The Circuit EP 138: Talking GPU Depreciation, OCP Takeaways
11 snips
Oct 20, 2025 The debate around GPU depreciation takes center stage, with insights on how AI might shorten GPU lifecycles. Jay highlights the physical versus economic lifespan of GPUs, emphasizing profitability trends in their early years. The duo also discusses NVIDIA's capacity to meet growing demand and the implications for future data centers. Networking advancements and the rise of neoclouds reveal new opportunities for vendors. Finally, they dive into the potential AI infrastructure bubble and its impact on financing and power constraints.
AI Snips
Chapters
Transcript
Episode notes
Physical Vs Economic GPU Lifespan
- GPU economic life and physical life diverge: cards can work 6–7 years but stop being profitable sooner.
- Rapid generation improvements can compress economic lifespan even if hardware still functions.
Pricing Curve Drives GPU Profitability
- Price curves matter most: most profitability for a new GPU rack occurs in the first two years under a steep downward pricing curve.
- Demand shifts (e.g., DeepSeek) can temporarily raise prices for older systems and extend profitability.
Nvidia Racks Won't Replace Most Capacity
- Nvidia's rack production is a fraction of global capacity so replacements can't be instantaneous.
- Even selling 60k racks would replace under ~20% of AI gigawatt capacity next year.