August 17, 2026, (Inside AI) — Microsoft has roughly 2.2 million AI chips installed in its global datacentres, according to internal documents seen by the Guardian. That figure sits far below the 1.8 million chips the company reportedly targeted for the end of 2024, and it is less than half the number some industry experts expected given Microsoft’s public statements about its AI build-out.
The apparent shortfall emerges amid a $280 billion expansion of Microsoft’s cloud and AI infrastructure since 2022. Datacentre capacity is usually measured in gigawatts of electrical power, but converting that figure into actual chip counts reveals a gap between what Microsoft says and what it operates. The company has said it added 5 gigawatts of capacity over two years, yet its sustainability reports suggest a much smaller operational footprint.
Shaolei Ren, a professor at the University of California, Riverside, told the Guardian that Microsoft’s audited sustainability data points to roughly 1.2 gigawatts of AI capacity in 2024. Even that lower figure would imply a need for about 4 million AI chips if Microsoft added 5 gigawatts in the past two years.
“According to their own metrics, Microsoft could be correct. But it isn’t clear what they mean when they say they have added datacentre capacity. They are giving insufficient context,” Shaolei Ren, Professor, University of California, Riverside
Read: Microsoft Plans Maia 300 AI Chip Unveiling in September, Report Says
“The sustainability reports are audited by a third party. They have more credibility than announcements.” Shaolei Ren, Professor, University of California, Riverside
Power promises collide with physical reality
Microsoft’s flagship AI development in the United States, a pair of datacentres in Wisconsin and Georgia called Fairwater, illustrates the gap. In April, CEO Satya Nadella said the Wisconsin site “is going live.” Satellite footage from Epoch AI indicates only part of it is operational. In May, Microsoft admitted to a Wisconsin newspaper that Fairwater was not yet online.
Ren said this pattern is common. A multi-gigawatt, multibillion-dollar investment can take three years to deliver just 300 megawatts of built capacity. Sources inside Microsoft told the Guardian the company’s total number of AI chips has “barely moved” over the past year.
Nadella himself hinted at the bottleneck during a podcast appearance late last year. He said the biggest problem was not chip supply but electrical power and finding datacentre shells close enough to power sources.
“If you can’t do that, you may actually have a bunch of chips sitting in inventory that I can’t plug in. In fact, that is my problem today. It’s not a supply issue of chips. It’s actually the fact that I don’t have warm shells to plug into.” Satya Nadella, CEO, Microsoft
A Microsoft spokesperson disputed the Guardian’s estimates without specifying which numbers were wrong. The company said it does not report volumes of specific chips and that the Guardian’s figures were “inaccurate, drawing the wrong conclusions from incorrect assumptions.”
Read: US Power Use to Beat Record Highs in 2026 and 2027 as AI Use Surges, EIA Says
Nvidia’s secret supply chain obscures the AI boom
The discrepancy highlights a deeper opacity in the AI industry. Nvidia, the dominant maker of AI chips, does not disclose how many chips it sells or to whom. Its customers, including Microsoft, do not reveal their chip inventories either. Without that data, outsiders cannot easily verify whether the AI boom is real or inflated.
Nvidia’s revenue of $215.9 billion in February suggests enormous chip sales. Last March, Nvidia CEO Jensen Huang said orders for its newest Blackwell chips from four top customers amounted to 3.6 million units. Microsoft has historically been one of Nvidia’s largest buyers. Yet the internal document indicates Microsoft has fewer than half a million Blackwell chips installed, far below the roughly 1 million it might be expected to hold.
Some of the gap may be explained by Microsoft’s partnership with OpenAI. The commercial terms are private, and some datacentre deployments tied to that unit may not appear in the documents reviewed by the Guardian. Another factor is chip inventory sitting idle while datacentres await power connections. The Guardian’s methodology, reviewed by Abdeltawab Hendawi at the University of Rhode Island and Ren, assumes roughly 80% of datacentre electricity goes to chips, with each server holding eight GPUs and drawing about 10 kilowatts. That yields a conservative estimate of 6.4 million chips for 10 gigawatts of capacity.
Microsoft has said it now operates hundreds of datacentres on five continents. The company’s own efficiency reports suggest 89% of electricity in new datacentres powers IT systems. But even with generous assumptions, the installed chip count falls short of what public statements imply. Nvidia did not respond to a request for comment.
