GPU Futures Would Hedge AI Compute Rental-Rate Risk
Silicon Data, which has raised $30.5 million, is building pricing and performance benchmarks that CEO Carmen Li says could underpin a risk-management market for rented AI compute. Li said CME’s planned cash-settled GPU futures, tied initially to Silicon Data’s H100 and A100 indices, would let server owners hedge falling rental revenue and compute customers hedge rising costs. She argues that chip depreciation alone does not determine value: residual value depends on the cash flow older hardware can still generate.
GPU futures would hedge rental-price exposure, not deliver chips
Carmen Li says Silicon Data’s benchmarks are intended to support a market for hedging the price of rented GPU capacity. CME’s planned products, she said, are scheduled for October 5 and would initially be cash-settled against Silicon Data’s H100 and A100 neocloud on-demand indices.
The contracts would not require anyone to deliver servers. In Li’s description, they address a different problem: the volatility of future rental rates.
A company that owns substantial GPU-server capacity is naturally long that exposure because its revenue depends on what it can charge to rent the machines. If rental rates fall, it could short futures to offset that decline. Physical delivery is beside the point, Li said, because the owner already holds the servers.
A compute customer has the inverse exposure. It pays rental rates and may want to lock in a cost for a specified period. That customer could go long the futures as protection against higher future rental prices.
If you are natural longs of GPUs, meaning you own tons of servers, obviously, your revenue is tied with the rental rates.
Silicon Data’s role in that arrangement is to provide the index on which cash settlement would be based. Li described the company’s broader aim as becoming an “independent referee” for the compute stack: developing indices and benchmarking services as more chips, tokens, and large-language models enter the market.
Silicon Data’s graphic says it tracks compute pricing and performance, creates industry benchmarks, and seeks to turn compute into an investable commodity. Li said the company plans to use the new funding to develop indices, products, and software applications.
Li sees risk management as the counterpart to AI infrastructure financing
Carmen Li framed the planned CME products as a risk-management layer alongside what she called the Nvidia “$500 billion story,” which she characterized as a financing-layer question. Financing and hedging are not interchangeable in her account: investors and companies with compute exposure need a way to discover prices and manage the risk attached to those exposures.
You can't have a market about half a trillion size without a way for the investor, the people with exposures, having a place for them to hedge and then price discovery.
That is the commercial case Silicon Data is making for its indices. The company says they are intended to support compute pricing and performance benchmarking, price discovery, and hedging for companies that own or rent compute capacity.
Its investor group reflects the market participants Silicon Data expects to serve. Silicon Data’s graphic listed Valor Equity Partners, CME Group, Samsung, DRW, and Sancus among investors. Li also named F-Prime, which she described as part of the Fidelity ecosystem, along with VanEck and Jump Trading.
Li said Valor’s Gavin Baker and his team were early backers of a number of neoclouds. More generally, she described the company’s investors as potential clients with positions on both sides of the market: the “natural longs” that own capacity and the “natural shorts” that consume it. Silicon Data wants its products to be aligned with and useful to both groups, she said.
Depreciation does not erase a chip’s economic value
Carmen Li rejected the idea that an older GPU’s depreciation necessarily means it no longer has economic value. Servers are machines that depreciate through their lifespan, she said, likening them to aircraft carriers and ships. But an asset’s residual value depends on the future cash flow it can still produce.
That distinction matters for older Nvidia hardware. Li said A100 and H100 prices had each risen about 20% since January and had been relatively stable over the preceding 20 days. Ed Ludlow noted that A100s appeared to retain high utilization and pricing despite being generations old.
Li’s explanation was utility. An A100 can still serve different workflows and smaller-model use cases, she said, allowing providers to keep charging whatever rate emerges where supply meets demand. She added that even L40s retain machine-learning use cases.
The relevant question, in Li’s account, is therefore not simply how old a chip generation is. It is what revenue the server can still earn from the workloads it remains capable of serving. Depreciation is real; it does not, by itself, settle the question of residual value.
Oversupply would show up in prices, expectations, and used-server values
Carmen Li identified three signals Silicon Data would watch for signs that AI infrastructure supply has begun to exceed demand: spot prices, the forward curve, and secondary-market server values.
Spot prices are the immediate meeting point of supply and demand. If they decline persistently, Li said, that could mean demand has weakened, supply is growing faster than demand, or both. She said that was not the current situation in Silicon Data’s market data.
The second signal is the forward curve—the relationship between short-term rental rates and prices for longer-term compute commitments. Li said the market was then in contango: customers were paying more to lock in longer-duration contracts than for short-term rentals. She did not regard that as evidence of oversupply. A sharply downward-sloping curve would be more concerning, because it would imply expectations that future rental rates would fall substantially.
The third measure is residual value as expressed through secondary transactions in servers. Falling resale prices, Li said, would indicate that market participants expect the future revenue from those machines to decline.
Together, the measures cover current rental prices, expectations embedded in longer-term commitments, and expectations reflected in resale values of existing hardware.