OpenAI Leaders Say Microsoft Isn’t Moving Fast Enough to Supply Servers
The Takeaway
• OpenAI says Microsoft hasn’t moved fast enough to supply it with servers
• OpenAI is discussing a deal with Oracle to build the world’s most powerful AI data center
• Microsoft promised OpenAI 300,000 Nvidia GB200 chips by the end of 2025
OpenAI and Microsoft soared to new heights by developing and sharing artificial intelligence, specialized servers and product revenue with each other. Microsoft’s cash has paid for nearly all of that work.
After marching in lockstep, their relationship is changing. Last week, after OpenAI CEO Sam Altman and Chief Financial Officer Sarah Friar raised $6.6 billion in capital, mostly from a slew of financial firms, they told some employees OpenAI would play a greater role in lining up data centers and AI chips to develop technology rather than relying solely on Microsoft, according to a person who heard the remarks.
Friar previously told some shareholders Microsoft hadn’t moved fast enough to supply OpenAI with enough computing power—hence the need to pursue other data center deals, which it had started to do, said a person who was involved in the conversation. For instance, OpenAI’s growing impatience with Microsoft recently prompted it to arrange an unusual data center deal in Texas with one of Microsoft’s rivals.
The companies previously included a provision in their contract that gave OpenAI room to work on such deals, despite agreeing to make Microsoft its exclusive cloud server provider, said a different person with knowledge of their agreement.
Taking more control of data center plans could help OpenAI stay ahead of rivals such as Anthropic and xAI, which aren’t locked into a single cloud provider.
The comments from OpenAI leaders last week reflected what’s been happening behind the scenes in recent months: While OpenAI lowers its dependency on Microsoft data centers, Microsoft has aimed to lessen its reliance on OpenAI technology as they increasingly compete in selling AI products and as Microsoft tries to lower the cost of running its products.
A Microsoft spokesperson did not have a comment on Altman’s and Friar’s remarks. An OpenAI spokesperson didn’t address those remarks and said its “strategic cloud relationship with Microsoft is unchanged.”
Now, as OpenAI moves forward with a plan to convert to a for-profit from a nonprofit, it is negotiating various financial details with Microsoft, such as the size of the stake Microsoft would have. Currently, in exchange for investing more than $13 billion in OpenAI, Microsoft has the rights to 75% of OpenAI’s future profits until the software company’s principal investment of more than $13 billion is repaid, and 49% of profits after that up to a theoretical cap.
Because of Microsoft’s outsize rights to future OpenAI profits under the current nonprofit configuration, it could end up with a significant portion of OpenAI’s shares after it turns into a for-profit. The nonprofit currently overseeing OpenAI also will get a sizable stake in the new for-profit firm—likely at least 25%, said a person who has spoken to company leaders.
The Microsoft-OpenAI partnership will still be critical for both firms. Microsoft can use OpenAI’s technology in perpetuity for its own products and OpenAI pays Microsoft to run its AI business in its Azure data centers. OpenAI pays a 20% commission to Microsoft when it sells access to AI models to other businesses through an application programming interface, and Microsoft pays OpenAI a 20% commission from reselling OpenAI models to Azure customers through a Microsoft API.
Musk-Related Concerns
But lately, Altman has sought to make the companies’ extensive computing partnership more flexible.
He’s become concerned about Microsoft’s ability to provide servers to OpenAI fast enough to stay ahead of new rivals such as Elon Musk’s xAI, according to four people working on data center projects for OpenAI. Altman’s urgency around securing more data center space picked up after Musk took a public victory lap about creating one the largest clusters of Nvidia AI chips in the world, these people said. Musk co-founded OpenAI with Altman and others.
Microsoft, for its part, effectively paid $650 million to hire a team of technologists from a startup called Inflection to try to develop software that might compete with OpenAI’s.
And Microsoft has become more cautious about paying for ever-bigger server clusters for OpenAI as the cloud giant aims to ensure it won’t take a loss on costly data centers that may not generate consistent revenue in the coming decades. Part of Microsoft’s concern is that clusters optimized for training AI models aren’t as cost effective when used for other purposes, such as powering AI apps, known as inference. It may be hard to predict the demand for large-scale training in the coming decade, which makes it risky to build expensive computing clusters that may sit idle for stretches of time.
One prominent example of how OpenAI’s relationship with Microsoft is changing involves a data center under construction in Abilene, Texas. OpenAI recently signed a deal to rent servers from the facility, which Oracle—a cloud rival to Microsoft—and other firms are developing and which they expect will be one of the largest in the world. Previously, Microsoft handled all data center affairs for OpenAI.
In the new deal, OpenAI negotiated directly with Oracle. Microsoft was informed about the negotiations and included in a press release about the deal, but in reality it has had relatively little involvement, according to two people working on it. Microsoft likely blessed the deal, given that Microsoft previously negotiated the right to block any arrangements OpenAI made with other cloud providers, according to someone who reviewed the contract.
Importantly for Microsoft, because Azure is the exclusive cloud provider to OpenAI, Microsoft is technically Oracle’s customer on OpenAI’s behalf, and can still count OpenAI’s use of the Oracle clusters as revenue for Azure, according to another person who spoke to a key Microsoft executive about the deal. (OpenAI said in a June post on X that it would “use the Azure AI platform on [Oracle] infrastructure,” without elaborating.)
Going Big in Abilene
OpenAI is now in talks with Oracle to lease the entire Abilene data center site, which could eventually grow to 2 gigawatts if Oracle is able to access more power at the site, according to one of the people involved in the deal. That amount of power could light up multiple big cities. The site is on track to draw a little under 1 gigawatt of power by mid-2026, meaning it could house several hundred thousand Nvidia AI chips.
It isn’t clear what role Microsoft would play in the proposed expansion in Abilene.
The changing relationship also means OpenAI could leave Microsoft out of more ambitious data center plans. The companies have long discussed how they could work together on bigger and bigger facilities, including a $100 billion supercomputing server cluster, which Microsoft code-named Mercury and OpenAI leaders referred to as Stargate.
During a question-and-answer session with employees last week, Friar said her team was already contemplating how to use debt financing to pay for such a cluster, implying it could happen with or without Microsoft’s hardware, similar to the Oracle deal.
To be sure, both companies were never sure they’d pursue Stargate together, given its proposed costs. OpenAI’s development of its own AI chip could bring down the expense of operating such a cluster, Altman told colleagues last week. The status of the chip couldn’t be learned, but OpenAI has discussed it with Broadcom, which develops AI chips with multiple companies, including Google.
‘Fairwater’ Disagreements
OpenAI currently rents most of its AI chips from Microsoft in the Phoenix area, according to The Information’s AI Data Center Database.
The companies are discussing the next phase of data center expansion, a project known as Fairwater. Microsoft plans to give OpenAI access to around 300,000 of Nvidia’s latest graphics processing units, the GB200s, from two data center sites in Wisconsin and Atlanta by the end of next year, according to the database and someone working on the plans. OpenAI will have access to a smaller number of GB200s earlier in the year in the Phoenix area as part of thean expansion of an existing Microsoft data center it already uses.
The companies have disagreed over certain aspects of the Fairwater data center design, according to two people working on the project. OpenAI has asked Microsoft to build a more advanced cluster of Nvidia GPU servers and to tweak the design to eke out more computing power, according to two people with direct knowledge.
Altman in recent weeks has also asked Microsoft to push up its timeline for the Wisconsin project, which is expected to be partially online by the second half of next year with GB200 chips, one of these people said.