Engineering brief
GPU capacity is becoming a traded market. Neutrality is the constraint
At a glance
The brief is enough
- Relevance
- Practical value
- Radar
- Time horizon
- This year
- Warnings
- Vendor pitch
A B300 hour rents for $5-6 but can generate $25-40 in inference revenue. SF Compute is betting that physical settlement, not cash indexes, will fix GPU market pricing.
Compute procurement and pricing are becoming financial markets, affecting capacity planning, budget, and infrastructure strategy.
Watch if you're budgeting GPU capacity, building a platform team, or considering buying or selling compute.
Summary
The claim. Evan Conrad runs SF Compute, a company trying to make GPU capacity tradeable like oil or power. The key move is physical settlement: actually delivering and operating the compute, not just trading contracts on an index. He says the market is early, with only 5% of SF Compute's capacity on the order book, but that liquidity on that slice is concentrated enough for customers to recover $25 for every $100 they spend buying and reselling capacity.
The imbalance. The strongest concrete signal is the pricing spread on Nvidia B300s. Conrad says an hour of B300 compute costs around $5 to $6, while running inference on it can produce $25 to $40 per hour at full utilization. That gap is either a temporary arbitrage or a sign that capacity is mispriced. For an engineering leader, it means the value of owning or reserving GPUs is no longer obvious; it depends on utilization, workload type, and market timing.
The market shape. SF Compute's order book matches orders with multiple constraints: start time, duration, interruption tolerance, storage, and cluster location. Contracts can span from an hour to three to five years. The forward curve forms a hump: near-term unsold capacity gets discounted, future capacity trades lower as buyers prefer shorter lead times. This is a market for breaking and rebooking leases, closer to power markets than to standard cloud reserved instances.
The neutrality constraint. Conrad argues Nvidia, hyperscalers, and GPU clouds cannot run a neutral marketplace because they own inventory or demand on one side. SF Compute deliberately does not own hardware to avoid manipulating the forward curve. That is a self-serving claim, but it is structurally plausible; commodity exchanges need independent clearing and settlement to attract liquidity.
The hard engineering. The underappreciated part is what physical settlement requires. SF Compute performs site visits, designs bills of materials, supervises rack rollout, and operates PDUs, switches, DPUs, UFM, and BMCs around nodes it does not own. It recently added the ability for customers to subdivide their InfiniBand fabric on the fly, which Conrad describes as making the data center programmable. That is the operational moat, and also the cost barrier.
The risk. SF Compute has closed several hundred million dollars in volume, but reported numbers are vendor-reported and partners are unnamed. Conrad warns against building cash-settled indexes before physical settlement exists, citing LIBOR manipulation. The argument is sound, but it also positions his company as the necessary layer. Teams should follow the mechanics, not the marketing.
My take
The interesting part isn't the order book. It's the claim that neutrality is a structural requirement, and Nvidia, hyperscalers, and GPU clouds all have incentives that disqualify them from running a fair market. That sounds self-serving, but it's consistent with how commodity exchanges work. I'd be careful treating SF Compute's numbers as proof; they are vendor-reported and the market is still tiny.
What I'd take seriously is the operational complexity underneath. If GPU capacity becomes tradeable, teams with interruptible workloads and fast checkpointing gain real pricing power. Teams with live inference and rigid SLAs will pay a premium for non-preemption. That changes architecture decisions today, even before the market matures.
What to do with this
- Map your GPU workloads by interruption tolerance and checkpoint recovery time to see which could run on tradeable capacity.
- Ask your cloud provider what preemption terms and storage tiers they offer, and price the difference against reserved contracts.
- Track the spread between spot and reserved GPU rates monthly; a widening gap is your signal to buy flexible capacity.
Watch the video
This video is blocked due to your privacy settings. To watch this video, please accept YouTube marketing cookies.
Related breakdowns
AI engineers are the whole team now, a leadership risk
AI engineers are told they are the team. A Moshi founder argues recommendation design, not model quality, decides whether AI tools get used.
Your Models Are Learning Chemist Sociology, Not Chemistry
Leash Bio turned screening data into an RL environment and found models learn chemist habits, not chemistry. Data provenance is now an engineering decision.
Physical AI's real test is napkin folding, not demos
Dyna Robotics says its robots fold thousands of napkins across multiple Din Tai Fung sites. The real signal is customer validation, not the world model demo.
Get TL;DW
Too Long; Didn't Watch.
A concise breakdowns of the AI and devtools videos that actually matter for engineering leaders.
Free. Weekly. No hype.
Video and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.