An investigative report by The Guardian finds a notable gap between Microsoft’s public statements about its AI infrastructure and the inferred, operational compute capacity. Internal documents obtained by the newspaper indicate about 2.2 million AI chips are deployed across Microsoft’s data centres—less than half of some expert estimates based on the company’s public disclosures.
Investments and public statements
Microsoft has spent roughly $280 billion on AI infrastructure since 2022, including more than $41 billion in the most recent quarter alone. The company’s annual and quarterly reports have given the impression that it built out 5 gigawatts (GW) of data-centre capacity over the past two years. A 2024 investor presentation cited an existing 5 GW of capacity and suggested that, with new investments, total capacity could reach as much as 10 GW.
Reaching 10 GW would require on the order of 6.4 million GPUs—graphics processors used to train and run AI models. The 2.2 million chips referenced in the leaked documents fall well short of that figure.
External audits and expert assessments
Shaolei Ren, a professor at the University of California, pointed out that Microsoft’s third-party audited sustainability reports portray a different picture: based on those reports the actual AI capacity in 2024 may have been around 1.2 GW. Ren noted that while announced capacity can appear achievable on paper in a single quarter, the physical commissioning and reliable provision of compute capacity are considerably more complex.
Sources within Microsoft told The Guardian that the number of AI chips has barely grown over the past year.
The Fairwater data-centre projects: plans vs. reality
The reporting examines Microsoft’s largest U.S. AI project, the Fairwater data-centre pair in Wisconsin and Georgia. Satya Nadella announced the Wisconsin site was live in April 2025, but satellite imagery and subsequent local reporting show only portions of the campus operating. Microsoft later acknowledged to a local paper in May that the facility had not been fully commissioned.
What was originally planned as a multi-gigawatt, multi-billion-dollar investment produced about 300 megawatts (MW) of capacity over three years, according to the article.
Hardware constraints and supplier limits
Analysts say part of the discrepancy can be attributed to OpenAI’s prioritized hardware requirements, which directly consume significant compute capacity for its research lab. Microsoft also appears to have received fewer Nvidia Blackwell chips than might have been expected from its historical purchasing patterns. Jensen Huang, Nvidia’s CEO, said in March 2025 that the four largest customers—likely Amazon, Oracle, Microsoft and Google—ordered a total of 3.6 million Blackwell chips. Earlier, Microsoft had been one of Nvidia’s largest customers, suggesting nearly one million Blackwells could have been deployed; the reporting indicates the actual number is less than half that.
Satya Nadella has himself pointed to infrastructure bottlenecks on a podcast, arguing the primary barrier is not chip manufacturing capacity but the lack of ready data-centre infrastructure to install and network those devices. Large inventories of chips are of limited use if physical infrastructure shortcomings prevent them from being brought online.
Lack of transparency and market implications
Microsoft told The Guardian the paper’s calculations rely on inaccurate data but did not specify which details it disputes. A clearer market view is hindered by limited transparency: Nvidia does not publish customer-by-customer shipment data, and major cloud providers do not disclose their actual inventories. This opacity makes it increasingly difficult for investors to distinguish between announced, contracted and actually commissioned compute capacity.
An AI assistant contributed to preparing this article; the final content was edited and verified by the newspaper’s journalist.
