AI spending keeps ballooning, yet figuring out how to price it remains a stubborn challenge. Nvidia reinforced the frenzy this week with a strong earnings report—even with tight supply, the company still projects roughly 70% revenue growth for fiscal 2028. At the same time, Nvidia is moving beyond pure chip sales, now offering financing to customers and taking a cut of rental revenue, leading Wall Street to increasingly compare it to a "central bank of the AI era." CEO Jensen Huang puts the shift more succinctly: "Today, compute itself is revenue."

Futures contracts: A new pricing tool for compute power

Wall Street is getting ready to assign tradable prices to at least some computing capacity. CME Group plans to launch futures contracts based on hourly rental costs of Nvidia's H100 and B200 graphics processors on October 5, pending regulatory approval. The contracts would be cash-settled against a benchmark index compiled by Silicon Data, a firm that tracks what companies actually pay to rent these chips. Currently, Silicon Data's H100 benchmark sits at roughly $2.68 per GPU-hour, while the newer B200 benchmark is around $5.66. Both have swung significantly over the past year, and their moves haven't always lined up with each other.

Compute is like hotel rooms: Wide price gaps

Compute rental is similar to hotel rooms—a Ritz-Carlton and a roadside motel both offer a bed for the night, but the prices are worlds apart. Chip type matters, but so do vendor, geography, lease length, network configuration, and availability. Futures contracts don't actually reserve chips; they're simply cash-settled based on benchmark price moves—even if hotels are full and prices spike, the contract won't get you a room.

Wall Street's mixed history of creating new markets

Wall Street's track record with novel derivatives is chequered. DRAM chip futures were proposed in 1989 but never launched; weather futures from 1999 are still around but remain niche. Bandwidth trading that same year tried to commoditize network capacity but never became the benchmark it aimed to be. Singapore Exchange tried DRAM futures again between 2002 and 2004, then gave up. Iron ore derivatives, however, took off around 2009–2010, riding the shift to index-based pricing, and ultimately succeeded by integrating into physical contracts. Whether GPU rental price futures can replicate iron ore's success is an open question. The reasons DRAM futures failed sound familiar today—a former exchange executive recalls: "The biggest hurdle was getting the whole industry to agree on what a standard chip even meant." Weather futures faced a different challenge: companies' actual risk exposures were too specific for standardized contracts to cover precisely. Bandwidth trading may be the closest reference. During the late-1990s fiber boom, Enron tried to commodity network capacity like energy, but the market never took shape. That era's fiber glut now serves as a warning for the current AI infrastructure boom—surplus capacity once crushed prices. Iron ore, by contrast, offers a success story. After decades of private annual negotiations, the market shifted to daily public indexes around 2009–2010, futures and swaps flourished, and those indexes eventually became widely used in physical deals. Compute futures, to succeed, would need to follow the same path.

Regulatory stance: CFTC supportive but cautious

U.S. Commodity Futures Trading Commission Chairman Michael Selig said: "Without a sound compute derivatives market, the U.S. cannot win the AI race." But the agency is also examining whether the market is standardized and transparent enough to support such contracts. The chairman wants the track built, but the agency is still verifying whether all parties agree on the car specs. CME still needs regulatory sign-off, and the CFTC's public comment period runs until October 20—later than the planned October 5 launch date. Competitor benchmark provider Yggdrasil Financial Technologies has warned the CFTC that privately compiled benchmarks could spark conflicts of interest or recreate a "LIBOR-style problem"—the former global lending benchmark exposed, after manipulation scandals, the risk of huge sums depending on a few people controlling a metric.

Investor perspective: The futures curve as a new signal

Even for those not trading, investors could glean valuable insight. Jessica Inskip, investment research director at StockBrokers.com, is particularly focused on how the futures curve reveals expectations for coming months. "Stocks and even commodities are two-dimensional—price goes up or down. The derivatives curve adds a third dimension: time. Compute can't be stored, and idle GPU hours are gone forever." That makes the futures curve a potential new gauge of AI buildout itself. "Nvidia's revenue and hyperscaler capex tell you what's been contracted; the compute curve tells you what's actually being consumed and what someone is willing to pay a year out." The divergence signal she watches: semiconductor stocks and spending plans keep climbing while H100 and B200 rental prices soften—potentially indicating capacity is running ahead of demand. But she stresses the signal only matters if genuine buyers and sellers participate. "Participation determines whether this is a real signal or just a sentiment index. If positions are all managed funds, then we've just built a noisier way to bet for or against Nvidia."