exponentialdave

NVDA just told you the depreciation debate is irrelevant and nobody’s repricing. Burry closed his fund after shorting NVDA, claiming companies overstate GPU useful life at 2-3 years. The market debates whether depreciation should be 3, 5, or 7 years. AWS just cut AI hardware depreciation from 6 years back to 5. Then NVDA drops this in their earnings transcript: A100 GPUs shipped in May 2020 are still running at full utilization in November 2025, powered by vastly improved software. That statement just invalidated the entire depreciation debate. Here’s what people are getting wrong. They’re modeling GPU economics like traditional data center hardware with fixed depreciation curves. CoreWeave uses 6-year depreciation cycles. Microsoft and Alphabet extended server equipment to 6 years. Everyone’s arguing about the right number. But NVDA just revealed the actual dynamic. The value isn’t depreciating on a fixed schedule. Software improvements are extending hardware utility faster than the hardware ages. CUDA optimizations, framework updates, and model efficiency gains mean a 2020 chip running 2025 software delivers more economic value than a 2020 chip running 2020 software. Think about what “full utilization” means for data center economics. These aren’t spare chips sitting idle. They’re generating revenue at capacity 5.5 years after purchase. The software moat keeps expanding, which keeps old hardware economically viable. This completely inverts the depreciation model everyone’s fighting about. The question isn’t “when does the hardware become obsolete?” The question is “how fast can software improvements extend hardware utility?” A100 resale prices saw appreciation in early 2024 during H100 supply constraints, proving the market values backward compatibility and software leverage over raw specs. The competitive dynamic here is brutal. NVDA controls the entire software stack. Every CUDA improvement, every framework optimization, every model efficiency gain extends the life of every GPU they’ve ever sold. Competitors need to match not just the current generation, but the accumulated value of 15 years of software development. Burry bet on hardware obsolescence. NVDA just showed him software determines economic life, not silicon generations.