SF Compute CEO Evan Conrad on physically settled GPU markets and programmable data centers
Article timeline
The articles in this story. Times are ET.
Latent.SpaceXLatent Space interviews SF Compute CEO on programmable GPU data centersAILatent Space's post promotes a conversation between swyx and Evan Conrad, CEO of SF Compute, about turning data centers into programmable infrastructure. The post compares the idea to Nicira's virtualized networking, which VMware bought for $1.3 billion after Cisco offered $600 million. The conversation covers GPU interruptions, hardware generations, the GPU order book, physical settlement before compute derivatives, and programmable InfiniBand.
- Latent SpaceBlogSF Compute explains how physical settlement makes GPU markets work
AIEvan Conrad, CEO of SF Compute, says the company has built the operating capacity of a cloud so buyers receive usable compute rather than a cash payment against an index. He says about 5% of total capacity now trades on its order book, and that the order book matches orders with constraints such as interruptibility, GPU generation and networking. Conrad argues that cash-settled compute contracts should derive from physical settlement, and that companies with large hardware or inference businesses cannot run a neutral exchange.