Compute is becoming an asset class. Here's what that means.

Hi {{first_name|Investor}}

In one week of August, NVIDIA ($NVDA) signed agreements with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize more than $500 billion of outside capital for AI compute. NVIDIA also offered to stand behind up to a quarter of what that hardware is worth when the financing ends.

Days later, CME Group ($CME) and Silicon Data set October 5 as the launch date for the first compute futures. They're cash-settled contracts on H100 and B200 rental prices, listed on NYMEX, the same exchange where crude oil and natural gas trade.

Put those together and something changes about how this whole buildout gets funded. Compute is turning into an asset you can finance, hedge, and trade, which is a thing that has happened to other physical inputs before, with results worth knowing about.

You've probably read that used GPUs are losing value fast. You've probably also read that GPU rental prices went up through 2026.

Both claims are true. They're describing different products.

Compute is two things wearing one name, and almost nobody separates them.

The hour is the service. A GPU-hour that goes unsold at 2pm is gone at 3pm, the same way an empty hotel room on a Tuesday earns nothing forever. You can't store it, you can't save it for a better market, and if nobody rents it, it evaporates. That's why compute behaves like electricity. It's also what the new futures contracts price.

The machine is the asset. A GPU is durable, movable, and resellable. It can leave one data center, get bolted into another, and keep earning for years. Ships work this way. Aircraft work this way. A megawatt-hour never will. The machine is what lenders secure their loans against.

Hold those two apart and the noise sorts itself out.

The bulls are quoting the hour. Rental rates rose through 2026 across every chip generation at once, including A100s that are 6 years old now.

The bears are quoting the machine. Published estimates of what a used GPU is worth disagree with each other by a factor of 6, running anywhere from 10% to 60% of original cost. One widely shared figure has H100s down about 73% from their 2023 peak. Other trackers put resale much higher.

Both sides think they're arguing about the same asset. They aren't, and that's why the argument never resolves.

On October 5, one of those two numbers becomes public for the first time. Anyone will be able to look up what an hour of compute costs, the way you can look up oil. The other number, what the machine itself is worth, stays a private argument between lenders, insurers and rating agencies who don't publish their assumptions.

That gap is where I think the next round of surprises comes from.

Three things are missing from compute today, and each one is a job a commodity market does.

There's no public reference price. Two companies buying identical GPU capacity can pay wildly different rates, and neither one knows who overpaid. A company renting H100s can call 3 vendors. It can't see the market.

There's no way to hedge. A company that signs a 3-year fixed-rate contract to rent out GPUs while funding those GPUs with floating-rate debt is exposed on both ends, with no instrument anywhere to offset it.

And there's no forward curve. Data center builders commit billions against demand curves they drew themselves, with no public strip of prices to show a lender what the market thinks their output is worth in 2029.

Oil had all 3 problems until 1983. Electricity had them until the late 1990s. Both got solved the same way, and the solving is what created the modern versions of those industries.

Compute isn't the first input to go through this. It's at least the fifth.

The record is uneven, and the failures failed for reasons specific enough to test against what launches in October. NYMEX's first electricity futures in 1996 died because the contract settled at delivery points no actual hedger was exposed to. Freight's first futures contract died of the same flaw while the index underneath it survived and became the acquisition target. Enron's bandwidth market, the closest cousin to compute, managed roughly 20 real trades before it disappeared.

Across all of them, the durable profits landed in the same seats: the exchange, the company that publishes the benchmark price, and the intermediaries who take a small piece of every trade. The people who owned the actual barrels and ships got the volatility.

Compute's version of that cast is already assembled and funded. Some of it is public, some of it isn't, and the conflicts in how it's structured are worth naming out loud. The report walks through each seat and scores it.

It's in the credit market, and it's the cleanest paper trail in the whole story.

CoreWeave ($CRWV) borrowed against its GPUs at roughly 14.1% in 2023. Lenders priced those chips like venture risk. By March 2026, the same kind of collateral carried an A3 investment-grade rating at SOFR plus 2.25%. By August, lenders were writing loans that run 2 years longer than the customer contracts behind them, which means somebody has now put a real number on what a used GPU rents for in year 4.

Credit repriced that machine from speculative gear to infrastructure in 36 months. That happened before any futures contract existed.

It comes up in every conversation I have right now.

I think we're in the approach to one. I also think the pop is further out than most of the warnings suggest.

Bubbles end when supply catches up, when revenue disappoints, or when credit breaks. None of the 3 looks close on today's evidence. Capacity is still rationed, the revenue is arriving, and the companies doing most of the buying generate more cash than any businesses in history.

What has changed is how the buildout is funded. The shift from cash flow to debt is the clock I'm actually watching, and it's the one number I'd track if I could only track one.

Open interest in week one, which counts the contracts still held after the close, and whether real hedgers show up alongside the trading firms.

Then the shape of the curve. If contracts for mid-2027 price above next month's, the market is saying the shortage eases slowly. If they price below, it's saying today's prices are the anomaly. Either way, it's the first public opinion on the scarcity debate that has driven 2 years of AI CapEx arguments, and it prints on day one.

One caveat on the date. The CFTC opened a comment period on compute derivatives on August 19, and the window runs past October 5, so the launch is still a plan rather than a certainty.

The report walks all 5 seats in the new market and names who's sitting in each: the company that publishes the price and who owns a piece of it, the exchanges racing to list the contract, the lenders holding GPUs as collateral, the parties backstopping what the hardware is worth at the end, and the fleets and silicon underneath all of it.

Every company gets scored on the same 2 gates, whether the volume is real and whether they keep a margin on it, then labeled proven, early signal, or theory. The scorecard at the back puts all of it on one page, including the names the test declined and the reasons why. There are 8 charts, the full framework map, a signpost calendar of dated events you can check yourself, and a list of the figures I couldn't verify along with where credible sources contradict each other. Full report available for insiders through the link below.