Last October, OpenAI and AMD announced a multi-year deal for 6 GW of GPU capacity, beginning with 1 GW in 2026. The line that got quoted most was OpenAI’s drive to secure “as much compute as we can possibly get.” That sentence is the current paradigm in seven words.

Today, securing compute means owning compute. Or at least committing to it for years, prepaying a fifth of the contract value upfront, and locking into one chip family, one region, one provider. Buyers do this not because it is the optimal economic structure, but because there is no other way to manage exposure to price or availability. The short end of the rental market has effectively disappeared. Premium-GPU conversations no longer start at less than six months, and the sweet spot providers prefer is one to two years. Meanwhile, H100 reserved-rental prices rose roughly 40% in five months between late 2025 and the spring. If you are a buyer, that is volatility you have no clean instrument against. If you are a seller, that is upside you couldn’t capture unless you happened to be selling spot during the move.

This is not only buyer behavior. In a recent piece on compute market structure, Moonrunner maps out the six participants and their incentives, and the hoarding equilibrium falls out cleanly. Frontier labs and hyperscalers want predictable capacity at scale. Neocloud providers and datacenter operators want to recover capex on hardware that depreciates fast, which they can only do by signing long contracts with creditworthy clients. Both sides have rational reasons to converge on bilateral, multi-year, OTC deals. The 20% upfront prepayment is not an accounting curiosity. It is partly working capital that neocloud suppliers use to finance the hardware they’re renting back to you, and partly a credit substitute for buyers who aren’t investment-grade, where the prepayment functions as collateral if utilization assumptions change. The buyer’s lock-up is the seller’s working capital and the seller’s credit cushion.

It is worth being precise about what the hoarding instinct bundles. “As much compute as we can possibly get” is part price exposure, part supply security, part strategic vendor positioning. A financial layer addresses the first cleanly, helps with the second by making short-tenor markets viable, and barely touches the third. Frontier players will keep locking in capacity for strategic reasons even with deep liquid futures. The argument is not that hoarding stops, but that buyers who currently have no choice but to over-commit gain one.

What is missing under all of this is two things, not one. There is no liquid market for short-tenor compute, so buyers cannot right-size flexibly or offload excess if their plans change. And there is no benchmark with hedging instruments on top of it, so participants cannot manage forward-price exposure except by, again, owning more capacity. With neither layer in place, in a market where compute is mission-critical and availability is scarce, physical possession becomes the default risk-management tool. Hoard or get caught short.

The two layers

Mature commodity markets do not have futures sitting directly on top of bilateral OTC. They have two coordinated layers underneath the long-term contract world.

The first is a liquid physical market. Spot prices, short-term contracts, clearing infrastructure that lets you transact the actual commodity at observable fair prices in flexible tenors. With this layer in place, you no longer need to pre-commit to secure access. You buy what you need when you need it, and sellers can move capacity without needing a single creditworthy long-term anchor. Compute does not have this yet in any serious form.

The second is a financial layer on top. Benchmark indices, futures, hedging instruments. This lets you manage price exposure without taking physical delivery, and lets sellers stabilize revenue without bundling commercial terms into multi-year commits. It also broadens who is bankable. Today, lenders underwrite against creditworthy long-term anchor tenants, and that model is working at scale. A futures market doesn’t replace it; it lowers the cost of capital at the margin and brings debt within reach of operators without IG-rated anchors. Small neoclouds, regional players, the long tail. The current bankability model isn’t broken; it is narrow.

The two layers reinforce each other, even if neither strictly requires the other. Cash-settled futures on a defensible benchmark don’t need a deep spot market to function: that is part of the point of cash settlement. But a liquid physical market makes the benchmark more credible and gives hedgers somewhere to translate financial positions into actual compute when they want to. The equilibrium that breaks the current hoarding pattern needs both layers maturing in parallel, even if one launches ahead of the other.

Listed futures also don’t eliminate risk; they reorganize it. Bilateral credit risk gives way to daily mark-to-market and initial margin set by the clearinghouse: a burden that falls hardest on non-IG buyers who currently use the prepayment as a credit substitute. The trade-off is worth it for most participants, but it isn’t free.

What it looks like when it’s working

Concretely: a frontier lab right-sizes its physical reservations to baseline usage, acquires variable load from a liquid short-term market, and hedges forward H100 or B200 price exposure against a benchmark contract. Capital that would have sat in unused reservations goes to compute it actually runs. A neocloud sells capacity into a liquid market across tenors instead of hunting for one anchor tenant, shorts the benchmark to stabilize revenue, and finds its lenders willing to underwrite at tighter spreads because the cashflows are now bankable.

This is not a stylistic preference. In mature commodity markets, financial volume is typically a multiple of physical volume, sometimes a high multiple, and that ratio is not parasitic. It is what funds the physical buildout.

It is also worth being explicit that this is a desirable state. Compute is becoming load-bearing infrastructure for the world economy, and that direction only intensifies. Critical infrastructure that runs on volatile cashflows, illiquid short ends and bilateral handshake deals is a bad combination: for the operators who cannot plan, for the buyers who overpay for security, and for the financiers who price the uncertainty into the cost of capital. Stabilizing these businesses through proper market structure brings more capital to the table, lowers the cost of building, and makes the whole stack less fragile. The financial layer is not a clever overlay. It is how a critical industry moves from improvised to operable.

The benchmark rolls with the hardware

GPUs are not a static commodity. NVIDIA ships a new generation roughly every year; AMD’s MI450 lands in the second half of 2026; custom silicon adds another set of references. That heterogeneity is sometimes raised as the reason a compute futures market cannot work. It is not.

Mature commodity markets handle this routinely. Oil has hundreds of recognized grades and two dominant pricing references. Sovereign debt has thousands of issues and a handful of benchmark bonds the world trades around. Compute would work the same way: a flagship benchmark anchored to the dominant production chip, spread products pricing the differentials between generations and grades, and the flagship rolling as the dominant chip rolls.

If anything, generational churn strengthens the case for separating exposure from possession. A two-year reservation is a price bet, an availability bet, and a generational bet rolled into one, and A100 and H100 reservations now expiring are already expected to flood the secondary market as teams upgrade. Owning the physical asset locks you into all three risks at once. A financial position lets you take views on each separately, and reset them as the market moves.

There is a real complication the oil analogy hides. WTI doesn’t go to zero; H100 has a path to functional obsolescence. A long-tenor position written against the rolling generation carries one-sided basis risk that crude grades don’t have. That argues for concentrating early liquidity at the front of the curve, where the dominant generation is stable, and for being honest that long-dated hedging on this market has to develop alongside settled views on which chip is the flagship.

What has to be true

This isn’t the first attempt at a tradable compute market. Earlier futures-adjacent products didn’t land at scale, for honest reasons: benchmarks weren’t credible enough to settle against, buyers weren’t yet feeling the volatility, and supply-side counterparties for two-sided flow weren’t operationally ready. The first two conditions are different now. The recent 40% move taught a generation of buyers what unhedged forward exposure feels like. Generational refresh is creating natural hedging demand at the short end. The third condition, operationally ready counterparties on both sides, is the part still being built.

None of this is here yet. It needs a credible benchmark, which is harder than it sounds: the four published indices for the same Hopper market currently disagree visibly enough to make the point. It needs clearing infrastructure for short-term compute. It needs an exchange willing to list, and counterparties on both sides who actually want to hedge rather than hoard. The pieces are visible. They are not assembled.

The companies stockpiling reservations today are not wrong. They are rational given what exists. The question worth working on is what becomes rational when something else does.