GPU rents rose 7.6% in a week while token prices fell, and DRW wants to trade the gap

Don Wilson went on Odd Lots, taped live in Chicago and published October 19, and described something the AI debt story keeps missing. The hourly price to rent an Nvidia GPU is going up. The price of the tokens those GPUs produce is going down. Wilson founded the trading firm DRW, and he wants that gap to become a listed market. CME has two GPU rental futures contracts waiting on a CFTC review that ends November 9. The same weekend, Scott Galloway blamed cheap tokens for CoreWeave's 15% bond yield. I think both men are describing the same squeeze from opposite ends, and only one of them has the price data.

Rent went up, not down

When Tracy Alloway asked about GPUs losing value over time, Wilson called it the trillion dollar question and gave a short price history. The hourly on-demand price of an H100 fell steadily until about Thanksgiving 2025. Then it turned up, and he credits a new Anthropic model that set off a surge in inference demand. More recently, he said, H100, B200 and B300 prices jumped again while the Silicon Data token index fell sharply.

The public numbers point the same way. Silicon Data's H100 rental index read $2.83 per GPU-hour on October 18, up 7.6% in seven days. In late August, Yahoo Finance put it at about $2.68, with the B200 at $5.66. A modest rise, but a rise. Wilson also has a stake in the answer. Joe Weisenthal introduced Compute Exchange and Silicon Data as part of what Wilson is building, so read his numbers as a market maker talking his book.

Galloway's 90% is about the other half of the trade

On Prof G Markets the same day, Galloway explained the CoreWeave selloff with token prices:

demand for compute or demand for these tokens is still skyrocketing but the cost of tokens has been cut by about 90%. So unless the demand is up 10x, you're running in place

Scott Galloway, October 19, 2026, source

His co-host Ed Elson went further and called Chinese open-weight pricing deliberate:

I feel as close to certain that your theory of Chinese AI dumping is true, that it is intentional

Ed Elson, October 19, 2026, source

Elson offers no evidence for the intent part, and I wouldn't lean on it. The 90% figure is plausible for tokens. The step from cheap tokens to falling GPU rent is where Galloway goes wrong, because the rental indexes went the other way this fall. A neocloud that rents H100s by the hour is charging more than it did in August. The business getting squeezed is the one in the middle, the one that rents GPUs and resells tokens. CoreWeave's real problem is the price of its $52 billion of debt, which Galloway also named. That is a credit story, and we covered it on October 15.

Who is renting all those GPU-hours

If rent is rising while tokens get cheaper, demand has to be growing faster than efficiency. Matt Garman gave a number for that in a BCG interview published the same day:

where it's now we're at, I think $25 billion in our AI business, just kind of three years in, which is that's much more growth than I would've anticipated back then.

Matt Garman, October 19, 2026, source

He also described what that looks like inside AWS, and that takes a lot of tokens per employee:

They have swarms of agents that they work with and that's how we build services and that's how we're getting 5X, 10X output from where we used to be.

Matt Garman, October 19, 2026, source

Jack Clark made the supply side of the argument at the IMF Per Jacobsson lecture in Bangkok last week. He said Anthropic's researchers now write close to ten times more code than in the company's first four years. He also said data center capacity sits mostly in the US, China and the Middle East, with supply chains so contested that prices keep rising. The only full recording we have is a subtitled re-translation, so I am paraphrasing him rather than quoting.

What a futures curve would do for CoreWeave

Weisenthal put the credit question to Wilson directly:

In a week like this, when you see the yield on a CoreWeave bond at 15%, what does that tell you?

Joe Weisenthal, October 19, 2026, source

Wilson's answer was that the bond market is pricing counterparty risk, not GPUs. A neocloud with contracts from the most creditworthy companies is fine. One selling on-demand capacity, or selling to weaker buyers, pays more to borrow. With futures, he said, it could buy GPUs, sell a strip of futures against them, and then rent on demand or on six-month terms to smaller customers who post cash upfront. He expects that to lower its cost of capital.

Weisenthal pushed back. Doesn't a hedge just let people borrow more against an asset that keeps losing value? Wilson said no, because every futures sale needs a willing buyer. He expects forward prices to settle below spot, since so many owners will want to sell forward. That is the honest version of the pitch. Futures won't stop GPU rents from falling. They tell lenders today what a B200 is worth to someone willing to commit to it, three years out.

November 9 and the index problem

According to TokenPost, CME had aimed for an October 5 launch. On September 21 the CFTC extended its review of both contracts by 45 days, citing "novel or complex issues." Each contract covers 730 GPU-hours and settles in cash. Wilson said the futures can list November 9 if the agency does nothing, and he noted dryly that this falls after the November 3 election. He said he couldn't explain the regulators' concern. The theory he had heard is that some indices aren't robust.

Alloway gave the skeptic's case:

So one of the other things that people say, the skeptics, is they say GPU or compute is not homogeneous enough for a futures market.

Tracy Alloway, October 19, 2026, source

Wilson answered that H100 and B200 spending alone is big enough to support a market. Other chips will trade as a basis against those two, the way thousands of electricity nodes price off a few hubs. I find that convincing for the hardware. I'm less sure about the index. Silicon Data samples neocloud on-demand prices. Big bilateral deals, like the frontier labs' offtake contracts, don't show up in it, and Weisenthal raised exactly that point.

The crunch spread

The most useful thing Wilson said was a name. Rent GPUs at the on-demand price, load an open-source model, sell tokens through a router like OpenRouter, and your margin is a refinery margin. Oil traders call it the crack spread. Soybean traders call it the crush. Wilson calls this one the crunch spread, and he said token futures are at the thought stage at DRW.

Put his rent data next to Galloway's token data and the crunch spread is narrowing from both sides. That would pinch companies like Fireworks and Together, which Wilson cited as proof the open-model market is already big. It also explains his answer when asked how DRW manages its own AI bill. Not every request needs a frontier model, and the goal is to route inference to keep token spend down. Every buyer who does that pushes token prices lower.

Does a GPU futures market save the neoclouds?

No, but it would show which ones need saving. A listed curve gives lenders a public price for a B200 three years out instead of a guess, and that helps operators with good contracts more than weak ones. It doesn't fix the crunch spread. If Galloway is right about tokens and Wilson is right about rent, the squeeze lands on the resellers in the middle, not on whoever owns the chips. The first market read comes no earlier than November 9, and the CFTC can still move that date.