the meter, unit substation rating class
IT load and PUE
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.From: campus MW at the meter, the chain of conversion efficiencies, the cooling design’s power fraction
Rack count
Racks = floor(IT MW × 1000 ÷ (one rack’s kW + its GPUs’ share of the fabric’s kW)).From: IT MW, rack power, network power per GPU
GPU count
GPUs = racks × GPUs per rack.From: rack count, GPUs per rack
Hall count
Halls = max(1, ceil(IT MW ÷ 45 MW per hall)).From: IT MW, assumed IT capacity per hall
Cooling power
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.From: IT MW, the cooling design's calibrated power fraction
Chiller count
Chillers = ceil(IT MW × 1.1 ÷ 4 MW per chiller, a ≈1,100-ton class unit).From: IT MW, chiller unit class
Dry cooler count
Units = ceil(IT MW × 1.1 ÷ 0.8 MW per dry-cooler unit).From: IT MW, dry-cooler unit class
Heat rejected by the cooling plant
Heat rejected ≈ IT MW × 1.05, adding back the plant’s own electrical overhead as heat.From: IT MW
Cooling tower count
Towers = warm-water design: ceil(IT MW ÷ 16 MW per tower); air/liquid design: ceil(IT MW × 1.3 ÷ 6 MW per tower).From: IT MW, tower unit class
When the towers run
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.From: cooling design, adiabatic threshold, 35 °C
Water per day
Water = WUE (L/kWh IT) × IT MW × 24 h, converted to cubic meters.From: WUE for this design, IT MW
Heat leaving the campus
Heat out ≈ meter MW: energy conservation means virtually every watt drawn eventually becomes heat.From: campus MW at the meter
Light speed in fiber
One-way delay per km = 1 ÷ (speed of light ÷ assumed fiber refractive index), ≈ 4.9 µs/km.From: speed of light, assumed fiber refractive index, n≈1.47
Hall-to-hall fiber strands
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).From: GPU count, fabric fibers-per-link
Training swing size
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.From: the rack build's GPU share, meter MW, the model's low-power floor between steps
Hot-day cooling response
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.From: outdoor temperature over the day, assumed chiller efficiency slope, adiabatic threshold, 35 °C
Ledger loss row
Loss = the power flowing into this stage × (1 − its assumed or published efficiency).From: MW reaching the stage, the stage’s assumed or published efficiency
Ledger conversion-loss row
Loss = the downstream load this stage feeds ÷ its assumed efficiency, minus that load itself.From: downstream load the stage feeds, the stage’s assumed efficiency
Ledger cooling/lighting overhead row
MW = the IT load × an assumed share of it, picked so the design lands in the cooling class’s published PUE band.From: IT load, assumed overhead fraction for this cooling class
Ledger network-power row
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.From: per-unit power, switch or module count from the fabric model
Ledger per-rack-component row
MW = a published or estimated power budget for this component in one rack × the number of racks the site’s power buys.From: per-rack power for the component, racks built
Unallocated ledger row
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.From: meter MW, every other ledger row, GPU-silicon MW
Racks built
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.From: IT MW, one rack’s kW, fabric power per GPU
Grid, campus and building inventory counts
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.From: meter MW or IT MW, assumed catalog unit size, redundancy margin
Rack and server counts
Count = the racks (or servers, or per-rack part multiples) the model’s power-and-fabric solve builds for this scenario.From: racks built, parts per rack
Silicon counts
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.From: GPUs or racks built, parts per GPU or per rack
Network inventory counts
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.From: GPUs built, NICs per GPU, physical port rate, per-tier port and fiber counts
Hall unit-substation secondary current
Current = 2.5 MVA ÷ (√3 × 480 V); the rating is already apparent power, so no separate power-factor term applies.From: unit substation rating
Hall busway per-rack current
AC case: current = rack kW ÷ (√3 × 415 V × assumed power factor). DC case: current = rack kW ÷ 800 V.From: rack kW, assumed power factor (AC only)
Hall rack liquid/air heat split
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.From: rack kW, liquid-cooled share for this accelerator/cooling combination
Hall air-side heat share
Air share = 1 − the liquid-cooled share used for the rack’s liquid/air split above.From: liquid-cooled share
Fiber housings for the campus fabric
Housings = total fabric strands ÷ fibers per housing (four EDGE8 1U housings’ worth, 4 × 144 fibers, per 4U enclosure).From: fabric strands, fibers per housing
Optical modules per GPU
Modules per GPU = the fabric’s total optical-module count (GPU-side plus switch-side) ÷ GPUs built.From: module count, GPUs built
Strands crossing to other halls
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.From: GPUs built, fibers per link
Rack power, built bottom-up
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.From: accelerator’s published GPU/CPU power, assumed per-rack scale-up/NIC/other power, assumed conversion efficiencies down the chain
Compute tray power
NVL72: tray power = (rack DC-bus power − NVLink scale-up power − busbar loss) ÷ 18 trays. DGX H100: tray power = rack power ÷ 4 servers.From: rack DC-bus power (or rack power), NVLink scale-up power per rack, busbar loss, trays or servers per rack
Tray busbar-clip current
Current = tray power ÷ 50 V busbar voltage.From: tray power, busbar voltage
Power-shelf or server-PSU loss per rack
Loss = power reaching the busbar or server rail × (1 ÷ conversion efficiency − 1).From: power reaching the busbar or rail, shelf or PSU conversion efficiency
Bus-converter or voltage-regulator loss, campus-wide
Per-rack loss = power passing through the stage × (1 ÷ assumed stage efficiency − 1); the campus figure sums this across every rack the scenario builds.From: power through the stage per rack, assumed stage efficiency, racks in the scenario
Count per rack, from the count per server
Count per rack = the published count per server (power cords per supply, or scale-out ports per NIC) × 4 servers per rack.From: published count per server, servers per rack (4)
CPU memory per Grace CPU
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.From: published rounded rack memory total, 36 CPUs per rack
Per-GPU HBM capacity, NVL72
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.From: NVIDIA’s published rack HBM total, GPUs per rack (72)
The copper-wall chart
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.From: each lane rate’s cited reach, the crossover estimate
Racks from an operator’s GPU counts
Racks of each kind = that accelerator’s GPUs ÷ its GPUs per rack (72 for an NVL72), with the kinds rounded so they add up to the campus total. A partly filled last rack counts toward power as the share it holds.From: the operator’s GPU counts by accelerator, GPUs per rack
Homes with the same power
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.From: meter MW, an average home’s draw, ≈1.2 kW
The same power at the rack’s voltage
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.From: meter MW, the rack’s voltage
Power that reaches the GPU dies
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.From: GPU package power, HBM share of package power, GPU count, meter MW
A watt’s slices
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.From: the power ledger, meter MW
Tokens a second, campus
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.From: GPU count, tokens a second per GPU, utilization
Tokens per kWh at the meter
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.From: meter MW, tokens a second, campus, training energy and lifetime tokens, from the calculator
Energy per reply
Wh per reply = joules per token × 500 tokens ÷ 3,600. That is meter energy, cooling included; the IT equipment’s part is that ÷ PUE.From: joules per token, a 500-token reply, PUE
Water per reply
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.From: energy per reply, PUE, WUE for this design
Time to first token, the example request
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).From: the illustrative request timeline

14Assumptions

Values the model chooses where no single published figure applies, and why. A figure labeled Assumed names one of these. Change one and the figures that rest on it change with it.

Power factor: 0.95
A common design value for large facility loads; the real figure depends on the equipment and its correction.
Fuel truckload: 30,000 L
A typical road tanker; used only to show the scale of refueling.
GPU core voltage: ≈0.8 V
Chipmakers do not publish core voltages for these parts; 0.7–0.9 V is the usual range for this class of process.
The model in the token math: Llama 3.1 70B (70 billion parameters, dense) with FP8 weights, 1 byte per parameter
The campus’s real models and precisions are not public, so the math uses a published open-weights model of a common serving size: Meta’s model card gives Llama 3.1 in 8B, 70B and 405B sizes, and NVIDIA publishes an FP8 version of the 70B model that halves its bits per parameter from 16 to 8. A dense model keeps the arithmetic simple: every weight takes part in every token. Mixture-of-experts models read only the experts each token uses; larger or smaller models move both figures in proportion. The 8 × 8 tile drawn beside it is schematic: not real weights, not a real layer shape.
Modeled campus size: set by the reader
When no real campus is selected, this page follows whatever meter power the reader sets in the scenario bar; it is not a published figure for a specific campus. A real campus preset starts from this page’s own estimate of its meter power, from the IT power reported for it (see that preset’s unknowns).
Typical campus interconnection voltage: 345–500 kV
Not one published standard; it spans the real campuses shown on this page, from Stargate Abilene and Project Rainier’s 345 kV interconnections up toward the 500 kV class used for the largest single-site loads.
Optical amplifier hut spacing: ≈80–100 km
A commonly cited rule of thumb for terrestrial DWDM amplifier spacing; the exact interval depends on the fiber’s loss budget and the specific route, not one publisher’s spec.
Map symbols on the regional map: representative, heights and sizes exaggerated
The substation yard at the edge of each campus, the transmission towers, the utility substations and the plant symbols (gas, coal, nuclear, wind, solar) are generic symbols of common US types placed at illustrative spots so they read at map scale. They are not any real plant, line or substation, and their layout, equipment counts and proportions are not published figures for these sites. Each remote campus draws its power from a drawn source: where its site facts report a source, that kind is drawn (on-site gas beside Stargate Abilene; Colossus 1’s temporary gas turbines, alongside a utility substation for its MLGW/TVA supply; Meta Prometheus’s two off-grid gas plants, its only supply; wind behind Abilene’s grid supply; Entergy’s new gas plants behind Meta Hyperion), and where a campus takes grid power with no named source, a utility substation beside it is fed by the state’s leading sources of generation in EIA’s 2024 state profiles (nuclear in Tennessee; gas and nuclear in Georgia; gas and coal in Wisconsin; coal and gas in Indiana). That is a representative picture of each campus’s supply, not surveyed lines: real campuses draw on many plants across a regional grid, and the plants and substations drawn are not the ones that serve them.
Free cooling depends on climate: more hours in cool, dry climates
A qualitative framing, not one published figure: the ASHRAE allowable ranges this page cites elsewhere are what let a site skip mechanical cooling for part of the year, and how many hours varies by site climate.
Site batteries carry the whole campus: pack power = the campus’s meter MW; no diesel generators
SpaceXAI lists a planned 3.3 GWh grid-connected battery pack for Colossus 2 and mentions no diesel generators, but does not publish the pack’s power rating or say how the campus rides through a grid outage. The model takes the simplest reading, that the pack alone backs up the full load, so the outage clock, the switchgear and the battery card can be drawn; the real design may differ. Third-party estimates of the battery already built there vary: Canary Media (09/11/2026) counted 720 Megapack containers by satellite and reported estimates of 720–1,400 MW, and a Memphis utility CEO’s figure of 2,000 MW behind the meter. The ≈1.1 GW this preset assumes falls inside that span, but none of these figures is the operator’s.
Campus interconnection voltage: 345 kV
This model fixes every campus’s transmission interconnection at 345 kV so the voltage staircase reads the same at every size. Real gigawatt campuses (Stargate Abilene) use this class too, but the largest ones need 500 kV or extra circuits instead, which this design does not switch to.
Substation breaker count and yard size: 6 dead-tank SF₆ breakers, ring bus, ≈200×150 m pad
Illustrative of a typical major transmission substation. This campus’s actual breaker count and footprint depend on the utility’s own design and are not published for a generic site.
Battery energy storage sizing fraction: 0.2× campus MW power, 0.4× campus MW energy (2 h)
No published rule ties campus battery size to campus MW; this model picks a round, illustrative fraction to show scale, not a specific vendor design.
IT load per data hall: 45 MW
A round figure representing a large modern hall; real halls vary with design and run both smaller and considerably larger.
Cooling power fraction of IT load: air 0.42, liquid 0.10, warm water 0.05
Calibrated so each cooling design, once the power chain’s own losses are added, lands in its publicly reported PUE band (air ≈1.5, liquid with chillers 1.10–1.20, warm water 1.05–1.15), not read off one specific facility.
Water use by cooling design: air 1.0, liquid 0.5, warm water 0.16 L/kWh IT
Illustrative figures chosen to fall within the water-use ranges reported for each cooling approach (near-zero for closed-loop liquid, higher for evaporative towers), not a single published number for this design.
Fiber refractive index: n ≈ 1.47
Typical single-mode telecom fiber; the exact figure varies slightly by fiber type and is not published for an illustrative route.
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
Only the R8-300 shelf’s published size and the CMD64’s 2RU height and 64 channels come from Ciena’s data sheet. The shelter, racks, router, which module sits in which slot, the port counts, the eight lit channels and the fiber routing are a representative arrangement, not any operator’s hut; a real terminal may use a different vendor, a fixed mux with no ROADM, or a dedicated transponder shelf. A fixed 75 GHz mux like the CMD64 suits 400G-class waves; the 800G waves this campus sends need about 150 GHz each, which in practice means a flexible-grid mux such as the 6500 RLS family’s colorless CCMD. The drawing keeps a CMD64-size unit. Fibers are drawn thicker than the ≈2 mm jumpers they stand for so they read at this distance, and the colors only tell wavelengths apart: they are all infrared.
Hall-to-hall route as drawn: ≈140 m of duct bank between the two detailed halls’ spine ends, under DR’s 500 m
The campus is an illustrative layout: its two detailed halls face each other about 140 m apart at the end where the spine rows and the duct bank sit, and with the spine rows drawn at that end, the fabric links that cross between them, risers and short in-building runs included, stay well inside the 500 m that NVIDIA gives for DR optics. That is a property of this drawing, not of any real campus; the plain hall blocks beyond, on a larger campus, are not routed or measured.
Duct-bank layout: 4-inch conduits, one cable per conduit, a spare row
Standard telecom civil-construction practice, not a specification for this particular campus; no published layout exists for an illustrative site.
Outage scenario timing: chiller restart ≈2 min, outage 15 min, grid-stable wait 5 min
Illustrative numbers chosen to show the shape of a recovery, not a published incident or a specific utility’s reconnection policy.
Hot-day cooling response slope: chiller ≈2.5%/°C above 25 °C; dry-cooler/adiabatic curve fit to the adiabatic threshold
No public dry-cooler or chiller derating curve was found for this exact design; the model picks a plausible slope to show the shape of the day, not a measured curve.
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%
No provider publishes a peak-to-trough demand ratio or an idle-power fraction for a production GPU fleet; these are illustrative choices consistent with general utilization discussions, not measured figures for a specific service.
Heat-export tie-in drawn in the heat layer: illustration, not built here
This campus exports no heat. The heat layer draws a ghosted supply/return pipe pair, a plate heat exchanger and a heat-pump skid only to show where a district-heating connection would attach; sizes, routing and placement are representative. The cited real schemes (Odense, Fortum, Stockholm) are separate facilities.
Main transformer fittings as drawn: representative
Oil-filled units of this class shed heat through radiator banks with fans (ONAF) and carry bushings, a conservator, a Buchholz relay, a breather and an on-load tap changer. How many radiator banks and fans, where the tap changer and cabinet sit, and the bushing shed counts vary by vendor; the drawing is a typical arrangement, not one product.
Cooling tower cell as drawn: counterflow cell, representative
The card does not name a tower type. The cell is drawn as a counterflow tower with air-inlet louvers on all four sides, a fan deck and an eased-inlet fan stack, following SPX Marley’s published descriptions of those parts; the type, cell size, rotor and fittings are a typical arrangement, not this campus’s equipment.
Substation equipment as drawn: representative
Larson’s 345 kV dead-tank SF₆ breaker listing documents six roof bushings with bushing current transformers and a grounded tank. The outward splay of the bushings, the mechanism cabinet, the center-break disconnect with arcing horns, the arresters’ grading rings, the capacitor-voltage-transformer stacks and every shed profile are typical forms, not this yard’s equipment.
Bulk fuel tanks as drawn: representative
The card gives only the total volume. The double-wall horizontal tanks, their saddles, vents, manway, fill box, gauge, ladder, platform, containment curb and manifold are a typical arrangement for above-ground diesel storage, not this site’s tank sizes or fittings.
Prefab e-house, control house and shelter drawings: representative
Prefabricated switchgear buildings, control houses and telecom shelters are real and common, but their layouts vary by vendor and no source here documents this campus. The ribbed wall panels, standing-seam roofs, door, landing and wall-pack HVAC positions, and the glazed gatehouse are drawn as a typical kit, not one product or site.
Unit substation size: 2.5 MVA
A catalog size in the 2.5–3 MVA class commonly used for 34.5 kV → 480 V campus distribution (industry trade coverage, not one standard); the model picks the round 2.5 MVA point and sizes N+ units from it.
Main power transformer efficiency: 99.6%
A round point inside the 99.5–99.7% range trade and manufacturer sources converge on for large (>100 MVA) power transformers; no single standards body publishes one required efficiency percentage for this class.
UPS module size: 1.25 MW modeled
The model divides IT load into 1.25 MW UPS modules. Real double-conversion modules of this class run 1.25–1.8 MW per cabinet (e.g. Eaton’s 9395XR at up to 1.5–1.8 MW); 1.25 MW is a round illustrative unit, not one product’s rating.
UPS double-conversion efficiency: 96.5%
A conservative point within the 96–97.5% online-mode range reported for this UPS class (Eaton’s 9395XR is tested to 97.5% online, up to 99% in eco/ESS mode); the model uses online mode at a slightly lower figure to leave margin.
Solid-state transformer efficiency: >98%
Navitas’s own claimed conversion efficiency for its SiCPAK solid-state-transformer power modules (medium-voltage AC to 800 V/1500 V DC), as exhibited and reported by trade press; SSTs for 2027-era racks are not yet a shipping, independently benchmarked product, so this is carried as a vendor claim, not a checked spec.
Busway, whips and 480 → 415 V transformers: busway and whips 99.5%; transformers 98.7%
The busway figure is a small, unpublished resistive-loss allowance for the overhead busway and rack tap-off whips; no vendor publishes a busway efficiency at this level of the chain. The 480 → 415 V dry-type transformers (or transformer PDUs) that feed it are taken at 98.7%: DOE’s minimum efficiencies for low-voltage dry-type units of this size are about 99% at its 35%-load test point, and efficiency runs a little lower at the heavier load an AI hall holds. About 1.3 MW of every 100 MW entering the IT path is lost here.
Campus cable & switchgear efficiency: 99.7%
An assumed loss allowance for the 34.5 kV feeders and switchgear between the main transformers and the halls; cable/switchgear loss at this length and voltage is small and not separately published by a single source.
800 V DC bus and battery efficiency: 99.7% bus, 99.8% battery float
Small assumed conduction/float losses for the DC bus and the batteries sitting on it; NVIDIA’s 800 V HVDC material describes the architecture but does not publish a bus- or battery-loss percentage.
In-rack DC-DC efficiency, 800 V → 50 V: 98.5% peak
Navitas’s own tested peak-efficiency figure for its 10 kW, 800 V-to-50 V DC-DC platform (98.1% at full load); the platform was still being evaluated by customers, not yet a shipping catalog product, as of its Feb. 2026 announcement.
Cooling and lighting overhead, share of IT load: air 42%, liquid 10%, warm-water 5%; lighting/controls/offices 1.7%
These shares are picked, not measured, so the fully-built design lands in each cooling class’s reported PUE band (air ≈1.5, chilled liquid 1.10–1.20, warm water 1.05–1.15) once the power chain’s own conversion losses are added on top. Google’s own fleet PUE (1.09 TTM) anchors the low end of that band.
Per-rack component power budgets (scale-up switching, NICs/DPUs, CPUs, drives/fans/management): from each accelerator’s own table
Chipmakers publish GPU, package and rack-level power, but not a line-item watts figure for NVSwitch/NVLink-switch trays, NICs/DPUs, CPUs or drives/fans/management individually; the model apportions a plausible share of the published rack total to each, checked against the published rack range in engine.test.ts.
HBM memory, share of GPU package power: h100 10%, gb200 13.2%, gb300 15%, rubin (pre-launch) 16%
No chipmaker publishes a line-item HBM watts figure separate from total package power; the model picks a per-accelerator share inside a 10–16% range (industry write-ups commonly cite roughly 8–15% for HBM3/HBM3e-class memory; Rubin’s 16% is a pre-launch extrapolation just above that), rising slightly with each generation’s wider HBM bus and higher stack count, and checked against the published rack power range in engine.test.ts.
Unit substation efficiency (cooling/building side): ≈99%
A round, conservative figure inside the loss reductions DOE’s 2024 distribution-transformer rule targets for this equipment class; no single spec sheet states one efficiency percentage for a 2.5 MVA 34.5 kV → 480 V unit substation.
Hall battery runtime: ≈5 minutes
Set by how many battery strings are installed, not a fixed design rule; five minutes is comfortably more than the generators need to pick up load (about ten seconds to start, under a minute to carry it).
Facility/rack water temperature rise: ≈10 °C
A representative supply-to-return delta for the loop; the real number depends on flow rate and load, which this model does not simulate hydraulically.
CDU approach temperature, facility to rack loop: a few °C (≈3 °C in this model)
A typical plate-heat-exchanger approach; the exact figure depends on the specific CDU and its flow rates, which this model does not size.
Standard switchgear, battery and busway design practice: automatic utility/generator transfer; Li-ion (LFP or NMC) UPS batteries; tap-off busway
Near-universal data-center design practice rather than a scenario-specific figure -- included for context, not computed by the model or unique to one vendor.
Solid-state transformer module size: ≈2.5 MW modeled
A round illustrative module size for a product class (800 V DC solid-state transformers for 2027-era racks) that is not yet shipping in a fixed catalog size; chosen to match the unit substation size this design already uses elsewhere.
Coolant distribution unit size: ≈1.25 MW modeled
A round point inside the 70 kW – 2.3 MW range vendors such as Vertiv and Motivair sell CDUs in; the model uses 1.25 MW as its illustrative per-unit size, not one product’s exact rating.
In-row cooling unit capacity: ≈60–100 kW each
A representative capacity band for row-based CRAH/in-row air-cooling units at this density; the model does not tie the count to one named product’s datasheet.
Rubin-generation rack power range: 170–240 kW
NVIDIA publishes no rack power for Vera Rubin NVL72. The low end comes from pre-launch trade coverage of announced specifications; the high end is GIGABYTE’s listed maximum, “up to 240 kW” per rack, on four 110 kW power shelves (Cisco’s Vera Rubin NVL72 data sheet lists the same four 110 kW shelves). The model’s ≈188 kW rack sits in the lower half because it runs the GPU at the lower of its two reported power points (1,800 W; see the Rubin GPU power assumption). Carried as an assumption, not a checked spec.
Rubin-generation NIC and NVLink figures: 1.6 Tb/s ConnectX-9 SuperNIC per GPU, NVLink 6
Vera Rubin has not shipped; NVIDIA has announced these figures ahead of launch, so they are carried as an assumption rather than a checked spec until a shipping datasheet confirms them.
CDU / in-row cabinet as drawn: 0.8 m row slot; doors, louvres, touchscreen and roof valve ports representative
The hall keeps one 0.8 m slot at each rack group. Real units are larger: Vertiv lists its CoolChip CDU 1350 at 900 mm wide and its CDU 2300 at 1200 × 1200 × 2400 mm. Door, display, louvre and pipe-port positions are not published for these units, so the drawn ones are representative.
CPO stand-in faceplate arrangement: representative
The counts on the stand-in switch follow the Quantum-X Photonics Q3450 (144 MPO connectors, 18 external laser modules, 4 UDQ4 liquid connections), and each MPO adapter is drawn near its real size. Where those parts sit on the panel, and the chassis outline, are not taken from a published drawing.
Unit substation as drawn: representative
Drawn as a close-coupled unit: an enclosed medium-voltage cabinet with the feeder entering from below, a liquid-filled tank with four radiator banks, and an enclosed low-voltage throat flanged to the bus duct. Sizes, radiator and gauge placement, and the duct supports follow no single product drawing.
Facility pipework as drawn: representative
Pipe sizes, long-radius elbows, flange spacing, trapeze hangers, the valve on each coolant-unit branch and the roof curb where the risers leave are drawn to read as real pipework; none is taken from a specific hall’s drawings. Blue and red keep the supply and return colours of the legend.
Fan wall as drawn: representative
Fan-array walls are a common way to move the air side of a hall; the cell size, the 4 × 14 count, the inlet bellmouths and finger guards here are drawn to read as such a wall, not taken from a specific product or hall.
Switch faces as drawn: representative arrangement, true-size parts
Chassis height, width and OSFP cage count follow NVIDIA’s published figures, and each module is drawn at the OSFP size; how the cages are arranged on the face, how many chassis sit in each rack, and how the jumpers are dressed are representative. Each module’s lid carries a small printed label naming the module class the scenario’s fabric uses (see the printed lid labels assumption). The face prints the model the card cites (Quantum-X800 Q3400, or QM9700) and a number at each cage; the numbering order, like the arrangement, is representative, and the storage/front-end switch, whose model is not named, carries numbers only. Rack trays number their scale-out cages the same way, and DGX H100 servers carry their product name on the bezel as text, without a logo.
Aqua multimode links at one leaf switch, as drawn: representative
NVIDIA publishes the parts: the twin-port MMA4Z00-NS switch module (2 × SR4 multimode, OM4 up to 50 m, tan pull tab), the single-port MMA4Z00-NS400 adapter module (850 nm VCSEL, aqua fiber), straight cables to two 400G adapters, and 1:2 splitter cables (MFP7E20 multimode, MFP7E40 single-mode) to four 200G adapters. Which leaf port carries them, which racks and tray slots the far ends land in, where along the cable the split sits, the cable route and the drawn thickness (about twice the published 3 mm so it reads at hall scale) are representative. These cables are drawn in addition to the hall’s counted links; nothing published here says how many production links run multimode or through splitters, so the fiber, module and power counts do not change because of them.
Busway hardware as drawn: representative
Tap-off units, joint covers, end caps and threaded-rod hangers are drawn to read as plug-in busway; no specific busway product was used for their shapes or spacing. The coloured band on each tap-off keeps the voltage legend colour.
Fiber frame bays as drawn: representative
Each bay is drawn as an open frame of 1U fiber housings with side cable managers, slack spools and a top entry trough; manager widths, housing count per bay and how the jumpers drape are not taken from a photographed installation.
Detector and sprinkler heads as drawn: representative
The card cites NFPA 75 for what must be protected, not for device shapes. The aspirating detector (enclosure, bar-graph display, LEDs, pipe inlet) and the pendent sprinkler heads are generic forms, not a specific product.
Containment doors as drawn: representative
Glass roof panels and sliding end doors with aluminium frames and an overhead track are one common way to contain a hot aisle; frame sizes and the door arrangement here are not taken from a specific product.
Pipeline-stage layout drawn on the hall floor: illustrative
The site draws one plausible way to lay four pipeline stages across four racks; real deployments place stages however the cluster’s job scheduler and topology dictate, not necessarily one stage per physical rack.
Rail-optimized fabric hop count: 1 hop, same rail
A property of the rail-optimized leaf-and-spine topology this site draws (GPU port n in every rack lands on the same leaf switch), not a measured figure for one campus.
InfiniBand switch hop latency: under ≈100 ns
NVIDIA does not publish a per-hop switching-latency figure for its InfiniBand switches; this is a commonly cited order-of-magnitude for cut-through ASIC switching, not a datasheet number.
Layout inside the pluggable module: representative
This research pass found no suitable public teardown of an 800G or 1.6T twin-port OSFP module and no published DSP die size; the 800G and 1.6T scenarios share this one layout. The footprint is drawn to the OSFP size and the layers are pulled apart vertically to show them. The parts the sources name (DSP, drivers, TIAs, photodiodes, modulators, lasers, the connector) are real; where each sits on the board, their sizes, and the power converters and controller are drawn as a typical module, not one product. The model uses one shared eight-lane DSP and PIC with separate eight-channel transmit and receive banks feeding two four-lane optical ports. Public 1.6T DSP documentation supports this lane architecture; the separate driver, TIA, photonics chip and their placement are representative, not a sourced teardown of the twin-port example. The placement follows ordinary board practice rather than a product: the DSP sits right behind the edge connector so the host lanes are short and straight; the point-of-load converters sit beside the DSP’s line-side edge, outboard of the bus to the driver and TIA, which keeps clear of them; the lasers sit at the photonic chip’s host-side edge between the pads of the lanes they feed; and each driver output reaches its modulator over a few millimeters, with no laser feed crossing it.
Exterior details of the pluggable module: representative
The OSFP MSA and NVIDIA’s MMS4A00 datasheet give the body envelope (22.58 mm wide, 13.0 mm tall), the dual-MPO channel orientation and 10.0 mm port pitch, and show a concave saddle ahead of a full-height nose. They do not dimension the saddle, the fin count or pitch (the MSA allows the heat sink to differ from its reference), the wall thickness, the receptacle bodies, sleeves, latch arms or bezel, so those are drawn as a typical module. The ferrule face (6.4 × 2.5 mm, 12 positions at 0.25 mm) follows the common MT ferrule, not a vendor drawing. A photo audit of maker product photos and drawings (NVIDIA MMS4X00, Lumentum, Coherent Corp, Accelink, Marvell, Innolight; research/osfp-module-styles-2026-10-02.md, 10/02/2026) found bare silver aluminum shells, lengthwise fins covering roughly 35 to 90% of the body where a module carries its own heat sink, and many flat-top modules with no fins that rely on the host cage’s riding heat sink. The finned top drawn here follows NVIDIA’s MMS4X00 drawing; it is one common style, not the only one.
Printed lid labels on pluggable modules: module class, reach and no part numbers
Each pluggable module drawn in the scenes carries a small printed label naming its class for the scenario: at the switch, OSFP 800G 2xDR4 for the 400G-per-GPU fabrics (NVIDIA’s MMS4X00 twin-port, "2xDR4") and OSFP 1.6T 2xDR4 for the 800G fabric (NVIDIA’s MMS4A00), and OSFP 800G 2xSR4 on the aqua multimode links at the GB200 leaf (NVIDIA’s MMA4Z00-NS); at the NIC, OSFP 400G DR4 for GB200 (MMS4X00-NS400), OSFP 800G 2xDR4 in the DGX H100’s twin-port cages and OSFP 800G DR4 for GB300 (NVIDIA lists a single-port 800G OSFP DR4 for the NIC). Vera Rubin’s module type is not published, so its label gives only OSFP 1.6T. The module level prints the label of the module on screen: on its switch side (the default) the same label as the hall’s switch faces (OSFP 800G 2xDR4, OSFP 1.6T 2xDR4, or OSFP 1.6T for Vera Rubin), on its NIC side the tray’s own (OSFP 400G DR4 for GB200, OSFP 800G DR4 for GB300; the DGX H100 and Vera Rubin print the same label at both ends), a third line giving the reach, and LPO in its LPO view (its LRO view keeps the plain print: no source found names LRO on a lid), the way vendors name linear-drive modules (STORDIS lists an S-OSFPC-16TDR4+L "1.6T DR4+ LPO" OSFP; Eoptolink files its OSFP 2x400G DR4 under "800G LPO OSFP"). The reach line reads "1310 nm · 500 m" for every tier drawn here: the OSFP MSA’s own DR4 color-code row (rev 5.22 sec. 3.8, Table 3-3: "OSFP 1310nm solutions for up to 500m ... DR4") and NVIDIA’s MMS4X00/MMS4A00 datasheets (both 1310 nm, up to 500 m) agree, and Vera Rubin’s own unpublished module is already drawn as this same DR4-style twin-port family, not a new claim. The coherent pluggable (Reed, 10/02/2026) shares this module’s own shell and lid field rather than a separately authored one (module-exterior-detail), so its label uses the same plate, ink and type sizes, printing OSFP / 800ZR / C-BAND DWDM instead. Each copper plug prints its class (DAC, ACC, AEC) on its lid. Label stock, type and layout are representative and encode nothing; real labels carry vendor part and serial numbers that are not drawn, and neither label carries a barcode or 2D code block.
Nameplates, voltage stencils and hazard signs: model values only; designations and sign artwork representative
Every figure printed on the power gear comes from this model: the voltage staircase (345 / 34.5 kV at the main transformers, 34.5 kV at the campus switchgear, 34.5 kV / 480 V at the unit substation, 415 V AC or 800 V DC on the busways, ≈50 V DC on the NVL72 busbar), the main transformers’ unit rating from the campus card and the unit substation’s 2.5 MVA assumption. Designations such as T1, E-HOUSE A and LINEUP 1 are representative names, not a site’s equipment numbers. The yellow high-voltage and arc-flash signs are generic and carry no incident-energy, boundary or PPE-category figures, which would come from a site’s own arc-flash study.
Pipe markers on the cooling water: ASME A13.1 style; wording and spacing representative
ASME A13.1 marks cooling water green with white letters and requires an arrow for the flow direction (summarised by Project Materials; the standard itself is not quoted here). The markers read CHW for the chilled-water designs and FACILITY WATER for the warm-water design on the hall headers, risers and CDU drops, and TCS (technology cooling system, the rack loop) on the rack manifolds. Marker lengths, letter heights and positions are representative, not a site’s pipe-marking schedule; the blue and red pipe colours remain this site’s supply and return legend, not an A13.1 colour.
Markings on the chip packages: device names only, text, no logos or part numbers
Packages that are visible carry the device name the scenario uses, laser-etched style: the GPU (H100, B200, B300 or Rubin) on the package substrate in the tray and on the GPU package level, Grace, Vera or Xeon on the CPUs, and ConnectX-9 and BlueField-4 on the Vera Rubin tray’s network packages; the HBM stacks keep their memory-type print. Where a package is covered by a heat sink or cold plate (the GB200/GB300 network chips, the NVLink switch chips, the DGX H100 network chips) it is left unmarked rather than printing on the cooler. Real packages also carry part numbers, date codes and logos, none of which are drawn.
Building names, rack labels and U numbers: representative designations
HALL A and HALL B on the halls, OPERATIONS on their office wings, SUBSTATION on the control house, FIBER ENTRANCE cast into the fiber vault lids, ROW n · Rnn on each rack in the hall and U numbers up the rack posts are the kind of identifiers sites use, drawn to help a reader find their way. The names and numbering order are representative, not any campus’s own scheme.
Drive, CDU and patch-lead markings: sourced or marked as modeled
GB300 trays mark their E1.S drives 7.68 TB: Lenovo’s GB300 NVL72 product guide (LP2357) lists E1.S NVMe drives of 7.68 TB and 3.84 TB and a 61.4 TB tray using 8 × 7.68 TB; other generations’ drives carry no capacity, which their sources do not give. Each CDU’s nameplate states the model’s ≈1.25 MW unit size and says it is modeled, not a rating. Scale-out patch leads carry MPO-12 tags (2× on the twin-port modules), the connector the module datasheets give; the tags themselves are representative.
Power of the single-port 800G DR4 module in a GB300 tray: not published; between 9 W and 17 W
NVIDIA lists the GB300 SuperNIC cages’ module as "NVIDIA single port transceiver, 800Gbps, OSFP DR4, MPO, APC 1310nm SMF, up to 500m" in its Quantum-X800 cluster documentation, with no power figure there. This research pass found no NVIDIA datasheet for it. The nearest published ratings bracket it: 9 W max for the 400G single-port DR4 (MMS4X00-NS400, four lanes at 100G) and 17 W max for the 800G twin-port (MMS4X00, eight lanes at 100G); the 800G single-port runs four lanes at 200G, so the card says only that it sits between them. Not a measured or vendor figure.
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
The module level has two sides (Reed, 10/03/2026). The switch side (the default, and what the top bar’s tab and an address open) is the module in the switch cages, the one the hall’s lid labels name for the scenario; the NIC side, opened from the compute tray’s module door or with ?module-side=nic, is the module in the tray’s own cages. NIC side: GB200 trays take NVIDIA’s MMS4X00-NS400 single-port 400G OSFP DR4 and GB300 trays a single-port 800G OSFP DR4 (NVIDIA’s Quantum-X800 transceiver list, platform NIC); each is drawn as ONE DR4 engine of the twin-port asset, the second engine’s lanes, modulators, photodiodes, lasers, pads, fibers and connector removed (four lanes each way, one MPO), because NVIDIA publishes no internal layout for either and a DR4 port is by definition four lanes each way. The DGX H100 carries the 800G twin-port at both ends (the server’s flat-top version) and Vera Rubin’s type is unpublished at both ends, so for those two the toggle names the end and draws the same module. Switch side: The module in the switch cages, the one the hall’s lid labels name for the scenario: NVIDIA’s MMS4X00 800G twin-port OSFP (2 × 400G DR4, 100G-PAM4, 17 W max) for the 400G-per-GPU H100 and GB200 fabrics, and NVIDIA’s MMS4A00 1.6T twin-port OSFP (2 × 800G DR4, 200G-PAM4, 33.5 W max) for GB300’s 800G fabric. Both are finned twin-port OSFPs with eight optical lanes each way on two MPO-12/APC connectors, so the level draws both with the same internals and changes only the captions, the DSP’s printed rate and the lid print; NVIDIA publishes neither module’s internal layout. Vera Rubin’s switch-side module type is not published: the level keeps the 1.6T drawing and its lid prints only OSFP 1.6T. The captions, the DSP’s printed rate, the lid print, the top bar’s tab text and the cards follow the side on screen.
One DSP or two in the twin-port module: one, as drawn; not published
NVIDIA’s twin-port datasheets describe two independent ports ("two transceivers inside", "two transceiver engines") but not how many DSP chips carry them. Single 8-lane DSPs are sold for exactly this use (Credo’s Cardinal 1650 lists 2x800G DR4; Marvell’s Ara brief shows 2xDR4), and one DSP per port is also possible. No public teardown settles it for NVIDIA’s module, so the drawing shows one DSP as a representative design.
The half-retimed (LRO) module, as drawn: representative
The OIF RTLR announcement, the LPO MSA FAQ and Credo’s Dove 850 release establish the architecture: a DSP retimes the transmit path only, the receive path is linear, and the host’s own signal processing does the receive work. No teardown of an LRO module was read, and Marvell’s Ara T product brief could not be retrieved. So the drawing keeps the full-DSP module’s package and position and marks its die top as a transmit retimer with the receive half hatched; a real transmit-only DSP is its own chip, with a size, pin-out and placement not shown here. The receive copper reuses the LPO layout’s direct routing from the TIA to the edge connector. Fewer heat arrows over the DSP and a dimmer shell glow mark its smaller heat load on the site’s log heat scale (heat-visual-scale): three of the full DSP’s six arrows, for an assumed 6 W against 14 W.
A silicon photonics module: one design of several
A DR8-class module can use directly modulated lasers, EMLs, or a silicon photonics chip lit by separate continuous-wave lasers. NVIDIA’s own 800G twin-port (MPO-16) datasheet names a 1310 nm EML laser; its 1.6T twin-port datasheet does not say. The level draws the silicon photonics kind because it shows the laser and the modulator as separate parts; the other kinds put both in one laser chip. How many lasers feed how many lanes is also drawn, not sourced: four lasers, each split to two modulators, is the arrangement drawn.
Inside a copper cable head: representative
What each kind carries is public: nothing in the signal path for a passive cable, one redriver for an ACC, a retimer in each end for an AEC. This research pass found no suitable public teardown of a head, so the card, chips and cable are drawn as typical, not as one product. Package styles are representative too: the ACC redriver as a small leaded QFN and the AEC retimer as a lidded flip-chip BGA with decoupling capacitors; their sizes, marks and placement are not taken from a named part. The card edge carries the QSFP112 MSA’s 38 contacts, drawn as 19 per face with grounds longest, then power, then signal (the MSA’s mating order); pad pitch, stagger lengths and which pairs use which face are simplified, not the MSA pin table. The strain-relief boot, crimp collar and flat pull tab follow exterior photographs; their shapes and sizes are representative. Each twinax pair is drawn as two insulated conductors with a drain wire inside a foil shield, the construction described in patent US20160073559A1; conductor gauge, insulation and shield sizes, the solder termination and the round pack inside the jacket are representative.
Discrete coherent electronics and optical assemblies: representative separate board packages
The drawing intentionally uses four distinct board-mounted assemblies: a packaged driver IC, a packaged TIA IC, a transmit optical assembly and a receiver optical assembly. The opaque analog packages do not share an optical carrier. This discrete board-package arrangement, its interfaces and its dimensions are design assumptions, not a verified 800ZR product layout. Sumitomo demonstrates distinct analog ICs assembled with optics inside CDM/ICR packages; Coherent sells a separate driver/TIA chipset for 400G ZR/ZR+, not an 800ZR-qualified board-package arrangement. These sources support separate electronic functions and ICs, but do not establish the exact physical packaging drawn. Integrated CDM, ICR and shared optical-subassembly designs are alternatives.
Inside the 800ZR module: representative
The functional chain is DSP to driver to modulator on transmit, and receiver photodiodes to TIA to DSP on receive. No cited teardown establishes this board layout. The order along the board follows from public packaging rules rather than a teardown: the OIF HB-CDM, micro-ICR and IC-TROSA agreements put each optical package’s RF interface at one end and its fibers at the opposite end, IC-TROSA groups its high-speed contacts to reduce skew against common coherent DSP ball maps, and Furukawa’s IC-TROSA drawing puts the driver and TIA at the RF end and the laser at the LC end. So the drawing places the DSP nearest the host connector, the driver and TIA beside its line-side edge with traces of about 6 mm, the optics next with every fiber port facing the front, and the laser toward the fiber end, its pigtail running straight back to the splitter; the line fibers pass the laser in the 2.5 mm channels beside its case. Exact positions, trace lengths, fiber routes and which laser end carries the pigtail are illustrative. Only the OSFP envelope and the published nano-ITLA case are to scale; the laser is a research example, not a verified part of a shipping module. The discrete board-mounted driver and TIA packages, separate optical assemblies, other component sizes, positions, routing and vertically separated display layers are illustrative. Package styles (QFN-style driver and TIA packages, a lidless DSP with underfill, stiffener and decoupling ring, photonic die edges and fiber-attach blocks), board passives and small controller packages are representative detail, not taken from a teardown. The card edge carries 30 pads a side (60 contacts) staggered for OSFP MSA Rev 5.0 sequenced mating (ground first, then power, then signal) with wider power pads at 15/16 and 45/46; which remaining pads are ground or signal is representative.
Coherent DSP interface drawing: schematic routing groups
Four illustrated host-path groups organize the view; they do not specify host electrical lane count, DSP ball count, a pinout or a one-to-one lane mapping. Separate host-side digital and line-side analog terminal banks identify different interfaces. Signals undergo transmit processing and DAC conversion, or ADC conversion and receive processing, inside the DSP package; no straight conductor or passive bypass through the chip is implied. Exact bank placement and converter implementation remain representative.
OSFP housing, pedestal and gap pad: representative
The lifted top housing is drawn as a die-cast cover with a closed-top heat sink: longitudinal fins under a flat top skin, open at both ends. OSFP MSA Rev 5.0 section 3.3 says the module allows airflow along its length and gives closed-top heat-sink examples; Cisco lists an integrated closed-top heat sink and 13 mm maximum height for its non-coherent OSFP 800G modules, and says its 800G ZR/ZR+ OSFP modules are mechanically compliant with the OSFP MSA. No cited source shows this 800ZR housing, so fin count, pitch and heights are illustrative. Under the lid, a machined pedestal and a soft gap pad over the DSP give heat from the die a drawn route into the housing; pedestal size, pad thickness and material are illustrative too. The lower case follows OSFP MSA Rev 5.0 section 3.2 in kind: a nose whose upper lip carries ventilation slots, and side walls that rise to at least 7.0 mm at the forward stop. Its cast pockets, wall bosses, latch pockets, front bulkhead and every position and size are representative.
OSFP shell, label and pull tab as drawn: representative
Published: the OSFP MSA (rev 5.22) recommends a label on the top face at the fiber end, about 15 × 20 mm (Fig. 3-4), allows a pull tab (Appendix B) and requires a color code on the pull tab or by another method (sec. 3.8). Its color table gives white for 1550 nm modules up to 80 km (example ZR8) and has no row for coherent modules such as 800ZR, so the white tab is the nearest entry, not a stated rule; vendors differ (Cisco’s non-coherent OSFP 800G modules use purple, orange and green tabs). Cisco lists 116 mm maximum length with pull tab for those modules, and the drawn tab stays within it. Reed’s call (10/02/2026): the coherent pluggable shares its shell and pull tab with the pluggable module (module-exterior-detail) rather than a separately authored shape, so the same finned shell with a flat label field, and the same rectangular tab, draw here too, recolored white; only the label wording (OSFP, 800ZR, C-band DWDM, no part number) differs. The finned shell is a representative choice, not a sourced one: the only 800ZR OSFP confirmed by photo in the 10/02/2026 audit, Marvell’s COLORZ III, is flat-topped with no fins and a white rectangular pull tab, and one maker is too few to call either style standard for coherent modules.
Duplex LC receptacle as drawn: representative
Cisco states that the media interface of its 800G ZR/ZR+ modules is a duplex LC connector. The molded body, bore depth, zirconia split sleeves, ferrule stub faces, board bracket and the 6 mm port pitch drawn here are illustrative; no cited teardown shows the receptacle inside an 800ZR module.
Coherent optical-component power and heat: well under half a watt each; no heat drawn
The modulator and receiver optics have separate illustrative bias/control feeds. Their own dissipation is small: an InP modulator’s bias and control take a few tenths of a watt at most (assumed ≈0.3 W), and photodiode bias milliwatts (assumed ≈0.05 W), against a couple of watts for the driver IC that swings the modulator and about a watt for the TIA. Both sit under the site’s half-watt drawing threshold (heat-visual-scale), so no heat stream rises from them; the driver’s and TIA’s heat is drawn at their own packages. The diagram does not specify bias topology or temperatures. Exploded gaps are display spacing, not thermal interfaces.
How much heat each part is drawn with: log scale: each 10× in power is 5× the motion; nothing under 0.5 W
Parts in one view span about a watt to about a kilowatt, so heat motion (how many pulses, how big, how bright) is drawn on a log scale: each part’s motion is its watts divided by the largest source’s in the same view, raised to the power log10(5) ≈ 0.7, so every 10× in power is the same step, 5× the motion. A 1 kW chip and a 10 W chip in one view differ about 25×; rank order always holds, and a part drawing more power never shows less heat than one drawing less. Parts under half a watt (photodiodes, modulator bias, a passive copper cable’s ID memory, fibers and connectors) draw no heat at all. Coolant and air loops are compared with each other, by the heat they carry, and each view keeps its total motion within a fixed budget, so a busy view is scaled down as a whole, its ratios kept. Watts come from the model where it has them (GPU, HBM share, CPU, NIC and DPU shares, rack and hall splits between water and air, campus heat) and otherwise from these order-of-magnitude allocations: pluggable module DSP ≈14 W, drivers ≈3 W, TIAs ≈2 W, lasers ≈3 W (from Semtech’s 23–25 W full-DSP, ≈16 W LRO and ≈10 W LPO module figures; an LRO DSP ≈6 W); 800ZR DSP ≈16 W, tunable laser 2.9 W (EFFECT Photonics’ published figure), driver ≈2 W, TIA ≈1 W, modulator bias ≈0.3 W, photodiode bias ≈0.05 W (module 24–25 W); CPO switch ASIC ≈550 W and each optical engine ≈13 W (NVIDIA’s 3.95 kW Q3450 and ≈9 W per port, over four packages), and in the Mach-Zehnder package each 6.4T engine ≈44 W (Broadcom puts a Tomahawk 6 CPO port near 3.5 W per 800G, 36.4% below the Tomahawk 5 CPO port, so ≈5.5 W per 800G port and eight per engine) with the switch chip drawn at the same 550 W; copper ACC redriver ≈2 W, AEC retimer ≈10 W, DAC 0.1 W per end; in the DGX H100 server, an NVSwitch ≈60 W, a DDR5 RDIMM ≈10 W and each power supply’s loss ≈70 W (10.2 kW at about 96% over six supplies), beside the Xeon 8480C’s published 350 W and a ConnectX-7 card’s 24.9 W typical; on a warm-water campus the evaporative towers trim about a fifth of the heat on the hottest afternoons. Only ratios within a view matter, so these shape the drawing, not any figure on a card. Pulse speed and the height a stream rises are not measurements.
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
In the Power layer every part that draws electrical power wears a soft halo around its footprint, on the same log rule as the heat drawing (heat-visual-scale): a part’s glow is its watts divided by the largest draw in the same view, raised to the power log10(5) ≈ 0.7, and the halo’s brightness and reach each take the square root of that, so every 10× in power is the same step and a part drawing more never glows less than one drawing less. Parts under half a watt (a passive copper cable end, photodiode and modulator bias) do not glow, and passive parts never do: fibers, MPO and cable assemblies as such, passive optics, heat sinks, cold plates, connectors, busbars, busways and stored-energy batteries. Equipment that passes power on glows by what it loses doing so: voltage regulators, bus converters, power shelves and supplies, UPS modules, transformers and switchgear. Where a level compares variants (the module’s full DSP, LRO and LPO, the CPO level’s two packages), the reference stays the full design, so a variant that draws less glows less. Watts are the heat drawing’s, the model’s power ledger for transformer, UPS and solid-state transformer losses, the fabric table for modules and switches, and these further order-of-magnitude allocations, used only for the glow: a module’s or 800ZR’s converters ≈2 W, an external laser module ≈5 W, a soldered LPDDR5X package ≈2 W, a PCIe switch ≈25 W, a CDU’s pumps about 1% of the heat it moves, fans about 3% of the heat their air carries, low-voltage switchgear about 0.1% of what passes through it; a tray’s NIC and DPU power, and its drive, fan and management power, are split evenly over the parts that draw them. Only ratios within a view matter. The halo’s slow breathing is decoration and stops with reduced motion; views without bloom draw it a little brighter, and Battery saver leaves out the faintest halos.
Layout of the CPO switch package: representative
NVIDIA publishes the counts (six optical subassemblies of three engines per switch chip, 18 fibers per engine); this research pass found no dimensioned drawing. Package and die sizes, the die’s underfill and surface finish, subassembly placement, cold-plate geometry and its drawn internal fin channels, circuit floorplans, the external laser modules’ enclosure, fins, pull tab and front-panel bezel, and physical fiber routing (drawn as flat 8-fiber ribbons with rounded bends) are representative. NVIDIA’s Hot Chips 2025 presentation, PDF page 14, identifies COUPE microlens surface coupling. The drawn MT-style connectors, fiber-array units, routing bends, and planar optical paths explain connectivity; they do not reproduce that coupling stack or its microscopic geometry. The exploded EIC/PIC gap and alignment guides are explanatory, not real long bond wires. NVIDIA describes socketed electrical connections for the three-engine optical subassemblies, not separate sockets for every engine.
CPO electronic die floorplan, as drawn: eight driver and eight TIA blocks on the EIC; modulators and photodiodes on the PIC; placement representative
Which circuits sit on which die is reported, not assumed, for the COUPE engines the ring view draws: SemiAnalysis (01/01/2026) reports that each engine in NVIDIA’s Quantum-X switch pairs a photonic die holding the modulators, waveguides and detectors with an electronic die holding the drivers, TIAs and control logic, hybrid-bonded with TSMC COUPE, and that the COUPE platform’s electronic die integrates the modulator drivers and TIAs. NVIDIA (Hot Chips 2025) and TSMC (its 2024 symposium release) document the stacked COUPE engine; NVIDIA has not published its electronic die’s floorplan. What stays assumed is the drawing: eight transmit-driver blocks and eight receive-TIA blocks on the EIC, their placement over the rings and photodiodes, and the generic die detail (a seal ring, one analog block per lane with standard-cell rows, a spiral inductor, a top-metal power grid, and a hybrid-bond pad array at a representative pitch on the underside), drawn on the visible top surface so it can be seen. No process-node sizes are claimed for the drawn engine. Published designs differ: Broadcom’s 6.4 Tb/s engine, as reported from its ISSCC 2026 paper 23.4, puts all 64 channels on one 7 nm electronic die driving Mach-Zehnder modulators; Daudlin et al., OFC 2021, demonstrates driver and TIA functions on one bonded EIC in a research prototype; Marvell’s 2021 light engine flip-chips a separate driver chip and a separate TIA chip onto its photonics chip; Ranovus Odin and Ayar Labs’ TeraPHY put the electronics and photonics on one monolithic die. The toggle draws a Mach-Zehnder package beside this one (see its own assumption).
The Mach-Zehnder package, as drawn: representative; counts Broadcom’s, proportions from its package images
The toggle on the CPO level shows a second shipping design in its own package, in the class of Broadcom’s 51.2T Bailly. What is published: a Tomahawk 5 switch chip with eight 6.4 Tb/s optical engines (Broadcom, 03/14/2024; SemiAnalysis, 04/15/2026: one PIC and one EIC per engine); 6.4T-FR4 engines with a Broadcom Fiber Connector; 128 front-panel 400G FR4 ports on duplex LC connectors; remote laser modules; edge-coupled fiber attach; a 4RU air-cooled reference system; and, from write-ups of Broadcom’s ISSCC 2026 paper 23.4, one 7 nm electronic die per engine carrying all 64 channels, segmented Mach-Zehnder drivers in three segments and TIAs that drive the switch chip directly. What is drawn is representative: the package and die sizes; eight engines as radial tiles, two per side, about 2.2 : 1 with their connectors, with the electronics at the switch-chip end and the connector at the outer end, in the proportions of Broadcom’s package images (its Tomahawk 6-Davisson slide as shown by ServeTheHome; Davisson has 16 engines, this drawing keeps Bailly’s eight); the electronic die stacked on only the electrical end of the photonic die, reached from the package through through-silicon vias (TSVs) in the photonic die (a few columns drawn in the detail; their count, pitch and size representative); the tile floorplan, with all 64 lanes each way as sixteen FR4 groups of four wavelengths, laser light on two fibers tapped to each group’s wavelength demultiplexer (that split is not published), multiplexers onto the 16 transmit fibers and demultiplexers off the 16 receive fibers; the electrode, pad, heater and photodiode positions; 16 transmit and 16 receive fibers per engine (128 ports over eight engines) drawn as two ribbons; the remote laser modules’ count, shape and allocation; the heat sink’s fins; and every flow path. The drawn Mach-Zehnder arm length is representative (real silicon Mach-Zehnder modulators are millimeters long). For comparison NVIDIA’s published package image (NVIDIA, 03/18/2025) shows its engines three to a subassembly along each side of the switch chip, as the micro-ring view draws them.
Rubin scale-out link selection: two 800 Gb/s physical links per GPU; illustrative optical fabric
NVIDIA publishes eight ConnectX-9 devices per four-GPU compute tray and 1.6 Tb/s aggregate scale-out bandwidth per GPU. This model selects two 800 Gb/s physical links per GPU and a Quantum-X800 InfiniBand optical fabric. That is an illustrative network configuration, not a claim that every Rubin installation uses this protocol, transceiver or cage arrangement. Internal placement and optical-module packaging remain representative.
1.6T optical module power: ≈25–30 W
A legacy hypothetical single-port 1.6T fabric option, not the current Rubin configuration (two 800G links per GPU). Two published reference points illustrate the range: NVIDIA’s own 1.6T twin-port module (MMS4A00) is rated 33.5 W max, and Broadcom’s Sian3 DSP targets 1.6T modules under 23 W. The hypothetical 1.6T table carries ≈27 W per module. The selected 800G fabric instead uses 17 W per NIC-side module and 33.5 W per twin-port switch module.
Bus-converter (IBC) efficiency: 97–98%
Neither NVIDIA nor the shelf/tray vendors publish a board-level efficiency figure for the fixed-ratio 50 V→12 V (or 54 V→12 V) bus-converter bricks. 97–98% is the range power-semiconductor vendors describe in general terms for this class of unregulated converter; it is not a disclosed number for a named accelerator board.
Voltage-regulator (VRM) efficiency: 90–92%
NVIDIA does not publish board-level VRM efficiency for these racks. The value is a mid-range estimate for a multiphase buck converter stepping 12 V down to under a volt at over a thousand amps, based on the general ranges power-delivery vendors (Infineon, Vicor, MPS) describe for this class of design, not a disclosed figure for a named board.
VRM phases per GPU: ≈20–30
Vendors describe multiphase GPU power delivery only in relative terms ("up to 12 phases" per module, stackable) and do not publish a phase count for a named accelerator board. 20–30 phases is a plausible total for a ~1.2–1.8 kW rail split across several parallel converter modules at typical per-phase currents for this class of design.
Two power feeds into a rack (A/B): A + B tap-offs
Feeding every rack from two separate busways, each on its own UPS side, is standard practice for concurrently maintainable (Tier III/IV-style) data center power, not a figure NVIDIA publishes for these racks. NVIDIA states only that the GB200 rack’s eight power shelves "provide N+N redundancy"; the split of shelves between the two feeds is drawn, not published.
DGX H100 bezel and rear panel as drawn: representative
NVIDIA’s user guide publishes the metal-foam bezel and its handles, the front power button, ID button and fault LED, six power supplies and C19/C20 cords, and four OSFP ports. It does not publish a dimensioned drawing, so the foam cell size, the handle and control positions, and the rear arrangement (supplies across the bottom, exhaust perforation, management ports) are drawn as representative, not measured from a unit.
NVLink cartridge and busbar hardware as drawn: representative
Four rear copper cable cartridges, a blind-mate backplane that the compute and switch trays slot into, and a rear busbar are published. The cassette construction, its cable windows and the twinax visible in them, the connector housings where each tray meets a cartridge, and the contact lands on the busbar are drawn as representative mechanics, not measured from a rack or taken from a vendor drawing.
Busway and tap-off units as drawn: representative
Overhead plug-in busway with tap-off units feeding each rack is generic data center hardware. No vendor page was fetched for a specific product, so the ribbed housing, hangers, outlet covers, joint pack, the tap-off box with its breaker handles and cord grip, and the rack-top grommets are drawn as a typical installation, not a named product.
Management switch front panel as drawn: SN2201-class: 48 × 1 GbE RJ45, 4 × 100 GbE QSFP28
NVIDIA’s DGX GB hardware guide lists two top-of-rack management switches without naming a model, and the DGX H100 guide names none. The face is drawn from NVIDIA’s own 1 GbE management switch, the SN2201, whose published ports are 48 RJ45 at 1 GbE and 4 QSFP28 at 100 GbE. That any given rack uses this model is an assumption; the port grouping, console jacks, link LEDs and patched leads are representative.
NVL72 tray and shelf front panels as drawn: representative
Published: four E1.S drives, the network ports and management ports on each compute tray front (ServeTheHome; NVIDIA DGX GB hardware guide), gold removal handles on the NVLink switch trays (ServeTheHome), and six supplies per power shelf (Flex, LITEON). Not published as drawings: the vent perforation, drive-sled latches and light pipes, label strips, supply fan grilles and handle shapes are drawn as representative. No public Vera Rubin tray-face photograph was found, so Rubin trays use the same graphite finish with champagne handles as a representative choice.
Coolant manifold hardware as drawn: representative
Published: vertical supply and return manifolds down the rear of the rack, and trays that blind-mate their liquid-cooling nozzles when pushed home (ServeTheHome, DGX GB200 NVL72). Not published as a drawing: the stainless finish, the coupler shape (body, knurled sleeve, latch ring and colored band), the rigid stubs, mounting brackets, bleed valve, floor ball valves and hoses. The coupler model is not identified; these parts are drawn as typical rack liquid-cooling hardware.
Unused rack units above the switches: 1U brush-strip cable manager and 1U blanking panels
Filling empty rack units with blanking panels so cold air cannot bypass the equipment is standard data center practice; no NVIDIA page shows what occupies this rack’s free units. The cable manager, panel count and styling are drawn as representative, not as a documented configuration.
DGX H100 power strips and cord routing as drawn: representative
Published: six supplies per server and C19/C20 power cords 1,200 mm long (NVIDIA DGX H100 user guide), and C19-class outlets on 415 V rack strips at this density (the power-distribution source on the strip card). Not published for this rack: the strip model, its outlet banks and colors, the breaker head and meter, which outlet feeds which supply, and the cord routing, all drawn as representative.
Management leads to the trays as drawn: one RJ45 copper lead per tray; routing representative
Published: NVIDIA’s GB200 SuperPOD reference architecture puts the management ports of the compute trays and switch trays on the out-of-band network and builds it from SN2201 switches, whose ports are 48 RJ45 at 1 GbE. Not published: how many management leads each tray takes, where its jack sits on the tray face, which switch port feeds which tray, the jacket color, or how the leads are dressed. The drawing gives each tray one lead, routes them from the upper switch up through the brush-strip cable manager and down a combed harness inside the right-hand post, and leaves the two trays pulled out for inspection unplugged. GB300 and Rubin racks reuse the same drawing; the reference architecture cited is GB200’s. No link speed is stated for the network itself, so the rate stays an assumption (bmc-network-speed).
Management/BMC network speed: 1–10 GbE class
NVIDIA’s hardware guides describe the rack’s out-of-band management switches (for BMC, sensor and power-control traffic) without stating a link speed; 1–10 GbE is the class commonly used for BMC/IPMI networks industry-wide, far below the fabric that carries model data.
NIC-to-cage cables as drawn: representative
Published: on GB200 the ConnectX NICs sit on a mezzanine board and DensiLink cables carry their lanes to the OSFP cages at the front (SemiAnalysis, which reports the same DensiLink approach from ConnectX-7 to the cages in DGX H100); on GB300 the four front cages sit on two OSFP boards of two ports each, the ConnectX-8 chips on two CX8 boards, and the parts list includes a riser cable (Lenovo LP2357). Not published: the cable path. One flat twinax assembly per port, its connector heads, the tie-bar clip, the standoffs, the GB200 and H100 cage boards and where each cable lands are drawn as representative, not taken from a vendor drawing.
Compute tray mechanical detail: representative
Public sources fix what each tray carries (GPU, CPU, NIC, DPU, memory and port counts) and how it is cooled, but no vendor publishes its mechanical drawings. Cold-plate bodies, coolant bosses and captive screws, internal manifold tube routing, voltage-regulator placement, memory-module screw positions, heat-sink heat pipes, fan-module rotors, busbar-clip fingers, hose finishes and bare-die appearance are drawn as representative hardware, not recovered from a product drawing. The circuit boards are too: solder mask and its glass-weave sheen, the routed trace groups, via fields and edge stitching, copper pours, silkscreen outlines and reference designators, the 0402/0201 capacitor and resistor clusters, test points and plated mounting holes follow common server-board practice around the published devices, but their layout is representative, not a vendor’s routing, placement or bill of materials.
DGX H100 internal layout as drawn: representative
NVIDIA’s DGX H100 user guide figures fix the arrangement: from the rear, the GPU tray on top, the motherboard tray under it and six supplies across the bottom; twelve fan modules over eight U.2 drives at the front; a midplane at the front end of both trays carrying power, PCIe, sensors and signaling; in the motherboard tray, rear to front, the riser slots and OSFP cages, the two CPUs with 32 DIMMs, then two network modules on an interposer board; in the GPU tray, rear to front, the four NVSwitches, the eight GPUs, then the midplane. Its topology figure shows three PCIe switches (one per CPU for drives and a storage NIC, one for the NVSwitches) and each GPU linked straight to its ConnectX-7. Not published: any dimension inside the chassis. Every depth, the midplane’s openings and connector shapes, the power distribution board and 54 V copper under the motherboard tray, the supply and fan-housing depths, the HGX board outline, where the bus converters and the three PCIe switches sit, the riser card outlines and the drawn sizes of the heat sinks are representative.
DGX H100 power-supply efficiency: ≈96%
NVIDIA does not publish a measured efficiency for the DGX H100’s power supplies. Describing them as “80 PLUS Titanium class” implies performance at or above that certification’s 94%-at-50%-load minimum; 96% is a plausible value within that class, not a disclosed NVIDIA figure.
GB300 GPU package power, modeled: 1,300 W
NVIDIA lists Blackwell Ultra at “up to 1,400W” per GPU and its GB300 NVL72 reference architecture sizes the rack at “up to 142 kW”. With every GPU at 1,400 W, this model’s bottom-up rack (CPUs, NVLink switching, NICs, conversion losses) comes to about 148 kW, above NVIDIA’s own rack figure. The two maxima do not both hold at once in this sum, so the model runs the GPUs at 1,300 W, which lands the rack at ≈140 kW, inside NVIDIA’s figure.
Rubin GPU package power: ≈1,800 W
NVIDIA has not published a per-GPU power figure for Rubin. 1,800 W is the figure early supply-chain reporting gave, and it continues the generation trend (700 W H100 → 1,200 W GB200 → 1,400 W GB300). Later trade-press reporting (Tom’s Hardware) says the shipping part can run up to about 2,300 W in a higher-power mode, so this is the lower of two design points, not a disclosed spec.
Coolant rise and flow rate across an NVL72 rack: ≈10 °C rise; ≈170–195 L/min
NVIDIA does not publish a single supply/return delta or flow rate for the GB200 NVL72’s coolant loop; the OEM/CDU specification governs the built system. The figures used here are one third party’s reported estimate for the rack’s heat load and a 10 °C design rise, and other integrators’ published CDU flow ranges differ from it.
Front-to-back air temperature rise, DGX H100 rack: ≈15–20 °C
NVIDIA does not publish a front-to-back air temperature rise for the DGX H100 reference rack. 15–20 °C is a common design range for air-cooled server racks at this power density, given a cold-aisle supply in the ASHRAE-recommended 18–27 °C band.
Vera Rubin rack mechanical representation: 18 compute trays, 9 switch trays; representative internal placement
NVIDIA now documents four GPUs and two Vera CPUs per compute tray, eight ConnectX-9 devices and one BlueField-4, and four NVLink switch ASICs per switch tray. Those counts are published. The model’s precise dimensions, routing and component placement remain representative; a pulled-out switch tray is one of the nine, not an additional tray.
Rubin modeled heat capture: idealized liquid share = 100%
NVIDIA documents a fully liquid-cooled, fanless and hose-free compute tray with rigid cooling connections. The energy model assigns all rack heat to the liquid loop as an idealization. This does not assert zero incidental convection or radiation, or that every rack part has a separately documented cold plate.
Vera memory population: up to 1.5 TB LPDDR5X per CPU
NVIDIA’s current product table publishes 1.8 TB/s NVLink-C2C and up to 1.5 TB LPDDR5X per Vera CPU. Installed memory population and exact module placement are representative; the capacity is a published maximum, not a guarantee for every system.
Rubin specification snapshot: 288 GB HBM4, 19.2 TB/s
The current NVIDIA Vera Rubin NVL72 product table, checked 9/29/2026, lists 288 GB HBM4 and 19.2 TB/s per GPU. Older launch material lists 22 TB/s. This build uses the current product table and preserves the discrepancy in its evidence notes instead of mixing specification vintages.
Rubin NVL72 power shelves as drawn: 8 × 1U, the GB200 layout
The rack level draws every NVL72 generation with the eight 1U, 33 kW shelves NVIDIA states for GB200 and GB300 (DGX GB200 user guide: “six air-cooled 5.5kW PSUs in eight power shelves”). Shipping Vera Rubin racks use fewer, larger shelves instead: Cisco and GIGABYTE list four 110 kW shelves (GIGABYTE: 3RU each). Eight 33 kW shelves under N+N would give only 132 kW, too little for the modeled 188 kW rack; the four 110 kW shelves cover it. The drawing keeps the GB200 layout; the shelf card states the reported shelves.
Grace-to-GPU NVLink-C2C bandwidth, GB200/GB300: 900 GB/s
NVIDIA’s Grace CPU page states 900 GB/s for the Grace-to-Grace link inside the CPU-only Grace Superchip, a different pairing than the Grace-to-GPU link this row describes. NVIDIA’s Grace Hopper Superchip page states the same 900 GB/s for its Grace-to-Hopper (CPU-to-GPU) link, one generation earlier. No GB200/GB300-specific NVIDIA page was found stating the Grace-to-Blackwell figure directly, so this assumes NVLink-C2C bandwidth is unchanged across those three pairings of the same interconnect generation.
GPU throttle temperature: ≈85 °C
NVIDIA does not publish a junction throttle point for these data-center GPUs; ≈85 °C is the commonly used ballpark for silicon of this class, shown only to place the die’s working temperature relative to it.
Tray cold plates, as drawn: lifted off the chips
In a real tray each cold plate sits on its package through a thin thermal interface, and the whole tray fits a 1U slot. The drawing raises the plates above the chips so the packages under them stay visible, which makes the drawn tray taller than its slot. The gap is explanatory, not a real clearance.
GPU package stiffener ring and substrate surface: one flat ring and clustered capacitors; sizes and placement representative
Large lidless flip-chip packages commonly bond a flat metal stiffener ring to the substrate top to limit warpage, leaving the laminate edge exposed. No public drawing gives this package’s ring width, thickness or corner radii, so the ring is drawn as a typical one, not measured from a product. No public capacitor map was found either: the clusters of two capacitor sizes, their count and positions, and the solder-mask pattern are illustrative.
HBM stacks, as drawn: height about 3x real; layer bands, top print and spacer representative
A real HBM stack is a thin molded block whose top sits about level with the GPU die, so one cold plate can touch both. The drawing makes each stack about three times taller than that so its DRAM layers can be counted as bands on one cut face; the band spacing and the TSV lines are illustrative, and the memory type printed on each top is a label for the reader, not a copy of a real part marking. On H100 the sixth site is drawn as a blank silicon spacer outlined in dashes, matching five working stacks; its exact construction is not taken from a teardown.
GPU die surface and floorplan, as drawn: illustrative x-ray; not a real floorplan
The visible top of a flip-chip die is its plain silicon backside, which is how the die is drawn. The line pattern shown in the data and heat layers is an x-ray illustration of the kinds of blocks a GPU die holds (compute clusters, an L2 cache band, memory interfaces along the HBM edges, the die-to-die interface along the seam). NVIDIA publishes no floorplan for these dies, so block count, sizes and positions are not taken from a die photo.
Activity on the GPU dies, as drawn: schematic: not a floorplan, not a scheduling trace
In the data layer the dies show work in progress: each tile of the illustrative floorplan (die-floorplan-drawing) carries a point of light that fires in short, staggered bursts on rolling waves, and comet streaks run on straight on-die paths from each HBM waterfall’s landing line to the band drawn as L2 cache, out from there to the tiles, and, on two-die packages, across the die-to-die seam. None of it is measured: NVIDIA publishes no floorplan, block positions or scheduling behaviour for these dies, and real traffic is not organised in visible waves. The pattern, its rhythm (tied to the waterfall’s pulse) and its paths are illustrative.
HBM traffic, as drawn: schematic arcs over the parts; the real route is buried
In the data layer each HBM stack shows its traffic as arcs that leave the stack’s top, cross its die-facing edge and drop into the die from that side, so the flow reads from the default view. The wires themselves cannot be seen from above: they leave the stack through its microbumps, cross the interposer (on Blackwell and Rubin, the silicon bridge under that die edge; on H100, the monolithic silicon interposer) and rise into the die’s HBM interface along the edge facing the stack. That buried route is drawn too, under the parts, and is the one the layout checks test. The arcs’ height, count and landing point are illustrative.
Interposer bridges and microbumps, as drawn: representative count, size and placement
TechInsights reports that Blackwell uses the bridge-based CoWoS-L variant instead of a monolithic silicon interposer. How many bridges it holds, their sizes and positions are not public; the drawing puts one long bridge under the die seam and one under each die-to-HBM edge in a dark organic interposer body. The microbump fields under each die and stack, their pitch and the underfill are illustrative.
Solder balls and bumps, as drawn: fewer, larger and farther apart than real; representative
The package is drawn with 676 board-side solder balls on a 3 mm pitch (2 mm on the smaller H100 package) and about a thousand bumps under the interposer on a 1.8 mm pitch so that single joints stay visible. Real large server packages use far finer pitches and many more joints; NVIDIA publishes no ball map or pitch for these packages, so counts, pitch and the reflowed ball shape are illustrative.
Layers between die and coolant: die → thermal interface → lid → thermal interface → cold plate or heat sink; H100: no lid
A reseller listing for the H100 SXM5 module describes a bare die with no heat spreader, the heat sink mating directly to the silicon, so on H100 the lifted plate is drawn without a skirt and stands for the heat sink’s flat base. No public source found states whether the Blackwell or Rubin packages carry a lid, and they may ship bare-die like H100; the chip level draws them with a representative lid for clarity, a flat plate with a straight skirt like the heat spreaders in public package photos, and it is drawn see-through, with an illustrative glow over the dies, so the heat path stays visible. The interface materials, their thickness and the plate’s dimensions are representative; the exact stack is proprietary to each vendor and cooling design.
HBM’s share of GPU package power: 10–16% assumed across the modeled generations
No chipmaker publishes a per-generation split of package power between the compute die and the HBM stacks; this range is the rough order of magnitude used across component-power discussions of GPUs in this class, and it should move with HBM generation and stack count.
Bytes per output token, text: a few bytes
A token ID is 2–4 bytes as the model emits it, or a few characters (≈4 bytes in UTF-8) once decoded to text; either way the network payload per token is tiny next to the compute behind it, which is the point the card makes.
BlueField-3 network cages on the tray front, as drawn: QSFP112 form, two per DPU; positions representative
NVIDIA’s BlueField-3 user guide lists QSFP112 as the network connector on every BlueField-3 DPU and SuperNIC model, dual-port on the DPU models. The drawing gives each BlueField-3 DPU two QSFP cages on the tray front, beside the drive bay. NVIDIA’s DGX GB documents say a compute tray has BlueField-3 DPUs but, in the sources read here, do not say where their ports sit on the tray front or which card model is fitted, so the cages’ positions and the claim that they face the front are an assumption. The Rubin tray carries BlueField-4, which is not drawn with these cages.
GPU package footprints and placement, as drawn: one package per accelerator at the tray and package levels; sizes and stack positions representative
NVIDIA publishes H100’s die area (814 mm²) and calls Blackwell’s two dies reticle-limit, but no package, interposer or HBM footprint or placement for any of these GPUs. The drawing uses 26 × 33 mm (858 mm², the reticle field) for each Blackwell and Rubin die; for H100, a die 26 mm wide and as deep as 814 mm² allows; HBM stacks of about 10.6 × 10 mm; for the two-die packages an 8.4 cm square substrate, a 6.2 × 5.9 cm interposer and four stacks along each long die edge; and for H100 a 5.5 × 5.8 cm substrate (one analyst’s estimate, scaled from photos of the SXM5 module), a 5.1 × 3.7 cm interposer and three stacks on each side, the sixth site a blank spacer. Rubin’s die size, package and stack placement are unpublished (two dies and eight HBM4 stacks are published), so it carries Blackwell’s arrangement forward. The tray and the package level are drawn from one description (src/scenes/gpu-package.js), so they cannot disagree. H100’s substrate is drawn about 5.5 × 5.8 cm, following one hardware analyst’s 2022 estimate (Locuza on X, scaled from an SXM5 module size the analyst marked uncertain); it is a representative size, not a measurement.
Rubin packaging technology: TSMC CoWoS-L (carried over)
NVIDIA has not stated Rubin’s packaging technology; this carries forward CoWoS-L, which trade analysis (Tom’s Hardware) reports TSMC uses for Blackwell’s two-die package, pending Rubin’s own teardown.
Rubin die area: ≈16 cm² total, two dies (carried over)
NVIDIA has not published Rubin’s die size. The figure shown carries forward the reticle-limited two-die total trade press reported for Blackwell, pending an independent teardown of Rubin.
Rubin NV-HBI die-to-die bandwidth: 10 TB/s (carried over)
NVIDIA confirms Rubin’s two dies are joined by NV-HBI but has not published its bandwidth. The figure shown carries forward Blackwell’s confirmed 10 TB/s, pending Rubin’s own datasheet.
HBM operating-temperature limit, other generations: ≈105 °C
Micron publishes 0–105 °C specifically for HBM3E (GB200/GB300’s memory). H100’s HBM3 and Rubin’s HBM4 do not have a confirmed public limit, so the same figure is carried over as a reasonable proxy for stacked DRAM of this class.
One scale-out fabric per accelerator in a mixed fleet: GB200 GPUs on a 400G fabric, GB300 GPUs on an 800G fabric, each non-blocking and sized for its own GPU count
An accelerator’s NIC speed sets the fabric it plugs into, and the model does not join two speeds in one non-blocking tree. How SpaceXAI actually connects its GB200s and GB300s is not public; Elon Musk ties the 110k GB200 figure to the number of fiber optic cables one central switch can take, which fits a separate fabric but does not describe it.
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
NVIDIA publishes a 45 °C coolant inlet and a ≈55 °C outlet for its warm-water racks, and ASHRAE a recommended 18–27 °C for supply air; the rest of each chain is this model’s own illustrative operating point, chosen so the heat tour, the cards and the Hot to cold chart all use the same numbers. Real plants move with load, flow rate and weather, and a die’s temperature depends on its package and cold plate.
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
Round numbers chosen to show the shape of one request. They are the same for every GPU choice, and are not a measurement of the selected hardware, a benchmark, or any particular service; real queueing, prefill and stream rates move with the model, the batch, the prompt and the load.
One serving copy per NVLink domain: one NVL72 rack (72 GPUs), or one 8-GPU server for H100
How many GPUs hold one serving copy depends on the model and the traffic: a copy can be smaller than a rack, several to a rack, or split so prefill and decode run on different GPUs. The request tour picks one NVLink domain so all its stops describe the same copy; the pipeline and data-parallel cards show a training layout instead.
Default tokens a second per GPU: 2,000 for an 8 TB/s GPU, scaled with HBM bandwidth
An illustrative serving rate, used until the reader sets their own in the calculator. Decode is usually memory-bound, so the default scales with each accelerator’s HBM bandwidth; real rates move a long way with the model, its precision and the batch.

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.