# Why AI Startups Will Disappear Without Compute Markets

Evan Conrad, San Francisco Compute | AI Engineer NYC 2026 | 17:49

Source: https://www.youtube.com/watch?v=Wk2bUUGFC_o
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/why-ai-startups-will-disappear-without-compute-markets
Published: 2026-10-22
Tags: finance, gpus, startups

## TL;DR
- Compute companies face one bundled risk problem that includes technical failures, financing, customer credit, delays, and uncertain demand.
- AI companies hit credit limits before power limits because GPU clusters require large deposits and long-term contracts.
- A physical compute market could let companies buy and sell capacity while keeping new AI labs from being shut out by rising deposits.

## Summary
Evan Conrad explains why San Francisco Compute operates GPU clusters while also running a live market for current and future capacity. He began with an AI lab that needed a $10 million deposit, raised only $500,000, and had to sublease its cluster to avoid bankruptcy. That experience exposed a connected set of problems: GPU capacity is sold out, deposits can consume a startup's cash, clusters arrive late or fail, and demand is hard to predict. Conrad argues that these are different forms of one risk problem. AI companies reach credit limits before power limits, which pushes the industry toward large clouds and established companies with strong balance sheets. San Francisco Compute's proposed alternative is physical settlement, where the market operator also runs the hardware. Conrad says this is difficult because operating the infrastructure can be more profitable than running the market, but without a functioning compute market, he expects deposits and contracts to become large enough to prevent new AI startups from forming.

## Key ideas
### A compute market must operate the hardware it sells
[00:01](https://www.youtube.com/watch?v=Wk2bUUGFC_o&t=1s)
San Francisco Compute defines its market as physically settled because it operates the GPU hardware behind the orders. Conrad compares this with a cloud provider, which both owns or controls the infrastructure and delivers the service. The company handles several layers of the system, including the supercomputer's design and bill of materials, the backplane fabric, bare-metal orchestration, virtual machines, and services above them. He says data center owners often have the physical space needed for a GPU cloud but lack the supercomputing expertise to run it well. The company offers this infrastructure to prospective clouds, venture firms with portfolio clusters, and AI companies that want residual value from their compute without operating a supercomputer.

### The company became a GPU broker because its AI lab could not carry the cluster
[02:09](https://www.youtube.com/watch?v=Wk2bUUGFC_o&t=129s)
San Francisco Compute began as Junlark, an AI company that wanted to build a product like Suno or Udio before those companies existed. Conrad says starting a lab then required a deposit of about $10 million, so the team raised $500,000 and bought a cluster instead. The cluster cost more to run than the company could support each month. Subleasing it was necessary to avoid bankruptcy, which forced the team into becoming a GPU broker. Conrad says the deposit has since grown from $10 million to $1 billion. The company's market therefore came from a direct financial problem, rather than from an abstract interest in exchange design.

### GPU startups face two separate build problems at once
[03:58](https://www.youtube.com/watch?v=Wk2bUUGFC_o&t=238s)
Conrad describes a compute market as having to solve two difficult problems together. It must create a market with a cold-start problem, and it must build a cloud without simply becoming another cloud provider. The market has to provide capacity even though buyers and sellers are not yet active. The infrastructure also has to work reliably across the hardware stack. Conrad connects this to the immediate situation for AI startups: compute is sold out, new clusters have lead times, and companies have no practical way to bring future capacity online now. A startup can spend its entire bank balance on a deposit before it has a working cluster.

### Technical, financial, and credit failures are one risk problem
[05:04](https://www.youtube.com/watch?v=Wk2bUUGFC_o&t=304s)
The cluster may arrive late because the lender wants to know the customer before financing deployment. That makes the moment when the customer signs the riskiest point in the operation, even though buying the cluster first would reduce delays. Hardware can also fail after delivery. If the operator has no spare liquidity or replacement nodes, a failed node must be returned and replaced through an RMA process, restarting the lead-time problem. Demand creates another failure mode. An AI company can overbuy and carry unused capacity, or underbuy and lose the market to a competitor. Conrad groups technical risk, financial risk, credit risk, and demand risk into one bundled problem.

### Credit is the bottleneck before power
[08:14](https://www.youtube.com/watch?v=Wk2bUUGFC_o&t=494s)
Conrad argues that AI companies reach credit limits before they reach power limits. Lenders do not know how to underwrite a two-person company claiming it will build superintelligence, so they do not know what the company's future asset is worth. GPU clusters therefore need long-term take-pay contracts, ideally with investment-grade customers. Those contracts create another problem: a customer that cannot exit a reservation cannot free capacity for a company whose product is suddenly growing. A new cluster then requires another deposit and another physical lead time. Venture capital intended for models or applications ends up financing the deposit sent to the cloud. The cloud uses that deposit to underwrite the customer.

### Liquidity would let compute move from shrinking demand to growing demand
[10:08](https://www.youtube.com/watch?v=Wk2bUUGFC_o&t=608s)
When a company reserves capacity and its demand falls, Conrad says it has no way to sell that capacity to another buyer. When demand rises, it has no way to buy more quickly. A compute market would need to address both sides by underwriting the value of the GPU cluster instead of only underwriting the customer's credit. Without that liquidity, the industry can consolidate around a small number of creditworthy clouds and model labs. Conrad compares this possible outcome to TSMC, with one buyer or a few buyers controlling access while other startups disappear. He presents the market as a response to specific financing and allocation failures.

### Physical settlement is necessary because GPU infrastructure is not standardized enough
[11:37](https://www.youtube.com/watch?v=Wk2bUUGFC_o&t=697s)
Conrad says the market operator cannot be separated from the person operating the cluster because GPU compute is not reliable or standardized enough. He compares brokering a cluster to selling a hotel room to someone else. The new occupant still needs room service, cleaning, and help with problems, but the broker cannot fix those issues if it does not operate the hotel. The same applies to GPU capacity. An operator must be able to handle failures, replace nodes, and manage the service being sold. Conrad identifies Dylan Page, a CNCF infrastructure group co-chair and former Lambda infrastructure employee, as someone who joined San Francisco Compute for this reason.

### The economics encourage compute markets to become clouds
[13:04](https://www.youtube.com/watch?v=Wk2bUUGFC_o&t=784s)
Conrad says there is a strong reason companies avoid physical settlement. Building the operating capacity for a compute market also means building most of a GPU cloud. If the cloud business can sign billion-dollar deals, the company has an incentive to abandon the market and sell capacity directly. San Francisco Compute chose physical settlement because Conrad believed the company's original goal of starting an AI lab would eventually be impossible without it. He presents this as an uncomfortable commercial choice. The infrastructure needed to make a market can be more profitable when used to become the concentrated provider that the market was meant to avoid.

### Rising deposits could eliminate new AI startups
[13:56](https://www.youtube.com/watch?v=Wk2bUUGFC_o&t=836s)
Conrad traces the entry price for a new lab from a $10 million deposit to $100 million and then to $1 billion. He says this cannot continue indefinitely because a new company cannot raise a $10 billion round before it has revenue, traction, or a product. If inference companies pass their credit problem to the application layer, they may accept only investment-grade customers, five-year token contracts, and large upfront deposits. That would make intelligence available mainly to established companies with credit. Conrad expects a future startup might need to commit $100 million to tokens just to compete with existing firms. Without a market that lowers this barrier, he expects consolidation and fewer new labs.

## Notable quotes
- "We make a physically settled order, physically settled forwards market." (00:01)
- "They're all one big bundled risk problem." (06:12)
- "It's not power. You hit credit way before you hit power." (08:19)
- "You have to operate the cluster like the hotel. You have to run the whole damn thing." (12:47)
- "If you don't fix this bottleneck, there aren't any startups." (15:35)

## Tools & references mentioned
- San Francisco Compute
- Junlark
- Suno
- Udio
- Nvidia
- Instinct
- Stability AI
- Character AI
- TSMC
- Dylan Page
- CNCF
- Kubernetes
- Lambda
- H100
- A100

## Who should watch
- You are building an AI lab or inference company and need to understand why GPU access requires large deposits and long contracts.
- You operate GPU infrastructure and are deciding whether to sell capacity directly, broker it, or build a market around it.
- You invest in AI startups and want to see how compute financing can push the industry toward established companies.

## Related talks

- [Why We Don't Need More Data Centers](https://aietalks.com/talks/why-we-dont-need-more-data-centers) (Dr. Jasper Zhang, Hyperbolic, 14:18)
- [Why Compute Still Trades in Group Chats](https://aietalks.com/talks/why-compute-still-trades-in-group-chats) (Affaan Mustafa & Alejandro Soumah Fernandez, Jump Trading, 20:37)
- [The Geopolitics of AI Infrastructure](https://aietalks.com/talks/the-geopolitics-of-ai-infrastructure) (Dylan Patel, SemiAnalysis, 18:29)
- [How to Build Your Own AI Data Center in 2025](https://aietalks.com/talks/how-to-build-your-own-ai-data-center-in-2025) (Paul Gilbert, Arista Networks, 23:00)
- [The Desktop Frontier](https://aietalks.com/talks/the-desktop-frontier) (Ahmad Osman, Osmantic, 18:02)
