Microsoft is facing questions about how many AI chips it actually has installed in its datacenters. A new investigation by the Guardian found a gap between the company's public statements and internal figures. The difference could be large.
Microsoft reportedly aimed to have 1.8 million AI chips installed by the end of 2024. Internal documents seen by the Guardian show it had 2.2 million chips installed by mid-2026. That number is far lower than some experts expected.
The chips in question are made by Nvidia. These chips power the artificial intelligence systems that companies like Microsoft rely on. Nvidia does not publish sales figures or say which companies buy how many chips.
This makes it hard for outsiders to check what tech companies say about their AI buildouts. Microsoft has spent about $280 billion on AI infrastructure since 2022. That includes more than $41 billion in the last quarter alone.
How the Numbers Were Calculated
Researchers estimated Microsoft's chip count using its own reports. Companies describe their AI capacity in gigawatts of power rather than a chip count. One gigawatt can power between 700,000 and 1 million homes.
Microsoft's public reports suggest it added 5 gigawatts of datacenter capacity in the last two years. Based on the company's own 2024 presentation, that could mean a total capacity of about 10 gigawatts today.
Dividing that power figure by the energy use of a typical AI chip suggests Microsoft should have several million more chips than it currently has installed. A professor at the University of California, Riverside, said Microsoft's sustainability reports paint an even lower estimate.
Shaolei Ren, the professor, said those reports suggest Microsoft's 2024 AI capacity was closer to 1.2 gigawatts. That would still mean Microsoft should have roughly 4 million chips today. The gap between expectation and reality remains large either way.
Datacenters Not Yet Online
Some of Microsoft's biggest AI projects also appear unfinished. The company's Fairwater datacenters in Wisconsin and Georgia were described as "going live" by chief executive Satya Nadella in April.
Satellite images from Epoch AI show only part of the Wisconsin site appears active. Microsoft later told a local newspaper the site was not yet fully online.
Ren said this pattern is common in the industry. A project that starts as a multi gigawatt plan can end up with only a few hundred megawatts built years later.
The investigation also looked at Nvidia's newest chip, called Blackwell. Nvidia's chief executive Jensen Huang said last year that its top four customers ordered 3.6 million Blackwell chips combined.
Microsoft has historically been one of Nvidia's biggest buyers. Based on past patterns, it should have close to 1 million Blackwell chips. The documents reviewed by the Guardian show less than half that number installed.
Nadella addressed a version of this issue on a podcast last year. He said the biggest problem was not a shortage of chips but a shortage of places to plug them in.
He said Microsoft may have chips sitting in inventory because there is not enough electrical power or building space ready to use them. Sources inside the company said its total chip count has barely changed over the past year.
Microsoft responded to the findings in a statement. The company said its infrastructure combines custom chips along with hardware from AMD, Intel and Nvidia across many generations.
Microsoft said it does not report the volume of specific chips in its systems. The company said the Guardian's estimates were based on incorrect assumptions but did not specify which numbers were wrong.
Nvidia did not respond to a request for comment on the findings. The exact number of AI chips Microsoft has remains unconfirmed outside the company.
