Stargate Project: The $500 Billion Bet That Will Define the Next Decade
Half a trillion dollars is not a budget. It is a declaration. When Donald Trump stood in the White House Roosevelt Room on January 21, 2025, flanked by Sam Altman, Masayoshi Son, and Larry Ellison, and announced the Stargate Project, he was describing a capital deployment larger than the GDP of Austria — directed entirely at building AI data centers across the United States. The number that actually matters: by late 2025, Stargate had already committed over $450 billion and reached 8 gigawatts of planned capacity, ahead of its own schedule. What was sold as a four-year promise started moving in months.
Most coverage of Stargate has treated it as either a geopolitical victory lap or a financial fantasy — rotating between breathless announcements and skeptical corrections. Neither reading is complete. The project stalled, burned through financing structures, generated corporate infighting that The Information described as a "chaotic mess," and then quietly rebuilt itself into something real. GPT-5.5 was trained at the Abilene, Texas site. Concrete poured. Racks went in. The story of Stargate is not the story of an announcement — it is the story of what happens when $500 billion meets the specific, grinding friction of power grids, debt markets, and partner disagreements.
This article traces that story in full: who is funding Stargate and how, what is actually being built and where, why the project collapsed and recovered within eighteen months, what it means for the US-China AI competition, and what the hardware running inside these facilities reveals about where the frontier of AI is moving next.
- The Structure: Who Owns What and Who Owes Whom
- The Physical Build: Sites, Scale, and What 8 Gigawatts Actually Means
- The Financing Collapse and Recovery
- The Hardware Layer: NVIDIA, AMD, and OpenAI's Secret Chip
- The Geopolitical Calculation: China, Export Controls, and Sovereign AI
- The Power Problem Nobody Is Solving Fast Enough
- Who This Is For
- Verdict
- Frequently Asked Questions
The Structure: Who Owns What and Who Owes Whom
Stargate LLC is incorporated in Delaware. Four entities sit at the table: OpenAI and SoftBank hold 40% each, having each committed $19 billion of equity. Oracle and Abu Dhabi's MGX sovereign wealth fund contributed $7 billion each. The remaining capital comes from limited partners and debt — a distinction that became consequential when the debt markets balked.
The division of responsibilities is cleaner than the ownership: SoftBank carries financial responsibility and Masayoshi Son chairs the venture. OpenAI holds operational responsibility and manages the actual AI workloads. Oracle builds and operates the physical data center infrastructure through its cloud platform. This arrangement explains why, when the joint venture itself stalled over governance disputes, construction at Oracle-managed sites continued anyway. The project that analysts were pronouncing dead in August 2025 had buildings going up in Abilene, Texas simultaneously.
SoftBank sold its entire remaining NVIDIA shareholding in November 2025 — roughly $30 billion in stock — to reallocate capital toward its Stargate equity commitment. That is how certain Son is about this bet.
Technology partners extend well beyond the equity holders. NVIDIA committed to a $100 billion investment in OpenAI to purchase AI processors for the data centers. AMD agreed in October 2025 to supply GPUs for 6 gigawatts of future deployment. ARM, Microsoft, and Oracle round out the inner circle of technology contributors. Microsoft, which previously held exclusive cloud rights over OpenAI, now holds right of first refusal on future infrastructure needs rather than a lock — a shift that quietly repositioned OpenAI from tenant to owner.
In December 2025, SoftBank completed a $41 billion investment in OpenAI, securing approximately 11% equity — one of the largest private funding rounds in corporate history. By March 2026, OpenAI raised an additional $122 billion, with Nvidia deepening its partnership across the new capital structure. The financing architecture that looked shaky in May 2025 had, by mid-2026, become one of the most heavily capitalized infrastructure programs in the private sector's history.
The Physical Build: Sites, Scale, and What 8 Gigawatts Actually Means
Eight gigawatts equals the electricity consumption of a small nation. That figure, reached by late 2025, represents planned capacity across confirmed Stargate sites — not a projection, not a pledge. The flagship campus in Abilene, Texas, alone targets 1.2 gigawatts across eight AI factory halls, with 400,000 NVIDIA GB200 Blackwell GPUs planned at full build-out. Two buildings were operational by September 2025, with GB200 racks already running early training and inference workloads.
By September 2025, five additional US sites had been announced through a process that reviewed over 300 proposals from more than 30 states: Shackelford County, Texas; Doña Ana County, New Mexico; Lordstown, Ohio; Milam County, Texas; and a Wisconsin facility Oracle is developing with Vantage. Combined with Abilene, these six sites push planned capacity toward 7 gigawatts and represent over $400 billion in committed investment over three years.
US Stargate Sites — Confirmed as of July 2026
| Location | Developer | Planned Capacity |
|---|---|---|
| Abilene, Texas (flagship) | Oracle / Crusoe Energy | ~1.2 GW |
| Shackelford County, Texas | Oracle | Multi-hundred MW |
| Milam County, Texas | SB Energy (SoftBank) | 1.2 GW |
| Doña Ana County, New Mexico | Oracle | Multi-hundred MW |
| Lordstown, Ohio | SoftBank | ~750 MW |
| Wisconsin | Oracle / Vantage | ~300 MW |
| Saline Township, Michigan ("The Barn") | Oracle / Related Digital | 1 GW |
The Michigan campus, announced in October 2025 and dubbed "The Barn" for a historic red building preserved at the entrance, carries a $7 billion price tag and sits on 250 acres with three 550,000-square-foot buildings. Michigan Governor Gretchen Whitmer called it the largest economic project in the state's history. That kind of statement, from a governor of a state that has watched auto manufacturing decline for four decades, tells you something about what Stargate means outside the tech press.
Internationally, Stargate has moved faster than its US counterparts. Stargate Norway launched July 31, 2025, powered by renewable hydropower — the permits that take years in Texas took months in Scandinavia. Stargate UAE, a 1-gigawatt Abu Dhabi cluster developed with G42, NVIDIA, and Cisco, represents OpenAI's first international deployment. Stargate Argentina — a $25 billion investment in Patagonia — is the largest AI data center project in Latin America. By April 2026, OpenAI announced it had surpassed its original 10-gigawatt US commitment, with more than 3 gigawatts added in the prior 90 days alone.
"No one in the history of man built data centers this fast." — Sarah Friar, OpenAI CFO, September 2025
Each site uses rack densities of 100–150 kilowatts per rack — roughly four to six times what a conventional enterprise data center runs. Direct-to-chip liquid cooling within closed-loop systems is mandatory at this density. At Abilene, closed-loop cooling means the one-time water fill per building equals approximately two Olympic-sized swimming pools, after which the water recirculates rather than evaporates. The number gets cited constantly in environmental discussions; it is worth knowing it refers to fill, not ongoing consumption.
The Financing Collapse and Recovery
Elon Musk posted "They don't actually have the money" on X within hours of the January 2025 White House announcement. He was not entirely wrong — he was just early. SoftBank's initial $100 billion commitment was supposed to deploy "immediately." By May 2025, Bloomberg reported that SoftBank had not formalized a financing blueprint, had not begun detailed negotiations with banks, and had yet to develop a template for the capital structure. JPMorgan, Apollo, and Brookfield had held preliminary conversations. None had committed.
Tariffs complicated the picture structurally. TD Cowen analysts estimated that Trump's tariff policies were raising data center build costs by 5–15%, with some component categories facing steeper increases. Server racks, cooling systems, and chips are all tariff-sensitive. Investors who had entered preliminary discussions began questioning whether the return math still worked at inflated construction costs.
By August 2025, Bloomberg reported that the Stargate joint venture itself had not hired any staff and was not actively developing any data centers. The Information later confirmed this characterization, citing three people involved in what they described as a "shelved idea." The three partners had been arguing over responsibilities, governance structure, and financing terms. OpenAI tried to build its own data centers independently. Lenders refused — an AI company burning $8.5 billion in cash annually with no path to profitability was not a credit a bank would extend alone.
The solution was structural rather than financial. OpenAI returned to Oracle as the primary partner rather than pursuing a three-way Stargate consortium. Oracle would borrow the money and lease capacity back to OpenAI; OpenAI retained design control. This two-way structure replaced the unwieldy tripartite governance. It also meant the 10-gigawatt goal OpenAI had set for Oracle and SoftBank by end of 2025 was missed — OpenAI filled the gap with emergency cloud contracts from AWS, Google Cloud, AMD, and Cerebras.
The project rebuilt faster than it collapsed.By September 2025, five new sites were announced. By late 2025, $450 billion in committed investment and 8 gigawatts of planned capacity were confirmed — ahead of the original schedule, despite the intervening chaos. SoftBank's first $10 billion tranche, borrowed from Mizuho and other Japanese lenders, was deployed. JPMorgan extended a $2.3 billion loan for Abilene in May 2025. The venture that had no staff in August 2025 had 6,400 construction workers at Abilene alone by late that same year.
The Hardware Layer: NVIDIA, AMD, and OpenAI's Secret Chip
NVIDIA's position inside Stargate is stranger than it appears. The company is simultaneously a technology vendor, an equity investor in OpenAI, and a procurement partner — in a circular arrangement that has attracted scrutiny from analysts monitoring for AI bubble dynamics. NVIDIA invested $100 billion in OpenAI for non-voting shares. OpenAI committed to spending that capital on NVIDIA chips across at least 10 gigawatts of Stargate data centers. NVIDIA CEO Jensen Huang negotiated last-minute terms personally. The net effect: NVIDIA has financial incentive to see OpenAI succeed and commercial incentive to sell it hardware at scale. Every Stargate gigawatt built is, in part, a NVIDIA revenue stream that NVIDIA helped finance.
NVIDIA's first delivery of GB200 Blackwell racks to Abilene began in June 2025. Each GB200 NVL72 rack connects 36 Grace CPUs and 72 Blackwell GPUs in a liquid-cooled, rack-scale design delivering 30x faster inference for trillion-parameter models versus its predecessor generation.
AMD entered the picture in October 2025, committing to supply GPUs for 6 gigawatts of future Stargate deployment capacity. This diversification is strategically meaningful: it reduces OpenAI's single-vendor exposure to NVIDIA and introduces competitive tension into procurement negotiations. OpenAI has also contracted with Cerebras, AWS Trainium, and other silicon platforms — the infrastructure portfolio that looked monolithically NVIDIA-dependent in January 2025 now spans multiple chip architectures.
The most consequential hardware development inside Stargate is one that has received the least coverage: OpenAI's custom AI chip, internally codenamed "Titan." Developed in a $10 billion partnership with Broadcom and fabricated on TSMC's 3nm process, Titan is an application-specific integrated circuit optimized for inference workloads. Mass production is targeted for the second half of 2026, with initial deployment expected by December. OpenAI secured an exclusive supply agreement with Samsung for HBM4 memory, confirmed in March 2026. A second-generation chip, Titan 2, is already in early development.
The economics behind Titan are more compelling than the engineering: training workloads run infrequently, a few large runs per model generation. Inference runs continuously — every ChatGPT query, every API call, every agent session. OpenAI's gross margins currently sit around 33%; internal targets are 46%. If Titan achieves even a 20–30% improvement in inference efficiency over equivalent NVIDIA hardware, the gross margin impact at $25 billion or more in annual revenue is measured in billions annually. Stargate is also, underneath its infrastructure story, a margin expansion program.
Tape-out costs for a leading-edge ASIC on TSMC's N3 process exceed $500 million for a single chip version, before software, peripheral systems, and iteration costs. OpenAI's chip team numbers roughly 40 engineers — smaller than Google's TPU team and Amazon's Trainium group by an order of magnitude. Whether 40 engineers can execute a chip program at this cost and complexity is the question the semiconductor industry is watching most closely about Stargate in 2026.
The Geopolitical Calculation: China, Export Controls, and Sovereign AI
Presidents do not attend data center groundbreakings. When Trump stood at the White House to announce Stargate, the act itself communicated something separate from the numbers: AI infrastructure had crossed into statecraft. The US framing of Stargate has been consistent — domestic compute capacity as national security, AI leadership as strategic advantage, export controls as a complementary instrument of the same policy.
China's position makes the stakes concrete. In April 2025, Trump banned the sale of NVIDIA's H20 chips to China — chips specifically designed to comply with earlier export controls — producing a $4.5 billion charge on NVIDIA's first-quarter results. China responded by banning its companies from purchasing H20 chips in September 2025 and ordering data centers less than 30% complete to remove installed foreign chips. The White House then banned NVIDIA's scaled-down B30A chips in November 2025. This is not competition. It is decoupling.
Stargate's international expansion reveals a second strategic dimension: the project is positioning OpenAI as the infrastructure partner of choice for democratic nations pursuing sovereign AI. The UAE deal was not charity — the UAE leveraged its geographic and financial position to secure a 1-gigawatt cluster and a strategic partnership with the US government. South Korea announced a $75 billion sovereign AI investment in June 2025. India's sovereign large language model launched in early 2026. Each of these programs creates demand for a trusted US-aligned AI infrastructure partner, which Stargate — and its "OpenAI for Countries" initiative — is built to supply.
China's counter-move is open source. As the Atlantic Council noted in January 2026, China's open-source AI strategy could prove more effective than raw compute investment for capturing global market share: free models, deployment-ready technologies, and no political preconditions. Several major US tech companies were already using Chinese large language models in their applications by 2026. Stargate builds hardware moats; open-source strategies route around them.
The Power Problem Nobody Is Solving Fast Enough
The single largest constraint on Stargate is not capital, not governance, and not chips. It is electricity. Ten gigawatts of AI compute requires power at a scale that strains regional grids designed around industrial-era demand profiles. Grid operators warn that AI will account for a rapidly growing share of load growth — global data center electricity consumption, approximately 415 TWh in 2024, is projected to reach 945 TWh by 2030, per peer-reviewed analysis published in 2026, representing a 15% annual growth rate that significantly exceeds global electricity demand growth.
Stargate Norway moved faster than US domestic sites for a specific, unglamorous reason: Norway's hydroelectric permitting process is faster than Texas's grid interconnection process. Power permits, transmission upgrades, and grid infrastructure take years in the US regardless of political support. Regulatory review of water usage, wetland impacts, and grid structure proceeds on statutory timelines that executive enthusiasm cannot compress. The Abilene site committed to closed-loop cooling precisely to reduce environmental review exposure. Sam Altman has said internally that he wants 250 gigawatts of AI compute capacity by 2033 — a number that would require approximately $12.5 trillion in investment and a transformation of the US power grid that has no precedent in the grid's history.
OpenAI's response to grid concerns has been specific rather than vague: the company committed to funding its own power generation, transmission upgrades, and grid infrastructure for each site without passing costs to existing ratepayers. Site-specific "Stargate Community plans" were published for each facility location. SoftBank and OpenAI each invested $500 million in SB Energy in January 2026, committing $1 billion to accelerate data center power development. Small modular reactors are planned for baseload power at select sites. These commitments are real — and insufficient at the speed Altman is describing.
Who This Is For
Investors in NVIDIA, Oracle, AMD, or Broadcom are already inside this story. Stargate is the primary revenue story for Oracle's cloud business; Oracle beat analyst earnings estimates in 10 of 12 quarterly reports from 2022 through 2025, with cloud revenue reaching $6.2 billion in Q3 2025 driven by Stargate-related expansion. For investors who own these companies without understanding the Stargate dependency, this article changes the risk calculus: Oracle's stock is 56% off its 52-week high as of early 2026, and a Bloomberg report flagged that free cash flow could stay negative for years if demand falters or timelines slip.
Enterprise technology buyers who have standardized on OpenAI APIs have an indirect stake in whether Stargate delivers its capacity commitments. You have just launched an AI-powered product on OpenAI's API and your provider is rationing compute because its infrastructure buildout is 18 months behind schedule. The bottleneck is not the model — it is the data center. Understanding Stargate's progress is understanding your supplier's ability to scale.
Policy analysts, national security researchers, and anyone working on AI governance at the international level are watching Stargate as the most tangible expression of the US position that compute capacity equals strategic advantage. The Chatham House analysis from February 2026 is the clearest articulation of what middle powers must now decide: align with the US-led AI infrastructure stack, align with China's open-source approach, or attempt to build genuine sovereignty through specialization and international partnerships.
Energy sector analysts have a distinct concern: Stargate's power demands will reshape electricity markets in every state where sites are located. The company's commitment to self-fund power generation rather than burden existing ratepayers does not change the physical reality that 10 gigawatts of new load will alter dispatch economics, transmission congestion patterns, and long-term capacity planning for grid operators across the country.
Verdict
Stargate is real. The governance was a mess, the initial financing structure failed, the joint venture had no staff and no active construction for months in 2025 — and the project still came out ahead of its own schedule on capacity commitments. That trajectory tells you something about the underlying demand pressure. When a project survives its own internal collapse and accelerates anyway, the force driving it is not hype — it is a procurement need that has no other supplier at scale.
The risks that remain are structural, not financial. Power grid timelines are not compressible by capital. TSMC capacity constraints are real and shared by every major AI lab simultaneously. OpenAI's chip team of 40 engineers is attempting a chip program that cost Google years and far larger teams. Any one of these constraints, extended by 18 months, reshapes the 2029 capacity picture materially.
For investors: the circular financing structure between NVIDIA, OpenAI, and the data center partners deserves scrutiny — not because the projects are not real, but because the valuation assumptions embedded in those cross-commitments require demand growth to materialize on a specific timeline. For everyone else: Stargate is the physical layer of the next decade of AI. The companies and countries that do not understand what is being built here will spend that decade reacting to decisions made by the people who do.
The question Stargate cannot yet answer is the one that actually determines its place in history: whether the models trained inside these buildings produce applications valuable enough to justify the grid transformation required to power them. The infrastructure is arriving. The killer application is not confirmed.
Frequently Asked Questions
What is the Stargate Project?
Stargate is a $500 billion AI infrastructure joint venture announced on January 21, 2025, created by OpenAI, SoftBank, Oracle, and Abu Dhabi's MGX to build AI data centers across the United States. SoftBank holds financial responsibility with CEO Masayoshi Son as chairman; OpenAI holds operational responsibility. The goal is 10 gigawatts of AI compute capacity by 2029, a target OpenAI says it surpassed in April 2026.
Who is funding the Stargate Project?
SoftBank and OpenAI each committed $19 billion in equity, holding 40% ownership each. Oracle and MGX contributed $7 billion each. Remaining capital comes from debt financing and limited partners, including a $2.3 billion JPMorgan loan for the Abilene, Texas site. NVIDIA invested $100 billion in OpenAI in exchange for non-voting shares, with OpenAI committing to spend that capital on NVIDIA processors for Stargate sites.
Where are the Stargate data centers being built?
The flagship site is in Abilene, Texas, operational since September 2025. Confirmed additional US locations include Shackelford County and Milam County in Texas, Doña Ana County in New Mexico, Lordstown in Ohio, Wisconsin, and a 1-gigawatt campus in Saline Township, Michigan. International sites include Stargate UAE in Abu Dhabi, Stargate Norway, and Stargate Argentina in Patagonia.
Did the Stargate Project stall or fail?
The Stargate joint venture stalled significantly in mid-2025. Bloomberg reported in August 2025 that it had no staff and no active construction. SoftBank had not finalized financing, and partner disputes over governance caused delays. OpenAI independently sought data center financing and was rejected by lenders. The project recovered by restructuring as a bilateral OpenAI-Oracle arrangement, with construction resuming and targets being met ahead of schedule by late 2025.
What GPUs are being used in Stargate data centers?
The primary hardware is NVIDIA's GB200 Blackwell GPU architecture, with the first racks delivered to Abilene in June 2025. AMD agreed in October 2025 to supply chips for 6 gigawatts of future capacity. OpenAI is also developing a custom AI chip codenamed "Titan," built with Broadcom on TSMC's 3nm process and using Samsung HBM4 memory, targeted for mass production in the second half of 2026.
How much electricity will Stargate use?
At full 10-gigawatt capacity, Stargate would consume electricity equivalent to a small nation's total demand. The Abilene flagship site alone targets 1.2 gigawatts — enough to power approximately 750,000 US homes. OpenAI has committed to funding its own power generation and grid upgrades at each site rather than passing costs to local utility ratepayers, including a $1 billion commitment to SB Energy for power development.
How does Stargate relate to the US-China AI competition?
Stargate is explicitly framed as a national security infrastructure program, combining massive domestic compute investment with US export controls restricting advanced AI chips to China. Through its "OpenAI for Countries" initiative, Stargate extends this strategy internationally — offering democratic nations a US-aligned AI infrastructure partnership as an alternative to Chinese AI infrastructure. China's counter-strategy focuses on open-source AI models that bypass hardware dependency entirely.
Is Stargate a good investment?
Stargate benefits Oracle, NVIDIA, AMD, Broadcom, and SoftBank directly, with Oracle's cloud revenue reaching $6.2 billion in Q3 2025 driven by Stargate-related expansion. The primary risk is the circular financing structure between NVIDIA, OpenAI, and data center partners, which requires sustained demand growth on a specific timeline. Oracle's free cash flow could remain negative for years if AI demand or Stargate timelines slip. The project is not a bet — it is already happening. The question is whether demand materializes at the speed the capital structure requires.