Over the past year, Microsoft and OpenAI have engaged in a challenging struggle over data center supply. This conflict culminated in Microsoft relaxing its exclusive agreement, allowing OpenAI to sign substantial cloud service contracts with competitors like Oracle, with projected payments to these rivals exceeding those to Microsoft itself by 2030. At first glance, it seems Microsoft is relinquishing billions in orders, but this tech giant's strategy is far more astute than it may appear.

According to a report by The Information, this shift does not signify a retreat by Microsoft. Insider sources revealed that CFO Amy Hood and her team repeatedly warned against blindly meeting OpenAI's insatiable expansion demands, as this could lead to overbuilding and leave Microsoft unable to recoup its investments. Simultaneously, CEO Satya Nadella is highly alert to external accusations of Microsoft "controlling" OpenAI, with investigations underway by the Federal Trade Commission and EU regulators.

Microsoft's concessions have yielded tangible benefits. Despite relaxing the exclusivity terms, it remains the sole provider of supercomputing cloud services necessary for training OpenAI's models, a critical component in AI development. Additionally, Microsoft continues to take a 20% share from OpenAI's soaring revenues while utilizing its model technology for its own Copilot products. OpenAI is expected to pay Microsoft around $135 billion in server lease fees by 2030.

Some Microsoft executives privately question whether Oracle can deliver on its promises to provide 4.5 gigawatts of data center capacity by 2032, a commitment that could be difficult to fulfill physically. Similar doubts have also emerged among investors. In contrast, Microsoft's cautious approach to capital expenditure may align more closely with the financial discipline expected of a mature tech company.

From Honeymoon to Rift: Exclusive Agreement Becomes a Bottleneck

The tensions between Microsoft and OpenAI surfaced in early 2024. At the time, Microsoft was establishing multiple dedicated data centers for OpenAI, including the Fairwater supercomputer project in Wisconsin, which incurred costs in the billions and required near 1 gigawatt of power. Microsoft aimed to operationalize the Wisconsin and Atlanta data centers by 2026 and to initiate a second facility in Wisconsin by 2028.

However, OpenAI CEO Sam Altman felt that this was insufficient. He requested that Microsoft provide additional gigawatt-level capacities by the end of the century, necessitating Microsoft to invest hundreds of billions in Nvidia chips and electrical infrastructure. An attendee revealed that Altman told an AI researcher that Microsoft's refusal to accelerate data center construction was the "biggest obstacle" to achieving general artificial intelligence.

Construction delays further exacerbated the situation. The Wisconsin project faced several months of delays due to uneven concrete pouring by contractors. At a mid-2024 meeting, Altman confronted Microsoft executive Noelle Walsh, who oversaw data center construction, questioning why timelines could not be shortened and even suggesting escalating the matter to Nadella. According to attendees, Walsh explained that the delays stemmed from construction factors beyond Microsoft's control.

On Microsoft's side, there were financial considerations. Hood's finance team raised concerns that unlimited expansion based on OpenAI's demands could lead to wasted resources. Azure executives privately believed that no cloud service provider could meet OpenAI's requested timelines for such massive compute power—a judgment that later served as a crucial basis for Microsoft's agreement to relax exclusivity.

Breaking Through and Compromise: Distribution of the $450 Billion Budget

The summer of 2024 marked a turning point. Altman requested Microsoft to permit discussions with other cloud service providers, and Nadella was sensitive to public accusations of "stifling OpenAI." Under Nadella's guidance, Hood's team reached a new agreement with OpenAI's CFO Sarah Friar: Microsoft retained preferential supply rights and could choose to meet OpenAI's needs or allow it to find another supplier.

Subsequently, OpenAI acted swiftly, signing its first 1 gigawatt contract for a Texas data center with Oracle in July 2024, followed by Microsoft approving a second round of negotiations in December, which resulted in the announcement of a staggering 4.5 gigawatts agreement seven months later. An OpenAI employee disclosed that the company has allocated a budget of $450 billion for server expenditures before 2030, with payments to Oracle and other competitors expected to surpass those to Microsoft.

The scale of these agreements is astonishing. Besides Oracle, OpenAI has signed a $22.4 billion cloud service contract with CoreWeave, rents servers from Google Cloud, and plans to utilize up to $100 billion investment from Nvidia to build its data centers. Oracle's stock surged 36% following the announcement of projected $400 billion in revenue from cloud services over four years, with OpenAI orders being a major driver.

However, Microsoft has not completely exited the scene; in addition to the anticipated $135 billion in revenue, it still exclusively provides the supercomputers required for model training—an essential aspect of AI development, much more crucial than inference deployment. A person involved in Microsoft's negotiations stated, "Every number is larger than we expected when we signed in 2023; now we need to figure out how to adapt to a rapidly changing world."

Astute Retreat: Microsoft’s Risk Management Logic

Microsoft shareholders exhibit differing views about this arrangement. Equity Armor Investments portfolio manager Joe Tigay admitted, "It would obviously be better if OpenAI only used Azure; I won’t lie." Yet, some investors appreciate Microsoft's cautious approach.

Gabelli Funds portfolio manager Hendi Susanto noted, "This allows Microsoft to manage the fluctuations in AI demand better; we can expect imbalances. Both Microsoft and Oracle are mature large software companies, and free cash flow is a primary driver of performance. From this perspective, Microsoft’s capital expenditure strategy appears superior."

Internally, Microsoft remains skeptical about Oracle's ability to meet its commitments. Two insiders revealed that some executives believe that Oracle's delivery of 4.5 gigawatts capacity by 2032 is "physically impossible," and potential costs could escalate due to Nvidia's next-generation chips being unpriced. In recent weeks, several technology industry analysts and investors have publicly voiced similar concerns.

In one meeting, Microsoft executive Jon Tinter asked Altman: If OpenAI requires hundreds of gigawatts of capacity—which would cost trillions of dollars—what should Microsoft do? According to attendees, the OpenAI team responded seriously: Microsoft could choose to build or allow a willing supplier to take on that responsibility. This nearly reckless ambition further convinced Microsoft that rather than bear astronomical financial risks, it would be better for competitors to take such risks.

A person involved in Microsoft's negotiations acknowledged that there were discontent on both sides regarding the partnership, but "those who truly understand the situation know it's the only feasible solution. Were there tense moments in conversations? Of course," this individual noted, "but you still need to figure out how to resolve the issues." Altman recently informed colleagues that OpenAI plans to use 250 gigawatts of data center capacity by 2033—equivalent to one-third of the current peak energy usage in the United States. In the face of such a massive appetite, Microsoft's retreat may not be a sign of weakness but rather a pragmatic move for self-preservation.