We scan new podcasts and send you the top 5 insights daily.
Evan Conrad says two VMs from one operator can differ by 40% in kernel-launch performance, and storage, WAN transit, preemption and chip generation also change value. Rather than wait for standardized contracts, SF Compute lets buyers and sellers attach multiple properties to orders, accepting wider spreads and arbitrage for flexible buyers.
Evan Conrad describes the order book as a market for breaking leases. Batch inference can accept interruption in minutes; live inference customers may need months of notice and face SLA penalties. Those interruption rights price differently, similar to power markets' day-ahead and live markets. The terms are the product.
Unlike oil, GPU compute is not a simple commodity. Its value is highly dependent on the specific software and workload being run, making it difficult to standardize and treat as a fungible asset. This presents a major obstacle to creating a liquid, tradable financial market for compute power.
The head of AI at Hudson River Trading highlights a practical barrier to creating a financial market for compute. For serious training, the minimum "lot size" is thousands of GPUs, not a small, fungible unit. This makes it difficult to standardize a contract and create liquidity, unlike commodities with smaller, interchangeable units.
Inference costs fluctuate based on regional electricity rates and idle GPU capacity across distributed providers. Smart order routing architectures that run prompt-level mini-auctions can undercut static aggregators by allowing compute providers to continuously adjust prices based on marginal cost.
SF Compute argues that for a true compute market to exist, it needs an independent operator that handles physical settlement. This means running the data centers to standardize quality and earn Wall Street's trust, much like an electrical grid operator.
A futures market for GPU compute is not viable yet because the product isn't fungible. The performance of an identical H100 chip varies significantly between cloud providers based on their proprietary software stack and operational excellence, measured by metrics like "goodput" and "MFUs."
SF Compute's order book has small volume but concentrated liquidity. Evan Conrad tracks whether a customer who spends $100 buying capacity can resell about $25 on average, and measures liquidity by whether it solves the commercial customer's problem rather than by counting makers. He prefers that customer-resale metric to high-minded maker counts.
Creating a true commodity market for AI compute is highly challenging because compute is not uniform. Variations in chips, data center design, and specific customer needs make it difficult to establish a standardized physical delivery product or a trusted index, which are prerequisites for a liquid, tradable market.
Evan Conrad says early H100s and GB200s broke or had firmware bugs, so sophisticated buyers may pay more for units deployed after a certain date to avoid first shipments. That cohort risk becomes another tradable property in SF Compute's order book, alongside chip generation and storage.
Evan Conrad argues financializing compute before physical settlement creates a market that is easy to manipulate and carries basis risk: the hedge price may not match actual deliverable capacity. SF Compute co-designed its order book with physical constraints; Conrad says cash settlement is good only if derived from physical settlement.