For two years, the people building artificial intelligence have told everyone else to step aside for the future — the workers, the teachers, the doctors, the exam-takers, the whole towns that rezoned and waited. This week the future’s own engine room confessed what the sales pitch always left out: it cannot find enough wall sockets. Microsoft, one of the loudest believers in the AI boom and one of the richest companies ever to exist, is sitting on advanced chips it cannot plug in, and its chief executive has said so in public. The machine age is not being slowed by a smarter rival or a hostile regulator. It is being slowed by electricity, concrete, and the unglamorous work of building a room with a working plug.
A Guardian investigation published Monday put a number on the gap. Internal Microsoft documents put the company at about 2.2 million installed AI chips as of the middle of 2026. Microsoft has said it added five gigawatts of datacenter capacity over the past two years — a buildout that implies millions more chips than that. It targeted 1.8 million chips by the end of 2024. Nearly two years and roughly $280 billion later, the count has barely crawled to 2.2 million, and insiders told the Guardian the total has barely budged over the past year. Microsoft called the estimates inaccurate without saying which figure was wrong.
When a company answers a spreadsheet with a shrug, the spreadsheet is usually right.
Satya Nadella — the Man Who Bought the Chips and Ran Out of Walls
Satya Nadella did not hide from the problem; he named it on a podcast last year. The bottleneck, he said, was not getting the chips but finding powered buildings to put them in — he has chips sitting in inventory he cannot plug in, because what he lacks is a finished, wired building to slot them into. That is a remarkable thing for the head of a three-trillion-dollar company to admit, and it is the most useful sentence of the year for anyone who has been told the AI future is inevitable. The revolution’s constraint is not intelligence. It is drywall.
The smartest machine in history still needs a wall outlet, and the walls are behind schedule.
Shaolei Ren — the Professor Who Read the Fine Print Microsoft Hoped Nobody Would
Shaolei Ren, a professor at the University of California, Riverside, did the arithmetic the company’s own filings invite. Microsoft’s audited sustainability reports, he said, point to AI capacity in 2024 closer to 1.2 gigawatts — a fraction of the five it has implied — and even that lower figure would still require around four million chips. His logic is hard to argue with: those filings are checked by a third-party auditor, while a press release is checked by nobody. The gap shows up again at the top of the line. Nvidia’s chief executive said last spring that the four biggest customers had ordered 3.6 million of its newest Blackwell chips, and Microsoft has always been among them. At a quarter share, Microsoft should hold close to a million Blackwells; the documents suggest it has installed less than half of that.
Believe the auditor, not the announcement. Auditors don’t get paid by the share price.
Fairwater, Wisconsin — the Town That Was Promised the Future and Got 300 Megawatts
Here the story stops being about spreadsheets and starts being about people. Microsoft’s largest American AI project, a pair of datacenters called Fairwater in Wisconsin and Georgia, was announced as the kind of investment that rewires a region. In April, Nadella declared the Wisconsin site live. Satellite footage showed only part of it running, and by May he had conceded to a local newspaper that it was not online after all. The project was sold as multi-gigawatt and multibillion-dollar. Three years in, roughly 300 megawatts have been built. The electricians, the construction crews, the towns that rezoned and waited — they carry the cost of a promise made on a timeline nobody could keep. By eWeek’s account, the hyperscalers are now fighting over who gets the electricians.
A datacenter’s first promise is to a town, not to a server. Towns remember.
Mustafa Suleyman — the Winners Will Be Whoever Can Afford the Plug
The scarcity is already deciding who gets the good version of the future. Microsoft’s AI chief, Mustafa Suleyman, has been saying since March that the next chapter of AI will be won not by the smartest model but by whoever can afford the compute. Premium, low-latency inference goes to the high-margin products — the Copilots, the enterprise tools, the healthcare software — while cash-strapped startups and consumer apps get throttled. Translate that from boardroom English: the “intelligence for everyone” pitch is quietly becoming intelligence for whoever can pay the power bill. The scarce resource in the AI age turns out not to be brilliance. It is a transformer and a building permit.
The future is abundant, provided you can afford the meter.
The Rest of Us — We Were Told to Brace for a Machine That Can’t Get Its Lights On
The takeaway is not that Microsoft is doomed; it is that the ceiling on this entire boom is physical, and it is lower than the sales pitch. For two years the same industry has told workers to retrain, students to brace, patients to expect a machine at the bedside, and towns to hand over their land for the coming abundance. It turns out the engine behind all of those warnings cannot be wired fast enough to keep its own lights on. Microsoft is racing to widen the bottleneck — silicon of its own, AMD hardware in its racks, a natural-gas plant behind the meter in Texas — but chips were never the real shortage. The real shortage is the quiet, expensive, human work of building the places those chips live.
The genie is real. It is just waiting on a building permit and a transformer.
Sources: The Guardian, eWeek, The News International.