Key Takeaways

  • Standard venture capital models write off high-end GPUs over three years, expecting rapid obsolescence to destroy their economic value.
  • Crusoe depreciates hardware on a six-year accounting schedule, matching real-world utility across changing workload types.
  • Hopper chips purchased in 2023 commanded higher rental rates three years later than when they first hit the market.
  • Packaging older compute through managed APIs allows infrastructure providers to monetize chips long after initial payback periods.

The Three-Year Paper Panic

When Crusoe started buying large fleets of Nvidia Hopper chips in 2023, venture investors warned co-founder Chase Lochmiller that the hardware would lose all value within three years. The conventional narrative assumed that every new architecture release would crush the pricing power of previous generations.

That math ignored how software developers actually consume hardware. “Here we are 3 years later and the prices being charged for utilizing hoppers is higher than the rates that were being charged 3 years ago when they were when they were brand new,” Lochmiller noted.

Standard accounting models often rely on a six-year timeline. “The way we depreciate the assets today is we use a six six-year depreciation cycle,” Lochmiller explained. “That's sort of like the standard across the industry.” The market mismatch happens because spreadsheet models treat hardware as a direct raw rental, assuming clients will only pay for peak frontier training speeds. In practice, secondary workloads create sustained market demand.

Why Older Silicon Stays Profitable

Frontier model training demands the newest, fastest chips on the market. Once those models enter production, the economic priorities flip. Running inference, fine-tuning smaller open models, and serving high-volume API traffic do not require top-tier hardware. They require cheap, stable, reliable compute.

When providers stop selling bare-metal access and start selling managed services, hardware age stops mattering to the end customer. “When you look at that business where it abstracts away the actual underlying chip, the underlying compute from the service that people are receiving, it opens up new monetization engines that can persist for a much much longer period of time,” Lochmiller said.

Application demand fills the capacity vacuum every time. As Lochmiller observed, “I think people underestimate the ingenuity of applications and application developers of turning compute capacity into value and into valuable services for the economy.” As long as software engineers build products that generate revenue, secondary compute remains a cash-producing asset rather than dead inventory.

The Margin Expansion Trap for Buyers

If older chips keep their value, founders buying compute should stop paying top-tier rates for raw hardware leases. Infrastructure companies that amortize silicon over six years can extract high gross margins by routing API queries to older nodes while charging standard platform fees.

For teams building AI products, this dynamic creates two distinct paths. If you need pure training runs on raw clusters, you pay for the newest generation. If you are serving user prompts, running batch tasks, or fine-tuning ten-billion-parameter models, locking into contracts on top-of-the-line clusters is pure waste. You can negotiate aggressive discounts by requesting previous-generation hardware or using API providers that handle the hardware matching behind the scenes.

What to Do With This

Audit your infrastructure spend this week by splitting compute costs into training clusters and inference capacity. For any workload that does not require frontier training speeds, migrate traffic away from flagship hardware instances and test your latency on previous-generation chips.