Nvidia can't build the compute futures market—because doing so means abandoning inference revenue
Every player holding computing supply has their own business to protect; only those who completely exit inference and cloud services to become purely neutral brokers can establish pricing everyone trusts.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Compute brokers must become cloud operators to meet buyer expectations
SF Compute began by trying to resell spare cluster capacity. Buyers kept asking for clean infrastructure, customer support, security—services only a full cloud provider could deliver. This taught them what they now call ‘physical settlement’: the buyer doesn't just need a matched price; they need actual runnable compute resources delivered. That means the supply side must look and operate like a cloud provider, not a middleman.
— Evan ConradThe brokerage market remains too thin to hedge effectively
SF Compute frankly admits its compute trading market is ‘still immature, shallow, and illiquid’. Only 5% of the company's total capacity appears on order books at any given time, and even that liquidity is highly concentrated. Their core measurement is direct: customers spend $100 to buy compute, then resell it back to the market and recover an average of $25. That number—not market-maker count or other vanity metrics—tells them whether liquidity actually helps customers.
— Evan ConradAbandon unified standards; use multi-dimensional order books instead
Compute capacity lacks agreed grades unlike oil. GPU virtual machines from different clouds vary by up to 40% in performance. Rather than wait for industry standards, SF Compute designed order books with multiple constraint dimensions: buyers specify time windows, storage configurations, and other attributes; the system matches them to sell orders. Liquidity becomes fragmented and spreads widen. But anyone who can arbitrage across compute types can capture the spread between them.
— Evan ConradEarly H100 and GB200 batches carried firmware-level defects later production did not
Initial H100 and GB200 units shipped with firmware bugs—sometimes in Nvidia's firmware, sometimes in motherboard vendors' implementations—that required patches or replacement. This means a single chip generation has quality tiers: sophisticated buyers pay more for ‘not the first batch off the line’ the way oil markets price different refining grades. That is another reason compute markets naturally resist reduction to a single standard unit.
— Evan Conrad$40-per-hour inference revenue versus $5–6-per-hour acquisition cost forces compression
A B300 running inference at full capacity generates roughly $40/hour in revenue; even at 50% utilization that is $25/hour. Yet the hourly cost to acquire that B300, amortized, is only $5–6/hour. Evan Conrad stated plainly: ‘One of these two numbers has to move’. That gap is the fundamental arbitrage driving capital into compute futures markets—whoever locks in supply earliest captures the difference.
— Evan ConradCash settlement works only when built on real physical delivery underneath
Conrad argues that financing too much of the derivatives market before physical settlement infrastructure matures creates an easily-manipulated market with severe basis risk—the price you hedge against may completely decouple from the price you can actually transact. He pointed to credit default swaps: at peak, the derivatives market was roughly 12 times the size of the underlying bonds. Cash settlement itself is sound, but only when derived from real, deliverable physical markets, not when it replaces them.
— Evan ConradNeutrality requires abandoning your other lines of business
Nvidia could create a neutral compute exchange; it chooses not to because buyers would distrust its pricing—it has inference revenue to protect. The same logic applies to Crusoe, CoreWeave, Lambda, and the hyperscale clouds: they earn too much from inference and cloud services. The moment they become neutral brokers, they voluntarily abandon that revenue. Conrad's conclusion: this would lower financing costs for the entire industry, but no player is willing to ‘kill its own business’ to make it happen.
— Evan ConradProgrammable networks let compute brokers resize clusters in real time
SF Compute's latest breakthrough: they virtualized the physical InfiniBand network topology. Previously, if a customer wanted to repartition InfiniBand segments, they contacted the provider and waited for manual configuration. Now customers dynamically split and recombine segments on the fly, scaling cluster footprint with order book demand in real time. This is what Conrad calls ‘making the data center programmable’—a paradigm shift comparable to cloud computing's emergence.
— Evan ConradIn their own words · checked verbatim
during COVID oil went negative because, uh, so like I will pay you to take oil from me because it is sitting there in the port.
Evan Conrad4:32
a customer who buys capacity from us and they spend a hundred dollars to buy that capacity. We’ll make back $25 on average from selling it back to somebody else.
Evan Conrad5:45
right now if you took a B300 and you trying to run inference on it, um, you can make $40 an hour or so, um, selling that as inference.… Meanwhile, you can buy B300s at like five to $6 an hour. One of these numbers has to change.
Evan Conrad20:04
if you try to build, um, too much financialization before you’ve, the industry has had time to build out the, um, the physical settlement. Um, I think what you have is you have, um, a market that’s very easy to manipulate and a market that has a lot of basis risk
Evan Conrad23:44
So our impression is that NVIDIA can’t make it, but benefits tremendously from it existing. The reason why they can’t make it is because they would then not be independent enough for people to trade on it.
Evan Conrad33:20
This is also why Crusoe and CoreWeave and Lambda, why the hyperscalers can’t make it, why Base 10 and Fireworks and Modal can’t make it because they have too much demand on that side.… You have to kill your business in order to make this. That’s why nobody does it.
Evan Conrad33:20
we allow customers to subdivide their InfiniBand fabric without having to ask us.
Evan Conrad34:30
Would you describe this as like you’re virtualizing the data center and making it programmable? Yeah, we’re virtualizing the data center and making it programmable.
Evan Conrad35:43
Figures
| Capacity on order books | 5% of total | 5:45 |
| Customer resale recovery rate | $25 average per $100 spent | 5:45 |
| B300 full-load inference revenue | ~$40/hour | 20:04 |
| B300 at 50% utilization revenue | ~$25/hour | 20:04 |
| B300 acquisition cost per hour | $5–6/hour | 20:04 |
| Credit default swap market size relative to underlying bonds | ~12x | 24:59 |
| Typical GPU cluster financing cycle | 3–5 years | 14:07 |
| SF Compute cumulative transaction volume | Hundreds of millions of dollars | 34:30 |
Glossary
- Physical settlement
- The buyer receives actual deliverable compute resources; infrastructure must be real and operationally capable.
- Cash settlement
- Only the contract price difference is paid; no actual compute resources change hands.
- Basis risk
- The risk that the hedging price and the actual transactable price diverge sharply enough to erase the hedge's value.
- Order book
- A system where buyers and sellers post bids and asks; a matching engine pairs them into transactions.
- Backplane
- The switching fabric, data plane units, network cards, and other interconnect infrastructure connecting GPU nodes.
- InfiniBand fabric
- The high-speed, low-latency network connecting GPU cluster nodes; more specialized than Ethernet, standard in supercomputers and large AI clusters.
How to listen
Investors, traders, and engineers tracking GPU compute trading, derivatives pricing structures, and the financialization of data center infrastructure.
The opening remarks about the AI Engineer conference venue and panel topics do not affect understanding of the compute market discussion that follows.