The economics of the AI boom just flipped: chips that were supposed to be worthless in three years are booking revenue into their ninth year of life.
The Summary
- CoreWeave signed contracts to rent Nvidia A100 GPUs through 2029 — chips launched in 2020 that short sellers bet would be obsolete by now
- A $500 billion Wall Street infrastructure push is now routing pension and insurance capital into GPU-backed investments, fundamentally changing how AI buildout gets funded
- The depreciation debate matters because it determines whether Big Tech's multi-billion dollar AI infrastructure spending is visionary or reckless
- If older chips stay economically viable longer, the unit economics of the entire AI cloud business improve dramatically
The Signal
CoreWeave CFO Nitin Agrawal told analysts Tuesday that customers are committing to rent A100 GPUs at "attractive prices" well into 2029. The company is also "largely sold out" of older Nvidia generations. That's nine years of productive life for hardware that critics said would be boat anchors after three.
This directly challenges the bear case on AI infrastructure spending. Short sellers and skeptics have argued that Nvidia's rapid chip improvements would render previous generations economically useless within 24-36 months. If true, companies like CoreWeave, Amazon, and Microsoft would face massive write-downs on their GPU purchases. The accounting implications alone would crater margins.
"We recently signed an A100 contract that extends into 2029 at an attractive price."
But CoreWeave's contracts suggest something different is happening in the real market. Older chips aren't becoming worthless. They're finding price-appropriate workloads. The A100 may not be optimal for cutting-edge model training, but it works fine for inference, fine-tuning, and a dozen other AI tasks that don't require bleeding-edge performance.
Meanwhile, Wall Street is building a $500 billion financing machine to keep GPU purchases flowing. The new structure routes pension and insurance capital into GPU-backed infrastructure investments. Big Tech companies are increasingly using debt to fund AI buildout, betting they can generate returns before the hardware depreciates.
This creates a fascinating circular dependency. The viability of the financing structure depends on the long-term value of GPUs. The demand for GPUs depends on the availability of cheap capital to buy them. And the whole system runs on the assumption that AI workloads will keep growing fast enough to absorb capacity from both new and old chip generations.
Key structural changes:
- Insurance and pension funds treating GPU infrastructure like real estate or energy assets
- Big Tech shifting from CapEx to debt financing for AI hardware purchases
- Secondary markets emerging for older GPU generations at tiered pricing
The CoreWeave data point matters because it provides real-world evidence that the depreciation assumptions built into these financing models might actually hold. If A100s can generate revenue for nine years, the economics of GPU-backed debt start to look reasonable. If they can't, the entire structure is built on sand.
The Implication
Watch the secondary market for older Nvidia chips. If prices stay stable and utilization stays high, it validates the long-depreciation thesis and makes the new financing structures look smart. If prices crater or CoreWeave-style contracts become outliers, the bears were right and there's going to be a reckoning.
For builders and buyers: this changes the math on when to upgrade. If older chips hold value, you can extract more revenue before moving to the next generation. That makes the total cost of ownership more predictable and the business case for AI infrastructure stronger. The winner isn't necessarily who has the newest chips. It's who can match workloads to the right hardware generation at the right price.