Key Takeaways
- Dips in compute demand reflect normal variance rather than market decay; DRW founder Don Wilson points out that fast-growing sectors never maintain a constant first or second derivative.
- Spikes in inference track software releases rather than steady linear curves; demand surged around Thanksgiving when Anthropic released a major model upgrade that drove immediate consumption.
- Neoclouds like CoreWeave pay high bond yields because financial markets lack hedging tools for hardware depreciation on chips like the Nvidia B200 over four to five years.
- Financial infrastructure like GPU futures would allow operators to sell forward strips against deployed hardware, enabling them to fund short contracts for early-stage clients without paying punitive credit costs.
Model Upgrades Break Predictable Demand Curves
Skeptics tend to treat any temporary plateau in GPU demand as proof of an AI slowdown. Wilson, speaking live in Chicago, dismisses that reading as a misunderstanding of how rapid growth behaves: “Any new thing that's growing this fast is not going to grow at a constant rate, it's not going to grow at a, you know, the first derivative isn't going to be constant, the second derivative isn't going to be constant.”
Instead of a smooth ramp, compute consumption moves in sudden jumps dictated by software upgrades. Around Thanksgiving, inference demand surged upward because Anthropic released a new model that provided a clear step up in usefulness. When an AI model becomes noticeably better overnight, companies immediately route more production traffic to it. Planning capital budgets around linear quarterly projections ignores how quickly a single software release can exhaust available chips.
Why CoreWeave Pays a Debt Penalty
The lumpy nature of compute demand creates severe headaches for neocloud operators like CoreWeave. Buying tens of thousands of processors requires massive upfront debt, and credit markets have punished these providers with high bond yields.
The bond market is not worried about whether the chips function. It is worried about what those chips will be worth when loans come due. As Wilson explained: “That is that and that is why these futures need to exist, because, you know, nobody knows what the value of a B200 will be four years from now, five years from now.”
If a neocloud rents its machines to an investment-grade hyperscaler on a ten-year deal, lenders sleep fine. But when an operator rents to smaller startups or sells on-demand capacity, lenders carry naked depreciation risk. If a newer chip makes the B200 obsolete, or if early-stage AI firms run out of cash, the hardware cannot cover the balance sheet. Wilson argues that these credit spreads expose an incomplete market: “Well, here, I mean I do think that it's highlighting the the massive need for risk management products.”
How a Futures Strip Lowers Capital Costs
Commodity markets addressed this dilemma long ago. Oil drillers and power plants do not shut down when spot prices fluctuate; they sell forward contracts to lock in their revenue. Wilson argues that compute requires the exact same financial plumbing.
Selling a forward strip of GPU futures allows a cloud operator to eliminate residual asset risk right away. The operator hedges future revenue through basis trades, shielding the balance sheet from hardware depreciation. Lenders no longer need to demand sky-high yields to protect against obsolescence. That shift reduces the cost of debt, allowing operators to offer flexible, short-term capacity to startups instead of demanding multi-year lock-in agreements.
What to Do With This
Audit your compute commitments this week. Pull every contract your team signed for reserved GPU clusters. If you signed a multi-year deal to secure discounts on current hardware, calculate the penalty for downsizing to six-month blocks. Shift your variable inference workloads to on-demand terms backed by upfront deposits, ensuring you avoid long-term depreciation penalties when new model architectures change your chip requirements.