Don Wilson on the GPU Market's Trillion Dollar Question

Don Wilson treats swings in AI demand as normal volatility, not proof that the market is slowing, and expects the GPU market to keep growing. He is less certain about NVIDIA’s share: he thinks it is unlikely to remain at its current level, but expects NVIDIA itself to continue growing. The distinction matters because a growing market can support continued expansion even as competitors capture a larger portion of it.

GPU prices have moved sharply in both directions, making future values difficult to predict and hedging potentially important. H100 on-demand prices fell and then rose around Thanksgiving after a new Anthropic model increased inference demand. More recently, prices for B200, B200 300, and H100 rose while the Silicon Data token index price fell. Wilson sees such movements as evidence of the uncertainty buyers and providers face, not as a reliable signal of a lasting trend.

That uncertainty is the case for GPU futures: no one knows what a B200 will be worth four or five years from now. CME’s initial contracts are expected to run for up to three years, with longer maturities possible if liquidity develops. Wilson expects contracts to focus on major chip types rather than every new generation, with each chip type drawing its own pool of liquidity. He argues that spending on H100s and B200s alone is large enough to support a market.

For neocloud providers, futures could help hedge GPUs already deployed while they sell on-demand compute or sign customer contracts backed by upfront collateral. Wilson expects providers that hedge to have a lower cost of capital, which could be rewarded by the market. He points to high CoreWeave bond yields as evidence of the funding challenges neoclouds face when serving customers who buy on demand or have weaker credit. Futures would not remove the underlying business risks, but could make GPU-price exposure more manageable.

Wilson rejects the concern that futures would let companies borrow to overbuild without reducing depreciation risk. A futures hedge requires a willing counterparty, and heavy selling of future GPU capacity would push prices down. The market’s existence would not guarantee that deployed GPUs retain their value. He also sees an adequate base of activity in open-source models and token sales by companies such as Fireworks and Together, alongside the spending on major GPU configurations.

Silicon Data focuses on on-demand neocloud compute, where providers are numerous and prices relatively transparent. Wilson says the company has collected substantial data and is working to make its index compliant with IOSCO benchmark practices; he advises buyers to examine the index underlying any futures contract. He could not explain the regulatory concerns about listing GPU futures, though he had heard the theory that some indices might not be robust. The CFTC comment period was extended; if the commission takes no action, GPU futures can be listed November 9, after November 3.

Wilson says other types of compute tend to move with standard GPU configurations, allowing traders to take positions on price differences, or basis, while specialized large training clusters may command a premium. He expects rising inference demand to lift the cost of training runs as well. Renting GPUs to run an open-source model and sell tokens could create a margin between input and output prices, analogous to oil crack spreads or soybean crush spreads; DRW calls this the “crunch spread.” Token futures remain only a possibility under discussion, rather than an established product.