Wall Street Wants to Trade GPU Hours Like Oil

CME Group and Silicon Data are racing to turn hourly GPU rental rates into cash-settled futures contracts — and Wall Street is already lining up ETFs to trade them before the Commodity Futures Trading Commission even approves the product.

Why now: AI companies rent compute the way airlines buy jet fuel — and neither can forecast next year's bill. Larry Fink said as much at the Milken Institute on May 5: "A new asset class will be buying futures of compute." Seven days later, CME and Silicon Data announced exactly that.

The bet:

• Silicon Data publishes daily GPU rental benchmarks across A100, H100, and B200 chips — the same indices SpaceX cited in its IPO prospectus.
• CME plans cash-settled futures based on those indices, pending regulatory review.
• Roundhill filed the Roundhill Compute ETF (GPUX) with the SEC to hold compute futures exposure; CNBC reports ProShares and Rex Shares filed similar proposals, including leveraged and inverse products.

The scale: Silicon Data CEO Carmen Li told CNBC she expects compute futures to become "larger" than oil — arguing AI energy demand will eventually exceed all other uses combined. BlackRock's Fink framed the shortage differently: "We're short power, we're short compute, we're short chips."

The friction:

• Unlike WTI crude, an H100 is not a standardized barrel. Silicon Data says more than 50 H100 configurations trade at different hourly rates depending on memory, networking, and data center location.
• The company normalizes daily spot prices to a "base H100 case" before calculating its index — a step oil markets never needed.
• Santa Clara University finance professor Seoyoung Kim told CNBC the CFTC "is going to want to know exactly what the product is" before signing off.

Who hedges, who speculates:

• Natural buyers: AI labs and neoclouds locking in training costs before the next RAM shortage hits.
• Natural sellers: Hyperscalers and data center operators with long GPU inventory fearing a price crash.
• Speculators: Traders betting on scarcity rents without touching a single chip — the same liquidity engine that made oil futures deep enough to matter.

Li welcomed the speculators: "You need natural hedgers. You need market makers. You need speculators."

The IPO connection: SpaceX's prospectus already treated Silicon Data's GPU rental data as disclosure-grade pricing. Google's $920 million monthly contract for 110,000 Nvidia GPUs shows why — at that scale, a few cents per GPU-hour is billions in margin.

What changes if it works: Compute stops being an opaque cloud line item and becomes a priced, hedgeable input — like fuel, grain, or electricity. The AI buildout's $650 billion capex story becomes a derivatives story too.

Convina's view: Fink called the trade. CME is building the plumbing. The real fight is over the benchmark — whoever defines the standard GPU hour will tax every model query, every neocloud contract, and every IPO prospectus that cites rental rates as proof of revenue. Silicon Valley spent a decade arguing AI was software. Wall Street just decided it is a commodity — and commodities get financialized before they get regulated.