COMPUTE ATLAS Supercomputer supply-chain graph
1153 systems 482 sites Sourced data

Blog/2026-10-02

· Compute Atlas

Stargate Abilene: 1.2 GW, four of eight live

Stargate Abilene is stated at 1.2 GW across eight buildings. Epoch AI counts four operating at 421 MW of IT load in July 2026, on a stack of leases, loans and on-site gas turbines.

Stargate Abilene is the flagship of the OpenAI, Oracle and SoftBank Stargate program: a campus on the Lancium Clean Campus site in Abilene, Texas, designed and built by Crusoe, leased to Oracle, and used by OpenAI for training and inference on NVIDIA GB200 racks. The headline number is 1.2 GW across eight buildings. The thing most worth knowing is that this is a design capacity, and almost every other number about the site, including who owns it, how much is running, and how it is powered, depends on whose statement you read.

As of the most recent public estimate we found, four of the eight buildings operate. Epoch AI’s data-center directory, with a status date of 28 July 2026, puts current IT power at 421 MW and projects 843 MW when all eight are done. Neither OpenAI, Oracle nor Crusoe publishes a measured draw or an installed accelerator count.

How it came to exist

Crusoe says construction began in June 2024 (Epoch dates land clearing to May 2024). The first phase was two buildings, 980,000 square feet and more than 200 MW, according to Crusoe’s March 2025 expansion announcement. That announcement grew the site to eight buildings, about 4 million square feet and 1.2 GW. Lancium supplies the site, grid interconnect and engineering, per Data Center Frontier’s account of the same announcement.

The money arrived in two steps. Crusoe’s 21 May 2025 release describes a $15 billion joint venture of Crusoe, funds managed by Blue Owl Capital and Primary Digital Infrastructure, sponsoring six more buildings. Phase two started in March 2025 and was expected to energize in mid-2026. Spyglass, a technology newsletter, reports that JPMorgan provided about $9.6 billion across two loans, including a $7.1 billion commitment, and that Crusoe and Blue Owl put in about $5 billion in cash. A January 2025 trade report puts an earlier JPMorgan loan at $2.3 billion, for a 206 MW first phase. These loan figures come from trade and newsletter reporting, not from the lenders.

On the hardware side, Oracle began delivering NVIDIA GB200 racks in June 2025, according to SoftBank’s 24 September 2025 release, which already calls the site operational. Crusoe’s own “live” announcement is dated 30 September 2025 and says the first two buildings were energized within a year of the June 2024 start.

Who owns, leases and pays

The reported chain has four layers, and none is documented in a single primary source.

  1. Owner and developer. The Crusoe, Blue Owl and Primary Digital venture owns the buildings and financed them with JPMorgan-led construction loans (Spyglass, trade press). Epoch’s directory instead lists Oracle as owner and OpenAI as user, which looks like a convenience label, not a statement about title.
  2. Tenant. Oracle leases the site for 15 years, per Spyglass.
  3. Hardware. Spyglass reports Oracle buying roughly 400,000 GB200 chips for about $40 billion and sub-leasing the compute to OpenAI. Treat that as a single-source estimate.
  4. Customer. OpenAI buys capacity from Oracle. The public record we could verify does not state the terms of OpenAI’s contract with Oracle.

Spyglass also notes that the Stargate entity itself has not invested in the site. So “Stargate Abilene” names a program, while the capital sits with Crusoe’s partners, lenders and Oracle. Oracle is the middle layer: it carries long building leases and power obligations on one side and a shorter customer contract on the other.

What is inside

Accelerators. Crusoe says each building is designed for up to 50,000 NVIDIA GB200 NVL72 GPUs on a single network fabric. The wording is ambiguous, because an NVL72 is a rack. NVIDIA’s specification is 72 Blackwell GPUs and 36 Grace CPUs per rack, one NVLink domain, liquid cooled, with 13.4 TB of HBM3E. Read as GPUs, eight buildings give 400,000 GPUs, about 5,600 racks. Spyglass’s 400,000-chip figure matches that reading. Read as racks it would be 28.8 million GPUs, which is not credible. Epoch counts differently: 100,800 B200 plus 100,800 B300 chips now, rising to a projected 403,200 by Q4 2026, or 509,000 H100-equivalents today. Its chip counts are an estimate, and its B300 label conflicts with operators who say GB200 (NVIDIA is the supplier either way).

Fabric and software. Crusoe says only that the building is on a single integrated network fabric and gives nothing about the vendor or topology. OCI is the operating cloud. We found no operator statement on InfiniBand versus Ethernet.

Cooling. Crusoe describes direct-to-chip liquid cooling on a closed, zero-evaporation water loop. Its live-announcement page describes the design as optimized for both liquid and air cooling. Epoch’s directory lists air-cooled chillers, which reject heat from the liquid loop. These are compatible. A cold-plate GB200 rack still needs a place to dump heat.

Power: grid. Crusoe’s expansion page states a 1.2 GW grid interconnection plus on-site resources. A GridStatus analysis places the interconnection on a new on-site substation tied to the Mulberry Creek 345 kV network, and notes the campus sits where ERCOT’s West and North zones meet.

Power: on-site gas. Here sources diverge. GE Vernova says Crusoe ordered 29 LM2500XPRESS aeroderivative turbines, about 35 MW each, nearly a gigawatt, the first ten for the Abilene site. GridStatus describes ten units at about 368 MW, roughly 31 percent of full demand, as backup and peak support. The Texas Tribune reports ten gas turbines of about 360 MW and 62 backup diesel generators, plus a pending application for 41 more turbines and 18 more generators under a major air permit. Crusoe’s CEO describes a roughly 350 MW gas plant serving the OpenAI and Oracle portion, as quoted by the AP. So the on-site plant is either a backup or a primary supply, depending on the speaker, and it ranges from 350 to 368 MW by source. GridStatus also analysed ERCOT dispatch data and concluded curtailed West Texas wind would not cover the facility’s expected usage, which cuts against a renewables-first description.

Schedule, and what slipped

The original plan had phase two energizing in mid-2026. Epoch’s directory records building 1 operational in September 2025, building 2 in December 2025, buildings 3 and 4 in May 2026, and buildings 5 to 8 structurally complete in July 2026 with operation projected for November 2026. A trade report says Crusoe topped out the eighth building in November 2025, so shells were ahead of fit-out and power.

The larger event was the expansion that did not happen. Bloomberg and Reuters reported in March 2026 that Oracle and OpenAI would not proceed with a planned 600 MW expansion lease. Oracle called coverage of problems at the site “false and incorrect”, said Crusoe and it were operating in lockstep, and said it had completed leasing for an additional 4.5 GW to support OpenAI. The reasons for the change were reported differently and none is confirmed by a primary source. Crusoe then announced in late March 2026 a 900 MW Microsoft campus next door: two buildings of 336 MW critical IT each, its own 900 MW on-site plant with battery storage, closed-loop non-evaporative liquid cooling, first energization targeted for mid-2027. AP puts the combined Abilene campus at 2.1 GW and ten buildings.

Cost is also moving. Epoch estimates $15.9 billion spent to date and $31.9 billion at completion.

What the other Stargate sites show

Epoch’s April 2026 survey found Abilene the only operating Stargate site. Shackelford County and Port Washington (Vantage), Saline Township (Related Digital), Doña Ana County (STACK and BorderPlex) and Milam County (SB Energy) were projected for Q4 2028, with Lordstown, Ohio, described as mainly a manufacturing site. SoftBank’s September 2025 release counted nearly 7 GW of planned capacity against a 10 GW commitment. The pattern across rows: a developer finances a shell and its power, Oracle leases, OpenAI is the end customer. Milam County differs, with OpenAI signing the lease directly with SB Energy. Where power is not yet on the grid the developer builds it: Project Jupiter plans a 700 to 900 MW gas microgrid, per El Paso Matters. Published capacities also differ by developer versus analyst, for example Shackelford at 1.4 GW from Vantage and 2.0 GW from Epoch.

What does not add up or is not public

  • No operator states an installed GPU count, a measured power draw, or which GB200 or GB300 parts are in which building.
  • Capacity scope changes between sources: 1.2 GW is facility power, 421 and 843 MW are Epoch’s IT figures. Dividing gives a ratio near 1.4, which is our arithmetic on Epoch’s numbers.
  • The June 2025 date marks first GB200 deliveries. The first building was announced live in late September 2025.
  • On-site generation figures (300, 350, 360, 368 MW) come from four different sources.
  • The loan, lease and chip-purchase figures are press-reported and no counterparty has confirmed them. The ownership label differs between sources.

What this dataset says

This section describes the dataset as published on 2 October 2026. The dataset changes as rows are added and corrected.

Stargate Abilene is status operational, confidence verified, unranked, with no Rmax, Rpeak, accelerator count or measured power. Its component rows name the NVIDIA GB200 NVL72 as accelerator (quantity blank), Crusoe as integrator and direct-to-chip cooling. Its one capacity row is 1,200 MW, scope facility, stage full_buildout, so the dataset has no operating-stage claim for the site and cannot express the four-of-eight picture above. Across the seven Stargate-named systems, this is the only one marked operational; Crusoe’s second Abilene campus and the five other US campuses are under construction at reported confidence, and Stargate UAE is also under construction. For a comparison with a tenant-landlord split elsewhere, Polaris Forge 1 lists 100 MW operating against 400 MW contracted. See build pipeline, power and cooling, data trust and announcement to first light for how stated capacity and delivered capacity diverge across the dataset.

Sources

deep-diveai-clusterspower-and-coolingacceleratorsprocurement