On the Latent Space podcast, Evan Conrad, CEO of SF Compute, laid out a model for GPU compute that looks more like a commodities market than a cloud contract: an order book, physical settlement so buyers get usable capacity rather than a cash payout, and a "programmable data center" where customers can carve up their own network fabric without calling a provider. The episode page is based on an interview recorded at AI Engineer NYC, and the transcript is automated with unverified speaker labels.
We are covering it because GPU rental prices and contract terms hit every team that trains or serves models. Every figure below is what Conrad said, on a podcast, about his own company; we found no independent confirmation, and we say so where it matters.
TL;DR: the claims and their status
| Claim | Who says it | Verified? |
|---|---|---|
| About 5% of total capacity sits on the order book, growing | Conrad | No |
| A buyer spending $100 recoups about $25 by reselling | Conrad | No |
| B300 inference could earn about $40/hour at full utilization, $25 at 50%; buying costs about $5-6/hour | Conrad | Illustrative; not a market quote |
| Two VM types from the same operator can differ about 40% in performance | Conrad | No |
| Several hundred million dollars of volume closed; a $245M compute sale | Show notes | No |
| Own cluster used to need $50-100M upfront | Episode intro | General estimate |
| Clusters are financed over 3-5 years; forward contracts can run 3-5 years | Conrad | General practice |
What problem is it solving?
Renting GPUs normally means signing a contract for a fixed period with a provider. If your needs shrink, you are stuck paying; if they grow, you may not find capacity. Large clusters are typically financed over three to five years, so providers want long commitments. Conrad's pitch is that an order book lets buyers and sellers match on attributes (GPU generation, memory, time window, interruptibility) so unused time can be resold instead of wasted.
That is the "hotel" analogy he uses: the company began by subleasing a room, then ended up running the hotel, meaning much of a cloud's operating stack. He says the company does not own hardware, because a neutral venue has to be independent enough for others to trade on it, and that NVIDIA, hyperscalers and neoclouds cannot build that market because of conflicting interests as buyers and sellers.
How the market works, per Conrad
Physical settlement first. The buyer gets compute, not cash. He argues that cash-settled derivatives should be derived from physical markets, and the host notes that indexes built on thin data can be manipulated, citing LIBOR. Conrad warns of "a lot of basis risk" when a hedged price has little to do with what capacity actually costs.
Matching by attributes. Orders carry constraints, and a matching engine pairs them with compatible sell orders, so liquidity can form before the industry agrees on standard grades.
Two kinds of capacity. Capacity available now often requires interrupting a running workload, while future capacity is easier to trade. He describes the idea of a market for breaking leases, since batch inference tolerates interruption with short notice while live inference typically carries an SLA and is priced differently.
Hardware cohorts. Early H100 and GB200 units had defects and firmware issues, so buyers may pay more for units deployed after a certain point. That is a useful reminder that a "GPU hour" is not a uniform product.
The price curve has a hump. He describes near-term unsold capacity as discounted, prices rising toward a normal level, and far-future capacity possibly worth less.
The programmable data center
The distinctive technical claim is that customers can subdivide their InfiniBand fabric on the fly without contacting the provider. Conrad says SF Compute manages the backplane: power distribution units, switches, DPUs, UFM, BMC and the node harness around bare metal. It designs clusters, works with OEMs on supply chains and project-manages powered sites through partners, but does not do the construction itself. He also names sovereign compute, helping countries set up their own clouds, as a target market.
If this works as described, a team could resize and repartition a cluster as a software operation. We have not seen it demonstrated, and neither the interview nor the show notes provide an independent test.
The B300 numbers, and why they are only a sketch
Conrad says a B300 running inference could earn about $40 per hour at full utilization and about $25 at 50 percent, against roughly $5 to $6 per hour to buy, and that "one of these numbers has to change." That is a statement about the spread between what compute can earn and what it costs, not a quote you can transact on.
Compare it with what we have tracked elsewhere: our coverage of record B300 rental prices, Nebius raising GPU rental rates and AWS capacity blocks price hikes shows rental rates moving up, not down. If a resale market makes unused capacity easy to sell, it could dampen waste but not scarcity. And the Gartner warning of a neocloud shakeout is the other side of the same coin: if demand softens, financed capacity gets stuck.
The performance variance warning
One detail deserves more attention than it got: Conrad says two VM types from the same operator can differ by about 40 percent in performance because of kernel launch overhead. If true, a market that treats "one B300 for one hour" as a fungible unit is missing something. Buyers need a benchmark, not just a spec. For your own workloads, run a fixed job on any new provider before committing, and record tokens per second and time to first token alongside price.
A buyer's checklist for any GPU marketplace
Whether you look at SF Compute's own site or any rival, these questions separate a useful market from a marketing deck:
- What exactly is delivered? GPU model, count, memory, interconnect and the date the hardware was deployed. Early units of a generation can carry defects, so cohort matters.
- What happens on interruption? Notice period, whether you get a refund or credit for lost time, and whether your checkpoint storage survives.
- Who is the counterparty? In a marketplace you may be buying from a third-party operator. Find out who is responsible if the node fails and how disputes are resolved.
- What can you resell, and at what fee? A resale option is only valuable if windows are liquid and fees are low. Ask for historical fill rates for the hardware you need, not an average.
- How is performance verified? If two VM types from the same operator can differ by 40 percent, ask for a benchmark harness or run your own before you commit.
- How is data protected? Bare-metal nodes that are repartitioned and resold must be wiped between tenants. Ask about disk and memory scrubbing and network isolation.
- What are the exit terms? For longer commitments, check whether you can hand capacity back and what it costs.
None of these are exotic. They are the same questions you would ask a traditional cloud, with the extra layer of "who sits on the other side of the trade." A market that answers them clearly is worth testing with a small job first.
What this means for what you build or pay
Finance mechanics only matter here if they change your bill. Three honest readings:
- Resale could lower the cost of over-buying. If the $25-back-per-$100 figure holds, a team that overestimates its needs loses less. It is an average, so your result could be far lower for popular windows.
- Interruptibility becomes a price knob. If your batch training or offline inference can pause, you could pay less. Live serving cannot.
- Standardization is not solved. Variance in hardware cohorts and VM performance means you still need to test.
If you are choosing a provider this quarter, ask for the interruption terms, the hardware generation and deployment date, the network fabric, and the resale or exit options in writing, and compare the total cost per useful token, not the hourly price.
Questions people are asking
Is this a futures exchange? Conrad separates physical settlement, which SF Compute does, from cash-settled derivatives, which he says should be built later on top of physical markets.
Why can NVIDIA not just build it? His argument is neutrality: a venue owned by a seller, or by a major buyer, will not be trusted by the other side.
Does a compute market lower prices? Not by itself. It can make idle capacity easier to find. Whether prices fall depends on supply, which our reporting on ASML's EUV constraints and Michael Dell's memory shortage warning shows is tight.
What is still unknown
- Independent data on order-book depth, resale rates and settlement quality.
- How the fabric subdivision works in practice and what the failure modes are.
- Pricing and terms beyond the examples in the interview.
- Whether other providers will join an open market.
Related reading
- B300 GPU rental prices at record highs
- Nebius raises GPU rental rates 20 percent
- AWS capacity blocks GPU price hike
- Gartner: neocloud shakeout
- ASML: AI demand exceeds EUV supply
- Michael Dell on the memory shortage
Figures are from a podcast interview published October 20, 2026 and are the guest's claims.