The birth of GPU futures
Why the financialization of AI compute is following the same path as oil, power, and LNG
We are in the middle of the largest capital expenditure cycle in the history of computing. The hyperscalers—Microsoft, Amazon, Google, Meta, Oracle—are spending hundreds of billions of dollars to build out AI data centers. These campuses cost $10-20 billion apiece, and the growth trajectory shows no sing of slowing. For now, these buildouts are financed on balance sheet or via corporate debt issuance. But if history is any guide, such enormous flows of capital inevitably attract hedging markets. Commodities as diverse as oil, power, freight, and bandwidth all went through the same cycle: bespoke bilateral deals at first, then forward contracts, then standardized futures.
A futures exchange for GPU compute may sound exotic, but it is exactly how financial markets have evolved every time a scarce, volatile, and capital-intensive input became critical to industrial growth. The conditions for a GPU futures market are already forming. I wouldn’t go as far as to say that a futures market is inevitable, but I do think it’s worth paying attention to the tailwinds.
Tailwind #1: Hyperscaler Capex and the Cost of Capital
The basic economic driver is simple. These data centers are financed like any other capital-intensive project. Weighted average cost of capital (WACC) matters. If a developer can shave even 50-100 basis points off their WACC, the savings are measured in the billions.
Right now, GPU procurement is lumpy, volatile, and opaque. Prices swing with each Nvidia launch, each export control, each supply chain hiccup. That volatility translates into financing friction. If lenders had a transparent forward curve for GPU compute, akin to what they have for oil or electricity, they could underwrite with more confidence. More confidence translates into lower interest rates.
This is the same structural force that drove the birth of electricity futures in the 1990s, LNG futures in the 2000s, and freight indices in the 2010s. Once an input becomes big enough to move balance sheets, hedging pressure builds until someone solves the plumbing.
Tailwind #2: Don Wilson and the Market Maker DNA
The second tailwind is institutional. Don Wilson, the founder of DRW, is behind Compute Exchange, an auction platform for GPU capacity. That matters because markets do not spontaneously self-organize. They need market makers willing to stand in the middle, quote both sides, and take heat while liquidity builds.
DRW did that in rates, energy, and crypto. Wilson knows the mechanics of building a market where none exists: auction design, clearing rules, designated market-maker incentives, and the crucial early tolerance for illiquidity. Compute Exchange can nurse a commodity from bilateral forwards into standardized futures.
Tailwind #3: Political Cover from the Trump Administration
The thrid tailwind is political. The Trump Administration’s AI Action Plan, while only advisory, explicitly recommends the development of a spot and forward market for GPU compute. That is extraordinary. Commodities usually grow in the shadows, then fight regulators once they’re big enough to matter. In this case, Washington is already blessing the idea.
That matters for two reasons. First, it creates political cover for an exchange to self-certify a futures contract with the CFTC. Second, it signals to institutional players—banks, insurers, project financiers—that compute is a strategic input on par with energy. Once you put compute as critical infrastructure into the policy bloodstream, you remove one of the biggest headwinds to futures adoption: fear of regulatory backlash.
Tailwind #4: The Staircase from Auction to Futures
Markets develop in a predictable sequence. Step one is an auction market, where capacity is allocated transparently and clearing prices are published. Step two is a spot market, where transactions can be referenced and indexed. Step three is a forward market, where parties lock in prices months ahead. Step four is a futures market, where contracts are standardized, cleared, and traded on margin.
That staircase is already under construction. Compute Exchange is the auction layer. Spot prints are beginning to emerge from brokers and secondary markets. Forward offtake agreements (essentially power purchase agreements for GPUs) are being signed by AI labs and second-tier clouds. Once you have those three precursors in place, a futures market is a natural extension. It gives everyone a way to manage basis risk and mark portfolios against a public curve.
Tailwind #5: The Role of MLPerf and Standardization
The hardest part of any new commodity market is standardization. Oil is oil, but compute is heterogeneous: H100ss vs A100s vs B100s, training vs inference, short vs long context. You need a unit of account.
Here, MLPerf provides a head start. It’s not perfect: benchmarks can be gamed, and workloads can drift. But it’s the closest thing to a shared performance language across the industry. MLPerf scores can serve as the backbone for contract definitions, much like sulfur content defines crude grades or calorific value defines natural gas. The point is not perfection, but rather convergence. As long as enough buyers and sellers agree to a benchmark, liquidity follows.
In practice, the first contracts will likely be narrow, e.g., US-West, H100 training-optimized instances, firm, no preemption. Over time, those contracts can broaden into baskets and indices. But without MLPerf or something like it, that convergence would be much harder.
Historical Precedent: Commmodities Once Thought Un-Tradable
Skeptics argue that GPU compute is too heterogeneous, too idiosyncratic to standardize. That is exactly what was said about electricity in the 1990s. Power is local, flows obey Kirchoff’s laws, and outages are stochastic. Yet today we have PJM and NYISO futures.
Bandwidth and dark fiber looked equally un-tradable in the 1990s. Each strand of fiber had unique latency, loss, and topology. Yet IRUs (indefeasible rights of use) and capacity contracts became securitized assets. LNG was bespoke, bilateral, and geopolitically charged, until JKM and Henry Hub LNG futures created liquid hubs.
Every commodity looks idiosyncratic at first. What matters is the economic incentive. Once capital intensity and volatility hit a threshold, markets emerge. Compute is no exception.
Who Would Use GPU Futures?
The natural longs are the AI labs, merchant GPU providers, and second-tier clouds. They face price spikes and availability risk; hedging reduces financing costs and smooths operations.
The natural shorts could be hyperscalers with surplus capacity, or brokers aggregating idle time across fleets. Even hardware manufacturers might short futures as a synthetic way to forward-sell production.
Speculators are obvious: DRW, Citadel, Jump. But also energy traders, because GPU futures are cousins of power futures. Eventually you get structured products: compute-power spreads, analogous to spark spreads in electricity. Insurance and reinsurance desks will use the same curves to price parametric outage covers. The ecosystem that forms around a futures market is often bigger than the market itself.
Why Now?
Three conditions make this moment different.
Persistent scarcity. Demand for compute is growing faster than supply. Even with Nvidia and AMD ramping, backlogs stretch into 2026. Scarcity is sticky. That creates a convenience yield, which is the economic gravity that sustains a term structure.
Scale of capital. A single hyperscale buildout can swing $20 billion of spend. You do not need 100 players to justify a futures market; you need a handful of whales with real hedging needs.
Policy tailwind. With Washington already blessing compute markets, the political cost of launching contracts is unusually low. That reduces friction for exchanges and clearinghouses.
Counterarguments
To be clear, there are real obstacles.
Heterogeneity risk: The contract spec could collapse under “cheapest to deliver” if not carefully designed.
Hyperscaler reluctance: The Big 4 might prefer opacity to transparency; they already enjoy purchasing leverage.
Technological deflation: If compute prices fall faster than expected due to architectural breakthroughs, the appetite for hedging could weaken.
Liquidity chicken-and-egg: Futures only work if enough participants trade. Building that initial pool of liquidity is always painful.
But none of these are fatal. Electricity had the same issues. So did LNG. The lesson from history is that the gravity of capital intensity and volatility eventually overcomes the friction of heterogeneity.
What Success Looks Like
The first GPU futures contract will not be perfect. It will be narrow, cash-settled, and tied to a specific index. It will probably look provincial: one metro region, one SLA band, one GPU class. That is how crude started (WTI vs Brent), how power started (PJM vs NYISO), how LNG started (JKM).
Over time, basis markets proliferate. You get regional contracts, SLA differentials, basket indices. You get optionality: caps, floors, spreads. You get structured finance products: securitizations of offtake agreements, GPU-lease ABS. Eventually, you get a full ecosystem: futures, options, swaps, indices, insurance, securitization. That is how every other commodity market evolved.
The endgame is a transparent forward curve for compute. That curve will reshape financing, risk allocation, and capital formation across AI.
Conclusion
The market for GPU futures will not arrive overnight. It will develop in fits and starts, through pilot auctions and thinly traded contracts. Skeptics will point to illiquidity, to heterogeneity, to failed first attempts. That is exactly what happened in electricity, bandwidth, and LNG. And yet, those markets all exist today.
The structural forces are too strong to ignore. Hundreds of billions in capex. Persistent scarcity. Policy tailwinds. Market-maker DNA. A staircase from auction to futures already in motion.
We are watching the birth of a new commodity market. GPU futures will emerge, not because they are elegant, but because they are necessary. And once they do, the financialization of compute will reshape AI infrastructure as profoundly as oil futures reshaped energy.
If you enjoy this newsletter, consider sharing it with a colleague.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:
Just something to keep in mind, but the market for (purely financial) commodities futures and the market for commodities futures that contain an actual delivery component operate in different ways. There is no reason to expect compute to be any different
It’d be interesting to see an MBA student or think tank analyst expand on this - this market seems inevitable and groundbreaking