# AI startups will die without a compute market _Evan Conrad told AI Engineer that AI startups face a credit wall that only a physically settled compute market can fix._ **Published:** 2026-10-22 **Source:** https://www.startuphub.ai/ai-news/technology/2026/ai-startups-will-die-without-a-compute-market --- Evan Conrad told [AI Engineer](https://www.youtube.com/watch?v=Wk2bUUGFC_o) that AI startups will not survive without a real market for compute, and his company San Francisco Compute is trying to build it as a physically settled forwards market, both the operator of the hardware and the order book at once. He calls it a grid for compute. Conrad said the idea came from trying to start a lab himself. His team first called the company Junlark and wanted to build something like Suno or Udio before those products existed. To do that realistically you needed a $10 million deposit on a cluster at the time, he said. They raised $500,000 instead and rented, then bought a cluster they could not afford to hold. They were underwater every month unless they subleased it. He walked NeurIPS with paper signs that said H100 and A100 at the same time because he made the poster too fast. That forced brokerage became the business. Today that entry ticket is not $10 million, he said, it is a billion dollars. That deposit is the symptom. Conrad argued the root is credit. Everyone is sold out, lead times are long, and lenders have no way to underwrite a two-person team that says it will build superintelligence. So the lender underwrites the customer, not the asset, and demands the startup's entire bank balance up front. Because CPU buyers were high-margin SaaS companies, clouds could sell on demand. If a SaaS cloud doubles prices you complain, Conrad said, if an AI cloud doubles prices you go bankrupt. AI buyers are price sensitive and will leave, so GPU clusters only get financed with long-term take-or-pay contracts, ideally with investment-grade credit. Without liquidity, a team that underbuys loses to a competitor and a team that overbuys ends up like Stability AI or [Character AI](https://www.startuphub.ai/ai-news/reviews/2026/character-ai-review-2026). Technical risk, financial risk and credit risk are the same bundled problem, and power is not the first wall. He said founders hit credit long before they hit power, and every inference company and new lab is bottlenecked by cash in the bank. The market fix has to be physical settlement, Conrad said, and that is why almost no one does it. Drawing on Dylan Page, co-chair of the CNCF infrastructure group and former Lambda infra, he said GPU compute is not standardized or reliable enough for the market operator to be divorced from the hardware. His analogy was a hotel. If you buy my hotel room for the night and want room service and cleaning, who do you call if I do not run the hotel. Brokering fails because you cannot fix the machine. But if you build the full operating stack, backplane fabric, bare-metal orchestration, VMs and services, you have built a GPU cloud. And a GPU cloud can sign billion-dollar deals, so the incentive is to stay a cloud and never become a market. That cold start is the trap. Conrad said San Francisco Compute only chose settlement because without it they could not start the lab they actually wanted. The alternative is consolidation, he warned. If lenders keep underwriting the customer, only a few big clouds and a few labs pass the filter and the market looks like TSMC, not a liquid exchange. If inference providers push the credit problem to apps instead, every token buyer needs Fortune 500 credit or a five-year prepaid contract, and startups that need $100 million in tokens to compete never get funded. The venture check that is supposed to build product ends up as a deposit check to a cloud. In a city where 823 SaaS companies headquartered in San Francisco are already ranked by software spend, per data tracked as of 2026 by Cledara, that concentration matters. If tokens become an inventory item you must prepay, distribution tilts to incumbents who already have credit, and the flywheel strengthens. Conrad said in three to five years the ante to start a company could be $10 billion before revenue or traction, which no seed round covers. Without a way to buy reserved capacity now or sell it later, the only startups left are the ones that did not need a market in the first place. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.