Nvidia H100 rental prices surge despite Blackwell launch: +40% in six months

The AI computing market has demonstrated unexpected resilience of older hardware: after the release of Nvidia Blackwell systems, the rental of H100 accelerators not only did not become cheaper, but has grown noticeably. Over six months, the average long-term rental rate has risen from approximately $1.70 to $2.

35 per GPU per hour, and available capacity based on Hopper for many providers has essentially run out. Market logic suggested the opposite. H100 is no longer the flagship: it is the previous generation of Nvidia accelerators on the Hopper architecture, while Blackwell is being promoted as the new foundation for the largest AI clusters.

Usually, after a generation change, older GPUs become cheaper because some clients move to faster systems, and more secondary supply appears on the market. But in 2026, this scenario did not work out. Analysts are recording two shortage signals simultaneously.

First, prices are rising themselves: an increase of approximately 40% over six months for a mature product looks atypical. Second, even at such rates, finding available H100s is difficult. Renters are trying to extend existing contracts at any price, in some cases immediately for several years ahead.

This means that for many companies, it is more important to guarantee access to computing than to wait for the market to cool down.

The reason is that Blackwell has not yet managed to replace Hopper in sufficient volume. New instances with Blackwell-generation accelerators are rolling out to market gradually, and significant portion of deployments are expected only closer to the middle of the year. While supply is limited, demand continues to grow faster than providers manage to introduce new racks and clusters.

As a result, the senior systems of the previous generation are not being freed up: they remain the workhorse for model training, fine-tuning, inference, and corporate AI services. This is especially noticeable on training tasks. Despite the appearance of Blackwell, many model training workloads in terms of price/performance still fit well on H100, especially if companies need not experimental mega-clusters, but familiar configurations of dozens or hundreds of GPUs.

For inference, the new generation is indeed more attractive, but in practice, hardware availability often matters more than theoretical performance. If the needed Blackwell cluster is not available in the coming months, business takes what can be deployed now. H100 has a strong practical position for this.

These accelerators are well-known to the market, frameworks are already optimized for them, cloud configurations and internal company pipelines are already tailored to them. Nvidia promotes H100 as a universal accelerator for training and inference of large models on Hopper architecture, and in real clusters, not only peak performance matters, but also supply predictability, compatibility, and clear economics. If Blackwell promises a multiple increase in efficiency on the heaviest workloads of large language models, then H100 remains an understandable and proven tool where migration to a new generation does not pay off immediately.

A separate factor is the very structure of the GPU rental market. The vast majority of deals do not go through public hourly prices, but through medium- and long-term contracts. When demand suddenly accelerates, cloud providers get the opportunity to choose clients, raise rates, and request longer commitments.

For startups and corporate teams, this means one unpleasant thing: scarcity is now expressed not only in price, but in the absence of slots. Even if the budget is there, the needed cluster may simply be unavailable.

The conclusion for the market is simple: in the AI boom, an accelerator is considered "outdated" only on paper. While new system production lags behind demand, the previous generation of Nvidia systems maintains almost strategic value and can become more expensive contrary to the usual logic of hardware updates. For cloud clients, this is a signal to book capacity in advance and be more cautious in planning rapid migration to Blackwell. For Nvidia itself, it is confirmation that demand for computing is still growing faster than infrastructure manages to be updated.