GPU rental prices rise
Reports say major cloud and tech providers have raised prices for renting NVIDIA GPUs, signalling that access to premium AI compute is becoming more expensive and concentrated among a few firms. Commentators describe this as a shift toward vendor pricing power over scarce GPU capacity. (xataka.com) (ecosistemastartup.com)
Renting Nvidia’s top artificial intelligence chips now costs more again, after a late-2025 dip gave way to fresh price increases in early 2026. (msn.com) SemiAnalysis data cited by multiple outlets shows one-year H100 lease pricing rising to $2.35 per graphics processing unit hour in March 2026, up from $1.70 in October 2025, a jump of nearly 40 percent in six months. (msn.com) Current list prices still vary sharply by provider. Lambda lists H100 instances at $3.99 per graphics processing unit hour, while CoreWeave lists H100 at $49.24 an hour for an eight-chip system, or $6.16 per chip. (lambda.ai) (coreweave.com) The gap is wider on newer chips. CoreWeave lists H200 at $50.44 an hour for eight chips, or $6.31 each, and B200 at $68.80 an hour for eight chips, or $8.60 each. Lambda lists B200 at $6.69 per graphics processing unit hour. (coreweave.com) (lambda.ai) These are the processors companies rent to train and run large language models. Nvidia’s H100 is the Hopper-generation chip that became the standard workhorse for artificial intelligence clouds after its 2022 rollout. (nvidia.com) (techcrunch.com) Nvidia then moved customers up the stack. The H200 adds 141 gigabytes of high-bandwidth memory and 4.8 terabytes per second of memory bandwidth, and Nvidia introduced the Blackwell B200 platform on March 18, 2024. (nvidia.com) (investor.nvidia.com) Cloud providers are charging not just for the chip, but for scarce capacity tied to power, networking, packaging, and ready-to-use clusters. CoreWeave’s public pricing page says its infrastructure is built around the latest Nvidia graphics processing units and offers on-demand and spot capacity by region. (coreweave.com) The result is a split market. Some specialist providers still advertise H100 access below $4 an hour, while premium managed platforms and hyperscale configurations sit materially higher for similar Nvidia silicon. (lambda.ai) (coreweave.com) That leaves startups and model developers paying closer attention to where they rent compute, not just which chip they want. In April 2026, the price of access depends as much on the landlord as on the graphics processing unit. (lambda.ai) (coreweave.com)
