The Programmable Data Center — Evan Conrad, SF Compute

Summary

• SF Compute’s core bet is that a tradable GPU market requires operating the infrastructure that delivers the compute. Conrad says the company spent three years learning why physical settlement matters: “these things aren’t as fungible as you think.” Even two VM configurations can differ by 40% in performance; financial contracts built ahead of reliable delivery risk manipulation and hedges disconnected from executable prices.
• SF Compute reports useful resale liquidity on the 5% of capacity currently listed in its order book. Customers spending $100 recover $25 on average by selling capacity to someone else, a practical measure of whether liquidity helps commercial users. Conrad separately reports closing several hundred million dollars of volume, while describing order-book volume as small.
• Swyx’s B300 inference estimates expose the episode’s sharpest pricing tension. Swyx estimates roughly $40 an hour of inference revenue at 100% utilization, or around $25 at roughly 50%, against rental costs of $5–$6 an hour. Conrad agrees it is a “huge overcharge”; buyers expecting higher compute prices may pay more to secure future supply.
• SF Compute wants a liquid forward curve that prices customers’ ability to surrender capacity. Buyers can purchase from an hour through three years; batch jobs might tolerate minutes of notice, while live inference may need one to three months. Swyx describes a “hump” in prices: unsold imminent capacity gets discounted like expiring milk, while shorter lead times can command a premium over distant delivery. Conrad also notes that some buyers may pay more for future capacity if they expect prices to rise.
• Conrad argues that independence is the condition for turning compute trading into cheaper financing. His impression is that NVIDIA, clouds and major inference buyers are not independent enough for people to trade on their venue. SF Compute avoids hardware ownership: “we have to be Switzerland,” enabling liquidity that lenders can finance against.
• The programmable data center gives asset owners an operating layer for their GPU investments. Customers can subdivide their InfiniBand fabric themselves as capacity needs change, while SF Compute operates the infrastructure and owns the SLA. Its sovereign-compute pitch combines local clusters with San Francisco demand and expertise: “We’re going to make it financeable.”

Deep dive

1. Physical settlement requires the operating capacity of a cloud

• Conrad’s origin story begins with subleasing a cluster, then acquiring responsibilities like a hotel-room seller being asked for cleaning, room service and security: “at a certain point, you’re just running the whole hotel.” SF Compute consequently built the operating capacity of a cloud to deliver compute reliably. “Physical settlement” borrows the futures-market distinction between trading a contract and actually receiving the oil or soybeans.

• The order book and delivery system were co-designed around compute’s constraints. Conrad compares this with power-market orders incorporating boiler cool-down times: a compute customer might accept interruption only with storage that enables quick recovery. Contract standardization remains “pretty far away.” The lesson was that “we just needed to operate the whole thing,” because consistent operations must precede a standardized product.

• Conrad initially calls the market “not very mature, deep, or liquid,” then distinguishes small order-book volume from concentrated liquidity within listed capacity. Only 5% of total capacity is on the book, with more moving onto it. Customers spending $100 recover $25 on average through resale. He reports closing several hundred million dollars of volume; his preferred liquidity test is “Is this solving the problem for the customer?”

2. Hardware differences prevent a single interchangeable compute product

• Swyx’s commodity-grading analogy meets a substantial engineering problem: one operator’s VM configuration can impose 40% performance degradation relative to another. Conrad locates the difference in GPU kernel-launch latency, which ultimately worsens MFU. Establishing equivalence therefore requires understanding what the infrastructure delivers, beyond identifying the GPU model.

• SF Compute matches orders across constraints instead of waiting for fixed standards. A buyer could accept any Blackwell at $6 an hour or less, matching against GB300s or B300s; another needs specific memory or NVLink domains. Swyx objects that exposing heterogeneity hurts a market maker. Conrad agrees that dimensions remain: flexibility creates arbitrage opportunities, while specificity costs more. Named queries are being introduced; query-based matching already operates.

• Storage and connectivity determine whether capacity is genuinely substitutable. Conrad points to WEKA or VAST systems, storage size, and whether data can move quickly enough to restart a workload elsewhere. That requires a fast transit path, typically a WAN. GPU clusters follow power availability, whereas traditional CPU clusters followed customers for latency, so newer locations may lack strong connections between sites.

• Even deployment cohorts could matter, though cohort-specific matching remains an idea under consideration. Swyx asks why ordinary burn-in would not remove defects. Conrad distinguishes faulty individual GPUs from firmware problems that may affect multiple systems: an early H100 issue could have been firmware in a Supermicro box rather than the GPU itself. Testing and replacing broken hardware would not resolve that; the firmware would need an update or a fix in new boxes.

3. The forward curve prices both expiring capacity and interruption rights

• Conrad’s “3 to 5 years” answer initially concerns how clusters are financed across the broader market. SF Compute’s order book supports arbitrary blocks, with purchases spanning an hour through “the next 3 years and anywhere in between,” technically finer too. Blocks currently require full payment upfront; longer-term leases do not. His priority is building a forward curve with substantial liquidity.

• Capacity already running a workload differs from capacity starting next month, which can trade without interrupting anything today. Batch inference might tolerate an hour’s notice, or even a minute or two; live inference might require one, two or three months because disruption can trigger customer SLA payments. Swyx’s realization—“the market for breaking leases”—captures why interruption rights command different prices.

• Asked whether the curve shows backwardation or contango, Conrad asks whether Swyx means future capacity priced above or below spot. Swyx says, “We see a hump,” and offers expiring milk as an analogy: imminent unsold compute must discount because “it can’t be sold in the past.” He suggests prices then rise toward their preferred level, while distant capacity might cost less because buyers value shorter lead times. Conrad acknowledges the explanation, then adds that buyers expecting prices to rise may be willing to pay more for later supply.

• Swyx estimates that a B300 could generate roughly $40 an hour selling inference at 100% utilization—“Any model, basically”—or around $25 at roughly 50%, versus $5–$6 an hour to rent it. Swyx says, “One of these numbers has to change”; Conrad agrees it is a “huge overcharge.” Buyers expecting compute prices to rise may consequently pay more for future delivery to lock in supply.

4. Cash settlement needs a credible physical market underneath it

• Conrad’s warning follows three years spent learning the delivery problem: financializing compute before physical settlement is established risks manipulation and substantial basis risk. The price used to hedge might have little relationship to the price at which a customer can actually obtain capacity. Making compute more fungible remains, in his framing, the industry’s job.

• Swyx proposes the familiar currency-and-rates model: participants agree on a fixing like LIBOR. Conrad’s immediate objection is that LIBOR was manipulated. His concern is that compute indexes can similarly rely on very small datasets with little correlation to the price of buying a new reserved contract. An agreed benchmark alone does not resolve that underlying measurement problem.

• Swyx says cash settlement could be a great future concept if derived from physical settlement. Conrad calls cash-settled contracts the kind of side bet seen in The Big Short, which may have little to do with the underlying thing. He says people are already making such bets in some areas, while Swyx cautions against derivatives on an underlying that does not yet exist.

• Conrad says SF Compute is a big fan of Silicon Data and Compute Exchange, describing them as early entrants. He characterizes their cash-settlement path and SF Compute’s physical-settlement path as mutually beneficial, and says they may partner.

• The surrounding ecosystem is developing distinct functions. Conrad identifies SemiAnalysis with efforts toward better standards, alongside residual-value insurance and deposit lending that fronts credit risk for clouds. He acknowledges limited familiarity with individual providers; his broader observation is an ecosystem forming around compute’s financial bottlenecks.

5. Independence supports a programmable data center business

• Conrad’s impression is that NVIDIA benefits from this market but cannot credibly run it; similar conflicts apply to Crusoe, CoreWeave, Lambda, hyperscalers, Baseten, Fireworks and Modal. They are not independent enough for people to trade on their venue, and without trading there is not enough liquidity for lenders to finance against the forwards market. That liquidity could lower industry financing costs. SF Compute therefore avoids hardware ownership. Building the venue requires giving up lucrative cloud and inference opportunities: “You have to kill your business in order to make this.”

• The operating stack spans site visits, bills of materials, OEM coordination to ensure timely delivery, cluster architecture, rack deployment and backplane management. Customers can now subdivide their InfiniBand fabric without asking the cloud to do it for them, supporting expansion and contraction as capacity trades. The ambition is to virtualize “all the stuff that you would have gotten if you had owned the hardware,” making the data center programmable.

• Owners can bring a data center, megawatts or powered land; SF Compute coordinates with partners to develop the facility and takes over supercomputing, while it does not self-perform builds. For investors seeking a GPU cluster’s residual value, Conrad offers operational responsibility—“We own the SLA”—and says SF Compute can immediately sell the capacity to someone else if they need to exit, so it does not hit their balance sheet.

• The business operates primarily in the U.S. but deploys globally with partners seeking to become local clouds. Sovereign-compute customers receive cluster setup, San Francisco expertise and San Francisco offtake. Conrad’s pitch connects that operating work directly to funding: “We’re going to make it financeable.”