Compute Markets: Why AI Startups Need Them

Compute markets are order books for GPU capacity where contracts settle with real hardware, and San Francisco Compute runs one alongside the clusters it operates. Evan Conrad explains the financing chain that makes them necessary and what breaks without them.

What compute markets actually are

Compute markets are order books for GPU capacity in which trades settle with real delivered hardware rather than cash. Evan Conrad, who runs San Francisco Compute, described his company at an AI Engineer talk as both the operator of the clusters and the market for them, which he compared to a grid for compute.

Physical settlement is the distinguishing feature. A buyer reserves cluster time for now or for a future date, and the seller is the party that racks, wires and repairs the machines. Conrad said the model resembles a hotel: if nobody operates the building, nobody can fix the room service.

Conrad said his team arrived at this business by accident. The company started life as an AI lab that raised $500,000, rented a large cluster, and went underwater every month. Subleasing that cluster kept it solvent, which made it, in his words, the first GPU broker.

Why AI startups hit a credit limit before a power limit

AI startups hit a credit limit before a power limit because lenders cannot value a young company's compute asset. Conrad stated that credit is the number one bottleneck for AI right now, not electricity, and that this constraint runs all the way up the ladder of company size.

The mechanism is straightforward. A cloud operator needs a long-term take-or-pay contract before financing a cluster, because the lender underwrites the customer's credit risk first. A two-person lab claiming it will build superintelligence gives a lender nothing to price.

Without that credit, the customer pays a large deposit instead. Conrad traced the minimum entry deposit from $10 million when his lab started, to $100 million, to roughly $1 billion by 2026. He said that trajectory cannot continue indefinitely.

The sold-out cluster problem starts with financing

Clusters are sold out because financing requires a signed customer before deployment, which pushes every new buyer to the back of a lead-time queue. Conrad argued that the order is backwards: the cluster should be built and landed first, then sold, which would remove most delay risk.

AI companies are more price sensitive than SaaS companies, and that changes the contract structure. If a SaaS business's cloud doubles its price, the company is angry but survives. If an inference company's cloud doubles its price, the company may go bankrupt, so it moves to whichever provider is cheaper.

Because customers will leave for a cheaper rate, on-demand GPU pricing cannot support cluster financing. That is why capacity is sold as long-term reservations, and why liquidity in those reservations matters when demand shifts suddenly.

Late hardware, broken nodes and demand that misses the mark

Every stage of the cluster lifecycle adds risk, and Conrad groups them into one bundled problem: technical, financial and credit risk treated as a single exposure. Late delivery is the first failure. Broken nodes are the second.

When hardware fails, a customer without spare liquidity has to return the node for replacement and wait through lead time again. A market with tradable capacity lets a buyer pull replacement capacity from someone else instead of restarting the procurement cycle.

Demand forecasting is the last failure. Conrad said AI margins leave little room for error, so a lab either overbuys and carries unused capacity or underbuys and loses the customer. Neither outcome is recoverable without a secondary market.

Why operating the hardware is part of the product

A market operator that does not run the hardware cannot guarantee delivery, so San Francisco Compute runs both. Conrad cited Dylan Page, a Kubernetes infrastructure veteran he said joined the company for this reason, and argued that brokering capacity alone does not work.

The incentive to skip the market is strong. Building physical settlement means building the full operating capacity of a GPU cloud, and once you have that, becoming a GPU cloud looks more profitable than running a marketplace. Conrad said almost everybody takes that path.

Consolidation risk: what happens if nobody builds the market

Without a functioning market, Conrad expects consolidation that resembles the chip industry: a few creditworthy buyers and a few large labs, with little room for new entrants. He framed this as a current trajectory rather than a certainty.

The alternative he described is risk pushed down to the application layer. If inference companies only sign investment-grade customers on five-year token contracts with large deposits, then access to tokens depends on credit, and tokens feed the whole economy.

Conrad's argument is that established firms already have credit, so they can buy tokens, ship products and improve their credit further. Startups without credit lose access to the same input, which narrows who can compete.

What a working market needs to get right

A compute market needs three things to function: underwritten asset value instead of customer credit, hard liquidity in reserves, and an operator who controls delivery. Conrad said the first two are what separate a real market from a listing board.

Underwriting the asset means valuing the GPU cluster itself rather than the startup renting it. That shifts the lender's question from whether a founder will succeed to what the hardware is worth over its useful life.

Liquidity means a reserved contract can be exited. Conrad tied this to demand spikes: when a company's product takes off, it needs capacity immediately, and a market without resale forces it to finance an entirely new cluster with full lead time.

FAQ

• What is a compute market? A compute market is an order book for GPU capacity in which contracts settle with real delivered hardware. San Francisco Compute runs one alongside the clusters it operates, so buyers can reserve capacity for now or a future date.

• Why are GPU clusters sold out? Clusters are sold out because lenders require a signed customer before financing deployment, which creates lead times for new buyers. Building hardware before selling it would reduce delay risk, according to Evan Conrad.

• Is power or credit the bigger bottleneck for AI companies? Credit is the binding constraint, Conrad said, because lenders cannot underwrite young AI companies and demand large deposits or long-term contracts instead. Power matters, but companies exhaust credit first.

• How much does it cost to start an AI lab? Conrad traced the minimum deposit from $10 million at his lab's founding, to $100 million, to roughly $1 billion by 2026. He said that trajectory cannot continue without excluding new entrants.

• What does physically settled mean for GPU contracts? Physically settled means the contract is fulfilled with delivered cluster time rather than a cash equivalent. The seller must operate the hardware to make that work.

• Why is brokering GPU capacity not enough? A broker cannot fix broken nodes, reschedule delivery or guarantee uptime. Conrad compared it to selling a hotel room without running the hotel, where guests have nobody to call when something fails.

• Does San Francisco Compute own its clusters? San Francisco Compute operates the GPU clusters and runs the market for capacity on them, which Conrad described as the reason physical settlement is possible.

• What happens to AI startups if compute markets do not develop? Conrad expects consolidation toward a few large labs and clouds, with new startups blocked by deposit and contract requirements. Access to tokens would then depend on credit rather than on product quality.

• Who is Evan Conrad? Evan Conrad is the founder of San Francisco Compute, an operator of GPU clusters and a market for reserved compute capacity. He spoke about compute markets at an AI Engineer event in October 2026.

• How does this relate to the Bootcamp do Dev Doido and Crazystack Typescript? The financing and infrastructure constraints described by Conrad matter to developers learning to build AI products, which is the kind of practical stack covered in the Bootcamp do Dev Doido and the Crazystack Typescript curriculum.

Turning a talk into written analysis

Conrad's argument rests on a chain of financing decisions that is easier to follow in writing than in an 18-minute talk. If you host conversations with operators, engineers or founders who explain how a system actually works, that reasoning can reach readers who will never watch the video.

Skala Blog turns a YouTube URL into a transcription and then an article draft, so the explanation your guest gave on camera becomes something searchable and quotable. The Bootcamp do Dev Doido and the Crazystack Typescript track cover similar ground for developers building on these platforms.

Evan Conrad's closing point was that the deposit ladder cannot keep rising without closing the door on new labs. A written record of that argument is one way to keep the door visible to the people who might walk through it.

Source video