Microsoft’s AI infrastructure appears to have a significant gap between its reported expansion and its installed chip count. The source report says Microsoft targeted 1.8 million AI chips in datacentres worldwide by the end of 2024, while internal documents seen by the source report indicate 2.2 million chips installed nearly two years later during a reported $280bn expansion.
The discrepancy does not establish that Microsoft has failed to build the infrastructure it has announced. It does show how difficult it is to assess the company’s AI capacity because Microsoft does not disclose the volume of specific chips in its infrastructure, while Nvidia generally does not report how many chips it sells or to which customers.
What the reported figures show
Microsoft has said it built AI infrastructure rapidly, including roughly $280bn invested in land, buildings and computational infrastructure since 2022. Its chief executive, Satya Nadella, also said last year that the company would double its global datacentre footprint by mid-2027.
Microsoft’s public statements indicate it added 5GW of datacentre capacity over the past two years and now has hundreds of datacentres on five continents. A 2024 investor presentation reportedly said 5GW was already installed, which could imply about 10GW now. The source report cautions that some of this capacity may serve non-AI cloud services.
Related coverage: Apple overtakes Nvidia to become the world’s most valuable company again.
Why power capacity does not equal working chips
The source report used power figures to estimate how many graphics processing units, or GPUs, Microsoft might have. A 10GW AI datacentre footprint could imply roughly 6.4 million chips after accounting for electricity used by computer systems rather than cooling and other equipment, based on estimates reviewed by University of Rhode Island professor Abdeltawab Hendawi and University of California, Riverside professor Shaolei Ren.
Ren said Microsoft’s sustainability reports point to AI capacity closer to 1.2GW in 2024. Even that lower figure, combined with an added 5GW of AI datacentre capacity, would imply roughly 4 million chips under the source report’s methodology. The calculation is only a broad approximation because chip models use different amounts of power and datacentres also contain networking, storage and other systems.
Projects, power and the OpenAI relationship
Ren said announcing or securing 1GW of power capacity in a quarter could be plausible, but bringing that capacity online for computing would be more difficult. The source report said people within Microsoft described the company’s total AI chip count as having barely moved over the previous year.
Microsoft’s Fairwater development in Wisconsin and Georgia illustrates the distinction between announced and operational capacity. Nadella said in April that the Wisconsin project was going live, while satellite imagery cited in the source report appeared to show only part operational; Microsoft told a Wisconsin newspaper in May that the site was not yet online. Ren said the project had initially been described as a multigigawatt investment, but only 300MW had been built three years later.
The source report said some deployments connected to Microsoft’s commercial relationship with OpenAI may not appear in the internal documents it reviewed because the precise terms of that partnership are not public. It also raised the possibility that some chips may have been purchased but not installed because suitable datacentre buildings and electrical power were unavailable.
What Microsoft says
Microsoft said the calculations were based on incorrect information and drew wrong conclusions from incorrect assumptions. It said its datacentres use custom silicon, AMD, Intel and Nvidia chips across multiple generations, alongside the networking, storage and systems infrastructure needed to operate at scale.
Microsoft does not report the volume of specific chips in its AI infrastructure. Nvidia did not respond to a request for comment cited in the source report, leaving the underlying supply and allocation picture difficult to verify independently.
No comments yet. Start the discussion.