The End of Startups

Author: Evan Conrad (co-founder, SF Compute) URL: https://evanjconrad.com/posts/end-of-startups Subtitle: The coming credit crunch One-line: The world’s compute is sold out, new capacity is financed against the credit of the buyer, and startups are not investment grade — so within six months most AI startups will not be able to buy compute unless the GPU hour itself becomes bankable collateral.

A short post (about 550 words), published October 22, 2026. It is also, openly, a pitch for SF Compute: the last two paragraphs describe the company’s product as the fix and close with a hiring line.

Key claims

• Compute is sold out, and new capacity is gated by credit. Getting capacity means signing up for a new cluster, and clouds and their lenders are increasingly unwilling to bring capacity online for tenants who are not investment grade.
• Startups are being asked to prepay. Non-investment-grade buyers are asked for larger and larger upfront payments; Conrad has seen it go as high as 100% of the contract up front, because “neither the lender nor the cloud could take on any additional credit risk.”
• The investment-grade layer is thinning too. The few investment-grade tenants are being downgraded as they take on debt, and they in turn sell tokens to buyers who are themselves not investment grade. Conrad calls this “the credit bottleneck.”
• Inference margins are what keep the chain creditworthy today, and they won’t last. A B300 rents for about $6/hr and earns about $30/hr selling inference. Conrad expects that margin to be arbitraged away and GPU prices to rise, which squeezes inference margins, worsens ratings, and pushes inference providers to demand investment-grade customers or large deposits.
• The result is a structural advantage for whoever already has credit, in any industry. Credit buys tokens, tokens build better products faster, and better products earn better credit. Application-layer competition demands spending tokens fast: “a Fortune 500 that can deploy $1b in tokens in a few days has an overwhelming advantage over a startup that first must raise $1b prior to their product existing.”
• The fix: make the GPU hour, not the cluster, the collateral. Today lenders underwrite only the offtaker’s (the AI lab’s) credit. A physically settled forwards market for compute lets a lender hold a fungible GPU hour on a liquid book instead of a non-fungible cluster that needs a servicer “that might blow up right at the moment they need it the most.” If the borrower fails, the lender has a real path to selling the collateral, which is what SF Compute is building.

Notable quotes

Within the next 6 months, most AI startups will not be able to buy compute. The world’s compute is sold out.

Startups are not investment grade. So, to get capacity, they are asked to pay larger and larger upfront payments.

When that happens, it becomes structurally harder to build a startup, in any industry. If you have credit, you get tokens, and if you get tokens, you build better products faster, make more money, and get better credit.

To solve this problem, we must get lenders comfortable with the GPU cluster as valuable collateral.

If a lender gets a GPU hour as collateral, the credit risk of the customer matters less; if the two person AI lab in a San Francisco Victorian fails to make their payments, the lender has a real path to selling the collateral. That makes startups possible at all, a path that seems increasingly difficult in AI.

How it connects

• Evan Conrad and SF Compute: the author and the company whose product the post argues for. Read it as an argument from an interested party.
• Compute Scarcity / Compute: the post’s premise is that compute is now allocated by balance sheet, not by price alone.
• Computer Chips Are the Most Magnificent Things Humanity Has Ever Made: the supply side of the same scarcity, seen from TSMC and ASML.
• Most Improbable Agent of You: Kevin Kelly’s opposite bet, published the day before, that falling cost per capability will push frontier-class AI onto devices individuals own.