How OpenAI’s $500bn data centre venture Stargate has shifted shape
Chief executive Sam Altman’s flexible approach to infrastructure projects is unsettling partners but boosting computing lead
OpenAI’s $500bn Stargate plan to secure computing power is being reworked and, in places, abandoned. Yet its willingness to strike whatever deals are necessary to secure capacity has given it an edge in the race to build AI infrastructure.
In recent weeks, the group has halted planned data centres in the UK and Norway, declined to expand its flagship site in Abilene, Texas and seen several senior figures tied to Stargate leave for rival Meta.
Originally announced in early 2025 by Donald Trump as a $500bn joint venture between OpenAI, Oracle, Abu Dhabi fund MGX and Japan’s SoftBank, Stargate was designed to finance and build dedicated data centres for OpenAI’s use. The idea was to pool capital to tackle the cost and complexity of AI infrastructure.
That concept has rapidly given way to a more flexible approach, with OpenAI increasingly relying on third-party providers, leasing capacity rather than building and owning its own facilities.
“I don’t know what ‘Stargate’ means at this point,” said a person who was involved in the data centre buildout at the outset. “I think it is a completely antiquated line right now.
The shape-shifting scheme embodies OpenAI’s broader efforts. From core technology to partnerships, the company places overlapping bets and abandons all but the most expedient. The approach owes much to chief executive Sam Altman’s “venture mindset”, said one person close to the company.
People familiar with the changes at Stargate said that, in practice, OpenAI has abandoned the joint venture in favour of a series of large bilateral deals. But OpenAI executives said the guiding principle remains to “build more compute”.
Oracle has played the most prominent role, agreeing last year to supply OpenAI with 4.5 gigawatts of computing capacity under a five-year, $300bn agreement.
The company has also struck partnerships with AMD, Broadcom, Nvidia, CoreWeave and Cerebras in deals that at one point totalled more than $1tn in commitments, some of which have since been scaled back.
Questions remain over whether it can afford its huge infrastructure spending. On Tuesday, a report that OpenAI had missed internal targets for revenue and user growth led to share price falls at companies linked to the start-up including SoftBank, Oracle and CoreWeave. OpenAI said in response that its business is “firing on all cylinders.”
OpenAI has cut costs by sidelining or revising Stargate plans. But the willingness to renegotiate or walk away from projects has unsettled partners and raised questions about OpenAI’s reliability as a counterparty.
It initially aimed to develop its own data centres under the Stargate banner as it faced two challenges: a complex supply chain, and wariness from lenders about funding a lossmaking start-up with no credit rating.
The joint venture — pooling cash from blue-chip partners and gaining the imprimatur of the US president — aimed to tackle both issues.
Since then, a surge in demand for AI tools has spurred chip, cloud and infrastructure groups into action and eased investor scepticism.
“We created enough demand in the market, and others came along, so we sidelined first party data centres,” said the person involved in Stargate.
OpenAI said: “Stargate is the umbrella for our compute strategy and we’ve done exactly what we said we would do: secure the compute OpenAI needs at unprecedented scale.
“We’ve expanded to a broad partner network across clouds, silicon, and infrastructure, capacity is coming online, and we’re ahead of schedule. Building at this scale is what allows us to put powerful AI into the hands of more people, businesses, and developers.”
OpenAI this month put on hold one data centre project in the north-east of England and rejigged another in Narvik, Norway. Both projects were badged as being part of Stargate in a tie-up with AI cloud start-up Nscale.
OpenAI annoyed the British government by blaming restrictive regulation and high energy costs for halting plans in the country. The UK AI minister Kanishka Narayan said that the “only thing that has changed [since] the moment of those commitments . . . has been the financing environment for OpenAI.”
UK-based Nscale had been investing heavily in anticipation of leasing capacity to OpenAI. Microsoft has stepped in at Nscale’s Narvik site. OpenAI said it still plans to access capacity in Norway via its broader Microsoft deal, instead of leasing directly from Nscale.
Last month, the AI lab ditched plans to expand its Abilene, Texas site, declining to take up a lease option with developer Crusoe.
The person involved in Stargate said OpenAI sourced cheaper options, including in Michigan, because “there is limited money in the world, whatever Sam [Altman] might say [and] the expansion was at an insane price.”
Microsoft has stepped in to take the extra capacity at Abilene. A person close to Crusoe said Microsoft was a preferable tenant. “It’s a better outcome, [they are] more creditworthy.”
Someone familiar with Microsoft’s thinking said its decision “helps out those who are feeling let down and misled by OpenAI on projects”.
“Stargate has now had three or four permutations, I don’t know what it is right now. I can’t tell you. Maybe it never really existed in the first place,” the person added.
When Stargate launched, the company set a goal of locking in 10GW of capacity — roughly equivalent to 10 nuclear power stations — at a cost of about $500bn by the end of the decade.
The company now aims to beat that target. It claims to have already secured more than 8GW of capacity, and forecasts spending more than $600bn by the end of 2030.
That would mean the lossmaking company far outspending rivals such as Anthropic and Elon Musk’s xAI. OpenAI’s commitments are more in line with Amazon, Google, Microsoft and Meta — which generate tens of billions in annual profit.
SoftBank, the main financial sponsor of the original joint venture, is working on building its own data centres but those are yet to come online. The computing capacity from one project in Ohio could eventually go to OpenAI after a tender process, said a person close to the situation.
“Stargate hasn’t disappeared but it has evolved and people can basically pick their own definition of Stargate now,” said another person familiar with the Japanese investor’s thinking. “In one way, basically any compute that involves Softbank or Oracle can be Stargate”.
SoftBank declined to comment.
According to the person involved in Stargate, the term lost its meaning when it was co-opted by OpenAI’s policy and public relations teams as a catch-all for infrastructure build-out. The idea that being the tenant of a data centre in Europe fell under the Stargate umbrella was “always fake”, they said.
Still, OpenAI’s infrastructure push could give it an advantage. Anthropic chief executive Dario Amodei has criticised rivals for their gung-ho plans. But with power constraints starting to weigh on its ability to meet fast-rising demand, Amodei signed off hundreds of billions of dollars of spending on long-term capacity this month.
OpenAI’s moves may be prescient, as long as the company can meet its lavish spending commitments.
In a note to sales staff earlier this month, OpenAI’s chief revenue officer Denise Dresser wrote: “We saw the exponential compute curve earlier, acted on it faster, and now have a real structural advantage.”