CME's 36-month GPU futures put a forward price on the compute bitcoin miners now sell

Silicon Data's index already gives a spot rent for Nvidia chips ^[3]. The listing adds a price for each month out to 36 ^[2]. For an older chip, most of what that curve says will be a forecast of decay. The H100 lost more than three quarters of its hourly rent between the 2024 shortage and late last year ^[1], and the B200 now earns about 2.1 times as much per hour ^[2]. At today's H100 rate and full use, one chip would bring in about $72,800 over three years if the price never moved ^[3]. The gap between that flat figure and the sum of CME's 36 monthly prices would be the market's estimate of how much H100 revenue newer chips take away.

Cryptopolitan's case is that the miners already own the underlying and could hedge their compute revenue with these contracts ^[9]. CoinShares describes TeraWulf, Core Scientific, Cipher Mining and Hut 8 as data center operators who also happen to mine bitcoin ^[16]. It expects AI to supply between 30% and 70% of listed miners' revenue by year-end, against about 30% now ^[7].

The contracts settle on Nvidia silicon ^[3], so they price GPU-hours. A miner that owns the chips and rents them by the hour holds that exposure directly and can sell it forward. Now take a miner that leases power and floor space, at a fixed multi-year rate, to a tenant who owns the chips. That miner has already sold its output forward. For it, the futures say something about the tenant's margins and about the rate it can ask at renewal. Cryptopolitan does not break down how the $70 billion of contracts ^[6] divide between those two models. CME is pricing Nvidia GPU-hours, not megawatts or rack space ^[3].

The size estimates around the launch are large, and some come from people who would profit from the market. Boston Consulting Group projects the compute market growing from about $360 billion in 2025 to nearly $2.3 trillion in 2030, according to the Financial Times as cited by Cryptopolitan ^[11]. That is roughly 45% a year ^[5]. Brett Harrison, chief executive of the derivatives platform Architect, estimates $10 trillion of notional a year by the end of the decade ^[12], about 4.3 times BCG's 2030 market ^[6]. Silicon Data is funded by DRW ^[3], whose founder Don Wilson has predicted that compute will become the world's largest commodity, according to the report ^[14]. "Compute has become the currency of the AI age," said Pete Keavey, CME's global head of energy and environmental products ^[13].

I see two ways this goes. CME's curve may become the single reference, with analysts covering converted miners citing an exchange price for H100 rent in 2028. Or volume splits between CME and ICE, which plans futures on another index ^[10], and investors get two thin curves that may not agree.

I think investors will use the curve before miners do. An analyst can set a company's projected H100 revenue against an exchange price the day the contracts list. A miner can hedge only when someone will take the other side three years out, and only if its revenue is priced by the GPU-hour. The counter-case: if most converted capacity sits in fixed leases, the curve values the tenants' chips and says little about a megawatt of hosting. The view is wrong if trading never gets past the first year, because then the 2028 and 2029 prices ^[2] have no volume behind them.