Evan Conrad explains the case for compute markets in this 18-minute AI Engineer talk, drawing on San Francisco Compute’s origins as an AI lab.

Why the lab became a broker

  • His team began with a music-model idea and about $500,000 raised, then needed to sublease cluster capacity to stay afloat.
  • The resulting business combines cluster operations with a live order book for present and future capacity.
  • “Physical settlement” means delivering usable computing capacity, backed by control of the machines and their supporting systems.

Five problems share a risk bottleneck

  • Sold-out capacity pushes customers toward new clusters with delivery lead times.
  • Lenders want long-term purchase commitments and credible customers before financing the hardware.
  • New labs lack credit histories, so deposits consume cash they could have spent building products.
  • Delays and broken nodes compound the problem when replacement capacity is unavailable.
  • Uncertain demand leaves buyers overcommitted or unable to obtain enough capacity when growth arrives.

A market needs an operator

  • Conrad argues that financing should account for the cluster’s resale or rental value across buyers, reducing dependence on one startup’s credit.
  • Buyers need a way to resell surplus reservations and obtain more capacity without commissioning another cluster.
  • Hardware, networks, orchestration and support must work consistently enough for capacity commitments to be credible.
  • His hotel analogy: selling someone a room reservation does little good if you cannot fix the room or provide service.

The incentive problem and the forecast

  • Building reliable delivery also creates a GPU-cloud business, with lucrative contracts that can pull operators away from running a market.
  • Conrad describes frontier-lab deposits rising from $10 million to $100 million and then $1 billion; those are his examples, not universal startup requirements.
  • He warns that inference providers could pass financing risk downstream through deposits and long-term token contracts, favoring established firms.

Caveat: Conrad sells the infrastructure he advocates. His claims about credit outranking power and startups disappearing are an operator’s thesis and conditional forecast, not established outcomes for every AI business.

“There is no rule that says that startups have to exist.” — Conrad, 15:23