U.S. Tells Prediction Market Kalshi to Take Down Its AI Chip Price Charts

The U.S. Commerce Department last month told Kalshi, a federally regulated exchange where users trade on event outcomes from elections to inflation prints, to take down its non-tradable price-reference charts tracking the implied hourly rental cost of Nvidia's B200, H200, and A100 AI chips, citing national security, according to a Semafor report citing people familiar with the matter.

The order hit the reference layer, not the trades. The charts were benchmarks only, built from Kalshi's underlying weekly and monthly markets on renting the same chips. Those underlying markets are still open. The forward-curve layer, used to negotiate compute deals and hedge price risk, is gone. In commodities, that kind of price index tells a refiner where oil is expected to trade in three or six months. In compute, it would have told a neocloud, a hyperscaler, or an AI lab where the market thinks hourly rental of a B200 will price by year-end. Compute here means renting AI-capable chips by the hour.

A non-tradable chart, with no contracts to settle and no money at risk, is an unusual target for a national-security order. It publishes where the market thinks hourly chip rental will price in the months ahead, then goes away. Commerce's reasoning is not on the record, and Semafor's reporting does not name a statute. Market participants told the outlet that one concern is that a thinly traded compute market could be manipulated to make it look like older-chip prices are falling, potentially moving the AI stocks that those chips anchor.

Kalshi launched the compute forward curves in July 2026 for the three Nvidia chips. The B200 and H200 are the current generation used in large model training. The A100 is an older workhorse still widely rented for inference and smaller jobs. CEO Tarek Mansour said in a press release that compute is "the new oil" and called for a derivatives market around it. The company pitched the benchmarks to the buyers and renters of large AI training capacity as a transparent price reference for a market that had been largely negotiated behind closed doors.

In parallel, Commerce pushed the CFTC to freeze approval of new compute-derivatives contracts for 60 days, and the CFTC opened a 60-day public comment period on compute derivatives in August. CFTC Chairman Michael S. Selig said in the agency's release that robust compute derivatives are necessary to "win the AI race." That phrasing puts compute pricing in the same strategic bucket as chip export controls and domestic fab capacity, not the commodity bucket of copper or wheat.

For the buyers Kalshi was pitching, the practical effect is that the same hedging logic still has to be done deal by deal, in private. For the broader market, the unresolved question is what the regulator thinks a national-security risk looks like in a non-tradable chart. Semafor's reporting points to manipulation risk as the working theory, but the order itself does not say so on the record, and Commerce has not published a public rationale.

Together, the take-down and the CFTC freeze treat compute pricing as a strategic asset rather than a commodity benchmark. Hourly chip rentals had been an opaque, bilateral market between hyperscalers and a long tail of neoclouds. A public forward curve would have made the implied future price of renting a B200 as visible as a Treasury yield. The take-down keeps the rental trades open but takes the reference layer off the table for now.

The next visible date is the close of the CFTC's public-comment window later this fall. What the agency does at the end of it will determine whether Kalshi's forward curves return in a different form, or whether the reference layer for AI compute pricing gets built somewhere else first.