The shortage of AI computing power has been an abstraction for most people: a line in an earnings call, a reason some chatbot felt slow. This week it turned concrete in three very different places at once, and the common thread is that silicon meant for everyone is being rerouted to whoever is training the largest models.
Start with the clearest signal. Hardware leakers report, and several enthusiast outlets have repeated, that Nvidia has stopped allocating its top Blackwell die, the GB202, to the GeForce RTX 5090, the fastest consumer graphics card on the market. All future supply of that chip would instead go to professional and data-centre cards. Nvidia has not confirmed it, so the claim should be read as credible chatter rather than fact. But the economics are not in dispute. The same wafer that becomes a 1,999-dollar gaming card can become a workstation accelerator that sells for many times more, and when supply is tight, the higher-margin bucket wins. Retailers are already behaving as if scarcity is coming: Micro Center has tied RTX 5090 sales to an ID check and a no-export pledge.
Second, look at how companies now buy the chips they cannot get. Industry researcher SemiAnalysis counted 323 Nvidia GPU providers in September, up from 209 less than eleven months earlier, and a separate tally puts the number of "neoclouds" above 300, a 55 percent jump in under a year. These are specialist renters that exist precisely because Amazon, Microsoft and Google cannot meet demand. Amazon's Andy Jassy has said plainly the company will not be able to serve all the demand it sees this year, or next. CoreWeave is essentially sold out. Into that gap steps everyone from Nebius to SpaceX, which now rents GPUs to Anthropic for 1.25 billion dollars a month. The spot price for a single Nvidia B200 has more than doubled since March.
Third, and strangest, the squeeze reached a laptop line. The Taiwanese contract manufacturer Quanta reportedly halted production of HP notebooks for about a week in September after the two sides could not agree on price, resuming only when HP executives flew in to renegotiate. The account is unconfirmed, and HP has not commented, but the logic is telling: servers for AI now make up as much as 80 percent of Quanta's revenue, so a stalled laptop contract is no longer worth absorbing. For two decades, PC brands could make their manufacturers eat rising costs, because those factories had nowhere better to go. AI gave them somewhere better to go.
None of these stories is enormous on its own, and two rest on leaks. Together they describe a single shift. The physical supply of computing, from the die on a gaming card to the factory floor to the cloud you rent by the hour, is being reorganised around the appetite of a handful of AI labs. Everyone else, consumers most of all, is quietly moving to the back of the queue.