The Programmable Data Center — Evan Conrad, SF Compute
At AI Engineer New York, on the sidelines of a new Compute Trading track, Latent Space recorded a conversation with Evan Conrad, CEO of SF Compute. The company, which began life as Junark trying to sublease a cluster, now presents itself as a compute market with what Conrad calls physical settlement: buyers receive usable machines rather than a cash payout tied to an index. The interview traces how the company got there and how far the market has actually progressed.
Conrad frames the expansion as a hotel problem. A customer asks for a room, then housekeeping, then room service, then security, and each request drags the operator deeper into the building until it is running the whole hotel. SF Compute's equivalent was discovering that delivering a GPU contract meant becoming something close to a cloud with access to the data center. Over the past several years the company has built out that operating capacity so it can hand over compute at a reliable service level, which is the substance behind the physical settlement label.
The term is borrowed from futures markets, where physical settlement means the barrel or the bushel is actually delivered. Conrad points to the delivery machinery that makes that possible in other commodities, and to the oddities that appear when delivery is real: crude futures famously went negative in 2020 when holders had nowhere to put the oil, and power order books carry constraints tied to how long a boiler takes to cool. Compute, in his account, inherits the same class of problems, so SF Compute designed its order book and its settlement process together rather than treating matching as a separate layer.
That shapes what can be expressed in an order. A buyer can state whether a workload may be interrupted and how quickly it needs to recover from storage, alongside the GPU generation, networking and software configuration required. Those attributes move both the contract terms and the price. Conrad is candid that the market is not yet mature, deep or liquid by commodity standards. Only about 5% of SF Compute's total capacity currently sits on the order book, with more being moved across steadily and new capacity arriving already book-ready.
He distinguishes volume from liquidity. The traded volume is small, but the liquidity within that small pool is concentrated. His customer-facing metric: for every $100 a customer spends buying capacity, SF Compute returns about $25 on average by reselling it to someone else. That resale figure is the number the company watches to judge whether the market solves a real commercial problem, rather than counting makers or other proxies for depth.
Standardized contracts, the subject of earlier conversations about engaging regulators such as the CFTC, remain distant in Conrad's view. The lesson of the last few years, he says, is that you cannot standardize a product you do not operate; the operations have to become uniform first. He offers a concrete illustration of how far from uniform things are: one operator can offer two virtual machine types where one is roughly 40% slower than the other, with the gap showing up in GPU kernel launch times and ultimately in lower model flops utilization. Commodity markets handle this with grades, and compute does not yet have settled equivalents.
SF Compute's answer is to skip standardization for now. Its matching engine accepts orders carrying multiple constraints, and buyers can specify ranges such as a window between two dates or a band of acceptable properties, which are matched against sell orders carrying compatible attributes. The stated reason is speed: liquidity can form while the underlying technology is still shifting. The stated cost is that liquidity fragments across many variants, widening bid-ask spreads. Conrad concedes the spreads are unpleasant but notes they create arbitrage for anyone whose workload can absorb many kinds of capacity, since such a buyer can effectively capture part of the spread.
The dimensions that matter are partly about time and partly about physical reality. Capacity available right now can only be sold by interrupting a running workload; capacity starting a month out can be traded freely and attracts liquidity more easily. Interruption is not a single property either. Whether a cluster sits on a Weco or Vast storage system, how large that storage is, and whether data can move quickly between clusters all change how gracefully a job survives preemption. Fast transit between sites, typically a wide area network, is often missing: GPU clusters are sited next to power rather than next to customers, which is a different placement logic from CPU-era infrastructure and leaves thinner interconnect between facilities.
Hardware cohorts are another axis Conrad raises. Early GB200 and H100 units had failure problems, so a sophisticated buyer may pay a premium for hardware deployed after a certain date rather than take the first racks off the line. This suggests a market where provenance and vintage carry real price, much as grades do in other commodities, and where an index blending all units could diverge from what an actual buyer pays for a specific configuration.
Independence runs through Conrad's argument for who should operate such an exchange. Companies with large hardware or inference businesses, NVIDIA chief among them, have incentives that make neutrality hard to credibly offer, he says, and he expects the chipmaker to benefit from a market existing without being the one to run it. The engineering he describes is unglamorous: visiting sites, designing clusters, running the backplane and making infrastructure programmable. One example he gives is the ability to subdivide InfiniBand fabric on the fly, partitioning a shared network so tenants can control their own topology and memory layout programmatically. That is the software-defined networking analogy the episode leans on, applied to GPU fabrics instead of routing tables.
Why it matters: teams that need GPUs but cannot commit $50-100m upfront are the intended beneficiaries of a market that sells capacity in smaller, resellable units, and the $25-on-$100 resale figure is the metric to watch for whether that promise holds. Developers should expect contracts to expose interruption tolerance, storage recovery speed and hardware cohort as first-class purchasing choices rather than afterthoughts, which means workload design decisions about checkpointing and portability now have a direct price. A likely trade-off, in my reading, is that fragmented liquidity and wide spreads persist until hardware generations stabilize. And the neutrality claim will be tested as larger players enter compute trading, since a market run by a participant in the underlying business faces the same conflict Conrad attributes to others.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
From GPU trading to physical settlement, independent markets, and programmable compute