San Francisco Compute's Evan Conrad wants to build a utility-grade compute market — starting with the biggest supercomputer ever

San Francisco Compute wants to be the grid operator for AI compute

Evan Conrad's pitch starts with an analogy that does real work. Before utilities existed, every factory ran its own generator, sized for peak demand and burning money during idle hours. Electrification scaled when utilities pooled demand across factories and standardized the supply side. Conrad argues compute is at that same pre-grid moment, and San Francisco Compute is positioning itself as the operator that builds the grid.

Why existing compute markets fall short

Most players entering the compute market today are trying to create index prices with cash-settled futures on top. Conrad's objection is structural: an index disconnected from actual hardware means you're trading a synthetic number, not compute. His analogy is blunt — it's like trading a coal plant for a nuclear plant and calling it a commodity market.

The deeper problem is that you don't yet have enough qualified operators to make a real market. Capital, land, and power are plentiful. Supercomputing expertise is not. A handful of operators like CoreWeave and Nebius are genuinely capable, but the overall floor is too low to support meaningful standardization.

“Right now, compute kind of looks like the early days before there was a power grid — every factory ran its own generator. Our view is that you need a market operator that has to be independent. That market operator also needs to do what's called physical settlement, meaning running the whole damn thing, all the way down to the data center itself. We're shooting for the biggest computer ever.”

The San Francisco Compute model

Conrad's solution has two parts. First, an independent market operator — independence being the prerequisite for Wall Street to trust the infrastructure. Second, physical settlement, meaning the operator doesn't just clear contracts but actually runs the compute, down to the data center level. In Conrad's framing, you can't standardize a market if you don't control the technology underneath it.

The hyperscalers aren't the target or the enemy. He frames them as the factories and the neo-clouds as their generators — useful relationships, but not a market. What's missing is a utility layer above them that lets new capital enter, derisk compute investments, and price GPU capacity forward. San Francisco Compute wants to be that layer, helping new operators become clouds and then placing them on the market.

Where the bubble actually sits

On the overbuild question, Conrad offers a specific mechanism. Clouds won't deploy a GPU cluster until they have a signed customer contract, and they require a large upfront deposit to pass through to the OEM and then to Nvidia. The AI lab paying that deposit doesn't have the capital yet, so it raises at a high valuation. VCs, scarred by missing OpenAI and Anthropic, accept the valuation. The risk, Conrad argues, lands with the venture capitalists — not distributed across a functioning market.

San Francisco Compute's solution is to make future GPU pricing knowable, which removes the need for the large upfront deposit, which in turn reduces pressure on AI labs to raise at stretched valuations. Conrad is essentially arguing that a liquid forward market for compute would compress the valuation inflation currently baked into the VC stack.

Build status

Conrad confirms the company is targeting "the biggest computer ever" and says a data center build-out is in progress, though specifics aren't disclosed. He describes San Francisco Compute as a "supercomputer developer" that helps capital-rich operators become neo-clouds, then finances them through the market it's building.

The bet is that every dollar of capital currently trying to convert itself into GPUs needs infrastructure to do so competently. If Conrad is right that expertise is the bottleneck, a credible operator-developer hybrid that can raise the floor across the market is the right place to be.