Microsoft has 2.2m AI chips installed. Its own capacity claims imply up to 6.4m

A Guardian investigation reports that internal documents show Microsoft with 2.2m AI chips installed across its datacentres worldwide ^[1], against a reported target of 1.8m by the end of 2024 ^[2]. The company cleared that target by roughly 400,000 chips ^[1], yet the Guardian says the total sits awkwardly beside what Microsoft has told investors about its AI capacity ^[3] - which is the part that matters if you are signing a multi-year commitment on the strength of a capacity announcement.

The arithmetic is worth following. Microsoft's annual reports and quarterly earnings suggest it added 5GW of datacentre capacity over the past two years, and it says it now runs hundreds of datacentres on five continents ^[6]. A 2024 investor presentation reportedly claimed 5GW was already installed, which would put the current total near 10GW ^[8]. Ten gigawatts of AI datacentres implies roughly 6.4m GPUs ^[9]. The 2.2m in the documents is about a third of that, a gap of some 4.2m chips ^[2]. The Guardian describes the figure as less than half what some experts had assumed ^[4], and an analyst who specialises in Nvidia told it: "They're low to me" ^[18].

The sceptical version of the numbers does not close the gap either. Shaolei Ren, a professor at the University of California, Riverside, told the Guardian that Microsoft's sustainability reports, which carry electricity figures, point to AI capacity nearer 1.2GW in 2024 ^[10]; adding 5GW since then would still require around 4m chips ^[11]. The documented fleet is roughly 55% of that ^[3]. Ren's objection is definitional rather than accusatory: "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. They are giving insufficient context" ^[12]. He added that the sustainability reports are audited by a third party and "have more credibility than announcements" ^[13].

Nobody outside the companies can settle this. Nvidia, one of the two most valuable companies in the world, treats its supply chain as one of the industry's most tightly held secrets ^[17], and with almost no exceptions does not report how many AI chips it sells or to whom, while its customers do not disclose their holdings either ^[14]. The Guardian's reading of the discrepancy is the operationally useful one: the newest datacentres may not be fully operational, or may be running without the chips they need ^[15].

For an operator, the distinction is not academic. A gigawatt is a claim about land, power and buildings; a chip is a claim about the thing a model actually runs on. Microsoft has put roughly $280bn into land, buildings and computational infrastructure since 2022, including more than $41bn in the past quarter ^[5], and Satya Nadella said last year the global datacentre footprint would double by mid-2027 ^[16]. Five gigawatts is four times the size of the largest datacentre park in Europe ^[7], and none of it obliges anyone to give a named region a named accelerator on a named date. Roadmaps and contracts built on announced capacity are borrowing credit from an asset that may not yet be populated.

Watch for two things. Whether Microsoft or any hyperscaler starts defining what "added capacity" means, given Ren's point that the context is missing ^[12]. And whether the audited sustainability disclosures keep telling a smaller story than the investor decks ^[13][10] - that divergence, not the headline gigawatts, is the number worth tracking.