The Stargate Project is a $500 billion joint venture backed by OpenAI, SoftBank, Oracle, and MGX, launched in January 2025 to build AI data centers across the United States. Starting with $100 billion targeted for immediate deployment, Stargate aims for 10 gigawatts of compute capacity over a four-year buildout. It is the largest concentrated AI infrastructure investment ever announced by private entities. (OpenAI. “Announcing The Stargate Project.” January 21, 2025)
What Is the Stargate Project?
Stargate is a newly formed company, not a product or a model. Its mandate is singular: build the physical infrastructure required to train and run frontier AI in production, on American soil.
On January 21, 2025, President Donald Trump announced the venture at a White House press conference alongside Sam Altman (OpenAI), Larry Ellison (Oracle), and Masayoshi Son (SoftBank). Trump called it “the largest AI infrastructure project in history.” The initial equity funders are SoftBank, OpenAI, Oracle, and MGX (the Abu Dhabi-based sovereign wealth fund). Technology partners include Nvidia, Microsoft, and ARM. (OpenAI. “Announcing The Stargate Project.” January 21, 2025)
The division of responsibility is deliberate: SoftBank holds financial responsibility and Son serves as chairman. OpenAI holds operational responsibility. Oracle serves as the primary cloud infrastructure provider, building and managing the data center campuses.
How Stargate Is Being Built: Sites, Silicon, and Scale
The flagship site is in Abilene, Texas, a roughly 875-acre campus built by Crusoe on the Lancium Clean Campus, about 180 miles west of Dallas. [Updated June 2026] Oracle describes it as “the world’s largest AI supercluster.” GB200 racks have powered training and inference workloads there since mid-2025, and as of spring 2026 about four of the planned eight buildings are operational. Oracle plans to deploy over 450,000 GB200 GPUs at the campus under a 15-year lease. At full build-out the site targets 1.2 gigawatts across eight halls, well short of the 2.1 GW once floated. In March 2026 Oracle and OpenAI scrapped a roughly 600 MW expansion, citing a power-delivery delay of about a year and a pivot toward Nvidia’s Vera Rubin platform; Microsoft has since taken the adjacent 900 MW of capacity through Crusoe. [Updated June 2026] (Epoch AI. “OpenAI Stargate: where the US sites stand.”) (Data Center Dynamics. “Oracle/OpenAI drop plans to expand flagship Abilene Stargate site.” March 2026)
In September 2025, OpenAI announced five additional U.S. sites: (OpenAI. “OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites.” September 2025)
- Shackelford County, Texas
- Doña Ana County, New Mexico
- Lordstown, Ohio
- Milam County, Texas
- Saline, Michigan
OpenAI broke ground on the Saline, Michigan campus on June 1, 2026, scaling it to a $16 billion investment, up from the $7 billion announced in 2025. [Updated June 2026] At the event OpenAI announced community commitments including $10 million for the Saline Recreation Center and roughly $1 billion in projected tax revenue over the lease. The partners pledged that infrastructure and energy costs would not pass to local ratepayers, and that the facility would rely on a closed-loop cooling system using about as much water as a typical office building, with power supplied entirely by DTE Energy. OpenAI also committed up to $45 million in Codex credits for more than 400,000 eligible Michigan students during the 2026-2027 academic year. (CNBC. “Stargate live updates: OpenAI’s Altman says ‘people are right to be anxious’ about AI.” June 1, 2026)
By June 2026 the partners put combined planned capacity above 8 gigawatts and committed investment above $450 billion within three years, which OpenAI frames as running ahead of the original 10 GW, $500 billion schedule. [Updated June 2026] A seventh U.S. site, Port Washington, Wisconsin (developed by Vantage and nicknamed Lighthouse), was added in October 2025, a roughly $15 billion, 902 MW campus targeted for 2028. (Epoch AI. “OpenAI Stargate: where the US sites stand.”)
The headline gigawatt totals are planned capacity under contract, not power energized and serving traffic today. Progress is uneven: Abilene leads with roughly 0.6 GW operational, while Doña Ana County in New Mexico still faces unresolved air permits, and the Lordstown, Ohio location is a server-assembly plant rather than a compute hall. Investigative reporting in early 2026 also indicated the original three-way joint venture never staffed up and that Stargate now runs as a set of bilateral deals, OpenAI with Oracle on the compute side and OpenAI with SoftBank and SB Energy on financing and power. [Updated June 2026] (Epoch AI. “OpenAI Stargate: where the US sites stand.”)
The Hardware Stack
The compute layer is dominated by Nvidia, but diversifying:
| Component | Details |
|---|---|
| Primary GPU | Nvidia GB200 (NVL72 rack configurations) |
| Abilene deployment | 450,000 Nvidia GB200 GPUs at full build-out (IntuitionLabs. “OpenAI’s Stargate Project: A Guide to the AI Infrastructure.”) |
| Nvidia investment | Up to $100B as 10 GW of Nvidia systems deploy (September 2025 letter of intent, which Nvidia states is separate from Stargate; reportedly may be scaled back, with no tranche confirmed funded as of mid-2026) |
| AMD agreement | 6 GW of Instinct GPUs (October 2025 deal with share warrants; first 1 GW of MI450 “Helios” in H2 2026) |
| Custom silicon | ”Jalapeño” inference ASIC with Broadcom on TSMC 3nm, unveiled June 24, 2026 (OpenAI. “OpenAI and Broadcom unveil LLM-optimized inference chip.” June 24, 2026) |
OpenAI plans to bring up its first gigawatt of Nvidia systems in the second half of 2026 on the Vera Rubin platform, which pairs the 88-core Vera CPU with the Rubin GPU. [Updated June 2026] (SiliconANGLE. “OpenAI to build five more Stargate data centers in the US with Oracle and SoftBank.” September 23, 2025)
Jalapeño signals OpenAI’s intent to reduce GPU vendor dependency over time, a pattern already established by Google (TPUs), Amazon (Trainium), and Microsoft (Maia). OpenAI says it designed the chip in roughly nine months and is targeting initial deployment by the end of 2026, expanding in later years; the performance claims are self-reported and not independently benchmarked. [Updated June 2026] (CNBC. “OpenAI unveils first chip as part of Broadcom deal in effort to ‘build the full stack’.” June 24, 2026)
The Power Problem: 10 Gigawatts and Counting
The Abilene campus alone, at 1.2 GW of draw, consumes enough electricity to power approximately one million four-bedroom homes. (Certrec. “Energy Demands for OpenAI’s Stargate Project.”) Scaling to 10 GW nationally requires more power than most countries use for entire industrial sectors.
Five of these 10 GW sites running on natural gas would require roughly 4% of America’s current natural gas production. (Mark Nelson (@energybants). “PROJECT STARGATE ENERGY ANALYSIS.” X, January 2025) OpenAI has stated it plans to use solar, battery storage, small modular reactors (SMRs), and carbon capture to power the network long-term, but widespread SMR deployment before 2030 remains speculative. OpenAI has pledged to “pay its own way” for grid infrastructure rather than relying on ratepayer subsidies. (Data Center Dynamics. “OpenAI pledges to ‘pay its own way’ to power Stargate data centers.”)
Environmental concerns are legitimate. Backup diesel generators produce nitrogen oxides, particulate matter, and carbon monoxide. Water use for cooling is substantial. Communities near proposed sites in Texas and New Mexico are already raising objections about air quality and grid load impacts. (Texas Standard. “Texas becomes the epicenter of OpenAI’s $500 billion Stargate Project.”)
Energy sourcing is being solved site by site rather than from a single national plan. Crusoe has lined up on the order of several gigawatts of natural-gas turbines across its fleet, and several inland sites run behind-the-meter microgrids: Shackelford County, Doña Ana County, and Milam County lean on on-site gas generation, Saline draws entirely from DTE Energy, and Port Washington is planned at roughly 70% renewables. The grid-versus-gas split matters for both the emissions math and the schedule, because turbine and interconnection lead times, not chip supply, set the pace at most sites. Reliability is not a given either: a multi-day winter-weather outage damaged liquid-cooling equipment at Abilene in early 2026 and strained OpenAI’s relationship with Crusoe. [Updated June 2026] (Epoch AI. “OpenAI Stargate: where the US sites stand.”)
Who’s Behind Stargate (and Who Isn’t)
The consortium matters as much as the capital. SoftBank brings capital and international access, though its share of the bill has strained its own balance sheet. To fund a $30 billion follow-on investment in OpenAI signed in February 2026, SoftBank took on a record $40 billion bridge loan in March 2026, its largest-ever dollar-denominated borrowing, underwritten by JPMorgan, Goldman Sachs, Mizuho, SMBC, and MUFG and due within twelve months. By June 2026 S&P had moved SoftBank’s outlook to negative, credit analysts were flagging a multi-year funding gap, and the firm had sold its entire Nvidia stake to raise cash. The short loan maturity is itself a signal: lenders are effectively betting on an OpenAI IPO inside a year. [Updated June 2026] (Bloomberg. “SoftBank Secures Record $40 Billion Bridge Loan for OpenAI Stake.” March 27, 2026) Oracle brings data center construction and large-scale cloud operations; the company began delivering the first GB200 racks to Abilene in mid-2025. OpenAI brings the demand: frontier model training and inference that requires dedicated, not shared, infrastructure.
Conspicuously absent from the equity table: Microsoft. That is notable because Microsoft had been OpenAI’s primary infrastructure partner since 2019. Stargate is OpenAI building compute it controls directly, a move away from exclusive Azure dependency. The relationship was formally rewritten in late 2025: OpenAI completed its conversion to a public benefit corporation in October 2025, Microsoft took a stake of roughly 27%, and Microsoft gave up its right of first refusal on OpenAI’s compute, which is precisely what lets OpenAI build off Azure. Microsoft still hosts large OpenAI workloads and OpenAI has committed to buying roughly $250 billion of Azure capacity, so this is diversification, not divorce. [Updated June 2026]
Also absent: Elon Musk, who publicly questioned whether OpenAI and SoftBank actually had the capital to execute the plan. Musk’s criticism targeted the funding structure, specifically whether SoftBank’s committed capital was real or contingent. (MIT Technology Review. “AI’s energy obsession gets a reality check.” January 28, 2025)
A related critique has only sharpened since: much of the announced capital is circular. Nvidia is investing in OpenAI, OpenAI is buying Nvidia systems, Oracle is borrowing to build sites OpenAI will rent, and SoftBank is taking on debt to fund its OpenAI stake. The dollars are real, but a large share of them loop back through the same handful of balance sheets, so the headline totals overstate how much genuinely independent capital is committed. [Updated June 2026]
Stargate vs. the AI Infrastructure Landscape
Stargate is not building in a vacuum. The major hyperscalers are each running their own infrastructure arms races:
| Company / Project | 2026 CapEx (est.) | Notable Initiative |
|---|---|---|
| Amazon / AWS | ~$200B | Project Rainier (Anthropic/Trainium buildout) |
| Alphabet / Google | ~$180–190B | Guidance raised at Q1 2026 earnings |
| Meta | ~$115–135B | Targeting >10 GW total capacity by end of 2026 (Prometheus, Hyperion) |
| Microsoft / Azure | ~$190B (CY2026) | Constrained by power and memory pricing |
| Oracle / Stargate | ~$50B (136% YoY increase) | Stargate-dedicated capacity: 4.5 GW under Oracle deal (OpenAI. “Stargate advances with 4.5 GW partnership with Oracle.”) |
Combined hyperscaler capital expenditure in 2026 was projected at $660–690 billion and, after the Q1 2026 earnings round lifted several of these guidance numbers, is tracking higher still. (Futurum. “AI Capex 2026: The $690B Infrastructure Sprint.”) [Updated June 2026] At that scale, Stargate’s $100 billion in year-one spending is aggressive but not anomalous. It is the concentrated nature of the investment (one entity, one AI workload category) that distinguishes it.
The DeepSeek Question: Does Compute Still Win?
One week after Trump announced Stargate, DeepSeek released R1, a reasoning model that matched or exceeded OpenAI o1 performance at approximately 20x lower inference cost. (MIT Technology Review. “AI’s energy obsession gets a reality check.” January 28, 2025) Nvidia lost $600 billion in market capitalization in a single trading session.
The implied challenge: if DeepSeek can achieve frontier-level reasoning with far fewer chips, does a $500 billion compute bet make sense?
The honest answer is: it depends on the question you’re asking.
DeepSeek demonstrated that algorithmic efficiency can compress the compute gap between well-resourced and modestly-resourced labs. But efficiency gains and raw scale are not mutually exclusive. They compound. A lab that achieves DeepSeek-level algorithmic efficiency and has 10 GW of dedicated compute has a different capability ceiling than one with only efficiency. The race is now on both axes simultaneously.
OpenAI’s own response was pragmatic: Sam Altman called DeepSeek “impressive” and indicated that efficient models would increase overall AI adoption, which increases infrastructure demand, not decreases it. The “Jevons paradox” framing holds that cheaper AI drives more consumption, which drives more infrastructure need. That is the working assumption behind Stargate’s continued investment.
Geopolitical Stakes: America’s AI Infrastructure Bet
The Trump administration framed Stargate explicitly as a national competitiveness initiative. The announcement accompanied America’s AI Action Plan and extensions of CHIPS Act tax credits, all oriented toward domestic manufacturing and reducing Chinese technological access.
China’s response has been a counter-initiative: a domestic compute consolidation program that clusters existing capacity, subsidizes chip production, and virtualizes infrastructure, informally called China’s “Stargate.” (Linkdood Technologies. “China’s ‘Stargate’ New Ambition on AI Supremacy Against the U.S.”) The Chinese government has accelerated AI infrastructure spending in direct response to the U.S. announcement.
Nvidia’s export control situation illustrates the hardware dimension of this competition. In April 2025, Trump banned sales of Nvidia H20 chips (previously designed to comply with export controls) to China, resulting in a $4.5 billion charge for Nvidia. China retaliated with export controls on gallium and germanium, materials essential for semiconductor production. By 2026 the fight had moved up the stack to the more capable H200: U.S. Commerce cleared case-by-case H200 sales to a short list of Chinese firms in spring 2026, but deliveries stalled when Beijing discouraged the purchases and held shipments at customs, so the “resumption” was on paper more than in racks. [Updated June 2026]
Stargate’s geographic concentration in the U.S. is not incidental. It is the infrastructure expression of a policy position: that AI capability is a strategic national asset, and that the compute underpinning it should be domestically controlled.
What Practitioners Need to Know
The Stargate project’s implications extend beyond OpenAI’s own products. Several downstream effects will shape enterprise AI decisions over the next 18–36 months:
- Inference pricing pressure: 450,000+ GB200 GPUs dedicated to OpenAI inference means significant capacity expansion. More capacity typically compresses pricing. Enterprises locked into current API pricing contracts should assess re-negotiation windows.
- Custom silicon timelines: OpenAI’s Jalapeño inference chip, unveiled with Broadcom in June 2026 on TSMC 3nm and slated for initial deployment by year-end, will inform competitive dynamics for model inference costs, following the Google/Amazon/Microsoft pattern.
- Grid as constraint: Power availability, not capital, is the near-term bottleneck. Data center construction pipelines are longer than capital timelines. The 10 GW target faces real physical limits that money alone cannot accelerate.
- Oracle Cloud relevance: Stargate dramatically increases Oracle Cloud Infrastructure’s strategic importance. Enterprises already in Oracle ecosystems gain proximity to Stargate compute. Those not in Oracle ecosystems may face access constraints as capacity is prioritized for Stargate workloads.
Frequently Asked Questions
Q: Is the Stargate Project the same as the Microsoft-OpenAI “Stargate” data center project reported in 2024? A: No. The 2024 reports described a Microsoft-OpenAI joint data center project, also internally called Stargate. The January 2025 announcement created a separate, independent entity: Stargate LLC, funded by SoftBank, OpenAI, Oracle, and MGX, with no Microsoft equity stake.
Q: How much of the $500 billion has actually been committed or spent? A: As of mid-2026, over $450 billion in investment has been committed across Stargate’s sites within a three-year horizon, up from the roughly $400 billion figure cited in late 2025, with about $100 billion targeted for deployment in the first year. [Updated June 2026] The $500 billion total spans four years. The Oracle partnership alone represents a commitment exceeding $300 billion in compute over five years starting in 2027.
Q: Does DeepSeek’s efficiency undermine the case for Stargate-scale compute investment? A: Not straightforwardly. DeepSeek demonstrated that algorithmic efficiency can close capability gaps, but more efficient AI typically expands total demand (Jevons paradox). OpenAI and its partners are betting that cheaper inference drives broader adoption, which drives more total infrastructure need, not less.
Q: What happens to OpenAI’s Microsoft Azure relationship now that Stargate exists? A: Microsoft remains a technology partner and still hosts significant OpenAI workloads on Azure; under the late-2025 restructuring OpenAI committed to buying roughly $250 billion of Azure capacity even as Microsoft gave up its exclusive compute right of first refusal. [Updated June 2026] Stargate does not eliminate that relationship, it diversifies OpenAI’s infrastructure independence, reducing exclusive Azure dependency over time.
Q: When will the full 10 GW of Stargate capacity be operational? A: The Abilene campus targets a 1.2 GW build-out across eight halls, with roughly four operational as of spring 2026 and the rest due through 2026 and into 2027. [Updated June 2026] The other U.S. sites are at various stages of permitting and construction. Full 10 GW capacity across all sites is a multi-year buildout with power availability as the primary constraint, not capital.
