Last reviewed
May 16, 2026
Sources
23 citations
Review status
Source-backed
Revision
v1 · 3,737 words
Add missing citations, update stale details, or suggest a clearer explanation.
The Abilene data center, also referred to as the Crusoe Abilene Stargate Campus or simply Stargate I, is an artificial intelligence training and inference data center campus located on a roughly 1,000 acre site outside Abilene, Texas. It is the flagship facility of the Stargate Project, the AI infrastructure joint venture announced in January 2025 by SoftBank, OpenAI, Oracle, and MGX. The site is developed by Crusoe Energy on land within Lancium's Clean Campus, leased for fifteen years and operated by Oracle as part of Oracle Cloud Infrastructure, and dedicated almost entirely to compute capacity for OpenAI. At full buildout the campus is designed to host eight data hall buildings drawing approximately 1.2 gigawatts of power and to deploy more than 450,000 NVIDIA GB200 Blackwell GPUs. The first two buildings were energized in mid to late 2025 and the campus was publicly announced as live on Oracle Cloud Infrastructure on September 30, 2025; OpenAI's CEO Sam Altman toured the site for a media event on September 23, 2025 and described it as the largest AI training facility in the world [1][2][3].
The Abilene campus is the site where, according to multiple press accounts, OpenAI completed the pre-training run for the model internally codenamed Spud and released on April 23, 2026 as GPT-5.5. It is the first piece of physical infrastructure delivered under the Stargate banner, and its construction timeline, partner roster, and 2026 commercial difficulties have become a reference point for the broader $400 billion to $500 billion Stargate commitment. While the eight building core project remains under construction, an additional 600 megawatt expansion proposed in late 2025 was scrapped in March 2026 after financing talks between Oracle and OpenAI stalled, and reporting in 2026 suggested that the spare Crusoe capacity at the site would be re-leased to Microsoft and Meta, with NVIDIA reportedly paying a deposit to retain a position at the campus [4][5][6].
The campus sits on approximately 1,000 acres at 5502 Spinks Road on the southeastern edge of Abilene, Taylor County, Texas, about 180 miles west of the Dallas-Fort Worth metropolitan area. The land is part of Lancium's Clean Campus, an industrial park that Lancium had previously zoned and pre-permitted for power intensive computing tenants located adjacent to high voltage Electric Reliability Council of Texas (ERCOT) transmission corridors. Crusoe Energy acquired development rights for the campus in 2024 and broke ground in June 2024 on what was originally announced as a 200 megawatt, two building site for an undisclosed AI tenant. The site was publicly tied to OpenAI and the Stargate Project on January 21, 2025 [7][8].
Abilene was chosen for several reasons that recur in subsequent Stargate site selections. Taylor County and the surrounding region of west Texas have unusually high wind capacity factors and have historically experienced curtailment of wind generation because supply outstrips local demand. The Lancium Clean Campus was already designed to absorb intermittent renewable generation through demand response, and the site has direct interconnection into the ERCOT grid through nearby substations. Texas state law also allows ERCOT to permit large new loads more quickly than most U.S. interconnection queues, and the state offers property tax abatements through Chapter 313 successor programs. Finally, west Texas land is inexpensive relative to comparable sites in Northern Virginia, the Pacific Northwest, or Phoenix, and the regulatory environment is unusually permissive toward on-site gas generation [9].
The campus is roughly rectangular in shape, with the eight data hall buildings arrayed in two rows of four. Each data hall building is approximately 425,000 square feet on a single story and is engineered around dense, liquid cooled racks rather than the air cooled hot aisle topology typical of older hyperscale facilities. Crusoe has described the construction as a "fast build" or "factory style" approach in which steel and concrete shells are erected on a roughly six month cycle and then fitted out in parallel, allowing several buildings to be at different stages of construction at any given time [7][10].
The construction of the Abilene campus has proceeded in overlapping phases since 2024. The table below summarizes publicly disclosed milestones through May 2026.
| Phase | Date | Milestone |
|---|---|---|
| Site selection | 2023 to early 2024 | Lancium Clean Campus rezoned and pre-permitted; Crusoe acquires development rights |
| Groundbreaking | June 2024 | First earthworks for a 200 MW two building campus |
| Stargate announcement | January 21, 2025 | Campus publicly identified as the first Stargate Project site |
| Expansion to 1.2 GW | March 18, 2025 | Crusoe and partners announce expansion from two buildings to eight buildings totaling roughly 1.2 GW |
| Oracle 15 year lease | mid 2025 | Oracle signs a 15 year master lease covering the full eight building campus |
| First GB200 racks | June 2025 | Oracle begins delivering NVIDIA GB200 NVL72 racks into Building 1 |
| Sam Altman site visit | September 23, 2025 | Altman and Oracle host a media event outside Building 2 |
| First two buildings live | September 30, 2025 | OpenAI and Oracle announce that Buildings 1 and 2 are operational on Oracle Cloud Infrastructure, running training and inference for OpenAI |
| Topping out | January 2026 | Crusoe announces the final building of the eight has been topped out |
| 600 MW expansion scrapped | March 6, 2026 | Bloomberg reports that Oracle and OpenAI have ended plans for a separate 600 MW expansion adjacent to the main campus |
| GPT-5.5 pre-training complete | March 24, 2026 | Reported pre-training completion of OpenAI's model Spud, released April 23, 2026 as GPT-5.5 |
| Expected full buildout | mid 2026 | All eight buildings expected to be in production; total capacity roughly 1.2 GW |
The September 30, 2025 "live" announcement was timed alongside OpenAI's September 23, 2025 announcement of five new Stargate sites and was used by the partners to demonstrate physical delivery against criticism that Stargate was primarily an investment narrative. At the time of the announcement only Buildings 1 and 2 were energized, with Buildings 3 and 4 in fit out and Buildings 5 through 8 in shell construction [1][2][11].
The campus was originally announced with two buildings totaling about 200 megawatts. Crusoe's March 2025 expansion increased the design to eight identical data hall buildings totaling approximately 1.2 gigawatts of IT load. Each building is designed to support up to roughly 50,000 GB200 GPU server nodes, although disclosed deployments per building are smaller than the design ceiling. The table below shows the publicly reported configuration; figures for Buildings 3 through 8 are based on Crusoe and Oracle disclosures and are subject to revision as fit out continues.
| Building | Approximate IT load | Disclosed GPU class | Status (May 2026) |
|---|---|---|---|
| Building 1 | ~100 MW | NVIDIA GB200 NVL72 | Operational since September 2025 |
| Building 2 | ~100 MW | NVIDIA GB200 NVL72 | Operational since September 2025 |
| Building 3 | ~150 MW | NVIDIA GB200 NVL72 | Fit out, expected 2026 |
| Building 4 | ~150 MW | NVIDIA GB200 NVL72 | Fit out, expected 2026 |
| Building 5 | ~150 MW | NVIDIA GB200 NVL72 | Shell complete, expected mid 2026 |
| Building 6 | ~150 MW | NVIDIA GB200 NVL72 | Shell complete, expected mid 2026 |
| Building 7 | ~150 MW | NVIDIA GB200 NVL72 | Shell complete, expected mid 2026 |
| Building 8 | ~150 MW | NVIDIA GB200 NVL72 | Topped out January 2026, expected mid 2026 |
| Campus total | ~1.2 GW | ~450,000 GB200 GPUs | Partial; full buildout mid 2026 |
The disclosed figure of "more than 450,000 GB200 GPUs" represents Oracle's contractual commitment under the 15 year lease rather than an instantaneous count. Public reporting and Oracle financial disclosures imply that the contracted ramp will continue through 2026 and into 2027, with later buildings likely to host successor NVIDIA GB300 NVL72 systems as the GB200 generation is superseded [3][12].
The Abilene campus draws primary power from the ERCOT grid through dedicated 345 kilovolt interconnections with substation upgrades carried out by Oncor Electric Delivery. Because Lancium had already designed the surrounding campus as a flexible load that could ramp with renewable supply, OpenAI and Crusoe have emphasized that the data center is structured to consume curtailed wind energy where possible. The site is, however, a constant load when running training workloads, and a substantial share of its energy in any given hour comes from the underlying ERCOT generation mix rather than directly from co-located wind [9][13].
For backup and grid stability, Crusoe installed a fleet of GE Vernova natural gas turbines on site rather than the diesel reciprocating engines that are standard in U.S. hyperscale data centers. The turbines are equipped with Selective Catalytic Reduction emission control and, according to GE Vernova, are roughly an order of magnitude cleaner than diesel generators on a nitrogen oxide basis. In March 2025 Crusoe also announced a separate joint venture with Engine No. 1 to procure 4.5 gigawatts of behind the meter natural gas generation for AI campuses, capacity that is expected to support both Abilene and additional sites [13][14].
Cooling at Abilene is built around closed loop liquid cooling, which is necessary at the rack density of GB200 NVL72 systems. Crusoe has publicly emphasized that the campus does not rely on evaporative cooling towers and instead uses non-evaporative dry coolers in series with the closed loop. The choice is presented as a response to chronic water scarcity in west Texas, where the Brazos River basin and Abilene's municipal supply have been under multi-year drought conditions. Independent observers have nevertheless noted that natural gas turbines and on site water make up cooling can still represent significant resource intensity, and Crusoe has stated that it intends to integrate post combustion carbon capture at the natural gas units, although as of May 2026 no commercial scale capture system was operational [13][15].
The Abilene campus is the result of an unusually layered partnership structure in which physical infrastructure, cloud operation, GPU supply, financing, and tenancy are split among separate companies. This layering is sometimes called the "Stargate layer cake" by industry observers and is summarized below.
| Layer | Partner | Role |
|---|---|---|
| Land owner | Lancium | Owner of the Clean Campus master site; provider of grid interconnection and demand response design |
| Data center developer and operator | Crusoe Energy (Crusoe Cloud / Crusoe AI) | Designs, builds, and operates the physical buildings, power, and cooling; holds long term ground lease from Lancium |
| Anchor lessee | Oracle | Signs 15 year master lease for the eight buildings; installs racks, networking, and Oracle Cloud Infrastructure software |
| Cloud platform | Oracle Cloud Infrastructure | Provides the cloud control plane on which OpenAI workloads run |
| Compute customer | OpenAI | Sole or near sole tenant of the Oracle capacity at the site; runs training and inference for ChatGPT and frontier models |
| GPU supplier | NVIDIA | Supplies GB200 NVL72 systems and reportedly took a deposit on Crusoe expansion capacity in 2026 |
| Financial sponsor | SoftBank, MGX, Crusoe debt syndicate | Funds construction; Crusoe raised approximately $15 billion in equity and debt by 2026, including an $11.6 billion package in 2025 specifically for Abilene |
| Investment partner | Engine No. 1 | Joint venture for behind the meter natural gas power |
| Governance | Stargate LLC | Coordinates capacity allocation across the Stargate Project |
The split between Crusoe as developer and Oracle as anchor lessee is unusual at this scale. In most U.S. hyperscale projects the cloud operator either owns the building outright or signs a long term build to suit lease with a single REIT. At Abilene, Oracle does not own the buildings but holds a 15 year operating lease that effectively underwrites Crusoe's project debt. OpenAI does not appear on the lease but is the underlying compute customer, with a take or pay style commitment to Oracle that supports the lease. This structure is what allows the partners to describe the same capacity using different and sometimes inconsistent numbers in press releases [4][8][16].
Once energized in September 2025, the first two buildings of the Abilene campus immediately began running OpenAI training and inference workloads on Oracle Cloud Infrastructure. OpenAI and Oracle have not disclosed a building level allocation but have stated that the early capacity is split between large model training jobs and inference for ChatGPT and the API. Industry reporting estimated the operational capacity at the site as of late 2025 at approximately 0.3 gigawatts of IT load, with the balance of the eight building campus expected to come online through mid 2026 [3][16].
The site has been most widely associated in the press with the training of GPT-5.5, OpenAI's model released on April 23, 2026. Reports from CNBC and other outlets indicate that pre-training for the model, internally codenamed Spud, concluded on March 24, 2026 at the Abilene campus. While OpenAI has not formally disclosed the training site for GPT-5.5, the timing and capacity of the Abilene buildout are consistent with the public accounts. GPT-5.5 was the first OpenAI flagship model whose pre-training run was reported to have occurred outside of Microsoft Azure infrastructure, and its release was used by OpenAI to mark the practical transition of frontier training to Oracle hosted Stargate capacity [17][18].
In parallel, the Abilene site is used for inference workloads serving ChatGPT and the OpenAI API in the central United States. OpenAI has described the campus as part of an emerging "compute fabric" across multiple Stargate and CoreWeave sites, and Oracle has highlighted Abilene as the largest single deployment of GB200 NVL72 systems in Oracle Cloud Infrastructure [2][16].
The Abilene campus is one of several U.S. sites being built under the Stargate Project umbrella. The original January 2025 announcement promised $100 billion deployed immediately and up to $500 billion over four years. By September 2025 the partners announced an additional five U.S. sites, taking planned capacity to roughly 7 gigawatts and total committed investment to more than $400 billion over three years. The table below summarizes the announced U.S. Stargate sites as of May 2026 and contrasts them with the Abilene flagship.
| Site | County / state | Lead developer | Cloud operator | Approximate target capacity | Status (May 2026) |
|---|---|---|---|---|---|
| Abilene | Taylor County, Texas | Crusoe Energy | Oracle | 1.2 GW (8 buildings) | First 2 buildings live Sept 2025; full buildout mid 2026 |
| Shackelford County | Shackelford County, Texas | Oracle | Oracle | Multiple hundred MW | Site work; expected 2026 to 2027 |
| Doña Ana County | Doña Ana County, New Mexico | Oracle | Oracle | Multiple hundred MW | Site work; expected 2026 to 2027 |
| Lordstown | Trumbull County, Ohio | SoftBank (SB Energy) | SoftBank / Oracle | ~ 1 GW planned | Ground broken; expected 2026 |
| Milam County | Milam County, Texas | SB Energy / SoftBank | SoftBank / Oracle | ~ 1 GW planned, fast build | Site work |
| Midwest site (Wisconsin) | Wisconsin | Oracle with Vantage | Oracle | ~ 1 GW planned | Site work; confirmed October 2025 |
| CoreWeave partnership sites | Multiple U.S. states | CoreWeave | CoreWeave / OpenAI direct | Additional ~ 1 GW | Various |
At the time of OpenAI's September 23, 2025 announcement of the five new sites, the partners stated that the combined Stargate Project, including Abilene, the new U.S. sites, and the existing partnership with CoreWeave, was on track to reach nearly 7 gigawatts of planned capacity and over $400 billion of investment over three years. Abilene at 1.2 gigawatts represents roughly one sixth of that planned capacity but accounts for the great majority of the capacity that is actually energized as of mid 2026 [2][11][19].
The layered structure first demonstrated at Abilene, with Crusoe or another specialist developer building behind the meter power and a hyperscale lessee operating the cloud, has been broadly replicated across the new sites. The Shackelford County and Doña Ana County sites are direct Oracle developments without Crusoe, while the Lordstown and Milam County sites use SoftBank's SB Energy unit as the powered shell developer and add additional financing from SoftBank. This diversification was, in part, a response to the fact that Crusoe's capital structure and natural gas heavy approach attracted criticism and could not be scaled fast enough to host all of Stargate's planned capacity on its own [2][20].
The Abilene campus has been politically and commercially celebrated and operationally contested in roughly equal measure. Supporters describe it as the largest single AI training facility in the world and as proof that the United States can re-industrialize around the AI supply chain. Critics have raised four sets of concerns: financing, energy and emissions, water and community impact, and partner discipline.
Financing and partner discipline. From late 2025 onward, financial press coverage focused on the difficulty of underwriting Stargate sized leases on top of OpenAI's projected cash needs. In March 2026 Bloomberg reported that Oracle and OpenAI had scrapped a separate 600 megawatt expansion adjacent to the main eight building footprint after financing talks dragged and OpenAI's compute scaling team, led by Sachin Katti, chose to redirect that demand toward later generation GPUs at other sites. Tom's Hardware reported in the same period that OpenAI had "effectively abandoned" first party Stargate buildings in favor of more flexible leased capacity, although Oracle subsequently issued a public statement that the core 4.5 gigawatt agreement covering Abilene and several follow on sites remained on track. Microsoft and Meta were reported in March and later in 2026 to be in talks to lease the freed up Crusoe capacity, with NVIDIA reportedly paying a $150 million deposit to retain a position at the site [4][5][21].
Energy and emissions. Crusoe's on site natural gas turbines have been the most prominent target of environmental criticism. Although the units are equipped with SCR and are described as roughly 90 percent cleaner than diesel generators on a NOx basis, they still emit carbon dioxide at scale when running, and several Texas based environmental groups argued during local hearings that the campus's net emissions are inconsistent with the partners' broader sustainability claims. Crusoe has stated that it intends to add post combustion carbon capture at the gas units, but as of May 2026 no operating carbon capture system existed at the site [13][14].
Water. The campus's choice of closed loop, non-evaporative cooling has been recognized by analysts as a meaningful reduction in water intensity relative to legacy hyperscale designs. However, the site still requires water for make up in the cooling loops and for fire suppression, and the surrounding Brazos basin has been in extended drought. The City of Abilene approved municipal water service contracts to the campus that include curtailment provisions during severe drought stages [9][22].
Community impact and traffic. During the September 2025 media event the site hosted thousands of construction workers on day shifts, and local press in Abilene reported significant strain on housing, child care, and road infrastructure in Taylor County. Crusoe and the City of Abilene negotiated road improvements as part of the property tax abatement package, and Crusoe has emphasized that the campus's permanent operational headcount once fully built out will be in the low hundreds rather than the thousands of construction era jobs [23].
Whether these criticisms ultimately constrain the campus remains to be seen. The fact that the core eight building project remains on the publicly stated mid 2026 schedule, despite the scrapped 600 megawatt expansion and the broader Stargate reorganization, suggests that Abilene specifically is delivering against its original plan even as the wider Stargate Project narrative has shifted.