Why AI Startups Will Disappear Without Compute Markets — Evan Conrad, San Francisco Compute

Highlights

• San Francisco Compute operates a physically settled forwards market for GPU capacity to bypass traditional credit bottlenecks in AI infrastructure.

• Early AI labs face crippling financial risks because lenders cannot underwrite artificial intelligence startups without multi-year take-or-pay contracts and massive upfront cash deposits.

• The required capital deposit to start a frontier AI lab escalated from 10 million dollars to 1 billion dollars over a span of three years.

• Without liquid compute markets that underwrite the asset value of GPU clusters rather than customer credit risk, the AI industry faces extreme corporate consolidation similar to TSMC.

• GPU brokers fail because they lack control over the physical data center infrastructure, backplane fabric, and bare-metal orchestration necessary to resolve hardware failures and service requests.

Timeline

1. 00:12 - 03:32

   The Origin of San Francisco Compute

   Supercomputer operations require precise control over the backplane fabric, bare-metal orchestration, and virtual machine services. Founders attempting to build early AI labs discovered that financing clusters without massive capital reserves leads directly to bankruptcy. Struggling to stay solvent after buying an expensive cluster forced the team to pivot from building models to brokering hardware, exposing the structural failure points in GPU cluster management.

   • San Francisco Compute operates a physically settled forwards market combining cloud hardware operation with a live capacity order book.
   • Initial operations began under the name Joonlark as an early generative audio lab that raised 500,000 dollars instead of the standard 10 million dollar deposit.
   • Financial survival required subleasing an underwater cluster, inadvertently turning the founders into the industry's first GPU brokers.
2. 03:32 - 10:29

   Credit Risk and the Death of Startup Funding

   The current AI market suffers from severe compute shortages driven by lead times and stringent credit requirements. Because lenders lack frameworks to evaluate artificial intelligence startups, they demand entire bank balances as deposits or mandate long-term off-take agreements. This creates a high-risk environment where cluster deployments face constant delays, hardware failures lack rapid node replacements, and rigid contracts prevent companies from adjusting capacity to actual demand.

   • Lenders cannot underwrite superintelligence or assess the asset value of two-person startups without immense financial collateral.
   • AI companies face severe price sensitivity and low profit margins compared to traditional SaaS businesses, causing bankruptcies if cloud costs double.
   • Credit bottlenecks emerge long before power constraints, forcing venture capitalists to fund multi-million-dollar cash deposits just to secure cluster allocations.
3. 10:29 - 13:59

   Physical Settlement Versus Traditional Cloud Brokers

   Without intervention, the AI industry will consolidate into a handful of major cloud providers and model labs akin to TSMC. Solving this requires combining market liquidity with direct physical infrastructure management. Standard GPU brokering fails because entities that do not own and operate the physical hardware cannot resolve backend failures or operational bottlenecks for customers who need agile capacity shifts.

   • Compute markets must underwrite the value of the underlying GPU cluster rather than the credit risk of unproven startup customers.
   • Brokering compute fails because third-party brokers cannot manage physical hardware issues, room maintenance, or service requests like a dedicated hotel operator.
   • Building physical settlement requires constructing the entire operating capacity of a GPU cloud, which tempts operators to abandon markets in favor of direct cloud revenues.
4. 13:59 - 17:41

   Preserving the Startup Ecosystem Through Compute Markets

   As the financial barrier to entry scales toward billions of dollars, small startups face complete exclusion from the AI landscape. If inference layers become bottlenecked by credit, only massive incumbent firms will retain access to artificial intelligence. Establishing open compute markets protects independent innovators, enabling diverse applications in medicine and engineering to emerge instead of corporate consolidation.

   • Rising capital requirements mean launching a competitive AI company will soon demand hundreds of millions of dollars in upfront token purchases.
   • Passing credit risk down to the application layer forces startups to sign rigid five-year contracts with Fortune 500 customers.
   • Compute markets exist to ensure independent startups can access intelligence resources and build breakthrough technologies without being crushed by credit barriers.