Evidence at a glance
A Large Contract Shows What the Market Is Selling
SF Compute is a platform seeking to make GPU capacity tradable. In September 2026, the company said it had sold two dedicated NVIDIA Blackwell B300 clusters to an unnamed AI company for a total of $245 million. The sale comprised two agreements, each worth $122.5 million and each running for 36 months. The example shows that compute trading is not limited to temporary, hourly rentals. It can also take the form of multi-year commitments for dedicated capacity.
Yet the size of a contract does not establish that a market is mature. A buyer is not purchasing an abstract GPU price index, but the ability to use resources with particular specifications at a particular time. The value of that promise depends on whether the cluster comes online as scheduled, whether its network can support the workload, and whether capacity can be restored after an interruption. For technical leaders, understanding this kind of market starts with delivery terms, not transaction totals.
Physical Settlement Pushes a Matching Platform Toward Cloud Operations
SF Compute treats physical settlement as part of its market design: the buyer receives usable compute rather than a cash difference calculated against an index. Fulfilling that kind of contract means bringing hardware and operating conditions into the promise itself. The company says it works on cluster design, supply-chain coordination, oversight of rack deployment, network and underlying infrastructure operations, and service-level commitments. It works with partners on data-center construction.
That creates a fundamental difference between this platform and a lightweight matching service. After orders are matched, the platform still has to deal with equipment configuration, whether workloads can be interrupted, how failures are recovered, and how quickly data can be restored from storage. These are not after-sales details that appear only once delivery is complete. They affect what buyers are willing to pay for and whether the platform can honor its contracts. In other words, tradability depends on operational capability underneath the market.
Without Hardware Standardization, Orders Need More Detail
SF Compute does not require every cluster to become the same product before trading can begin. The approach described in the interview lets buyers place orders with conditions such as timing and GPU or compute range, then uses a matching engine to find sell orders. That can accommodate capacity differences that already exist in the market, rather than waiting for all clusters to be fully standardized before attempting to trade them.
The cost is that the more specific an order becomes, the fewer buyers and sellers may be able to match it. GPU generations, memory, network interconnects, and software configurations are not fully interchangeable. Even similar resources can fall into separate markets depending on whether workloads may be interrupted and what recovery is required. More detailed specifications make delivery promises more precise, but can divide liquidity into smaller pools and widen bid-ask spreads. The question is not whether to standardize, but how far to do so while balancing faithful delivery against market depth.
The Ledger Records Rights to Capacity Over Time
The Abel ledger records compute balances as a number of nodes available during a particular time window, rather than as one undifferentiated balance. According to the company’s description, money and compute rights are posted together in the same transaction when capacity is bought or sold. What is being recorded is closer to a right to use capacity at a future time. Reserved capacity that goes unused can be transferred to another buyer, and commitments can even be made and resold before a cluster comes online.
The mechanism addresses a mismatch between the time required to finance a cluster and the time at which a customer needs capacity. In the interview, Evan Conrad described cluster financing as requiring commitments of three to five years, while saying the order book covers capacity from the next hour out to three years in the future. Those horizons should not be conflated: the first concerns the financing cycle, and the second the range of tradable capacity. Conrad also said customers recover an average of $25 through resale for every $100 spent on capacity. That is a company or executive disclosure, and the public material provides no independent audit. The research material also says capacity blocks currently require full prepayment, so resale may reduce the cost of idle capacity without removing the risks of tied-up capital or failing to find a buyer.
Programmability Does Not Mean Liquidity Has Arrived
The phrase “programmable data center” has a concrete example in the interview. Conrad said customers can dynamically subdivide an InfiniBand network without contacting a supplier each time they need a configuration changed. This gives users some control over the network and resembles the broader idea behind software-defined networking, which separates control logic from data forwarding. However, the public material does not provide detailed information about the relevant APIs, permission boundaries, performance effects, or availability. It would be unwarranted to infer that customers can control every aspect of network configuration.
The market figures also call for restraint. Conrad said about 5% of total capacity was on the order book at the time of the interview, with that share still growing. He also described cumulative trading as being in the hundreds of millions of dollars, without specifying the measurement basis or time period. A 5% order-book share does not mean that no trading is taking place. It does show a gap between listed capacity and the platform’s total capacity. For an evaluator, the practical order of operations is to verify the specific GPU and network, interruption terms, recovery process, SLA, and prepayment requirements, then determine whether resalable capacity actually fits the workload. A more programmable data center is one foundation for trading, not proof that liquidity and delivery risks have been solved.