Microsoft may have significantly fewer advanced AI chips running in its data centres than its public statements about artificial intelligence capacity suggest, according to a Guardian investigation published this week.
The newspaper reports that internal documents show Microsoft had targeted having 1.8 million AI chips installed globally by the end of 2024. Nearly two years later, in the middle of a $280bn (roughly €260bn) expansion of its AI infrastructure since 2022, the company has 2.2 million chips installed, according to the Guardian's sources. That is, the paper notes, less than half the number some experts had anticipated given the scale of Microsoft's spending.
Why the numbers matter
AI chips, most made by Nvidia, are the components that power the training and running of large AI models such as those behind Microsoft's Copilot products and its partnership with OpenAI. Hyperscale cloud companies including Microsoft, Google, Amazon and Meta have committed hundreds of billions of dollars to acquiring and installing them, making chip counts a closely watched proxy for how fast a company can actually expand its AI capability, rather than merely announce it.
An analyst specialising in Nvidia's chip supply told the Guardian the numbers were lower than expected. "They're low to me. They're less than I expected Microsoft would have," the analyst said.
“"They're low to me. They're less than I expected Microsoft would have."”
Microsoft disputed the newspaper's calculations, saying they were based on incorrect information, but did not specify which figures it believed were wrong, according to the Guardian. Sources inside the company told the paper that its total number of installed AI chips has "barely moved" over the past year.
A power problem, not just a chip problem
The discrepancy fits a pattern Microsoft's own leadership has acknowledged. Chief executive Satya Nadella said in a recent interview alongside OpenAI's Sam Altman that the constraint on Microsoft's AI growth is now electricity supply rather than chip availability. He described having advanced processors sitting unused because data centres lack the electrical infrastructure, sometimes called "warm shells", to connect them, saying it is not a supply issue of chips but a matter of not having facilities ready to plug them in.
That admission is consistent with the Guardian's finding that announcing gigawatts of planned power capacity is far simpler than actually bringing that capacity online for computing. One analyst quoted by the Guardian, cited under the surname Ren, said it may be possible to secure or announce a gigawatt of power capacity within a single quarter on paper, but bringing that capacity online and using it for computing within the same quarter is considerably harder.
The picture is complicated by a genuine industry-wide bottleneck. Nvidia's newest Blackwell-generation chips have been effectively sold out through much of 2026, according to market analysts, with advanced chip packaging capacity at Taiwan's TSMC booked many months in advance. That squeeze affects all large cloud providers, not just Microsoft.
Why it matters for Europe
For European readers, the story carries direct relevance. Microsoft has pledged $30bn (about €28bn) for AI infrastructure and operations in the UK between 2025 and 2028, and has separately committed to large data centre projects across the EU, including in Ireland, Germany, France, Spain and the Nordics. If the pace of chip deployment inside Microsoft's global network is genuinely slower than advertised, European customers and governments relying on Azure's AI capacity, from public sector digitisation programmes to corporate cloud contracts, may face longer waits than official announcements imply.
The findings also arrive as EU regulators scrutinise the market power of a handful of US cloud and AI providers under the bloc's Digital Markets Act, and as European energy grids grapple with the same electricity constraints Nadella described. Whether Microsoft's AI ambitions are being slowed by chip supply, power capacity, or a mix of both, the underlying message for the region is that the eye-catching capital expenditure totals cloud companies announce may not translate into computing power as quickly as headlines suggest.
This article is free to read. It always will be — no paywall, no account, no tracking.
