H100 rental prices skyrocket, even A100 is rising; the theory of old GPU depreciation debunked?

Author | Eric

According to Bloomberg's latest A100 and H100 rental price indices (corresponding to Silicon Data’s A100 Rental Index and H100 Rental Index), GPU rental prices increased in February. The H100 showed a strong performance with consecutive monthly increases, indicating that AI demand is accelerating for the second time. Even so-called 'old cards' are still cash cows.

The rental price of H100 continues to rise, while A100 also sees a slight increase

As of February 24, the Silicon Data A100 Leasing Index on Bloomberg was at USD 1.40, while the H100 Index stood at USD 2.43, rising consecutively for several months and approaching the peak seen at the end of March 2025.

Notably, H100 currently commands a premium of about 74% over A100, but the price of A100 is not worthless as an 'outdated' product. This indicates that computing power buyers are paying a premium for immediate computing resources, driving up the utilization rate of older cards. In situations where new cards are in short supply, computing demand providers are snapping up older cards.

The argument that 'old cards are worthless' has been debunked

A100 was launched in May 2020, when NVIDIA stated it had entered full production and started shipping. H100 debuted at the GTC conference in March 2022, with the first partner products rolling out in October 2022.The first batch of A100 customers has already surpassed the 4-5 year depreciation period for servers, and H100 is also nearing its depreciation period, yet both continue to perform strongly.

This effectively refutes the views of those who previously bet against AI. Nearly six years after the release of A100, its rental price is increasing rather than collapsing. If A100 were truly just a rapidly depreciating, short-lived research experiment, then the clearing price for this scarce 'immediate GPU supply' should approach zero.

However, the reality is the opposite: long-tail demand remains robust, especially when the latest chips are in short supply, allowing companies to use them for inference, fine-tuning, and as a transitional solution.

To some extent, this alleviates market concerns over high capital expenditure by major players in 2026-2027

During this earnings season, the market has been alarmed by the escalating capital expenditures of major tech firms for 2026 and 2027. However, the strong leasing performance of the previous generation of GPUs has improved valuation models, proving that older cards can still maintain high utilization and monetization capabilities while new card production ramps up.

This means that hyperscalers' technological transformation no longer relies on the ROI of a single generation of AI chips, operating more like a rolling model where products from different generations can generate profits.

who benefits the most?

$NVIDIA (NVDA.US)$and other companies in the AI semiconductor supply chain are the first to be affected.because a tight leasing market typically indicates strong end-user demand and willingness to pay for performance-based tiering.

Cloud platforms are quick to follow.The rise in leasing prices has increased monetization opportunities for both new and older computing clusters. This benefits not only large cloud platforms such as $Microsoft (MSFT.US)$、 $Amazon (AMZN.US)$、 $Alphabet-C (GOOG.US)$and$Oracle (ORCL.US)$ , but also those firmly focused on GPU computing power within the Neocloud camp, such as $CoreWeave (CRWV.US)$ and $NEBIUS (NBIS.US)$ 。

173K Views