Key Takeaways

  • Compute rental markets have flipped into contango, meaning forward contracts for 2027 delivery cost more than near-term capacity.
  • Atreides Management portfolio companies booking 4,000 Nvidia Blackwell GPUs for mid-2026 faced steep price jumps when trying to add another 4,000 in September 2027.
  • Spot prices for GPU rentals rose daily throughout October, disproving assumptions that supply growth would drive down near-term compute costs.
  • CoreWeave trapped itself in fixed, long-term customer agreements that sit far below current market clearing rates.
  • Shorter-term compute sellers are capturing rapid margin expansion, recovering hardware payback periods in under twelve months.

Future Compute Is Getting More Expensive

Tech hardware usually gets cheaper over time. You pay top dollar for the first run of silicon, and two years later the same capacity sells at a discount. In commodities, when prices fall over time, traders call it backwardation. Today, the artificial intelligence compute market runs in reverse.

Carmen Lee, founder of Silicon Data, made waves at GTC by declaring that compute has entered contango. In a contango market, future prices trade above spot prices. Buyers booking hardware for two or three years out are paying more, not less, per hour.

Baker saw this shock directly across his investments. “Just the number of companies we have in our portfolio who are like, okay, we got a cluster of 4,000 Blackwells in May or June of 2026, and we want to bring on another cluster of 4,000 in September of 2027,” Baker explained. “And they are shocked by how much more they have to pay in the future.”

Despite headline capital spending announcements from hyperscalers, available supply cannot match appetite. When buyers try to guarantee physical power, cooling, and chips for late 2027, vendors price in growing scarcity.

The Cost of Locking In Early

The price dynamics have created a massive split between infrastructure operators. Early providers chased the comfort of stable, multi-year revenue. That choice backfired.

Baker pointed to CoreWeave as an example of operational excellence hurt by poor pricing strategy. “CoreWeave is great. They run GPUs well. They bring them out. They are very, very good,” Baker said. “They just sadly locked in a vast amount of compute in long-term contracts that are now way under market.”

By locking customers into multi-year agreements at older rates, operators surrendered upside during an explosive demand squeeze. Meanwhile, companies that held capacity or sold on spot markets cleaned up. Sarah Friar observed that critics laughed at their heavy compute commitments, but they could resell that capacity at three, four, or five times their cost.

Daily spot data backs this up. “October has kind of been a strange month because the price to rent essentially any kind of a GPU has gone up every day,” Baker noted. “So the spot market for compute is getting tighter.”

Infrastructure providers selling on floating or short-term durations are recovering capital expenditures in less than a year. Those who committed to flat fees for three years are stuck subsidizing their customers.

What to Do With This

Audit your compute agreements and avoid signing flat, three-year reserved capacity deals if you run infrastructure. If you buy compute, lock in firm 2026 and 2027 capacity quotes this week before quotes adjust higher again.