Everything About AI
NextFin News - Wall Street's race to turn artificial-intelligence computing power into a tradable commodity hit a regulatory wall this week. On Sept. 21, the US Commodity Futures Trading Commission extended its review of CME Group's planned compute futures by 45 days, pushing the earliest possible launch well past the exchange's Oct. 5 target date and into mid-November. The letter, signed by DJ Hennes, acting director of the CFTC's Division of Market Oversight, cites "the need for additional time to analyze the Products in light of novel or complex issues" — a formulation that signals the regulator is not prepared to bless a derivatives market built on benchmark prices that do not yet exist in any standardized, manipulation-resistant form.
The postponement is the most concrete evidence yet that the infrastructure for financializing AI compute is running ahead of the plumbing. CME can build the exchange, the clearinghouse, and the contracts. What it cannot manufacture is the underlying price.
The Postponement and What the Letter Actually Says
The CFTC acted under Commission Regulation 40.3(c)(2), extending the review period for New York Mercantile Exchange Submission No. 26-370 and Submission No. 26-370S — the filings covering the proposed Silicon Data H100 Rental Index Futures and Silicon Data B200 Rental Index Futures — for an additional 45 days after the conclusion of the initial review period. The extension runs until the end of Nov. 9, 2026. The letter was addressed to Maureen Guilfoile, managing director and chief regulatory counsel at CME Group, and directs questions to the division's staff.
Two details in the letter matter more than the delay itself. First, the CFTC explicitly tied the extension to a pending Request for Comment on the Listing of Compute Derivatives Contracts, published in the Federal Register on Aug. 21 with a comment period closing Oct. 20. The regulator is not just reviewing two contracts; it is using these filings to build a framework for an entire asset class. Second, the extension mechanism is procedural, not substantive — Regulation 40.3(c)(2) allows the division to take more time on products that raise novel or complex questions, but it is not a rejection and does not preclude approval.
CME and Silicon Data, the benchmark provider backed by global trading firm DRW, announced the products in May 2026 as a "first-in-class compute futures market." The contracts are financially settled, each covering 730 GPU-hours — one GPU running for one month — with monthly contracts listed out 36 months. They trade under the codes GPU1 (H100) and GPU2 (B200) and track neocloud, or non-hyperscaler, indices of hourly on-demand rental costs published by Silicon Data. The Oct. 5 launch date was always conditional on regulatory review. Nov. 9 is the end of the review period, not a launch date: nothing lists automatically when the clock runs out, and the CFTC can still raise objections or request changes.
Why Regulators Are Hesitant: The Benchmark Cannot Be Anchored
The core problem is that GPU rental prices are not oil prices or Treasury yields. Crude has a deep, liquid, executable spot market with thousands of daily transactions and transparent grades. GPU capacity is priced through private, bilateral contracts across a fragmented market of hyperscalers, neocloud providers, brokers, and enterprise buyers. There is no single executable market price, and the reported rates are often indicative rather than transaction-verified.
A cash-settled futures contract must settle against a benchmark that is representative, manipulation-resistant, and durable. The CFTC's "novel or complex issues" language maps directly onto all three tests. In a market where most capacity is privately contracted, a trader holding a large futures position could benefit from influencing the underlying index — or at least from exploiting the thin, non-executable nature of reported rental rates. Cash-settled energy and commodity contracts face similar risks, but those markets have deep spot markets that anchor settlement. Compute does not.
There is also hardware obsolescence. A GPU rental index built on H100 pricing today will look materially different in 18 months, when newer architectures displace it. The index provider must decide how to handle product transitions — rolling the benchmark forward, weighting multiple generations, or letting contracts reference aging hardware whose rental market has thinned. Each choice changes what the contract actually hedges, and the CFTC has no settled precedent for how a derivatives contract should age out of the hardware it references.
The volatility of the underlying market makes the hedging need real — and the benchmark risk acute. According to Silicon Data's NeoCloud GPU rental indices through Aug. 14, 2026, H100 rental prices recorded a maximum drawdown of 45.3% and a maximum run-up of 50.5% between September 2024 and August 2026, with annualized realized volatility of 26.2%. B200 rental prices swung 19.7% down and 42.4% up between August 2025 and August 2026, with 30.2% annualized volatility. Even the older A100 moved 26.8% down and 23.9% up over the same period. These are not small moves: providers and renters on the wrong side of an H100 swing absorbed a near-50% adverse price change. That is precisely the exposure that makes a hedge attractive — and precisely the volatility that makes a benchmark hard to trust.
The Second-Order Question: A Reference Price That Creates Reality
Here is the implication the market is not fully grappling with. If a CFTC-regulated compute futures contract launches and gains liquidity, the benchmark it settles against will not merely reflect the GPU rental market — it will begin to define it. Enterprise buyers negotiating private contracts will start referencing the futures price. Lenders underwriting data center projects will model against it. The index becomes the market, not the other way around.
That is exactly how crude oil, natural gas, and electricity derivatives evolved: the futures curve became the reference price for physical deals, and the physical market reorganized itself around the financial benchmark. For compute, the question is whether a benchmark built on a thin, fragmented, privately priced spot market can carry that weight without distortion. The risk is not that the futures market fails to launch. The risk is that it launches successfully and then becomes the price — a price that reflects the liquidity and positioning of financial participants more than the economics of actual GPU rental.
The scale of the exposure explains why everyone is pushing ahead despite the uncertainty. With roughly $1 trillion in AI capital expenditures being deployed, an enormous stock of hardware is being built whose economic return depends entirely on the rate at which its services can be rented. That creates a genuine, balance-sheet-level risk factor — the rental rate for GPU capacity, which lands in the P&L of every participant in the AI economy, as an expense for those who consume compute and as revenue for those who supply it.
The Competition: CME vs. ICE and the Framework Fight
CME is not alone. Intercontinental Exchange is developing competing compute futures products based on Ornn's Compute Price Index and NativX's COIL Index. Multiple benchmark providers are healthy for competition but complicated for regulation: if the CFTC approves a framework for compute derivatives, it will effectively be choosing which methodology defines the market. The CFTC's Request for Comment asks about benchmark reliability, market manipulation, settlement methods, liquidity, and how a standardized unit of compute should even be defined. The outcome will shape not just CME's product but the entire category.
The industry's demand-side argument is straightforward. AI companies are signing multi-year, multi-billion-dollar compute contracts and need a way to hedge price risk. Data center operators and cloud providers face the mirror image of that exposure. Without a hedging vehicle, both sides are simply betting on the direction of GPU rental prices.
As the backbone of the digital economy, compute is the new oil of the 21st century. Every AI model trained, every transaction cleared, and every byte of data processed runs on compute, which is becoming a fast-emerging asset class in its own right. Investors need a trusted futures market to provide transparency, liquidity and effective risk management - all of which fall squarely into CME Group's wheelhouse.
Carmen Li, chief executive of Silicon Data, framed the problem in market-structure terms when the partnership was announced.
Compute markets today are still highly fragmented, with pricing that can vary dramatically across providers, regions and contract structures. At Silicon Data, we built our benchmarks to bring consistency, transparency and real-time visibility to GPU markets that have historically lacked standardized reference pricing.
Don Wilson, founder and chief executive of DRW, which backs Silicon Data, went further.
It has been clear to me for some time that compute will become the largest commodity in the world. The exponential growth in spending on data centers as we move towards that reality has been hampered by the lack of a hedging vehicle.
Cyclical Delay, Structural Problem
The right read is a cyclical regulatory delay layered on top of a structural problem. The 45-day extension is a routine procedural move — the CFTC routinely extends reviews of novel products, and extensions do not imply rejection. Extensions of this kind are common for first-of-their-kind filings, and the division's letter leaves the door open. That is the cyclical leg: the launch will likely happen, just later.
But the structural leg is real and unresolved. The benchmark problem — fragmented pricing, manipulation risk, hardware obsolescence, the absence of a deep executable spot market — does not go away on a calendar. It gets solved only if the market matures: if enough transparent, executable rental transactions accumulate that an index can be anchored to real trades rather than indicative quotes. Until then, any approved contract carries a settlement-integrity risk that no amount of review time fully eliminates.
The strongest counter-thesis is that none of this matters. Futures markets routinely launch on thin underlying markets and bootstrap liquidity through financial participants who have no intention of physical delivery. Crypto futures, volatility products, and single-stock futures all proved that a tradable contract does not need a perfect spot market to succeed. If speculators and hedgers provide enough volume, the benchmark becomes liquid enough to be reliable — the tail wags the dog into shape.
That argument has force, but it carries a specific failure mode. It works only if the index methodology is transparent and the underlying rental data is auditable against actual transactions. If the benchmark is opaque, or if reported rates cannot be verified, then the contract is vulnerable to exactly the manipulation the CFTC flagged — and the "tail wags the dog" dynamic becomes a manipulation opportunity rather than a liquidity solution. The falsifying signal is concrete: if, within six months of launch, fewer than half of the reported benchmark rates can be tied to executed, verifiable rental transactions, the counter-thesis fails and the manipulation concern becomes the dominant risk.
What to Watch
Three signals will tell the story from here. First, the Oct. 20 close of the CFTC's comment period and the substance of feedback from exchanges, benchmark providers, cloud companies, and trading firms. Second, whether the CFTC raises additional objections after Nov. 9 or allows the products to proceed — the end of review is not an approval, and the division can still request changes to contract terms or settlement methodology. Third, whether ICE's competing products advance on a parallel track, which would indicate the regulator is building a category-wide framework rather than adjudicating one exchange's filing.
For AI companies and cloud providers waiting for a hedging tool, the practical takeaway is patience. A compute futures market is coming, but the first version will be narrow: a few hardware generations, cash settlement, and a benchmark whose representativeness every user must test against their own actual rental contracts before relying on it as a hedge. A company running 50 H100s around the clock pays roughly six figures a month in rental costs, but the futures contract covers a standardized neocloud index — and if that company's actual rates are hyperscaler-negotiated or region-specific, the hedge will not track its real exposure.
The deeper lesson is that Wall Street can build a market for anything — but it cannot manufacture the underlying price. Compute futures will trade eventually. The question is whether they will price compute, or merely price the market's guess about compute.
Explore more exclusive insights at nextfin.ai.
