CME GPU compute futures launch: your AI bill gets a price

GPU compute futures are set to start trading today. CME Group and Silicon Data scheduled two contracts, the Silicon Data H100 Rental Index Futures and the Silicon Data B200 Rental Index Futures, for an October 5 listing on NYMEX, pending regulatory review. You will never trade one. But the price of a rented Nvidia GPU-hour now has a public reference, and that number sits under every AI invoice you pay.

What actually happened

CME announced the contracts in August. Each one tracks a Silicon Data index of hourly GPU rental costs, one for the Nvidia H100 and one for the Blackwell B200. Each contract represents about a month of rent for one GPU. They settle in cash against the index. Nobody delivers a server to your door.

Silicon Data builds the benchmarks from rental prices across cloud providers, hyperscalers, colocation data centers, and private rental markets. CME's Pete Keavey called compute "the currency of the AI age." Silicon Data CEO Carmen Li said the contracts give the market "a public, tradable reference price" for compute.

The limits matter. The contracts track two specific chips, not the full cost of an AI cluster. Power, cooling, networking, and the vendor's margin stay outside the index, as MLQ's write-up points out. And CME had not named liquidity providers or published every contract detail when it announced the launch.

Why GPU compute futures matter for your business

Hedge funds and cloud operators will use these contracts. You won't. What you get is the index: a public, daily number for what an H100 hour costs. That is new. Until now, the GPU price under your AI vendor's per-token rate was a guess.

Here is how we use it:

1. Watch the trend, not the tick. If the H100 and B200 indexes fall for months while your per-seat or per-token price stays flat, you have a renewal argument.
2. Read the futures curve as a forecast. Prices for contracts a year out tell you what the market expects compute to cost. Plan your 2027 AI budget against that, not against a sales deck.
3. Keep contracts short when the curve points down. A three-year AI commitment signed at today's rates is a bet that compute won't get cheaper. Most of the time, that's a bad bet.
4. Know what you're actually buying. If you rent GPUs directly for fine-tuning or self-hosted models, you now have a benchmark to check your provider's quote against.

The bigger point: compute is turning into a commodity with a price, like electricity. Commodities get cheaper to buy when you can switch suppliers. Build your stack so you can.

Sources: CME Group and Silicon Data press release, CME compute futures, CNBC.