Ask the people who spend the largest compute budgets in AI what they worry about, and customers are not on the list. Nick Rubin, in a post on X on Tuesday, described a dinner he co-hosted with Liz Wessel for six people responsible for buying, deploying, and running computing capacity at major AI labs and infrastructure companies. He does not say who sat at the table, so the account rests on his word alone. It is still a rare look at how the buyers think.
The first finding is that nobody is short of buyers. Almost everyone present felt they could sell all the capacity they managed to bring online. Rubin writes that traditional capacity planning has lost its purpose, because demand forecasts can miss by 10x or more and supply caps growth regardless. A sharper forecast would change nothing. The only question is how much working compute can be switched on, and how fast.
That moves the contest to money. Rubin calls capital the biggest bottleneck, along with power. Customers will pay extra for capacity that exists now, but a provider that wants to build for them has to commit to a particular GPU generation a year or two out, while the chips, the customer, or both may change. Prepayments can be huge, and many of the customers are young companies. Without firm commitments, lenders hold back. Even the talk about solar came down to cash: batteries make cheap on-site power more reliable and also raise the upfront bill, so the option that repays its owner quickest can beat the better engineering.
The third point cuts against a common assumption about hardware. The operators told Rubin that older GPUs keep their usefulness and value longer than the pace of new releases would suggest. That is good news for anyone financing a cluster, since a lender needs to believe the machines will keep earning for years. Rubin adds a limit. Where power is scarce, swapping a working chip for a newer one that does more with the same electricity can pay off. But NVIDIA’s newer systems tie the whole rack together, so an upgrade turns into a project covering power and cooling, not a simple parts swap.
Running the machines is the fourth problem. Cabling, cooling, failed hardware, slow repairs, and local permits all stand between a plot of land and a working cluster. One operator said they would put their own staff on site at a provider’s facility to keep capacity running, since they wanted it that much.
The fifth is the public. One attendee watched a major project nearly stall at a city hearing, after every legal approval had come through, because residents turned out against it. Others worried that a new administration could change rules for projects they are funding today. Rubin reports that nobody had a plan for reversing the mood. The best hope was visible benefits, such as progress against disease, arriving soon enough.
Our read: this describes a business where the scarce skill is no longer winning customers. It is closing the gap between a signed contract and a humming cluster. Rubin’s own conclusion is that firms good at financing, power access, operations, and local relations will carry that edge into each new hardware generation. That is a claim from one host about one dinner, not a survey, and it comes without figures on budgets, prices, or utilization. Even so, it matches where big AI deals have been heading, toward money and electricity.
For startups that depend on rented GPUs, the practical consequence is that a provider’s balance sheet and its record of keeping clusters running now deserve as much scrutiny as the price per hour.
Reported by Nick Rubin, in a post on X, on 20 October 2026.
