Silicon Data data shows that in the non-Hyperscaler market, the rental rates for A100, H100, H200, and B200 are approximately $1.65, $2.72, $3.29, and $5.61 per GPU hour, respectively. H100 has rebounded significantly from $2 at the end of last year, while B200 and H200 have also strengthened. Based on a 36-month forward rental curve, the estimated residual value of H100 has risen from $14,000 in October last year to approximately $20,000 currently (residual value is a model estimate, not the actual transaction price).
The rental trend confirms Jensen Huang's logic of assetizing computing power: once GPUs stably generate inference rental income, they can be used as collateral for financing, similar to airplanes and ships, using future cash flow. On August 10th, Nvidia, together with Apollo, BlackRock, and others, promoted a financing round for over $500 billion in AI infrastructure, with some loans secured by equipment. Nvidia may provide guarantees of up to 25% of the residual value. If this model is successful, it will reduce Neo-cloud financing costs and, in turn, support GPU demand.
However, it's important to note that Silicon Data's own secondary market data shows that the listing price of an H100 around 3 years old is only 20%-30% of its peak new product price, indicating significant technological depreciation. The stable rental rates of the A100 over six years only suggest that older models still have value in low-cost scenarios such as inference and fine-tuning. A more accurate assessment is that the economic lifespan of GPUs may be longer than the previously assumed 2-3 years, with strong inference demand offsetting some of the technological depreciation. This is crucial for the valuation of collateral in the $500 billion funding plan.
