Key Takeaways

  • Tech buyers assume GPU compute follows standard hardware depreciation, but forward price curves for compute are actually upward sloping.
  • CoreWeave continues to sign three-year rental contracts for older Nvidia A100 chips, proving older hardware does not quickly become worthless.
  • Securing new server capacity from providers like CoreWeave now pushes delivery dates out to the fourth quarter of 2027 or early 2028.
  • Procurement delays run 20 to 26 weeks, forcing companies to commit capital months before they know actual inference or training volume.
  • Next-generation chips could push hourly costs higher rather than lower, with Nvidia Rubin hardware potentially reaching $20 to $30 per GPU hour.

The Myth of Cheap Future Compute

Silicon Valley assumes hardware always gets cheaper. Moore's Law taught two generations of software founders that if you wait three years, computing power doubles while costs drop by half.

Kush Bavaria found the exact opposite while building Ornn. In financial markets, backwardation describes a setup where future delivery costs less than spot prices today. Founders assumed compute markets would mirror that pattern.

“When we started the company, a lot of people thought the forward curve for chip prices is actually backwardated,” (play from 10:23) Bavaria said. “In the sense that three years out from now or five years out from now, the price for a GPU hour is going to be less.” (play from 10:30)

That price collapse has not happened. Instead, older silicon continues to command long-term leases at stable rates. “CoreWeave has also said their A100s are still being rented out for a three-year period, and the H100s are also being rented out for a very long time,” (play from 17:37) Bavaria noted. When three-year-old hardware commands multi-year commitments, the idea that chips quickly depreciate to scrap value falls apart.

Meanwhile, next-generation silicon is not arriving with discount tags. Bavaria estimates that upcoming architectures will expand the top of the pricing ladder: “I would not be surprised if their Rubin costs are like $20, $30 per GPU hour.” (play from 11:07) When top-end hardware gets more expensive and older chips stay fully booked, compute prices climb along the forward curve.

The 26-Week Procurement Trap

Because capacity remains tight, procurement schedules have detached from ordinary business planning cycles. If a growth-stage company wants to buy direct capacity today, delivery dates stretch into the next decade.

“If you want to go to CoreWeave and buy something, you are probably quoted Q4 2027 or Q1 2028 at this point,” (play from 8:23) Bavaria said. Even when companies buy through shorter cycles, hardware lead times stretch across five to six months.

That delay puts software teams in an impossible spot. You must lock in spending today for capacity that lands half a year later. If your product takes off, you get starved of compute. If your product pivots or usage flattens, you hold balance-sheet liabilities for silicon you cannot fill.

“You need to hedge your risk somehow,” (play from 14:42) Bavaria argued. “If I am buying today, I do not actually know what prices are 22 to 26 weeks from now.” (play from 14:45)

Commodity businesses in oil, wheat, and freight solved this decades ago with forward contracts and paper exchanges. Compute buyers have operated like 19th-century merchants, buying physical assets on multi-year terms and hoping the market does not turn against them.

What to Do With This

Audit your model serving bills from the last three months and split your compute into base load and swing capacity. Sign long-term commitments only for the 40 percent of capacity that never drops off, then force your team to price swing capacity through spot instances or flexible cloud aggregators instead of signing another two-year reservation.