Nvidia CEO Jensen Huang Just Put a Nail in Michael Burry's Bear Case

Nvidia's older A100 GPUs, introduced in mid-2020, are still being rented under contracts extending through 2029, while rental rates for newer H100 chips reportedly rose 22% month over month. The article argues this undermines Michael Burry's claim that AI GPUs should be depreciated over 2-3 years rather than the commonly assumed six years, which he says inflates AI-company profits. Nvidia CEO Jensen Huang characterized its compute hardware as durable, fungible and revenue-generating; the article also highlights Nvidia's 28x earnings valuation as attractive.

Analysis

The accounting debate is more consequential for GPU owners than for NVDA. A longer depreciable life lifts reported EBITDA-to-net-income conversion and debt-service capacity at leveraged GPU clouds such as CRWV, but it also implies that installed fleets can remain productive longer, reducing the urgency of replacement purchases. For NVDA, the key economic question is not physical durability but whether older generations retain sufficient price/performance to cannibalize new-system demand.

Reported rental-rate moves are an imperfect signal: spot scarcity can raise pricing even while long-duration contracted rates, utilization, and realized returns on invested capital deteriorate. The relevant 1-3 month catalyst is hyperscaler and neocloud disclosure of depreciation-life changes, GPU utilization, and committed-capacity backlog. A broad extension of useful lives would be optically positive for cloud earnings, but could trigger investor concern that 2027-28 accelerator unit growth is being pulled forward.

Consensus is likely too binary on depreciation. Older GPUs can remain economically useful for inference, fine-tuning, and lower-priority workloads while still losing frontier-training share to Blackwell/Rubin. That segmentation is structurally favorable to NVDA's software ecosystem and secondary-market liquidity, but it is less favorable to pure-play GPU renters if legacy supply expands faster than incremental inference demand. The article's valuation framing should not be relied upon without independently reconciling forward EPS, stock-based compensation, and customer concentration.

Key Decisions for Investors

• Maintain a tactical long NVDA only on evidence that next-generation backlog conversion remains intact through the next earnings print; use a 3-6 month horizon and size against the risk that extended fleet lives reduce 2027 replacement demand. Falsify on a material reduction in data-center revenue guidance or management commentary that legacy capacity is displacing new deployments.
• Prefer NVDA over CRWV as a relative-value expression over 6-12 months. NVDA monetizes both new hardware and ecosystem lock-in, while CRWV bears utilization, power-cost, refinancing, and residual-value risk; reassess if CRWV demonstrates sustained contracted utilization and improving free-cash-flow conversion after interest expense.
• Set an alert for disclosed GPU depreciation-policy changes at major cloud customers and CRWV. Treat a shift toward longer lives as a potential near-term earnings tailwind for operators, but not automatically bullish for NVDA until order-growth and legacy-versus-new workload mix are confirmed.
• Do not trade NFLX from this signal; it has no identifiable earnings or competitive linkage to accelerator useful-life assumptions.