The CFTC is asking for public input on AI compute futures. That is not a signal of progress. It is a confession of ignorance. They do not know what they are regulating. And neither does CME.

This is not a critique of intent. It is an observation of structure. AI compute—GPU cycles, teraflops, training time—is not a commodity in the classical sense. It has no uniform grade, no transparent spot market, no physical deliverability that aligns with the existing legal framework. The attempt to wrap it in a futures contract is an act of financial engineering that will succeed or fail based on one variable: the integrity of the index.

Let me be clear. I have seen this movie before. In 2017, I audited over 50 ICO smart contracts. Twelve had critical reentrancy vulnerabilities. The market did not care about the code; it cared about the narrative. The same principle applies here. The narrative is "AI compute is scarce and needs hedging." The reality is that the underlying asset is not standardized, its supply is controlled by a single company, and its price is subject to technological obsolescence that no traditional commodity has ever faced.

We do not ride the wave; we engineer the tide. But engineering a tide requires understanding the fluid dynamics of the market. So let us examine the mechanics.

Context: The Birth of a New Derivative

CME Group, the world's largest derivatives exchange, is reportedly planning to launch AI compute futures in October 2025, pending CFTC approval. The product would allow participants to hedge or speculate on the price of AI computing power. The CFTC, in a rare preemptive move, has issued a request for public input, signaling that the regulatory framework for this novel asset class is still being constructed.

The context is straightforward. AI compute demand has exploded, driven by the proliferation of large language models and generative AI. NVIDIA's H100 and upcoming B200 GPUs are the gold standard, with rental prices fluctuating wildly based on supply constraints and the ebbs and flows of venture capital. Cloud providers like AWS, Azure, and Google Cloud offer compute instances at varying rates, but there is no centralized price discovery mechanism. This opacity is precisely the opportunity CME seeks to exploit.

But the opportunity is not without precedent. CME has successfully launched futures on Bitcoin, Ethereum, and even micro contracts. Each required a different approach to index construction. Bitcoin futures use a reference rate from multiple exchanges. Ethereum futures follow a similar model. AI compute, however, lacks a comparable set of transparent, liquid spot markets. The data must be sourced from private data centers, cloud service providers, and potentially NVIDIA itself. This is not a trivial data aggregation problem; it is a structural conflict of interest.

Core: The Five Pillars of AI Compute Futures

Pillar One: Index Construction – The Achilles' Heel

Any futures contract requires a reliable benchmark. For AI compute, the natural unit is the dollar per GPU-hour for a specific GPU type (e.g., H100). But GPU types proliferate. An H100 is not an A100, and a B200 will render both obsolete. The index must either be a composite or a single-grade contract. A composite index introduces complexity and basis risk. A single-grade contract faces a limited lifespan as the underlying GPU becomes obsolete. The only way to make this work is to have a dynamically updating index that reflects the current market basket of GPUs, weighted by usage. But that requires continuous data collection from a small number of suppliers. And those suppliers have no incentive to provide accurate, timely data that could be used against them in price negotiations.

Based on my experience auditing the data feeds of centralized lending protocols in 2020, I recognize this pattern. The data source is the single point of failure. In DeFi, oracles were the weak link. Here, the oracles are the private pricing feeds of hyperscalers. If the index is not transparent, auditable, and resistant to manipulation, the contract will be a tool for insiders, not a hedging instrument for the industry.

Pillar Two: Liquidity Dynamics – The Bootstrap Paradox

New futures contracts face a chicken-and-egg problem. Without liquidity, hedgers cannot execute large orders without moving the price. Without hedgers, speculators have no reason to provide liquidity. CME will likely use market maker programs to bootstrap volume, but that is a temporary solution. The real test is whether commercial participants—the actual consumers and producers of AI compute—will use the contract. Cloud providers may not want to hedge; they already have natural long positions and can adjust pricing dynamically. AI companies, especially startups, may lack the sophistication or capital to engage in futures hedging. The initial liquidity may come from hedge funds treating it as a pure beta play on AI hype. That is a fragile foundation.

Consider the historical parallel: CME's Bitcoin futures launched in 2017 with high initial volume, but much of that was speculative. It took years for institutional hedgers to emerge. AI compute futures may face a similar timeline, but the underlying asset is more complex and the user base more concentrated. The liquidity risk is not just about low volume; it is about the quality of that volume. A market dominated by speculators provides poor price discovery for the physical market.

Pillar Three: Participant Behavior – The Asymmetry of Power

The natural shorts are data center operators and cloud providers who want to lock in revenue. The natural longs are AI companies who want to lock in costs. But the asymmetry is stark. Cloud providers have pricing power. They can simply raise prices. AI companies, especially those reliant on venture capital, are price takers. The futures market may become a one-way bet, with chronic contango or backwardation depending on the supply-demand imbalance. That is not a healthy market; it is a subsidy mechanism.

Furthermore, the participation of NVIDIA is critical. As the dominant supplier of AI GPUs, NVIDIA could choose to support the futures market by hedging its own revenue or by providing data for the index. Alternatively, NVIDIA could ignore the contract, or even undermine it by offering bilateral long-term contracts that bypass the futures curve. The outcome depends on NVIDIA's strategic calculus. If NVIDIA sees the futures market as a way to stabilize its own cash flows, it will participate. If it sees it as a threat to its pricing power, it will not. The market's success hinges on this decision.

Pillar Four: Regulatory Risk – The Commodity Definition Trap

The CFTC's public input request is a warning shot. The commission will scrutinize the index methodology, the potential for manipulation, and the suitability of the contract for commercial hedging. The critical question is whether AI compute qualifies as a "commodity" under the Commodity Exchange Act. If it does, the CFTC has jurisdiction. If not, the product may need to be structured differently. The definition of a commodity is broad, but there is a subtlety: the underlying asset must be capable of being delivered. For AI compute, delivery is a conceptual nightmare. You cannot deliver a GPU hour the way you deliver a barrel of oil. The settlement will almost certainly be cash-settled, which means the contract is a derivative on a price index, not a true commodity future. This has implications for margin, tax treatment, and the legal status of the product.

The CFTC's opinion is likely to be positive, given the agency's historical openness to innovation. But the public comment period could reveal significant opposition from consumer groups or industry incumbents who fear that financialization will increase compute costs. The regulatory timeline is uncertain, but a delay beyond October 2025 is possible. The key signal to watch is whether the CFTC publishes a proposed rule or a concept release. A concept release would indicate a longer deliberation period.

Pillar Five: Technological Decay – The Moore's Law Problem

This is the most overlooked factor. Compute prices do not follow a mean-reverting process like oil or wheat. They follow a deterministic downward trend driven by Moore's Law. The cost per unit of compute has historically declined by 30-50% per year. A futures contract that prices compute for delivery in two years must account for this. Traditional commodity futures are influenced by storage costs, convenience yields, and seasonal patterns. AI compute futures are influenced by the release of a new chip. The curve may be in structural contango, with spot prices high due to scarcity and forward prices declining due to anticipated technological improvements. But this is not a natural contango driven by carry costs; it is a reflection of technological progress. The pricing model must incorporate this. If it does not, the contract will be systematically mispriced.

Consider the H100. Its rental price has dropped from over $30 per hour in early 2024 to about $15 per hour in mid-2025, a 50% decline. This is not a cyclical correction; it is a structural shift driven by the anticipation of the B200. A futures contract that does not embed this decay will always be in contango, but the contango will be too steep for naive models. The margin requirements will need to adjust dynamically, and the risk of a sudden jump in backwardation (if a new chip is delayed) is real. This is a derivatives market that requires a new risk model, one that incorporates semiconductor roadmap data as a pricing factor.

Contrarian: The Decoupling Thesis

The contrarian view is that AI compute futures will not behave like any commodity we have seen. They will behave more like a derivative on a technology monopoly. NVIDIA controls over 80% of the high-end AI GPU market. The price of compute is effectively set by NVIDIA's pricing strategy, chip allocation, and product roadmap. A futures market that depends on spot prices from cloud providers is at the mercy of NVIDIA's decisions. The index will reflect NVIDIA's pricing power, not the true equilibrium of supply and demand.

Furthermore, the market may decouple from the underlying physical reality. Just as Bitcoin futures are traded based on sentiment and macro flows, AI compute futures may become a proxy for AI enthusiasm. The futures price may diverge significantly from the actual GPU rental market, creating arbitrage opportunities for those with access to physical compute, but increasing the noise for everyone else.

The real risk is not price volatility; it is the structural obsolescence of the index itself. If the index is based on a specific GPU, say the H100, and a new chip like the B200 is released, the H100 price could drop 50% overnight. The futures contract would be left with a decaying asset. The margin requirements would need to be recalculated, and the hedging effectiveness would be destroyed. This is not a risk that can be managed with standard risk models. It is a technological event risk.

Collateral is just debt wearing a mask of trust. In this case, the collateral is the reliability of the index. If the index fails, the trust evaporates.

There is also a geopolitical dimension. Export controls on advanced GPUs to China create a bifurcated market. The futures contract is likely to be based on Western pricing, but the existence of a parallel market in China (with lower prices due to restricted supply) could create arbitrage opportunities and distort the index. The CFTC will need to address this, perhaps by excluding Chinese data or by creating a separate contract for the Asian market. The complexity is high.

Takeaway: The Tide Must Be Engineered

The market will not price AI compute; it will price the narrative of compute scarcity. The true arbitrage is between engineering capital and financial capital. Those who understand the engineering—the silicon design, the cluster architecture, the thermodynamic limits—will have an edge. Those who only see the financial flows will be left with a derivative that is only as good as its underlying assumption.

We do not ride the wave; we engineer the tide. But the tide of AI computing is not a smooth sinusoidal function. It is a step function, driven by the release of new chips. The futures contract must be designed to survive those steps. Otherwise, it will be a footnote in the history of financial innovation.

The question is not whether CME can launch the product. It is whether the market can build a consensus on what "compute" means. Until that consensus exists, the futures contract is a structure built on sand.

From a macro perspective, the timing is favorable. The Fed is in a rate-cutting cycle, which lowers the cost of carry for futures positions. Liquidity is abundant, and institutional investors are hungry for new asset classes that offer uncorrelated returns. AI compute futures could fill that gap, but only if the product is robust enough to withstand the inevitable technological shocks.

My advice to clients is simple: wait. Let the index be tested. Let the first year of trading reveal the flaws. The early adopters will be the market makers and the speculators. The real hedging will come later, once the basis converges and the index gains credibility. The opportunity is not in the first trade; it is in the first correction. When the market panics over a chip announcement, the futures will show their true character. Then, and only then, will we know if this is a tool or a toy.

I will be watching the data sources. I will be watching the open interest. I will be watching NVIDIA's conference calls. The signals are there. The market is not a teacher; it is a mirror. And this mirror will reflect the engineering of the index, not the promise of AI.

Collateral is just debt wearing a mask of trust. Trust the index, and you trust the data. Trust the data, and you trust the oligopoly. The only way to engineer the tide is to build a market that can survive the ebb.