The Programmable Data Center — Evan Conrad, SF Compute

Full analysis

Evan Conrad, CEO of SF Compute, wants to turn GPU time into a traded commodity, like oil or pork bellies, with contracts you can buy from one hour to five years out. The claim that should stop you: a single NVIDIA B300 chip earns about $40 an hour running inference at full tilt, but you can rent that same chip on the spot market for $5 to $6 an hour. One of those numbers is lying.

What's actually being decided here. Nothing you have to act on this week. This is a read on where GPU rental prices are heading, and whether the way you pay for compute is about to change shape. Easy to undo: you're not signing anything. But the arbitrage number is a signal worth pricing into any multi-year capacity plan. What sets the clock is that gap closing, and nobody controls when.

The Compute Pragmatist. That $40-versus-$5 gap is the whole episode. It cannot persist. Either inference prices fall hard as supply floods in, or spot GPU rental climbs toward what the chip can actually earn. My money is on inference prices falling, because that's where competition is savage and switching is a config change. If you sell AI features priced off today's cheap token costs, model for those costs dropping further, which is good for your margins and bad for anyone reselling raw inference. The people squeezed are the thin-margin inference resellers who bought capacity betting prices would hold. They won't.

The Skeptic. Only 5% of SF Compute's capacity is actually live on the order book. Customers who spend $100 recoup $25 reselling unused time. That's a thin market calling itself an exchange. Conrad warns that compute price indexes are manipulable like LIBOR was, where a dozen banks rigged a benchmark governing trillions. Fair warning, but his own order book is exactly the kind of thin, early market where a few big players set the price. Sovereign partners? "Announcements expected within weeks." Unnamed. I've heard "weeks" before. The physical-settlement discipline is genuinely smart. The liquidity to make it matter isn't here yet.

The Builder. The one shipped feature worth your attention: elastic InfiniBand partitioning. InfiniBand is the high-speed wiring that lets a cluster of GPUs act as one big machine. Until now, resizing that fabric meant emailing your cloud provider and waiting. SF Compute made it an API call. If you run multi-tenant inference or spin training jobs up and down, that's real: carve a cluster into pieces on the fly, hand them back when done. Everything else here is market structure you can't touch yet. This you could use, if you're at SF Compute's scale, which most readers aren't.

The Enterprise Buyer. The pitch to CFOs is the actual product: a liquid forward curve lets lenders finance GPU reservations against a real price, which lowers the cost of locking in multi-year capacity. That's the prize. But I can't sign a procurement contract against a benchmark that 5% of capacity trades on. Conrad's LIBOR warning cuts against his own sell: if you're using any third-party compute index for budgeting or vendor negotiation today, you're pricing off thin data that doesn't reliably track what you'd actually pay. Wait for depth before you treat any compute index as a number you can plan against.

Where they split. The Pragmatist sees a real signal in the arbitrage and expects inference prices to drop. The Skeptic sees a market too thin to trust any price it produces, including the one that makes the arbitrage look clean. Both can be right: the direction (inference gets cheaper) holds even if SF Compute's specific order book never gets liquid enough to matter. The second fault line is Conrad's own contradiction. He's the clearest voice warning that compute indexes can be rigged, and he's running the earliest, thinnest one. His defense is physical settlement, meaning you get actual compute, not a cash payout against a number. That discipline is the right instinct. It doesn't manufacture volume.

What it hinges on. One belief: does the $40-to-$5 gap close by inference prices falling, or by spot rental rising? If you buy AI by the token, you're on the winning side either way, because more supply and fiercer competition push the price you pay down. The thing to verify before you act on any of this: ignore SF Compute's order book until a lot more than 5% of capacity is trading on it. A forward curve nobody trades is a chart, not a market.

Prediction: The gap between B300 inference yield and B300 spot rental price will narrow mainly through falling inference prices, with at least one of the major inference providers (OpenAI, Anthropic, Google, or a large reseller) cutting published per-token prices on a frontier model by 25% or more before the next major model-pricing cycle around Q2 2027.

Confidence: Medium — direction is well-grounded; which number moves and the exact cut size are less certain.

Why: A B300 earning $40 an hour on inference while renting for $5 to $6 is a gap too large to hold, and the two forces closing it are not symmetric. Spot rental prices rise slowly because new GPU supply keeps arriving, while inference prices fall fast because token pricing is where providers compete hardest and switching costs a customer almost nothing. Every pricing cycle for the last two years has pushed per-token costs down as capacity caught up to demand, and more B300-class supply is landing now. The opposite outcome, spot rental climbing toward $40, would require GPU scarcity to worsen sharply, which runs against the supply build-out everyone is financing.

Revisit by 2027-06-30: We're right if a major inference provider cuts published per-token pricing on a frontier model by 25% or more. We're wrong if no such cut lands and spot GPU rental instead rises above $15/hr for B300-class capacity.