Blackwell GPU rents spike
Hourly rental for Nvidia Blackwell GPUs climbed to $4.08 — roughly a 48% rise from $2.75 two months earlier — according to a compute‑pricing index. Analysts and coverage link the jump to growing agentic AI workloads that consume extra inference cycles because they plan, call tools and retry, making compute costs a direct product-design concern (alltoc.com).
Renting Nvidia’s newest Blackwell graphics processors now costs about $4.08 an hour, up 48 percent from $2.75 two months ago. (techmeme.com) The price jump comes from the Ornn Compute Price Index, a benchmark that tracks graphics processor hour pricing across cloud and on-premise markets and was added to Bloomberg Terminal on April 2. (prnewswire.com) A graphics processing unit, or graphics chip, is the rented engine behind many artificial intelligence systems, and Blackwell is Nvidia’s latest data-center generation for training, fine-tuning, and inference, which is the step where a model answers a user request. (nvidia.com) The recent pressure is tied to “agentic” artificial intelligence tools, which do more than return one answer: they break tasks into steps, call outside tools, check results, and often retry, which uses more inference time per user session. (blogs.nvidia.com) That extra work is showing up in rental markets even for older chips. One March 2026 report said one-year Nvidia H100 rental pricing rose to about $2.35 an hour from roughly $1.70 in October 2025, even after Blackwell systems started arriving. (stockminded.com) Blackwell hardware is built to handle bigger artificial intelligence jobs with more memory and faster data movement than the prior Hopper generation. Nvidia says its DGX B200 system is designed for training, fine-tuning, and inferencing in one platform. (nvidia.com) Some cloud marketplaces still show lower spot prices for a Blackwell B200 than the index level, which suggests the $4.08 figure reflects a broader market benchmark rather than the cheapest available listing at any moment. Spheron, for example, recently advertised B200 spot access from $2.25 an hour and on-demand access at $6.03 an hour. (spheron.network) Nvidia and its cloud partners are also pitching newer Blackwell Ultra systems as a way to cut the cost of these agent-style workloads. Nvidia said last month that SemiAnalysis data showed Blackwell Ultra could deliver up to 35 times lower costs for agentic artificial intelligence tasks, with Microsoft, CoreWeave, and Oracle Cloud Infrastructure deploying GB300 NVL72 systems. (blogs.nvidia.com) For companies building chatbots, coding assistants, and research tools, the bill now depends less on one flashy demo than on how many steps each answer takes. The market price for a Blackwell hour is turning that design choice into a line item. (techmeme.com)
