CharlieKey takeaways from your favorite podcasts
← 20 October 2026Latent Space

The Programmable Data Center — Evan Conrad, SF Compute

Evan Conrad argues compute markets only work if built on physical settlement, with SF Compute operating the full cloud stack as a neutral party so financing and trading can sit on top of it.

Watch on YouTube ↗

Charlie's summary

  1. Physical settlement first: you must operate the whole cloud before a compute market can work

    SF Compute began by subleasing a cluster. Like a hotel seller pulled into room service and security, it ended up running the whole hotel. Evan says the company spent years building a cloud's operating capacity so it can deliver compute reliably, which it calls physical settlement. It co-designed settlement and the order book so the real constraints of compute, such as interruptions with fast-recovery storage, are built into orders.

    ▶ 2:28
  2. Don't standardize contracts early: a constraint-based order book creates liquidity while the technology shifts

    Evan thinks standardized contracts are still far off. Operators can offer virtual machines with a 40% performance gap in GPU kernel launches. SF Compute instead lets buyers attach properties to an order, such as time range and storage, and matches them against sells. The cost is wider bid-ask spreads. The upside is arbitrage for buyers who can use many kinds of capacity. Evan's liquidity test: a buyer spending $100 gets back about $25 by reselling.

    ▶ 6:32
  3. Compute has a day-ahead market and a live market, like power, and interruption is priced by workload

    Evan compares compute to power markets, which have a day-ahead market and a live market with different price dynamics. Capacity starting later can be traded easily, while selling capacity in use means interrupting a workload. Batch inference can be cut off with a minute's notice. Live inference needs months, because breaking an SLA is costly. So the two are priced differently, and the order book effectively becomes a market for breaking leases.

    ▶ 15:45
  4. The forward curve is a hump, and inference revenue sits far above rental price

    Evan sees a hump: unsold near-term capacity drops in price like expiring milk, then prices rise, then later capacity is worth less because buyers pay for short lead time. The anomaly is that a B300 can earn about $40 an hour on inference, or about $25 at 50% utilization, while renting for $5 to $6 an hour. “One of these numbers has to change.” Some buyers will pay more to lock in future supply.

    ▶ 18:20

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.