Carmen Li told AI Engineer that buying GPUs should work like buying stocks, with a neutral marketplace that routes orders, verifies supply and clears trades without owning inventory itself.
The pitch is efficiency.
Li is chief executive of Compute Exchange and of Silicon Data, which she calls the Bloomberg for compute. Compute Exchange started in 2024 when Li saw five new clouds and three hyperscalers and bet the supplier count would explode. It now lists more than 100 verified providers and 80,000 GPUs across regions excluding China, and promises 24-hour turnaround on requests. A forward contract, which is an agreement to deliver compute at a future date for a price set today, is the core product alongside reserved contracts and token packages. A token forward, which is a guaranteed deployment of a specific model with fixed quantization and precision for a set period, aims to replace volatile on-demand API pricing. The platform also runs a secondary market for used servers, from 490s to B200s, where Li said B200s have traded above MSRP.
On Compute Exchange, a buyer either browses inventory or files an RFQ, which is a request for quote that states model, count, interconnect, CPU, storage and region. Li said the venue sees more than $1 billion RFQed each month, ranging from one node for one month to 20,000 nodes for three to five years, and routes each request to providers likely to have capacity. For existing reserved capacity the team verifies UUIDs and runs Nvidia DGX, Inference X and MLPerf checks, and for future delivery it reviews purchase orders and permits. The exchange also does KYB and KYC on buyers, and Li said it will score both sides and hand out regional provider awards. If a buyer over-commits, the platform will buy back the contract, which Li framed as flexibility that brokers who hold inventory cannot offer without conflict.
What had to be true before this pitch could land was a fragmented supply base and a pricing history. Li put the count at more than 300 private cloud providers globally plus token factories and API routers, up from a handful in 2024. Silicon Data built indices from that activity for CME to settle futures that Li said will start trading in a few weeks on the H100 near-cloud on-demand rate and on B200 prices. Using a Treasury-curve method, her team plotted a term structure where last year longer dates were cheaper and this year the whole curve shifted up into contango, which is a market condition where future prices sit above near-term prices, and then flattened. Implied forward rates from that curve point to higher prices over the next 36 months, though Li said consensus is not a forecast. The same history feeds residual-value curves that compare discounted cash flows for A100, H100 and B200 servers against the five-year straight-line depreciation many firms still use for accounting, with A100s now tracking below that line and H100s and B200s recently above it, a move Li compared to used cars spiking during COVID.
That last point matters because Li wants physically delivered futures, which are contracts that settle by delivering the underlying resource, to sit beside cash-settled hedges. She used the DRW Trading analogy from her years as an energy options market maker where WTI settles with barrels and Brent settles in cash, and said Compute Exchange is building the WTI for compute. Elsewhere, cloud pricing already leans on commitments, as AWS Savings Plans offer lower rates in exchange for a consistent usage commitment measured in dollars per hour for one or three years. Li said block trades tied to Silicon Data indices are already happening to hedge years four through six of a server financing when only the first three years are covered by resale contracts. What still has to happen is proof that a futures market can find liquidity and that verification plus permits actually prevent defaults when a seller fails to deliver on Jan. 1, 2027 or any later start date.
