Compute Is An Asset Class

On August 10, 2026, NVIDIA said it wants to help pull more than $500 billion of outside capital into a group of new financing platforms. The partners are Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, and the goal is building AI factories. A day later, CME Group and a firm called Silicon Data said GPUs would get their own futures contract, the kind of instrument usually reserved for oil, wheat, and electricity. A few months earlier, ICE had already said it planned to list futures on a different compute price index, built by a company called Ornn. Two announcements in one week, a third already in motion, and one question none of the headlines quite answered: what is a GPU, to a finance team?

The easy answer is still that a GPU is equipment. You buy it, you write it down over a few years, you replace it. That answer is getting harder to hold. The same conclusion from The Price of Electricity Belongs in Your AI Forecast is that the AI bottleneck was always electricity, the physical floor sitting under every layer above it. This piece is about the layer on top of that floor. Is compute becoming something you can price and finance, like power or oil? Or is it still a piece of IT that only holds together because the company that sold you the chip is quietly backing what it will be worth later?

Compute Has a Price Tag

The dollar figure is the loud part of NVIDIA's announcement. The structure is the part that matters.

NVIDIA signed memoranda of understanding, not a signed check, with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to build independent compute financing platforms. The stated goal is to mobilize more than $500 billion of other people's money over time. That number is a target for what these platforms hope to raise and put to work, not NVIDIA spending, not a single fund, and not NVIDIA revenue. Every dollar of it still depends on final agreements that have not been signed yet.

Inside that structure sits the detail that matters most. NVIDIA has said it may cover part of the gap if a chip is worth less later than the lender hoped, up to 25 percent of that deal, decided project by project. The real credit work stays with the lenders: who the customer is, what demand looks like, how much the chips are used, what the cash flow can support, and what the equipment will be worth later. CEO Jensen Huang has said NVIDIA's share of that gap is significantly lower than in other compute financing deals already in the market, though NVIDIA has not published a detailed comparison to back that claim.

Huang's pitch is the more interesting sentence. He wants AI factories financed the way power plants and toll roads are financed, as productive infrastructure with a revenue stream, not as one-off IT purchases a CFO writes off in a few years. That is a real ambition. It is also an ambition that still needs a chip vendor to guarantee part of the downside before institutional capital will commit. None of this financing shortens an interconnection queue or clears a permitting fight, the two constraints already covered in The Price of Electricity Belongs in Your AI Forecast.

Computer Chips Now Have a Public Price

If NVIDIA's MOUs are about who provides the money, the futures announcements are about something narrower and more useful: whether anyone can actually see the price.

Right now, renting an H100 for an hour costs whatever your provider quotes you. Two buyers of identical capacity can pay meaningfully different rates and never find out, because there is no published reference price the way there is for oil or power. CME Group and Silicon Data, a GPU market-intelligence firm backed by the trading firm DRW, announced on August 11 that they plan to list two cash-settled futures contracts on October 5, pending regulatory review: one tracking Silicon Data's hourly rental index for NVIDIA's H100, and one tracking its index for the newer Blackwell B200. Each contract represents roughly a month of rent for a single chip, and both are set to trade on NYMEX, the same exchange that lists crude oil. Silicon Data's CEO, Carmen Li, has framed the goal as building a public, tradable reference price for compute, the same role an index like WTI plays for oil. CME's global head of energy products, Pete Keavey, has described the aim as turning compute into a standardized, tradable commodity instead of a fragmented set of private quotes.

A second effort is further along on the pricing side and earlier on the exchange side. Intercontinental Exchange has said it plans to launch GPU futures based on the Ornn Compute Price Index, or OCPI, which Ornn built to be transaction based: it prices off trades that actually happened, not list prices or surveys. OCPI already runs on the Bloomberg Terminal, and Ornn has published forward curves for the H100, H200, and B200. The company was co-founded by Kush Bavaria and Wayne Nelms, who met at MIT, and it raised roughly $33 million in a seed round led by a16z in June 2026. "Compute has grown into a trillion-dollar market, yet it still lacks the pricing and risk-transfer infrastructure that every other major commodity relies on," Bavaria has said publicly. As with the CME contracts, the ICE futures are a plan, not a live market yet. Both efforts are waiting on regulatory review before anyone can actually trade a single contract.

Kalshi has floated GPU-linked event contracts and its own implied forward curve. Worth watching, but not a settled number yet. Some of the biggest claims about the size of this market are already ahead of any real trading.

Compute Now Trades Like Oil

Kush Bavaria made the clearest public case for this shift on Peter Diamandis's Moonshots podcast, episode 278, recorded the same week as the NVIDIA news. His argument, in paraphrase rather than a direct quote: compute is already a trillion-dollar-scale market with none of the pricing infrastructure every other major commodity has. Oil has a public spot price and a forward curve. Power has one too. Compute, until now, has mostly had private negotiations between whoever happens to be buying and whoever happens to be selling that week. If GPU-hours get an index built from real trades, a forward curve on top of that index, and eventually a cleared futures market, buyers and sellers can hedge the way an airline hedges jet fuel or a utility hedges power, instead of guessing.

Worth hearing directly if you want more than a paraphrase. The episode is on Apple Podcasts.

What Is the Chip Worth Later?

Here is the part that never makes it into a press release: what is this thing actually worth once the financing term ends?

Every leased or financed asset lives or dies on its residual value, the number a lender is betting the equipment will still be worth once the loan matures. Real estate has a residual that mostly holds. Aircraft have a residual that airlines and lessors have modeled for decades. GPUs have a residual nobody has priced through a full hardware cycle yet, because the category is simply too new.

Amazon gave the clearest public tell on this question, and it has nothing to do with this week's news. Effective January 1, 2025, Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five, citing faster technology development, especially in AI and machine learning. The change cost the company about $1.4 billion in additional depreciation and roughly $1.0 billion in net income for fiscal 2025, mostly inside AWS. A year earlier, Amazon had gone the other direction and extended some of that same equipment's useful life from five years to six. The company that runs more compute than almost anyone on earth changed its mind twice in two years about how long its own hardware would hold value, and this time the correction ran toward shorter, not longer.

That is the quiet warning behind every strong rental-rate chart circulating right now. Strong rental prices today do not guarantee resale value tomorrow. A GPU can be fully booked at a premium rate the week before a faster chip makes it half as valuable, and the loan that financed it does not know the difference.

That is what makes NVIDIA's residual-value backstop the most telling detail in the whole announcement. The company selling you the chip is also, on some deals, guaranteeing part of what it will be worth later. That is not how a mature asset class usually works. Nobody expects an oil producer to guarantee the resale value of the tanker hauling its crude, and no one expects a power company to backstop the residual value of a turbine it sold. A seller guaranteeing the asset's future worth is a sign the market has not yet agreed, on its own, what that worth actually is.

If you have to guarantee the residual, it is not quite an asset class yet.

That is not a dismissal of everything happening this year. It is a description of where the market sits today: real price discovery forming on top, and a real credit question still unresolved underneath.

What This Means for Finance

If you work in finance at a company, here is the simple version.

A public price is useful. Two companies buying the same GPU-hour can finally check their quote against something other than a salesperson. A forward curve lets a company lock in a compute cost the way an airline locks in jet fuel, instead of eating whatever the market does next quarter. Capital that used to sit out because AI factories looked like one-off IT bets can start underwriting them as infrastructure with a revenue stream, which is the whole premise behind NVIDIA's financing platforms. And once an index and a forward curve exist, a finance team gets a market estimate of what a chip might be worth later, instead of the straight-line depreciation guess most companies are using now.

The risks are just as plain, and none of them are resolved. An MOU is not a closed fund. Every dollar of that $500 billion figure still depends on agreements that have not been signed. NVIDIA backstopping part of the residual value on chips it is also selling is a circular arrangement: useful for getting deals done today, not proof the market has priced the asset on its own. Amazon's six-to-five-year correction shows that a chip can age faster than a loan. That is the gap NVIDIA's guarantee is meant to paper over, project by project, until enough loan cycles actually clear. The new futures markets will also be thin at first, and your company's real token or API bill is not the same thing as an H100 rental index. That mismatch is called basis risk. A hedge built on the public curve can still miss what you actually pay. Every index in sight is priced off a small number of NVIDIA chip generations, so the whole emerging market is concentrated around one vendor's product roadmap. None of this financing gets a data center connected to a grid that has no capacity to give. And a new, thinly traded market is also a market someone can speculate in well before it is doing the job it is supposed to do.

This Goes in the Forecast

None of this means a company should hire a GPU futures trader next quarter.

Most corporate treasury desks will not be trading H100 futures anytime soon, and that is the right call for now. The contracts are new, thin, and still pending regulatory approval. Treasury already watches FX, interest rates, and sometimes energy. A public compute price is the same kind of number. You do not need to trade it to use it.

The useful move is smaller. Treat the emerging compute curve as a watch item for the AI forecast. If the forward price of GPU-hours is rising, the assumption that your token or API line is a falling software cost is already out of date. That is the same conclusion from The Price of Electricity Belongs in Your AI Forecast, reached from the power side. That piece was about the physical input. This is the public price forming on top of it.

Three Questions

Before your next AI budget review:

1. Who buys compute at your company?
2. What move in the public price would change your forecast?
3. Who owns the watch?

Not a trading strategy. A named owner for a number that used to live only in private quotes.

Thoughts From First Principles

Markets form around things that are scarce, jumpy, and important enough that somebody eventually has to put a price on the risk instead of just eating it. Oil got there. Power got there. Bandwidth got there in its own smaller way. Compute is entering that club this year. Stay Naive has argued before, in Intelligence Got Cheap. Here's What Just Got Expensive., that scarcity moves rather than disappearing once an input gets cheap. This is what it looks like when that scarcity finally gets a price.

Hold two things at once. The market forming around compute this year is real progress: real money, a real public price, a real forward curve. Those are things a real market eventually needs. And the guarantee sitting under NVIDIA's part of the deal is a real signal that the hardest question, what a GPU is actually worth once it stops being the newest chip, is still unanswered by anyone except the company that sold it to you. Both can be true. Refusing to pick a side too early is the whole point.

Reflection Point

If someone put the forward curve for GPU-hours in front of you tomorrow, would you know whether your AI forecast is priced for this year's market, or for a regime that already ended?