Blackwell GPU rents spike
Reported rental rates for Nvidia’s Blackwell GPUs jumped to $4.08 per hour, a 48% rise from $2.75 two months earlier, driven in reports by demand for agentic AI workloads. Coverage links the price movement to tighter compute economics for inference and training at scale. ( )
Renting Nvidia’s newest Blackwell graphics processors now costs about $4.08 an hour on the spot market, up 48 percent from $2.75 two months earlier. (techmeme.com) The price move came from the Ornn Compute Price Index, which tracks negotiated graphics processor rental prices across cloud and on-premise markets and was added to the Bloomberg Terminal on April 2, 2026. (prnewswire.com) The Wall Street Journal report summarized by Techmeme said the jump was tied to rising demand for “agentic” artificial intelligence, software that does multi-step work and keeps calling models in the background. (techmeme.com) Blackwell is Nvidia’s latest data-center platform, built for both training and inference, the stage when a model is already trained and starts answering users in real time. Nvidia said in March 2024 that Blackwell could cut large-language-model inference cost and energy use by as much as 25 times versus its predecessor. (nvidianews.nvidia.com) Nvidia’s own product material pitches Blackwell as infrastructure for “AI factories,” with GB200 NVL72 linking 72 Blackwell graphics processors in one rack for trillion-parameter inference and training. (nvidia.com) That matters for rental pricing because newer artificial intelligence products are running more inference after launch, not just spending on one-time training runs. Nvidia said Blackwell was designed for “extreme-scale AI inference,” and its March 2026 materials cited cloud deployments for low-latency, long-context agentic coding workloads. (blogs.nvidia.com; blogs.nvidia.com) The squeeze is showing up beyond one index. Edgen, citing Ornn data and the Wall Street Journal, reported that CoreWeave raised prices by more than 20 percent and lengthened contract terms as capacity tightened. (edgen.tech) The same report said Anthropic’s Claude application programming interface availability fell to 98.32 percent in March, below the 99.99 percent uptime target many enterprise customers expect. (edgen.tech) Blackwell’s appeal is speed per watt and speed per dollar. Nvidia said HGX B200 can deliver up to 15 times higher inference performance than the Hopper generation on very large models, while MLPerf results published in November 2024 showed roughly 2 times per-graphics-processor gains on some training tasks. (nor-tech.com; developer.nvidia.com) For now, the headline is simple: the chips built to make artificial intelligence cheaper to run are getting more expensive to rent as more companies try to use them at once. (techmeme.com; prnewswire.com)
