Nvidia H100 rental prices climb despite Blackwell launch: up 40% in six months

The AI compute market has shown surprising resilience in older hardware: after the launch of Nvidia Blackwell systems, H100 rental prices not only failed to fall but rose noticeably. Over six months, the average long-term rental rate climbed from roughly $1.7 to $2.35 per GPU per hour, and many providers have effectively run out of Hopper-based capacity.

Market logic suggested the opposite. The H100 is no longer the flagship: it is the previous generation of Nvidia accelerators, built on the Hopper architecture, while Blackwell is pitched as the new foundation for the largest AI clusters. After a generational shift, older GPUs usually get cheaper, because some customers move to faster systems and more secondhand supply reaches the market. In 2026, that scenario did not play out.

Analysts see two signs of scarcity. First, prices themselves are rising: an increase of about 40% in six months looks unusual for a mature product. Second, even at those rates, free H100s are hard to find. Renters try to extend existing contracts at any price, in some cases several years ahead. For many companies, guaranteeing access to compute matters more than waiting for the market to cool.

The reason is that Blackwell has not yet replaced Hopper at the needed scale. New Blackwell-generation instances are reaching the market gradually, and a significant part of deployments is expected only closer to mid-year. While supply is limited, demand keeps growing faster than providers can bring new racks and clusters online. As a result, last-generation systems are not being freed up: they remain the workhorse for model training, fine-tuning, inference and enterprise AI services.

This is especially visible in training. Despite Blackwell, many training workloads still fit H100s well on price-to-performance, particularly when companies need not experimental mega-clusters but familiar configurations of tens or hundreds of GPUs. For inference, the new generation is genuinely 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, businesses take what they can launch now.

The H100 has a strong practical position here. The market knows these accelerators well, and frameworks, cloud configurations and internal company pipelines are already optimized for them. Nvidia promotes the H100 as a universal accelerator for training and inference of large models on the Hopper architecture, and in real clusters what counts is not only peak figures but also predictable supply, compatibility and clear economics. Blackwell promises a multiple-fold efficiency gain on the heaviest large language model workloads, but the H100 remains a familiar, proven tool where migrating to the new generation does not pay off immediately.

Another factor is the structure of the GPU rental market itself. Most deals do not go through public hourly prices but through medium- and long-term contracts. When demand accelerates sharply, cloud capacity providers can pick their customers, raise rates and ask for longer commitments. For startups and corporate teams, this means one unpleasant thing: scarcity now shows up not only in price but in a lack of slots. Even with a budget, the cluster you need may simply be unavailable.

The takeaway for the market is simple: in the AI boom, an accelerator is "obsolete" only on paper. While new system output lags behind demand, the previous Nvidia generation keeps almost strategic value and can get more expensive, defying the usual hardware-refresh logic. For cloud customers, it is a signal to reserve capacity early and to be more careful about plans for a quick migration to Blackwell. For Nvidia itself, it confirms that demand for compute is still growing faster than infrastructure can be updated.