1 Chart That Destroys The Nvidia Bear Thesis

Nvidia H100 cloud-GPU rental prices rose 22% over the past month to $3.28 per hour, despite the chip being roughly three years old and Nvidia launching its newer Rubin/R100 platform. The pricing rebound suggests older Nvidia accelerators retain meaningful demand and residual value, countering investor concerns around rapid chip depreciation and AI-infrastructure oversupply. With revenue nearly doubling in its latest quarter and the shares trading at about 29x earnings, roughly in line with the S&P 500, the article argues Nvidia remains undervalued relative to its growth outlook.

Analysis

The relevant signal is not a one-month rate increase itself, but that installed H100 capacity appears to retain monetization value through a major architecture transition. That weakens the impairment/rapid-obsolescence case for GPU lessors and hyperscalers, potentially extending useful lives and reducing the need for accelerated depreciation charges. For NVDA, the direct revenue impact is limited because secondary-cloud rental rates do not accrue to Nvidia; the read-through is higher customer ROI, more durable residual values, and therefore greater willingness to finance additional capacity.

The second-order beneficiary is the leveraged GPU-cloud cohort—CoreWeave (CRWV), Nebius (NBIS), and Applied Digital (APLD)—where utilization and realized GPU-hour pricing drive debt-service coverage far more than they do at NVDA. However, a spot rental index can be distorted by regional supply, premium networking configurations, and short-term capacity tightness; it is not sufficient evidence that normalized utilization or contract pricing has inflected. The market should demand confirmation in quarterly disclosed utilization, backlog conversion, and capex guidance before capitalizing this into long-duration earnings.

Near term, this supports NVDA multiple resilience into Rubin availability and hyperscaler earnings, but it does not eliminate the substitution risk from custom ASICs. The key test over the next 1-3 months is whether older-generation rates remain firm after new-platform supply ramps; if they do, the AI stack looks less like a single-product replacement cycle and more like a tiered compute market. Over 6-18 months, durable used-GPU economics could lower the cost of AI inference, expanding demand but also shifting incremental spend toward lower-cost capacity and cloud operators rather than exclusively frontier training hardware.

Key Decisions for Investors

• Maintain/add NVDA on weakness rather than chase the rental-rate headline; use a 3-6 month horizon into next earnings. Thesis requires data-center revenue and gross-margin guidance to remain intact despite Rubin ramp costs; reduce if management signals a material mix-driven margin reset or hyperscaler capex deceleration.
• Watch-list long CRWV versus short a broad software proxy (IGV) only after CRWV reports sustained utilization and contracted backlog conversion. A durable older-GPU pricing floor improves asset coverage and refinancing optics, but the position is not actionable without updated leverage, utilization, and customer-concentration data.
• Pair trade: long NVDA / short AMD over 3-6 months if post-Rubin channel checks show older Nvidia generations retaining rates while AMD accelerators require discounting to win workload share. Falsifier: verified MI-series availability gains accompanied by comparable rental-rate strength and material customer deployments.
• Set an alert for a reversal in H100 rental rates below the prior-month level while new-platform availability rises. That combination would indicate supply normalization rather than structural asset durability and would re-open downside risk for GPU lessors and NVDA's terminal-margin assumptions.