Guardian Investigation: Microsoft Has 2.2M AI Chips Installed — Far Below What Its Own Capacity Claims Imply

How many AI chips does Microsoft actually have running? It sounds like a simple question with a simple answer. It is neither — and that is exactly the problem.

On August 17, The Guardian published the results of a months-long investigation into Microsoft’s AI infrastructure build-out, and the numbers it surfaced from internal documents tell a story that is hard to square with the company’s public narrative. Nearly two years into a roughly $280 billion expansion in land, buildings, and computational infrastructure — including more than $41 billion in the most recent quarter alone — Microsoft has approximately 2.2 million AI chips installed in its datacentres worldwide.

That figure is not small in absolute terms. But context makes it startling: Microsoft had reportedly targeted 1.8 million chips by the end of 2024, meaning the count has “barely moved” over the past year according to sources within the company, even as capital expenditure shattered records. And it is less than half the number that independent experts say Microsoft’s own capacity claims should imply.

The math that doesn’t add up

The core of the Guardian’s finding rests on a unit the industry prefers to keep vague: gigawatts. Datacentre operators talk about “AI capacity” in power terms, and Microsoft’s annual reports and quarterly earnings suggest it has added roughly 5GW of datacentre capacity over the past two years, on top of an internal 2024 presentation that already claimed 5GW installed — implying as much as 10GW total today, spread across “hundreds of datacentres on five continents.”

Working with Shaolei Ren of UC Riverside and Abdeltawab Hendawi of the University of Rhode Island, the Guardian walked through a standard conversion from power to silicon. If 80% of a datacentre’s electricity reaches the IT equipment (Microsoft’s own sustainability reporting suggests up to 89% in its newest facilities), 10GW leaves about 8GW for compute. A server with eight H100-class GPUs draws around 10kW. Divide it out and 10GW of capacity should host roughly 6.4 million GPUs — against the 2.2 million the internal documents show.

Ren, who has studied Microsoft’s audited sustainability reports, goes further: those reports suggest Microsoft’s AI capacity in 2024 was closer to 1.2GW, not 5GW. “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,” he told the paper. “They’re giving insufficient context.” An analyst specializing in Nvidia put it more bluntly on seeing the installed count: “They’re low to me. They’re less than I expected Microsoft would have.”

‘Warm shells,’ not chips

Satya Nadella has an explanation of sorts — one he offered unprompted on the All Things AI podcast late last year. Microsoft’s binding constraint, he said, is not silicon supply but powered buildings. “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.”

The site-level evidence backs him up. Microsoft’s flagship US AI project, the Fairwater campus pair in Wisconsin and Georgia, was declared “going live” by Nadella in April. Satellite imagery tracked by Epoch AI appears to show only part of the Wisconsin site operational, and in May the company admitted to a local newspaper that Fairwater was not yet online. What began as a multi-gigawatt, multibillion-dollar announcement has, three years in, delivered roughly 300MW of built capacity by Ren’s count.

The internal documents also suggest Microsoft holds fewer than half the Blackwell-generation GPUs one would expect. When Nvidia CEO Jensen Huang said last March that his top four customers — widely understood to include Microsoft — had ordered 3.6 million Blackwell chips, simple arithmetic would put Microsoft’s share near a million. The documents indicate well under that installed.

Microsoft pushes back

Microsoft’s response to the investigation was categorical but detail-free: “The estimates the Guardian has shared with us are inaccurate, drawing the wrong conclusions from incorrect assumptions.” The company declined to specify which numbers were wrong or why, noting that its datacentres “combine custom silicon, AMD, Intel and Nvidia chips across multiple generations” and that it “does not report on the volume of specific chips in its AI infrastructure.” Nvidia did not respond to a request for comment.

There are legitimate caveats. The OpenAI partnership may place some Microsoft-funded capacity outside the documents the Guardian reviewed. Mixed-generation fleets, the 89%-efficiency figure, and power oversubscription all bend the math. But as the Guardian notes, the discrepancies run in one direction — and the company offering the correction is also the only party that knows the real number.

Why this matters beyond Microsoft

The deeper problem the investigation exposes is structural. Nvidia, now one of the two most valuable companies in the world, does not disclose how many chips it sells or to whom. Its hyperscale customers do not disclose how many they own. Capacity is announced in gigawatts that may mean powered land, constructed shells, or energized racks. The result is that the single most important physical input to the AI economy — compute — is effectively unauditable from the outside.

That opacity cuts both ways. If installed capacity genuinely lags announcements, the “AI bubble” narrative gains ammunition: hundreds of billions in capex is buying land and buildings faster than it is buying inference. If the chips exist but sit in inventory awaiting “warm shells,” then the bottleneck is power and construction — a slower, more fixable problem, but one that still means reported capacity overstates delivered compute today. Investors, energy planners, and policymakers are currently forced to guess which world they are in.

For a company that has staked its future on being the AI infrastructure layer for OpenAI and enterprise customers alike, the gap between 2.2 million installed GPUs and the ~6.4 million its capacity claims imply is not an accounting quibble. Until Microsoft — or regulators, or its audited sustainability filings — offers numbers precise enough to reconcile, every gigawatt headline it publishes should be read as an aspiration with a construction schedule attached, not a statement about silicon that is actually running models today.