Don Wilson on the GPU Market's Trillion Dollar Question

Don Wilson · DRW

GPU futures would let neoclouds hedge and lower their cost of capital, which CoreWeave's 15% bond yield shows is needed.

A 15% yield on a CoreWeave bond signals a massive need for risk management products. Neoclouds that sign contracts with the most creditworthy companies are fine, but many GPU buyers are less creditworthy, so neoclouds selling to them face a higher cost of capital. With GPU futures, a neocloud could buy GPUs, deploy them, sell a strip of futures, and sell on-demand compute or six-month contracts to weaker customers who post a large share of contract value upfront. The market will reward firms that do this with much lower cost of capital.

“need for risk management products. And so, you know, think about if you're a neo cloud and, you know, you sign contracts with the most creditworthy companies in the world, then you're fine, right? But there are a lot of other consumers of GPUs that are not the most creditworthy companies in the world.”at 4:17

Don Wilson · DRW

Silicon Data's neocloud compute index is transparent and well built, meeting the IOSCO benchmark other providers lack.

Silicon Data has collected massive amounts of data focused on on-demand compute from neoclouds. Since there are many neoclouds but few hyperscalers, this index covers the market well and is transparent about on-demand compute. Don says it was built to be IOSCO compliant, the benchmark that other index providers fail to meet. He argues traders should check the index they trade futures on, and he favours this one.

“Silicon Data, that team has done an amazing job of collecting massive amounts of data... there is a lot of transparency about on demand compute... they've done a very good job to create an index that, you know, will be [IOSCO] compliant. That's the benchmark.”at 7:11

Don Wilson · DRW

Nobody knows what a B200 will be worth in four or five years, so long-dated GPU futures are needed to hedge it.

Whether GPUs hold value is the trillion-dollar question. Nobody knows what a B200 will be worth four or five years from now, and a single GPU can depreciate as chips become more efficient and users move on. Futures that trade far out let buyers and sellers hedge this residual value, and the price swings in both directions over short periods show a hedging instrument is needed.

“that is the trillion dollar question... nobody knows what the value of a B200 will be four years from now, five. And, years from... and a hedging instrument is badly needed.”at 8:47

Don Wilson · DRW

CME's initial GPU futures will run out three years, extending further as they become more liquid.

Don says the CME's futures contracts are going to go out three years initially. As they become more liquid, they will extend further out. Long-dated contracts are needed to hedge GPU values that nobody can predict four or five years ahead.

“I think initially the CME's futures contracts are going to go out three years. And as they become more liquid, they'll go out more.”at 9:13

Don Wilson · DRW

GPU rental prices swing sharply with inference demand; H100 prices rose after Anthropic's new model lifted demand.

H100 on-demand hourly prices fell steadily through late 2024. Around Thanksgiving they turned up sharply after Anthropic released a newest model that was a massive step up, which drove a surge in inference demand. Prices for B200 and H100 then surged again, while the Silicon Data token index fell. Don concludes that compute prices are volatile in both directions, which is why a hedging instrument is badly needed.

“all of a sudden there was a ton of demand for inference. And, you know, more recently we've seen... another surge upward in price.”at 10:07

Don Wilson · DRW

Forward GPU compute contract prices will fall because heavy hedging supply meets limited rational buyers.

Futures need a willing buyer and seller, so the idea that producers could sell overpriced contracts to infinite demand doesn't hold. Some rational buyer must take the other side. Because so much capacity needs hedging by selling forward, Don expects forward contract prices to settle lower, with investors needed to earn a return by buying those contracts.

“But there has to be a rational buyer on the other side. And so I think what we'll see is that those forward contracts are... in fact going to be at lower prices just because there's so much hedging demand to sell in the forward space.”at 13:29

Don Wilson · DRW

One futures market can cover H100 and B200 demand; other chips are highly correlated, so basis trades will emerge.

The money going into H100s or B200s alone is large enough to support a single futures market. Other accelerators such as AMD, Cerebras, Etched, Trainium and TPUs are highly correlated with these, so traders can trade the basis between them. Premium configurations, such as large training clusters, will carry a premium, but if inference demand rises, training runs get more expensive too, so it is all basis between related products.

“the amount of money going into, you know, just [H100s] or just [B200s] is large enough to support a single futures market. Now in addition to that, these other things are very correlated.”at 16:26

Don Wilson · DRW

AI's rapid growth won't be constant, so the recent Anthropic wobble is expected volatility, not a slowdown.

Something growing this fast cannot grow at a constant rate, and its first and second derivatives will not be constant. Given that, the recent jitter in Anthropic's revenue growth, which came around its investor meeting ahead of its IPO, is expected volatility. Don says he is not reading anything more into it, and suggests people may have simply taken vacations in August and used fewer tokens.

“That's growing this fast is not going to grow at a constant rate... the first derivative isn't going to be constant. The second derivative isn't going to be constant. So, to me, this is just kind of expected volatility.”at 3:15

Don Wilson · DRW

Open-source model use has exploded, and a rich hosting ecosystem already exists beyond the big closed labs.

Open-source model usage has exploded, so the market for compute is not just Anthropic and OpenAI. Companies with billion-dollar valuations were created to run open-source models on GPUs and sell the tokens, such as Fireworks and Together. Don argues this is already a rich enough ecosystem to support a futures market beyond bilateral offtake deals.

“Look at the growing use of open source models. I mean, it's exploded. And you're seeing companies with many billion dollar valuations that have just been created to put open source models on GPUs and sell the tokens... And so we already have a really rich ecosystem.”at 14:42

Don Wilson · DRW

Profit from renting GPUs to sell open-source tokens resembles a refining crack spread, which he calls the crunch spread.

Renting GPUs at the on-demand price, loading an open-source model, and selling tokens through OpenRouter can be profitable or not depending on the spread between token revenue and compute cost. That margin is akin to what an oil refinery earns, a crack spread, or a soybean crush plant earns, a crush spread. Don calls this the crunch spread at DRW. He says token futures will eventually be possible to trade this margin, though the variety of tokens will draw objections.

“that profit margin that you can generate is akin to what an oil refinery does, right? ... Or what a soybean crush plant does.”at 18:07

Don Wilson · DRW

Perpetual futures serve a useful purpose: they can make ETFs more efficient and pass through stock borrow fees.

Perps are futures contracts. Oil ETFs roll front-month futures with slippage, and a Brent-linked perp bought by an ETF would be a more efficient exposure, though not a hedge for oil producers, who need to hedge far out the curve. Retail holders of hard-to-borrow stocks, paying 100% annualized to short, could get borrow fees passed through via a perp. On leverage, a quarterly future can have massive leverage and auto-deleveraging just as a perp can, so a perp could be listed with reasonable leverage and no auto-deleveraging. Don argues a future and a cleared swap carry the same risk to society, so the distinction is mostly nomenclature.

“My view is that [perps] serve a useful economic purpose. They are futures contracts.”at 24:03