Blackwell GPU Rentals Spike
Hourly rental costs for Nvidia Blackwell GPUs have jumped to about $4.08 — roughly a 48% increase from $2.75 two months earlier — reflecting tighter short-term access to usable AI compute. Market reports link the spike to growing demand for agentic AI workloads that need steady GPU time rather than one-off experiments. (alltoc.com) (intellectia.ai)
Renting Nvidia’s newest Blackwell graphics processors now costs about $4.08 an hour, up from $2.75 two months ago. (techmeme.com) That price jump comes from the Ornn Compute Price Index, a benchmark based on negotiated GPU rental transactions rather than posted list prices. Ornn said on April 2 that its index tracks rental pricing for chips including Nvidia’s H100, H200, and B200 across providers, configurations, and locations. (prnewswire.com) A Blackwell GPU is Nvidia’s latest data-center chip for artificial intelligence work, and the B200 version carries 192 gigabytes of high-bandwidth memory. Nvidia says Blackwell systems are built for training, fine-tuning, and running large language models at much higher speed than the prior Hopper generation. (nvidia.com) Cloud providers have been rolling Blackwell out gradually, not all at once. CoreWeave said its Nvidia HGX B200 instances became generally available in May 2025 after an initial February 2025 deployment for a customer using more than 10,000 Blackwell graphics processors. (coreweave.com) The recent squeeze is tied to “agentic” artificial intelligence systems, which are tools that keep working through multi-step tasks instead of answering a single prompt and stopping. OpenAI’s Responses application programming interface is marketed for “agent-like applications” with built-in web search, file search, computer use, and code tools, and Anthropic has separately promoted computer-use and agent products for longer-running tasks. (openai.com) (anthropic.com) Those workloads tend to hold onto graphics processors for longer stretches, because the model is browsing, clicking, calling tools, and revising work across many steps. OpenAI’s computer-use documentation describes models that click, type, scroll, and inspect screenshots, while Anthropic describes computer use as fitting models into ordinary software environments rather than custom one-shot tools. (developers.openai.com) (anthropic.com) The market is also fragmented, which makes “the” Blackwell price hard to pin down. A March 21 pricing guide compiled from provider listings put single-GPU B200 rentals at $5.98 an hour on RunPod and $6.08 on Lambda, while CoreWeave’s published B200 access started with an eight-GPU cluster priced at $68.80 an hour. (deploybase.ai) Nvidia is selling Blackwell not just as a single chip but as tightly linked systems. Its GB200 NVL72 rack connects 72 Blackwell graphics processors and 36 Grace central processing units with one NVLink domain, which Nvidia says acts like a single massive GPU for trillion-parameter model training and inference. (nvidia.com) That helps explain why short-term rentals are getting tighter even as more hardware ships: buyers are not only chasing raw chip counts, but also specific configurations, interconnects, and memory footprints that let larger models run continuously. Ornn said posted list prices often diverge from what large buyers actually pay, and that configuration, provider, and location all change the clearing price. (prnewswire.com) For now, the clearest signal is the hourly rate itself: usable Blackwell capacity got more expensive in early April 2026, and the market is treating steady-access AI compute as scarcer than it looked two months ago. (techmeme.com)
