The Great GPU Shortage – Rental Capacity – Launching our H100 1 Year Rental Price Index (AI column)

Interesting. article From SemiAnalysis about the aggressive GPU rental market, which entered the clinch and prices soared due to rising demand. The value of the annual contract for the H100 increased by 40per cent1.70 GPU hour in October 2025 $2.35 by March 2026. On-demand capacity is essentially chosen for all types of GPUs, and Blackwell clusters that go into operation in the middle. 2026 In many cases, they are already busy in advance. Even find a cluster 64 Hopper GPU has become non-trivial.

The authors directly link the new wave of pressure on the market with a sharp increase in demand against the background of agentic loads on top of models. That is, the bottleneck is now formed not only by the race of laboratories in the training of new models, but by production-engineering, where the model ceases to be a “chat” and becomes the executor of work processes. In the classic chat UX, the economy was relatively simple: one user request, one reference pass, one response. In the agent world, one task is decomposed into a call tree: the orchestrator builds a plan, generates subagents, who simultaneously search for sources, read code, go to tools, return intermediate results, after which a separate circuit makes synthesis, criticism and verification. For example, from the outside, the task may look like “deal with the incident” or “prepare the PR”, and under the hood it is already a swarm of subagents.

As a result, the demand in the market is growing not only by the number of users and not only by the number of companies, but rather by how many tasks are delegated to agents. The same developer, analyst, or researcher can now run not a single query to a model, but dozens of coordinated passes—with long context, retros, checks, and parallel branches. Multi-agent loads have an unpleasant feature for infrastructure: they increase not only the total consumption of tokens, but also peak competition, and this is a direct pressure on the GPU fleet.

Interestingly, this demand is uneven. The authors write that some inference workloads, especially larger MoE scenarios, work better on fresh systems like the GB300 NVL72, while workouts and parts of workloads where performance costs are important continue to live well on the H100. In other words, the new agent layer isn’t just “eating Blackwell” – it’s also supporting demand for the Hopper, extending the economic life of the GPUs already delivered. Therefore, providers now do not have a price war for download, but a seller market with long contracts, prepayment and a choice of who to give capacity at all.

A few conclusions can be drawn from this: 1Planning of capacities according to the model “how many chats per day” no longer works – you need to count fan-out on subagents, the depth of verification cycles and the level of parallelism (task-maximum) 2A unit economy should be considered not on request, but on completed tasks: an agent can be expensive in tokens, but cheap relative to the cost of a person or the cost of delaying a business process. 3If the ROI from AI tools really remains multiples above their cost, as SemiAnalysis writes, then the demand for computing will be weakly price sensitive for a long time. So, the growth of GPU consumption is not a temporary surge of hype, but a new reality.

TLDR; GPUs used to consume mostly models, but now GPUs consume business processes broken down into swarms of agent subtasks. And it is multi-agent loads that may be the main hidden driver of the new computing deficit.

#AI #Engineering #Management #SystemDesign #Architecture #Processes #Software #Productivity #Agents #Economics