Picture a lender in New York or London, Evan Conrad said. An AI lab stops paying for its computing, the cloud company that rented it the machines then misses its own loan payments, and the lender ends up repossessing a supercomputer. "And the lender has no idea what to do with that supercomputer, not really, nor do they know what that's going to be worth in the future."
Conrad is co-founder and CEO of SF Compute, which operates GPU clusters for their owners and runs a market where compute contracts are bought and resold. GPUs are the chips that train and run AI models, and a cluster is many of them wired together into one large machine. In a highlights episode of The Cognitive Revolution, published October 24, 2026, he told hosts Nathan Labenz and Prakash Narayanan that the thing holding back AI's build-out is "not power, it's not chips, it's entirely credit."
Why lenders bet on the customer
Narayanan asked why a recent SF Compute offering came with three-to-five-year contracts. Conrad said the term comes from the lender. Because a lender can't value a repossessed cluster, let alone run it, it is left to bet almost entirely on the credit quality of the customer at the end of the chain: the AI lab.
That is a problem, in Conrad's account, because almost no AI lab has good credit. "They're all new companies," he said. "We're still trying to figure out if Anthropic and OpenAI are going to be around or which one is going to win." A two-person startup trying to found a new lab has even less to show. To get a cluster at all, he said, such a company has to pay a very large deposit and sign a long contract so that everyone is comfortable it will stick around.
The conversation turned to take-or-pay contracts, in which a customer commits to pay for capacity whether or not it uses it. Such a promise reassures a lender when a company like Meta or Google stands behind it, not a ten-person startup funded for 18 months. Conrad agreed. Lenders and clouds want an "investment-grade" customer, meaning one whose credit rating marks it as unlikely to default. But he said a provider is unlikely to land Meta or Google unless it has a tremendous amount of capacity and is willing to "basically not be a cloud." Clouds that want to sell more widely end up serving customers "a little bit less than investment-grade."
SF Compute's own deals show the kind of terms involved. On September 10, 2026, the company announced that it had sold an unnamed AI company $245 million of capacity on two dedicated NVIDIA B300 clusters, under two 36-month agreements.
The workaround: pay it all up front
For a lab that just wants to buy capacity, Conrad said, there is one way around weak credit: raise so much money that you can pay for the contract in cash instead of over time. "If you can just pay for a three-year contract entirely up front, like, it doesn't matter. Anyone will work with you."
That, he argued, is why AI startups raise "astronomical rounds" before they have revenue, a product or traction. He called the result survivorship bias. Many labs try to raise money, but the only ones that end up existing are those that secure extremely large rounds. Investors ask which lab or inference provider will win (inference is the work of running a trained model for users). "And the answer is the one that you fund," Conrad said.
What SF Compute wants to change
"We started SF Compute because they weren't going to fund us," Conrad said. The company's aim is to make a GPU cluster into collateral whose value a lender understands. Then the lender would worry less about the customer's credit, a startup would not need to pay the big deposit, and it would not need a giant funding round just to get machines.
Without that, Conrad said, AI will "look like TSMC," the dominant contract chipmaker. Anyone starting a new chip fabrication company today, he said, had better bring a brand new technology, because the capital spending required would run to something like $50 billion.
Narayanan asked the practical question: if a startup needs a thousand GPUs, how much less does it have to raise with SF Compute in the picture? "So today you're going to need to do roughly the same," Conrad said. Building up the value of the collateral will take several years, he explained, because lenders only get comfortable after seeing proof over a long period. He called that the reason almost nobody else attempts it at scale.
Keeping capacity for sale
SF Compute customers can resell capacity they have bought, matching buyers and sellers through an order book, a running list of bids and offers for specific hardware and time windows. Conrad described two ways the company keeps that market active. Some capacity has been sold specifically so that resale happens if GPU demand goes down. Other capacity carries a rule borrowed from economics.
He traced that rule to Henry George, the economist behind the land value tax, and to an offshoot of it known as a Harberger tax. SF Compute's version, as Conrad put it: "we'll sell you the capacity, but somebody else has to be able to buy some portion of that capacity from you." He acknowledged the approach is "very controversial," and said the point is to discourage hoarding, because if one customer hoards all the capacity, no one else can get it.
The company's research platform goes further, he said. Jobs wait in a queue, and customers can pay for priority but cannot buy all of it. A small researcher running an occasional small job gets moved to the front. SF Compute's announcement of the platform describes the same tilt: the queue favors customers with less capacity already in use, a customer's first node gets priority ahead of paid tiers, and priority does not reserve capacity.
Conrad said these rules keep SF Compute liquid. "Right now, basically every GPU cloud is sold out. San Francisco compute is not sold out," he said, because the company has "locked in our ability to not sell out for years."
The crunch he predicted
Narayanan reminded Conrad that on another podcast he had predicted a credit crunch within three months, and asked where it was. Conrad said he thought his timing was "almost exact," though he did not remember when he had made the prediction. As evidence, he pointed to debt at CoreWeave, a GPU cloud provider, getting more expensive, while allowing that "there's a lot of complexity in that one." He added that new customers are very hard to close unless they are investment grade, and that many customers are being asked to pay 100% of their contracts up front.
His forecast followed from that. For possibly the next year or two, he said, venture capitalists should be prepared to "write some really big checks" for their portfolio companies or encourage them to sell. He expects "a lot of consolidation of AI labs and inference companies." To get their next cluster and grow, he said, they will be asked to pay the entire contract up front, and many won't be able to. Some will go to investors for big rounds and be refused, because at some level of scale, he said, "they'll probably just not be able to stomach the risk." Others, in his view, will be forced to sell to a buyer with great credit that does not face the same demand.
Labenz called SF Compute's effort "quite a financial and GPU alchemy project" and said he hoped it works, because he would like a future that still has startups, one where "we don't collapse into like five big tech companies, you know, for all future time." Conrad replied: "That is also my concern."