Microsoft says it's building AI infrastructure faster than almost anyone on the planet. Its own internal numbers suggest otherwise.
A Guardian investigation found that Microsoft targeted 1.8 million AI chips installed across its data centers by the end of 2024. Nearly two years and $280 billion later, internal documents show it actually has 2.2 million installed, roughly half what industry watchers expected given the scale of Microsoft's public claims.
Doing the Math on Gigawatts
Microsoft doesn't disclose chip counts directly. Companies like it describe AI capacity in gigawatts instead, and Nvidia, which makes the chips, doesn't report sales figures or customers either.
That secrecy is exactly what makes the math worth doing. Microsoft's public reports suggest it has added 5 gigawatts of data center capacity over the past two years, on top of roughly 5 gigawatts it says it already had installed by 2024. That implies a total of about 10 gigawatts, which, after accounting for cooling systems and non-chip server components, works out to somewhere around 6.4 million AI chips.
Shaolei Ren, a professor at UC Riverside, cross-checked that math against Microsoft's own sustainability reports, which are audited by a third party and cover electricity usage separately from financial statements. Those numbers put Microsoft's actual 2024 AI capacity closer to 1.2 gigawatts, implying a chip count nowhere near the public figures suggest.
"According to their own metrics, Microsoft could be correct. But it isn't clear what they mean when they say they have added datacenter capacity," Ren said. "They are giving insufficient context."
A Flagship Project That Isn't Actually Online
The clearest example is Fairwater, Microsoft's largest U.S. AI project, split across Wisconsin and Georgia. CEO Satya Nadella said in April the Wisconsin site "is going live."
Satellite imagery reviewed by Epoch AI suggests only part of the building is operational. In May, Microsoft told a Wisconsin newspaper the site wasn't actually online yet.
Ren said this pattern is common across the industry. A project announced as multi-gigawatt and multibillion-dollar often has a fraction of that built years later, sometimes as little as 300 megawatts.
There's also a shortfall in Microsoft's holdings of Nvidia's newest Blackwell chips specifically. Nvidia's CEO said the company's top four customers ordered 3.6 million Blackwells combined. If Microsoft's share tracked its historical position as one of Nvidia's largest buyers, it should hold close to a million. Internal documents show less than half that.
Microsoft Pushes Back, Without Specifics
Microsoft disputes the Guardian's calculations but hasn't said which figures are wrong or why. A spokesperson said the company doesn't report chip volumes and that its infrastructure blends custom silicon with AMD, Intel and Nvidia hardware across multiple generations.
Nadella has offered his own explanation elsewhere. On a podcast last year, he pointed to power, not chip supply, as the real constraint.
"You may actually have a bunch of chips sitting in inventory that I can't plug in," he said.
That framing suggests Microsoft may genuinely own more chips than it has installed, sitting unused because there isn't yet the electrical infrastructure, what Nadella called "warm shells," to run them.
What This Means for Miami
Investors and enterprise buyers in Miami's growing AI and fintech sector have largely taken hyperscaler capacity claims at face value when pricing AI-linked deals and forecasts. This investigation is a reminder that gigawatt figures and chip counts from even the largest cloud providers aren't independently verified, and the gap between announcement and operational reality can be substantial.
For South Florida's smaller data center developers competing for AI tenants, there's a different lesson. A facility that's actually online and delivering usable compute today has real value the hyperscalers' own numbers can't always back up, whatever the headline capacity figures suggest.
