Buying GPUs Should Work Like Buying Stocks — Carmen Li, Compute Exchange
Carmen Li, CEO of Compute Exchange and Silicon Data, makes the case for treating GPU procurement like exchange trading. She covers the neutral, verified marketplace model, its reserved, forward, token and secondary server products, and the RFQ routing system. She then examines Silicon Data's forward curves for H100 and B200, the market's expectation of rising prices, and how used GPUs are holding value well above standard depreciation.
Buying compute, Carmen Li argues, should look less like phoning hundreds of sales contacts and more like trading on an exchange. The GPU market has expanded quickly, with over 300 new cloud and private providers, a crowd of token factories and API routers, and buyers who each need capacity at different times and scales. Finding the right capacity at a clear price, with verified counterparties and settled terms, is slow when every deal is a bilateral negotiation.
A neutral exchange, by design
Compute Exchange positions itself as a neutral marketplace for GPU procurement, and its defining rule is that it owns no GPUs. The analogy is to Nasdaq not owning the stocks it lists. Once an exchange holds inventory, it has a conflict of interest with the providers and buyers it serves, and it could cherry-pick clients or hoard relationships. Staying out of the asset lets the exchange treat every transaction individually and match each buyer to the most suitable resources.
Trust is built on verification on both sides. Providers are vetted, and the platform currently lists over 100 verified providers and more than 80,000 GPUs across all regions outside China. For existing capacity, the team checks UUIDs, inventories and real performance through benchmark suites including MLPerf, so that what is promised is what is delivered. For future capacity, such as GPUs needed from January 1, 2027, it reviews purchase orders and permits as part of due diligence. On the buyer side, Know Your Business and Know Your Customer checks confirm that a client can pay on a scheduled basis, which matters to a new cloud deciding whom to sell to. The company commits to a 24-hour turnaround.
Lessons from energy markets
Li began her career as an energy options market maker at DRW Trading in Chicago, and she draws a direct parallel. In oil, WTI is a physically delivered contract used by those who need to procure the commodity, while Brent is cash settled and serves financial hedgers managing price volatility. Compute needs both kinds of instruments. Compute Exchange is building a WTI-style physically delivered layer, with its forward markets serving as a precursor to physically delivered futures.
Reserved, forward, token and secondary markets
Reserved contracts cover existing capacity, with a minimum of one node for one month and terms running out to three, four or five years depending on availability. There is no on-demand offering. If a buyer overcommits on quantity or duration, the contract can be sold back to the exchange, which limits the downside of a bad forecast.
Forward contracts let buyers lock in capacity that does not yet exist, starting at a future date.
Token contracts are a newer idea, described as a dedicated deployment layered on top of reserved bare-metal clusters. Buyers receive guaranteed SLAs for specific models, with quantization and precision fixed for a period. The aim is to address the price volatility of on-demand token APIs.
The third market handles secondary server sales. It began about a year ago when people asked the exchange to take off their hands large holdings, such as 20,000 H100s sitting in a data center. The market now covers everything from 4090s to B200s and AMD hardware, and secondary B200 transactions have at times cleared above MSRP. Li compares this to used cars trading above sticker price during the COVID shortage.
Requests for quote and routing
Buyers can browse listings and chat with providers directly, or submit a request for quote for something not yet listed. The platform reportedly sees more than a billion RFQs a month, with requests ranging from a single node to 20,000 nodes. Behind the scenes, inventory tracking shows roughly which providers have which capacity. A request for 300 B200s in US East starting January 1 is routed to the handful of providers likely to have it. They respond through the platform with pricing, specifications, SLAs and, for existing servers, benchmark results. Price is not the only criterion, and the exchange aims for as much transparency as possible before introducing the parties to one another. It does not keep the relationship. Providers and customers are both scored, and year-end awards are planned for regional and responsibility categories.
Li also demonstrated the platform, which is free to sign up for. Users can filter by vendor, GPU model, deployment type, GPU count, interconnect, CPU platform, storage, system software stack and region. They can add head nodes, describe their use case and submit an RFQ, and selecting several regions is encouraged for flexibility.
Forward curves as a market signal
The second half turns to Silicon Data, which Li also leads and describes as the Bloomberg of compute. It produces indices that CME will use to settle futures, starting with an H100 on-demand rate and B200 prices, with trading set to begin in a few weeks.
The implied forward curves tell a clear story. Last year the H100 curve sloped downward, meaning longer contracts were cheaper per GPU-hour. Starting early this year, the whole curve shifted up and moved into contango. Longer terms no longer carry a meaningful discount, and the curve has since flattened and kept rising. B200 followed the same path. The implied forward rates, derived with a methodology similar to the Treasury futures curve since no futures trade yet, suggest the market expects prices to rise over the next 36 months. Li is careful to note that consensus is not the same as being right.
For her, the shape of the curve shows that GPUs behave like any commodity with a term structure, and that risk along that structure can be hedged. Block trading is already happening around the indices. Consider an operator with servers on the balance sheet, financing obligations for five or six years and reserved contracts covering only the first three. Years four through six are exposed, and a swap or block trade can hedge that gap. Even the first three years carry counterparty risk if a resale contract defaults, which is another place a hedge helps. Indices can also trade as a floating rate, and some banks use Silicon Data as an appraisal when they hold GPU servers on their books.
What a GPU is worth over time
Silicon Data’s residual value model is a discounted cash flow. Take an A100, which has been in service for six years. If a three-year lease is signed at a given rate, discounting those cash flows back to time zero, under assumptions about utilization and cost, gives the asset’s implied value. Residual values are computed for normalized A100, H100 and B200 servers globally and compared with straight-line depreciation, the five-year convention most operators use. The A100 tracks the accounting schedule. The H100 has risen sharply, and the B200 has moved above its original MSRP. Li notes that NVIDIA’s September investor day deck cited Silicon Data as the only data provider it referenced, a sign that these valuations are gaining recognition in the industry.