Blackwell GPU price spike

Renting Nvidia’s newest Blackwell graphics processors now costs about $4.08 an hour, up from $2.75 two months ago, according to the Ornn Compute Price Index. (techmeme.com) A graphics processing unit, or GPU, is the chip that does the heavy lifting for artificial intelligence systems, and Blackwell is Nvidia’s latest data-center line. Nvidia finance chief Colette Kress said in August 2024 that demand for Blackwell was “well above supply,” and in November 2024 the company said demand was expected to exceed supply for several quarters. (cnbc.com) (crn.com) The pricing jump is showing up as Blackwell systems spread through cloud providers that rent computing by the hour. Lambda lists Nvidia B200 SXM6 instances at $6.69 per GPU-hour, while Nebius says Blackwell platforms are still in pre-order rather than general on-demand pricing. (lambda.ai) (nebius.com) The new buyers are not just training giant models once and parking the hardware. Nvidia said in March 2025 that Blackwell Ultra was built for “agentic AI” jobs such as coding assistants and long-context systems, which keep GPUs busy serving many back-and-forth requests with low delay. (blogs.nvidia.com) That helps explain why rental prices can rise even as Nvidia pitches Blackwell as cheaper per task. A faster chip can lower the cost of each answer, but if more companies rush to deploy agents at the same time, the market price for scarce hours can still climb. (blogs.nvidia.com) (prnewswire.com) Blackwell supply has also been constrained by a complicated rollout. Nvidia said in April 2025 that Blackwell chips had started production at Taiwan Semiconductor Manufacturing Company’s Phoenix plants, while mass production of Nvidia supercomputers in Texas was still expected to ramp over the next 12 to 15 months. (blogs.nvidia.com) Some of the largest systems are only now becoming broadly available. Nvidia said CoreWeave put Grace Blackwell GB200 NVL72 systems online at scale in April 2025, and Oracle said it was offering Grace Blackwell systems across public cloud, dedicated regions, and customer-owned deployments. (blogs.nvidia.com) (blogs.oracle.com) The Ornn index itself is part of a newer market that treats compute more like a traded input than a fixed cloud menu price. Ornn said on April 2 that its benchmark for GPU-hour pricing had been added to the Bloomberg Terminal so investors and operators could model revenue, cost, and collateral against changing rental rates. (prnewswire.com) For companies building artificial intelligence products, the immediate number is still the hourly bill. As long as Blackwell demand stays ahead of available capacity, premium compute is likely to keep behaving less like commodity cloud and more like a scarce reservation book. (techmeme.com) (crn.com)