Compute procurement should use a neutral exchange where buyers can compare verified supply instead of calling hundreds of salespeople.
2
Compute Exchange verifies existing GPU capacity, performs due diligence on forward capacity, and checks whether buyers can pay.
3
Silicon Data's forward curves suggest that H100 and B200 prices may rise, while used H100 and B200 servers are holding value above standard depreciation curves.
Summary
Carmen Li argues that buying compute should look more like trading stocks than managing hundreds of sales relationships. Compute Exchange owns no GPUs, which lets it work with competing providers without favoring its own inventory. The platform verifies existing capacity through inventory, UUID, performance, and benchmark checks. For future capacity, it reviews purchase orders and permits. It also verifies buyers through KYB and KYC checks. Buyers can request reserved GPU capacity, forward contracts for GPUs or tokens, or used servers. Li then shows how Compute Exchange routes requests for quote to providers with likely matching inventory. In the second half, she describes Silicon Data, which produces compute indices and forward-looking asset valuations. H100 and B200 forward curves have shifted upward, and implied forward rates indicate that market participants expect higher prices. Her data also shows H100 and B200 residual values exceeding straight-line five-year depreciation, with B200 values above MSRP.
Compute procurement needs an exchange because supply has become fragmented
Li says Compute Exchange started in 2024 as GPU supply expanded across new clouds, hyperscalers, token factories, and API routers. She compares this with stock trading: buyers can call brokers, but an exchange gives them a place where liquidity, settlement, clearing, and price discovery are organized. In compute, a buyer may otherwise need to contact hundreds of salespeople. For a reserved requirement within the next 48 hours, Li argues that going to a marketplace of verified resources is more efficient.
A neutral marketplace should not own the assets it matches
Compute Exchange does not own GPUs by design. Li compares GPU ownership by the exchange to Nasdaq owning stocks. She says ownership would create conflicts with both providers and buyers because the platform could favor its own inventory, select certain clients, or hoard relationships. Instead, the company treats each transaction separately and tries to match a customer with the most suitable resource. Its focus is on reserved compute, forward contracts, and token packages.
Verification covers both physical capacity and the ability to pay
Li says Compute Exchange has more than 100 verified providers and more than 80,000 GPUs in supply, with a stated 24-hour turnaround. For existing reserved capacity, it checks GPU UUIDs, inventory, and performance, and runs Nvidia DGX, InferenceX, and MLPerf tests. For future capacity, it reviews purchase orders and permits. It also performs KYB and KYC checks on buyers so providers can assess whether customers can meet scheduled payments.
Compute contracts can reserve future hardware or dedicated token capacity
The platform offers reserved contracts and forward contracts for GPUs and tokens. A GPU forward contract covers capacity that will become available later. A token forward contract is closer to a dedicated deployment, where a customer reserves a bare-metal GPU cluster and receives a dedicated token deployment on top. Li says the deployment can include fixed model SLAs, quantization, and precision for a defined period. The design addresses volatility in on-demand token API pricing.
Used GPU servers have become an active secondary market
Li describes a secondary server marketplace that began after owners asked her to take large quantities of H100s out of data centers. The market now includes used and refurbished hardware, including B200 systems, and some transactions have taken place above MSRP. Compute Exchange supports reserved GPU capacity from one node for one month through multi-year periods. Customers can also sell back excess commitments when their quantity or duration changes.
Requests for quote let buyers compare providers through one workflow
A buyer can submit requirements such as 300 B200 GPUs in US East starting January 1, and the platform routes the request to providers that appear to have matching capacity. Providers return pricing, specifications, and SLAs. Existing servers can be benchmarked. Li says price is only one consideration, so the platform puts the responses in one place and lets the buyer choose. The platform does not retain the relationship between buyer and provider.
Forward curves show a market expecting higher GPU prices
Li presents Silicon Data's forward curves for H100 and B200 capacity. The previous year's H100 curve sloped downward, which meant longer contracts were cheaper per shift or hour. Earlier in the current year, the curve moved upward and became contango, with less discount for longer commitments. The curve then shifted higher and flattened. Li says the implied forward rates, calculated from the term structure using a method similar to Treasury futures curves, show a market consensus that prices could rise over the following 36 months. She adds that this consensus can be wrong.
GPU residual values can differ sharply from accounting depreciation
Silicon Data estimates the value of GPU servers through discounted cash flow, using assumptions such as utilization and operating cost. It calculates residual values for A100, H100, and B200 servers and compares them with the five-year straight-line depreciation commonly used in accounting. Li says A100 values follow a downward curve but remain different from the accounting curve. H100 prices have risen, and B200 values have moved above MSRP. She also connects these valuations to financing and hedging decisions for servers held on balance sheets.