Imagined 5 GW AI factory — Gulf Coast

The story

An imagined AI factory on empty land on the US Gulf Coast: 8 halls, 388 identical pods of 96 GB200 NVL72 rack systems, 2,681,856 GPUs. The grid arrives at 400 kV; gas turbines and batteries carry it through outages. Each pod's racks are cooled 70 % by liquid (CDUs to dry coolers) and 30 % by air. At a summer afternoon the engine finds 5,001 MW of IT, PUE 1.17 and 11,146 million tokens a second from a 70B model — and 21.6 million tonnes of CO₂ a year on today's grid mix.

The numbers

engine on the representative hall (hall.json) at 15:00 local, mid-July, scaled by the pod count; rounded. Token figures are engine estimates, not benchmarks. Energy and CO₂ use the tariff and grid factor in the references below.

The representative hall

Notable

• 388 pods × 96 GB200 NVL72 rack systems (120 kW each from the catalog) = 2,681,856 GPUs.
• Per pod the engine sees 12,888.4 kW IT, 15,100.5 kW facility, PUE 1.17 and 28,726,819 tokens/s at the design afternoon.
• 24 CoolIT CHx750 CDUs per pod carry 70 % of the heat as liquid; Vertiv Liebert CW CRAHs take the rest.
• A year at 0.065 USD/kWh: $3336 M for 51,324,840 MWh; 21,556 kt CO₂ at 420 g/kWh.
• Catalog list prices cover 5 of 8 part types per pod; the GB200 NVL72 has no list price in the catalog, so the BOM leaves it out.

What would break (engine)

• Substation trip (utility lost for 15 minutes): Rides through. lowest UPS runtime 9 min. (Substation (400 kV, imagined) lost power (+6 more); UPS bank A on battery · 10 min left (+3 more))
• CDU loop leak at a rack manifold (6 L/min): Outage. 1 devices lose power or trip (72 GPUs); 1.4 L released before isolation; 254.9 M tokens lost. (Leak at Manifold R001 · 0 L released; Manifold R001 isolated after leak; nvl72-r001 throttled to 20 %; nvl72-r001 thermal shutdown)
• Trunk cable cut (GCX-1 ↔ HOU-IX): GCX-1 ↔ HOU-IX cut — traffic on the protection path, +0.6 ms

What we assumed and why

• Scale: 5,000 MW IT target from the brief (IMAGINED-GIANTS.md §1); the pod count is that target ÷ the engine's pod IT load.
• Rack system: catalog record rack_system/nvidia/gb200-nvl72, 120 kW IT, 72 GPUs, 70 % liquid-cooled (catalog thermal data).
• Cooling: cdu/coolit/chx750 CDUs (750 kW, catalog) in loops of 3 per row, N+1; heat rejected to dry coolers; crah/vertiv/liebert-cw CRAHs (120 kW) on a chilled-water plant for the air share.
• Power: 2N from a 400 kV substation through ATS (grid / genset bank), transformers, switchboards and MW-class UPS banks (generic nodes: no MW-class UPS in the catalog), busway taps per rack; standby genset banks of Cat 3516C (2 MW standby, catalog); on-site gas turbines + batteries (sized as the genset bank in the model).
• Climate: a July day at 29 °C mean with a 9 °C daily swing and 70 % RH (typical July climate for the region); the numbers are read at 15:00 local, the warmest hour.
• Grid factor 420 g CO₂/kWh and tariff 0.065 USD/kWh: dc-world/reports/giants-references.md §2 / §3.
• Tokens: every rack serves a 70B chat model at 70–95 % load (engine estimate from GPU memory bandwidth and FLOPS; labelled estimated, not a benchmark).
• Water: closed-loop dry coolers, no consumptive water (evaporative trim not modelled).
• Estate: the DR and peer sites, link capacities and carriers are part of the imagined design, not real contracts; latency is estimated from distance.
• References for scale, cooling and power patterns: §1.1 Stargate Abilene (scale, DLC + CDUs, PUE 1.12–1.15); §1.2 Hyperion & Prometheus (GW campus, 500 kV, gas backstop); §1.3 Colossus (on-site turbines + BESS, 100k GPUs in one fabric); §2.1 ERCOT grid factor ~400 g/kWh and ~$0.065/kWh (dc-world/reports/giants-references.md).

The estate

Open it on the earth: cut a link, lose a site

Files

• estate.xlsx — Sites and Links (DC Sim template v1.4)
• hall.xlsx — the representative hall, device by device
• giant.json — the record with derived numbers and assumptions

Public reference numbers: mega-campus references. CC BY 4.0. Imagined — a fictional design built from public numbers.