Why the Future of AI May Run on a Commodity Contract
U.S. policy and derivatives trading expertise are converging to financialize GPU compute
I recently wrote about GPU futures, and a reader DM’d me a bunch of questions. This post addresses those questions. What follows explains how GPU futures might arrive and why various market participants would be interested in GPU futures.1
Editor’s note (added Aug 15th 2025): Since originally publishing this post, I have confirmed that the AI Action Plan is advisory, not binding like an executive order or statute. Its reference to forward/spot GPU markets is therefore a policy signal rather than a formal commitment. I still view it as meaningful momentum, in that it lowers the stigma for agencies and market participants to pursue the idea, but it does not guarantee implementation.
Wihout a way to lock in future GPU prices, developers of AI data centers must rely on equity-heavy capital stacks or expensive debt. This slows the buildout of data centers and keeps compute scarce. The fix might come from a surprising place: a GPU futures contract. Backed by U.S. policy and derivatives-market expertise, it could turn GPUs into a bankable commodity, dropping project financing costs by hundreds of basis points and changing the economics of AI infrastructure overnight.
Two developments in the past year suggest that path dependency for such a market is already taking hold:
The Trump administration’s AI Action Plan explicitly endorses the creation of forward and spot markets for GPU compute. This is not some throwaway bullet point. It’s a policy signal that the U.S. government wants market-based allocation of compute.
Compute Exchange, co-founded and funded by Don Wilson, the founder of DRW, one of the most formidable derivatives trading firms, is building an auction platform for GPU compute. If there’s anyone who knows how to take an illiquid, quirky OTC market and turn it into a liquid, standardized derivatives complex, it’s Wilson.
Put those together and you get the outlines of a market structure sequence we’ve seen before with other commodities: auction prices —> spot markets —> forwards —> listed futures2. Once that wheel starts turning, the structural incentives are hard to reverse. That’s what path dependency means: the early decisions lock in the trajectory.
If that trajectory completes, the GPU futures curve will lower the cost of capital for data center developers, in some cases by hundreds of basis points.
Why policy and Don Wilson matter
The AI Action Plan’s endorsement matters because it lowers regulatory risk. When the White House says, in print, “we want forward and spot markets for GPU compute,” it’s giving cover to agencies like Commerce, CFTC, and DoE to treat the idea as legitimate policy. It also tells institutional buyers, including hyperscalers, sovereign wealth funds, AI labs, etc., that the government sees market-based pricing as the preferred allocation mechanism.
Wilson’s involvement matters because DRW has a track record of turning niche assets into globally traded products. There is a playbook: standardize the trade unit, create transparent price discovery, introduce forwards, then migrate to listed futures. The capital and market-making expertise to seed liquidity are already there.
That’s how we got deep and liquid markets in electricity, emissions, oil, even freight. And it’s how GPU compute could follow.
The sequence: auctions to futures
The journey from bespoke deals to a bankable futures curve happens in stages. Here’s the likely timeline, and what each stage unlocks for data center financing.
Stage 0 - Precursors (now)
Market reality:
GPU capacity is allocated through ad hoc auctions, one-off RFQs, and bilateral deals. Spot prints exist, but they’re inconsistent. Forward deals are bespoke, with messy specs (chip class, location, interconnects).
Financing reality:
Project finance is tricky because merchant compute revenue is too volatile. Lenders demand long-term take-or-pay leases from hyperscalers or major AI labs. This forces developers into equity-heavy capital stacks or expensive mezzanine debt.
Path dependency seeds:
Policy endorsement signals that compute pricing belongs in public markets.
A credible auction venue begins to create regular, posted pieces.
These early steps are habit-forming. Once operators, investors, and CFOs start using a publised reference price in models, it becomes painful to rip out later.
Stage 1 - Reference indexes emerge (T + 1-2 quarters)
Market structure:
Auctions become standardized. Consistent lot sizes (e.g., normalized TFLOP-hours), delivery windows, and KYC’d providers. A daily spot index appears, along with short-term forwards, for a few distinct GPU grades (say, H100-class, B200-class) with clear quality differentials.
Financing unlocks:
Index-linked CPAs (Compute Purchase Agreements): offtakers sign 2-4 year deals priced as “Index +/- spread”.
Lenders start to see partial revenue certainty, allowing blended contracted/merchant revenue models.
Equipment ABS pilots emerge, with cash flows indexed to spot prices + floors.
Once this index exists, it gets baked into budgeting, covenants, and valuation models. That’s the first layer of path dependency.
Stage 2 - OTC forwards standardize (T + 2-4 quarters)
Market structure:
The industry converges on standard forward contracts with clear definitions for:
Grade (chip type, memory size, interconnect speed)
Location basis (latency zones, metro delivery points)
Availability profile (reserved vs burstable)
SLAs and delivery assurance
Price reports begin publishing forward curves; dealers quote two-way markets.
Financing unlocks:
Hedged debt service coverage ratio (DSCR): Lenders can now size debt against a minimum hedge ratio (say, 50-70% of output hedged 12-24 months forward).
Term loans and mini-perms price off forward coverage.
Sale-leaseback transactions on GPU fleets use forward curves to set residual values.
By this point, legal boilerplate, accounting treatment, and risk systems have institutionalized the market. That infrastructure is expensive to change. This is another lock-in.
Stage 3 - Listed futures launch (T + 4-8 quarters)
Market structure:
An exchange lists cash-settled GPU futures contracts on the reference index, with monthly expiries and a clearinghouse. Basis contracts for major locations (e.g., US-East premium) appear.
Financing unlocks:
Capital efficiency: Futures require margin, not full collateral, freeing balance sheets during construction.
Bankable merchant revenue: Lenders tolerate a higher share of merchant exposure if it’s hedged on a liquid exchange.
Compute ABS scale: Securitizations reference exchange prices, giving rating agencies confidence.
Cross-commodity hedging: Developers can lock a compute spark spread: AI workload revenue minus GPU futurs, power futures, and cooling load. Compare with crack spreads in the oil markets.
Now the futures curve becomes the benchmark. Everything keys off it: loan covenants, equity research, securitizations, even executive comp.
Stage 4 - Market deepens (Year 2-3)
Market structure:
Options on GPU futures (calls, puts, collars) give flexible downside protection.
Quality differentials (e.g., memory-rich vs standard) trade explicitly.
Roll markets3 let you manage the transition from one chip generation to the next.
Financing unlocks:
Portfolio financings for multi-site campuses hedge across strips and basis contracts.
Tighter credit spreads: Mezz and private credit see lower risk when projects carry listed hedges + options.
Residual value insurance keyed to the forward curve lowers equity requirements.
By this point, the entire ecosystem, including brokers, clearing firms, auditors, and insurers, is specialized around the GPU curve. This is too embedded to reverse territory.
Step 5 - Integration with corporate finance (Year 3+)
Market structure:
Compute-indexed notes: corporates issue debt with coupons linked to GPU indexes.
Treasury adoption: Large AI companies run formal compute-hedging programs.
Financing unlocks:
WACC compression: Transparent, liquid curves let sponsors shift from equity-heavy builds to true project finance.
Secondary equity liquidity: Exits and valuations become about basis and roll risk, not opaque pricing guesses.
Why some companies won’t want this, and why they may have no choice
If this market seems inevitable, it’s worth asking: who might not want it, and could they actually avoid it?
Reasons to resist
Hoarding as moat: Hyperscalers like controlling GPU supply to squeeze competitors. A transparent market erodes that advantage.
Price secrecy: Public prints reveal what they pay, when they buy, and sometimes how much capacity they’re taking.
Loss of buyer power: Smaller players can bid against them; spreads compress.
Volatility risk: Listed markets can spike against them in a crunch, raising procurement costs.
OEM politics: Nvidia and others benefit from opaque, relationship-driven pricing, and big buyers may want to preserve that.
Forces that will pull them in anyway
Policy endorsement → procurement pressure: Once the AI Action Plan greenlights forward/spot markets, federal buyers can start sourcing through them. To win those contracts, you play by those rules.
Financing reality: JV partners, leaseback financiers, and campus co-developers will demand hedgeable merchant revenue, meaning participation in the curve.
Risk management: Even hyperscalers need to hedge flex capacity, chip transitions, and disaster scenarios.
Market gravity: Once liquidity builds, staying outside means flying blind on where prices are clearing. This is an information disadvantage most won’t tolerate.
Network effects: As more budgets, covenants, and securitizations reference the GPU curve, the opt-out lane narrows to zero.
The likely path: hyperscalers resist in Stages 1–2, dabble OTC in Stage 3, and are fully visible in Stage 4–5. They end up participating not because they love the idea, but because the market has become infrastructure in its own right.
The bottom line
We’ve seen this movie before. In oil, in electricity, in emissions, in freight: you start with messy bilateral deals, then auctions, then spot indexes, then forwards, and then one day there’s a liquid futures contract that the entire industry keys off.
When that happens in GPUs, the futures curve will do more than give traders a new toy. It will change the cost of capital for AI infrastructure, turning merchant compute from a financing headache into a bankable commodity. Once that shift happens, it’s very hard to go back.
The path dependency is already forming. The only real question is how fast we get from auctions to a curve you can build a $1B data center on, and who gets pulled into the market along the way, whether they want to be there or not.
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:
follow me on Twitter,
connect with me on LinkedIn, or
send an email to dave [at] davefriedman dot co. (Not .com!)
The timeline I present here is a very compressed one. It is compressed because Don Wilson has said he sees the market for GPU compute futures eclipsing the size of the oil futures market (~$750 billion+ notional) within a decade. Yes, he is talking his book to some extent, but he also is much more knowledgeable about futures markets than I am, so I am implicitly deferring to his judgment in this post.
Two short case studies might help illustrate this.
Case Study 1: Oil Futures in the 1980s —> From Opaque Deals to Capital Markets Backbone
Starting point: In the late ‘70s, most crude oil moved through long-term supply contracts between national oil companies, majors, and refiners. Pricing was opaque, often linked to official posted prices or confidential term deals.
Catalyst: The NYMEX launched light sweet crude oil futures in 1983. At first, majors and OPEC members dismissed it as irrelevant paper barrels.
Path dependency moment: Once independent refiners and traders began using NYMEX settlement prices in term contracts, the futures curve became a reference price in lending covenants and project finance. Bankers could underwrite debt against hedged protection.
Result: By the early ‘90s, ignoring NYMEX prices meant flying blind on where the market cleared. Even OPEC began watch, and occasionally gaming, the futures curve. Today, crude oil futures aren’t a niche. They’re the market’s spine.
Relevance to GPUs: Just as oil finance shifted from bilateral secrecy to benchmark-linked lending, GPU futures would give developers and lenders a transparent reference to size debt against merchant compute.
Case Study 2: ERCOT Power Markets, Early 2000s —> The Birth of a Roll Market
Starting point: Texas power generation used to be a patchwork of bilateral deals and vertically integrated utilities. Prices for forward delivery were ad hoc, and project finance for merchant plans was scarce.
Catalyst: ERCOT introduced nodal pricing and congestion management in the early 2000s. Forward markets developed for major hubs, and then exchange-traded power futures followed.
Path dependency moment: Once merchant power plants could lock in spark spreads (electricity price minus fuel cost) on exchange, lenders began underwriting merchant revenue streams. Roll markets became active because plants needed continuous hedges across seasons.
Result: The roll market itself became a liquidity center. Power traders made money not just on outright prices but on curve shape changes, and plants could maintain multi-year hedge coverage.
Relevance to GPUs: In GPUs, roll markets would let data center operators manage generation shifts, meaning operators could move hedges forward as chip generations turn over, and give financiers the comfort that revenue is locked in beyond a single expiry.
In commodity trading, a roll market refers to the market activity and pricing involved when traders “roll” a position from one futures contract expiry to another, usually to maintain exposure without taking physical delivery.
Great post, Dave. Would be great to trade notes on this, I'm starting on my compute markets 101 piece and have been talking to all the companies in the space.