Frontier GPU rents spike
Hourly rental prices for Nvidia Blackwell GPUs have jumped sharply, with a market note reporting a rise to $4.08 per hour—up 48% from $2.75 two months earlier—which reflects rising demand for agentic AI workflows. Nvidia also published details about MiniMax M2.7 as a packaged runtime for scalable, multi-step agentic applications across its stack. (intellectia.ai (developer.nvidia.com)
Renting Nvidia’s newest Blackwell graphics processors now costs about $4.08 an hour, according to the Ornn Compute Price Index cited by The Wall Street Journal. (techmeme.com) That is up 48 percent from $2.75 two months earlier, the market note said. The same report said artificial intelligence companies are rationing access to some products as demand for “agentic” systems rises. (techmeme.com) A graphics processor, or GPU, is the chip that powers most modern artificial intelligence training and inference. Renting one by the hour is the cloud-era version of leasing factory time, and the price moves when demand outruns available machines. (prnewswire.com) Blackwell is Nvidia’s current top tier for many large artificial intelligence workloads, and major cloud providers spent 2025 rolling it out. Amazon Web Services said in May 2025 that its P6-B200 instances had reached general availability, while Google Cloud said its A4 virtual machines would use eight Blackwell B200 GPUs. (aws.amazon.com) (cloud.google.com) The demand spike is tied to a shift in how companies use models. Instead of one prompt producing one answer, agentic systems run multi-step jobs like coding, research, and office tasks, which keeps GPUs busy longer and pushes up compute bills. (developer.nvidia.com) Nvidia underscored that shift on April 11, 2026, when it published a technical post about MiniMax M2.7. The company said the model is aimed at “agentic harnesses” and other complex use cases including reasoning, machine learning research workflows, software engineering, and office work. (developer.nvidia.com) Nvidia’s post described MiniMax M2.7 as available across its stack, including Nvidia NIM microservices and Nvidia Dynamo, which package models and serving software so developers can deploy them at scale. In plain terms, the company is selling not just chips but the runtime layer that keeps long, multi-step agents running across many GPUs. (developer.nvidia.com) Ornn said this month that its compute price benchmark is now carried on the Bloomberg Terminal, a sign that GPU time is being tracked more like an infrastructure commodity. If Blackwell rents stay near current levels, the cost of building and serving agentic artificial intelligence products will stay tied to access to frontier chips. (prnewswire.com)
