GPU Rental Market Heats Up, Validating "Computing as an Asset" Thesis
Recent market data from Silicon Data reveals a broad-based increase in rental rates for key AI computing hardware within the non-hyperscaler segment. Current hourly rates per GPU stand at approximately $1.65 for the A100, $2.72 for the H100, $3.29 for the H200, and $5.61 for the B200. Notably, the rental price for the H100 has rebounded significantly from its level of around $2 per hour late last year, while the newer B200 and H200 models are also commanding strong demand.
Perhaps more telling is the estimated residual value of these assets. Financial models based on forward rental curves now project the 36-month residual value of an H100 to be around $20,000, up from an estimated $14,000 in October last year. It's critical to note that this is a model-derived valuation based on income potential, not a current secondary-market transaction price. This upward trend lends initial market support to the "computing power as an asset" concept frequently discussed by NVIDIA CEO Jensen Huang.
From Aircraft to GPUs: Stable Cash Flow Fuels Mega-Financing
The core logic hinges on a classic principle of asset financing: once a GPU cluster can generate stable, predictable inference rental revenue, it acquires the key attribute of bankable collateral—reliable future cash flow. Industry reports indicate that NVIDIA, alongside major financial institutions like Apollo and BlackRock, is advancing a specialized financing initiative for AI infrastructure that could exceed $500 billion in scale.
The innovative aspect of this model is that a portion of the loans would be secured directly by the AI computing equipment itself. To bolster lender confidence and reduce financing costs, NVIDIA is reportedly considering providing guarantees for this collateral, potentially covering up to 25% of the model-derived residual value. If successful, this structure could dramatically lower the capital barrier for emerging AI cloud providers. The influx of affordable capital, in turn, would likely fuel continued demand for NVIDIA GPUs, creating a potential virtuous cycle.
The Underlying Risk: Technological Depreciation Remains a Formidable Force
Behind this narrative of financialization, sobering data presents a counterpoint. Silicon Data's own secondary market listings show that a three-year-old H100 GPU currently trades at just 20-30% of its original peak price. This starkly illustrates that technological obsolescence remains a powerful and real force in the fast-evolving AI hardware landscape.
Conversely, the older A100 GPU has maintained relatively stable rental rates over a six-year period. This likely only proves that for cost-sensitive workloads like inference and fine-tuning—which don't always require cutting-edge peak performance—older generation hardware retains economic utility and value.
Critical Takeaway: Reassessing Economic Life vs. Financing Security
A more nuanced industry insight emerges: the economic useful life of AI GPUs may be longer than the previously assumed 2-3 years. Robust and growing inference demand is partially offsetting the depreciation driven by pure technical obsolescence. This assessment is vital, as it sits at the heart of the proposed $500 billion financing plan—the long-term stability of the collateral's value.
The future trajectory of AI computing's financialization will depend not just on the slope of the rental curve, but on the prolonged tug-of-war between the curve of technological depreciation and the hardware's ability to generate cash flow. Financial institutions will need to find a delicate balance between embracing innovation and upholding fundamental risk discipline.