The Programmable Data Center — Evan Conrad, SF Compute

What it claims

Evan Conrad, CEO of SF Compute, argues that GPU compute cannot be safely financialized until physical delivery works. He says cash-settled indexes built on thin data are easy to manipulate and carry basis risk against the real price of new reserved capacity, and he compares this to LIBOR. SF Compute's answer is to run the cloud operations itself, without owning hardware, and to match heterogeneous orders instead of waiting for standard contract grades. Conrad says SF Compute has closed "several hundred million dollars of volume", but that the market is "not very mature or deep or liquid": about 5% of the company's capacity sits on the order book, with more being moved over.

How it works

Orders carry multiple constraints, such as a hardware range, start and end time, and other properties. A matching engine fills them against sell orders that have those properties. Conrad's stated trade-off is less concentrated liquidity and wider bid-ask spreads. In return, buyers whose workloads run on many hardware types can take the spread.

Named queries ("SKUs") are new, and customers can define their own. Conrad's example is a bid for any Blackwell at $6 an hour or less, filled by GB300 or B300.

The dimensions that change price, per Conrad:

• Current or future capacity. Selling capacity now means interrupting a running workload. Capacity a month out trades freely.
• Interruption tolerance. Batch inference can take a notice period of minutes to an hour. Live inference needs one to three months because it pays SLA penalties, so the two lease types are priced differently. Conrad compares this to power's day-ahead and live markets.
• Storage and recovery. Whether a Weka or VAST system is present, its size, and whether a WAN links clusters. GPU sites sit near power rather than customers, so many lack fast transit.
• Hardware cohort. Early H100 and GB200 units had firmware faults across whole batches, Supermicro firmware in one case. A cohort feature is planned but not built.
• Operator configuration. Two VM types from one operator can differ by 40% in GPU kernel-launch time, which lowers MFU.

On delivery, SF Compute does site visits, bill-of-materials and OEM supply-chain work, cluster architecture, rack rollout supervision, and backplane operation (PDUs, switches, DPUs, UFM, BMC). It does not self-perform data center builds. Customers can now repartition their InfiniBand fabric on the fly without filing a request.

Evidence

All figures are Conrad's, given verbally:

• A customer who spends $100 on capacity recovers $25 on average by reselling unused time. This is the company's main liquidity metric.
• Contracts run from one hour to three years. Clusters are financed over three to five years. Blocks are paid fully upfront, but longer leases are not.
• The forward curve is hump-shaped. Prices dip near delivery because unsold time expires, rise further out, then ease for distant dates. Some buyers instead pay more for far-dated capacity to lock in supply.
• A B300 earns about $40 an hour selling inference at full use, or about $25 at roughly 50% utilization. It can be bought for $5 to $6 an hour. Conrad calls this gap an arbitrage that has to close.

Limits

This is an interview, and the transcript is automated with unverified speaker labels. There are no published contract specifications, API documentation, or pricing tables. Standardized contracts and CFTC engagement are described as "still pretty far away". Conrad names no partners and says announcements were expected "in a week or two". His claim that NVIDIA, CoreWeave, Crusoe, Lambda, hyperscalers, Baseten, Fireworks and Modal structurally cannot run a neutral exchange is his own argument. It is not independently supported.