CME plans to settle GPU futures against the index of a startup that has raised $35.2m

Carmen Li puts early-2026 inference prices between roughly $1.20 and $1.55 per million tokens ^[3]. That is a 29% range measured off the low end ^[1]. A buyer consuming a trillion tokens pays $1.2m at the bottom and $1.55m at the top, a difference of $350,000 ^[2].

The exposure she describes and the contract CME plans are not quoted in the same unit. Her volatility figure is a price per million inference tokens ^[3]. The planned products, as cryptobriefing describes them, are cash-settled futures and options referencing Silicon Data's benchmarks ^[6], and the flagship benchmark is named for one chip generation, the H100 Rental Index ^[2]. A company whose invoice arrives per million tokens and hedges with an H100 rental contract is holding two different prices.

Cash settlement also makes Silicon Data's published print the payout. The indices aggregate hundreds of thousands of data points ^[2]. The company's own venue, Compute Exchange, has reported hundreds of millions of dollars in notional volume since launch, alongside more than 1,000 registered users spanning AI developers and financial institutions ^[4].

Li has put the potential derivatives market for compute at $6 trillion to $30 trillion, against the hundreds of billions already going into data center capital expenditure ^[8]. Read the exchange's reported notional as $600m and the low end of her range is 10,000 times that ^[3]. She has raised $35.2m in total, $4.7m of seed in 2025 and a $30.5m Series A in August 2026 ^[5]. She started the company in April 2024 ^[1]. That leaves about 30 months between incorporation and the targeted listing ^[4].

Maybe CME, which cryptobriefing calls the world's largest derivatives exchange ^[7], lists the contracts in the fall and GPU owners turn up to sell next quarter's rental revenue forward. Then the index becomes the reference price for compute, and Silicon Data licenses data off it for years. Or the contracts list and barely trade, because the firms with compute exposure would rather sign term capacity they can consume than collect a cash difference. Or approval takes longer than the fall 2026 target, which the company has already named as a condition ^[6].

I would expect the second, at least through the first year of listing. The counter-argument is a good one. A buyer who cannot get capacity at any price still has a price to manage, and the step from a $4.7m seed to a $30.5m Series A inside roughly a year ^[5] suggests someone with a balance sheet already tested that demand. What would show me wrong is open interest. I mean real positions thirty days after listing, held by firms that own the hardware and not only by market makers quoting both sides.