Blackwell GPU prices spike
Hourly rental rates for Nvidia’s top‑end Blackwell GPUs have climbed sharply in recent weeks, making premium compute noticeably more expensive. Reports say the hourly rent hit $4.08 — roughly 48% higher than $2.75 two months ago — a rise attributed largely to demand from agentic AI workloads. (alltoc.com, intellectia.ai)
Renting Nvidia’s newest Blackwell graphics processors now costs sharply more on the spot market, with Ornn’s benchmark at $4.08 an hour on April 13. (techmeme.com) That is up 48% from $2.75 two months earlier, according to the Ornn Compute Price Index item cited by The Wall Street Journal and aggregated by Techmeme and Bens Bites on Monday, April 13. (techmeme.com, news.bensbites.com) Ornn said last week that its Compute Price Index tracks graphics processor hour pricing across cloud and on-premises markets, and that the benchmark is now distributed on Bloomberg Terminal. (prnewswire.com) A graphics processor is the chip that does the heavy math for artificial intelligence, and renting one by the hour lets companies avoid buying servers upfront. Nvidia’s Blackwell line is its current top tier for those jobs. (nvidia.com, docs.nvidia.com) Blackwell systems are built for training models and serving answers after launch, and Nvidia says a DGX B200 server uses eight Blackwell graphics processors in one box. Nvidia markets that system as delivering three times the training performance and 15 times the inference performance of DGX H100. (nvidia.com, docs.nvidia.com) The demand spike is tied to “agentic” artificial intelligence work, where software chains together many model calls and tools to complete a task instead of generating one reply. Nvidia said Blackwell cloud systems were launched to handle “reasoning” workloads that generate many extra tokens and need more memory, compute, and fast chip-to-chip links. (news.bensbites.com, blogs.nvidia.com) Cloud providers are already selling Blackwell at a wide spread. Lambda listed Nvidia B200 capacity at $6.69 per graphics processor hour on April 13, while market trackers published in April showed Blackwell offers ranging from roughly $2.25 to well above $6 depending on contract terms and provider. (lambda.ai, getdeploying.com, tech-insider.org) CoreWeave, one of the first providers to put Blackwell in general cloud availability, said in February 2025 that its GB200 NVL72 instances were aimed at generative and agentic artificial intelligence workloads. Nvidia said the same launch made Blackwell generally available in the cloud. (coreweave.com, blogs.nvidia.com) The price jump does not mean every customer pays $4.08 an hour, because spot prices move with supply, contract length, and how much hardware a buyer reserves. It does show that the cheapest immediately available Blackwell capacity has gotten harder to find as more companies chase the same top-end chips. (prnewswire.com, lambda.ai, getdeploying.com) For startups and model builders, that means the newest Nvidia compute is becoming less of an impulse rental and more of a budget line item, even before software, storage, and networking are added. (lambda.ai, getdeploying.com, coreweave.com)
