If you read the financial press this week, the AI infrastructure boom looks unstoppable. One analysis makes the case for ASML, the Dutch company whose lithography machines every advanced chip depends on, citing forecasts that annual AI infrastructure spending climbs from 800 billion dollars this year toward 1.8 trillion by 2050. Others tout the "picks and shovels" plays: Vertiv and Eaton for the power and cooling data centres suddenly need, a roster of hardware suppliers with "direct data center exposure," even Google selling its custom chips to South Korea. The story sells itself. Demand for compute is insatiable, so buy the companies that feed it.
A far less comfortable account ran the same week, and it is worth taking seriously precisely because it attacks the premise everyone else is building on. In a lengthy investigation, the writer Ed Zitron argues that the "insatiable demand" narrative is largely an illusion, and that a huge share of the GPUs Nvidia has sold are not running in data centres at all. They are, he contends, sitting in warehouses or in buildings that have no power.
The load-bearing example is Microsoft. The company has repeatedly told investors it added roughly a gigawatt of capacity per quarter, language that reads as AI capacity. Yet Bloomberg reported that of Microsoft's roughly 12 gigawatts, only about 2 gigawatts is AI-specific. Zitron estimates Microsoft has spent around 265 billion dollars in capital since 2022 but has only about 50 billion dollars of GPUs actually in service. Chief executive Satya Nadella has admitted to "a bunch of chips sitting in inventory" he cannot plug in. If the most experienced hyperscaler on earth is struggling to energize what it bought, the argument goes, everyone is.
Scale that up and the numbers get vertiginous. Across the hyperscalers and the newer "neoclouds," Zitron counts more than 374 billion dollars in construction still in progress, and estimates that over 200 billion dollars of purchased GPUs are yet to be installed. Nvidia and Broadcom, on his math, have sold something like 561 billion dollars of AI chips since 2023, which would imply that roughly half of it has never been switched on. The chips were sold. Whether they are producing anything is another matter entirely.
The mechanism he points to is a distortion at the demand end. Two companies, OpenAI and Anthropic, account for the bulk of hyperscaler AI revenue and have signed compute commitments running past a trillion dollars. Their willingness to sign is what fills the backlogs that everyone cites as proof of demand. Strip out those two contracts, Zitron argues, and the genuinely diverse, paying market for AI compute is small. That would make Nvidia's revenue less a thermometer of broad demand than a record of a few very large bets placed years ahead of the buildings meant to house them.
None of this is settled, and it comes from a committed skeptic rather than a neutral referee. The chips could yet find workloads; the labs could keep growing into the capacity; the warehouses could empty. But the value of the piece is the question it forces. If public capacity figures are, as he puts it, weasel-worded to the point of uselessness, then investors and reporters have been reading GPU sales as a proxy for compute demand when the two may have drifted far apart. The optimistic case for the whole sector rests on that link holding. The uncomfortable possibility is that it already snapped, and the receipts are stacked in a warehouse somewhere waiting for power.