Indeed, Wall Street is already springing into action. CME Group CME.O plans to launch compute futures this month, linked to Silicon Data’s indices. Intercontinental Exchange ICE.N is planning its own version, using data from Ornn.
Much as oil drillers want more clarity on returns from their big up-front investments in opening a new spigot, data-center builders could benefit from hedging against future price drops. And just as an airline wants to be able to protect against rising costs to fuel its planes, Anthropic could use a way to guard against sudden spikes in rental prices. Given that the Claude maker is paying the equivalent of $50 billion per gigawatt of computing capacity to xAI, compared to much lower implied rates charged by server farm operator CoreWeave CRWV.O, a transparent benchmark could also help when negotiating oddball bilateral deals.
One complication is that, unlike grain or oil futures that involve delivering the underlying product, an hour of GPU usage cannot be held in storage somewhere. ICE’s planned futures are instead cash-settled, meaning that money simply changes hands at the end of the contract, rather than the good itself.
Furthermore, no two chips are quite alike. AMD AMD.O also makes GPUs. Intel INTC.O and ARM design CPUs. Alphabet GOOGL.O, OpenAI, Amazon and Microsoft are all lighting up their own specialized, in-house silicon. Nvidia, though, is easily the industry benchmark, with an estimated market share of over 80%. A custom chip used by one or two customers may be technically powerful, but holds little value for a broad market of buyers who would need to rejig their systems to use it. The H100, by contrast, is practically plug-and-play.
Little wonder, then, that CME and ICE’s contracts are tied to rental prices for Nvidia’s chips, for now. That’s a coup for Huang, who has pushed to remain the de facto industry standard. Those efforts span everything from investing in upstart cloud providers, helping to ensure there are plenty of operators supplying its wares to the market, all the way to a $500 billion plan to provide customer financing from private lenders like Apollo Global Management APO.N and Blackstone BX.N.
Becoming the anchor for benchmark compute prices can further that plan. Sure, big technology companies like Microsoft can fund AI infrastructure with cash flow or unsecured debt. But AI labs like Anthropic or smaller cloud operators like CoreWeave, without the benefit of stellar credit ratings, need help. Lenders have already agreed to some loans backed by GPUs, and Nvidia’s silicon is the most dependable collateral around. Yet making this common practice without the lure of a contract to serve a blue-chip customer like Meta Platforms META.O is tricky. A particular issue is that the long-term value of an H100, when Huang’s company regularly releases new and speedier chips, is uncertain.
Futures are a neat way for lenders to hedge against that risk. Assuming there’s $3.6 trillion of AI infrastructure investment from 2026 through 2030, BCG analysts estimate that a reliable forward curve of compute prices could help reduce borrowing costs by about $116 billion in total, or roughly $26 billion a year.
Of course, as Alphabet or Amazon try to broaden adoption of their own chips, competing futures could badly fragment the market. Moreover, an H100 hour in Texas is not necessarily equivalent to – or as valuable as – one in Virginia or Europe. Big, bilateral contracts struck at ad hoc terms might render spot prices of surplus capacity less meaningful. And, most importantly of all, this all redounds largely to the benefit of Nvidia specifically. That’s unlike oil, where barrels of the same grade are essentially equivalent, whether provided by Chevron or Exxon. If compute is the new crude, Huang is a baron like no other.
NVDA•
