SF Compute's Evan Conrad on building a programmable data center and an independent GPU compute market
Latent Space · youtube · 2026-10-20
Latent Space's swyx interviews SF Compute CEO Evan Conrad at AI Engineer New York about the compute market SF Compute is building—and why it needs an independent operator.
Key points:
- From selling a room to running the hotel: GPU compute trades like a commodity, but delivering it is messy—SF Compute had to build its own cloud operations to fulfill trades.
- What moves compute prices: interruptions and workload migration, storage, and hardware generations/SKUs; batch vs. live inference require different contract terms.
- Physical vs. cash settlement: trading a price disconnected from actual GPUs carries real risk; SF Compute emphasizes physical delivery plus order-book liquidity.
- Why NVIDIA can't build this: NVIDIA and big clouds have conflicting incentives as both suppliers and competitors, lacking the neutrality an independent market needs.
- Vision: turning data centers into programmable GPU clusters, plus the rise of sovereign compute.
More from Venture
- SpaceX seeks access to 80,000 US cell tower sites; Polymarket puts carrier odds at 29% — Polymarket · 2026-10-23
- Datadog offered nearly $4B for cloud security startup Upwind, scoop says — steph_palazzolo · 2026-10-23
- Investor: If AI is a bubble, it's the weirdest one ever seen — robleclerc · 2026-10-23
- Bindu Reddy: I'd trade my Meta stock for Anthropic at $1.5T or OpenAI at $1T — bindureddy · 2026-10-23
- A Year of Reddit Scraping + LLM Citations: Dead Site Goes From 1 to 50 Google Clicks/Day — doooyle · 2026-10-23
- Polymarket AI prediction crashes to 47% after weeks near 90%, $849K volume — patience_cave · 2026-10-23