TL;DR
Internal documents seen by the Guardian show Microsoft running 2.2m AI chips, roughly two years after it reportedly aimed for 1.8m by the close of 2024. Set against the data centre capacity the company says it has added, that figure is far too low. Microsoft says the calculations rest on incorrect information but has not specified which figures are wrong.
The arithmetic that does not reconcile
Microsoft’s own filings indicate 5GW of capacity added over two years, on top of what an investor presentation reportedly claimed was already 5GW installed in 2024. Ten gigawatts of AI data centres would imply something nearer 6.4m graphics processors.
Shaolei Ren of the University of California, Riverside reads the picture differently from the sustainability reports, which are separately published and third-party audited. Those suggest capacity closer to 1.2GW in 2024 — and even on that conservative basis, adding 5GW since would call for around 4m chips. His objection is to the vagueness: it is unclear what the company means by added capacity, and announcements carry less weight than audited figures.
An analyst covering Nvidia put it plainly — the numbers are lower than they expected from Microsoft. Sources inside the company say the chip total has barely shifted in a year.
Announced is not the same as operational
Fairwater, the paired Wisconsin and Georgia development that is Microsoft’s largest US AI project, illustrates the gap. Nadella said in April that the Wisconsin site was going live; satellite imagery analysed by Epoch AI suggests only part of it runs, and in May the company conceded to a local paper that it was not yet online. Securing a gigawatt of power on paper within a quarter is achievable, Ren notes. Bringing it into service in the same quarter is not.
The OpenAI partnership may account for some deployments outside the documents reviewed. Blackwell holdings look thin too: Jensen Huang cited 3.6m units ordered by Nvidia’s top four customers last March, which on Microsoft’s historic share should mean close to 1m — it has under half that installed.
What UK buyers should take from this
Nvidia discloses neither volumes nor customers, and its customers disclose nothing either. UK organisations planning capacity on the assumption that hyperscaler expansion announcements translate into available compute are relying on figures nobody can audit. Announced capacity, contracted capacity and running capacity are three different things.
Looking forward
If the biggest buyer in the market is this far behind its own briefing, the constraint on enterprise AI is physical build-out rather than model capability — and that constraint will surface as queue times and regional availability long before it appears in any earnings call.