Blackwell GPU rents jump

Renting Nvidia’s newest Blackwell graphics processors now costs about $4.08 an hour, up from $2.75 two months earlier, according to the Ornn Compute Price Index. (techmeme.com) The jump works out to roughly 48% in about eight weeks, based on the prices cited Monday, April 13, 2026, by The Wall Street Journal through Techmeme and other aggregators. (techmeme.com) A graphics processing unit rental market works like short-term apartment leasing for computing: startups and labs rent chips by the hour instead of buying servers outright. Blackwell is Nvidia’s current top data-center architecture, which the company says is now in full production. (nvidia.com) Blackwell systems are built for training and serving large artificial intelligence models, especially “reasoning” systems that generate many extra tokens before answering. Nvidia said in February 2025 that those workloads need more memory, faster chip-to-chip links, and more total compute than earlier chatbot-style systems. (blogs.nvidia.com) Cloud providers started widening access to Blackwell over the last year, but supply has stayed concentrated in a small group of specialist operators. Nvidia said CoreWeave was the first cloud provider to make GB200 NVL72 Blackwell systems generally available in February 2025. (blogs.nvidia.com) CoreWeave said in May 2025 that it expanded general availability of Nvidia HGX B200 instances after an initial February 2025 deployment for one strategic customer that used more than 10,000 Blackwell graphics processors. (coreweave.com) Lambda said in August 2025 that its on-demand Blackwell instances started at $4.99 per graphics processor hour, a posted price that shows how quickly frontier-grade capacity moved into premium cloud tiers. (lambda.ai) Nvidia says Blackwell’s second-generation Transformer Engine adds 4-bit floating point support, a lower-precision format meant to speed inference while keeping accuracy high enough for production use. The company also says fifth-generation NVLink lets 72 Blackwell graphics processors in one rack act like a single giant system. (nvidia.com; blogs.nvidia.com) CoreWeave says each HGX B200 instance packs eight B200 graphics processors with 180 gigabytes of HBM3e memory each, plus InfiniBand networking for cluster-scale jobs. Those specs help explain why buyers pay a premium when they need the newest chips for training or low-latency inference. (coreweave.com) The price spike points to the same bottleneck the artificial intelligence industry has wrestled with since 2023: demand for top-end chips is arriving faster than new capacity. As more companies ship agent-style products that think longer and call more tools, hourly Blackwell rents are becoming a live measure of that squeeze. (techmeme.com; blogs.nvidia.com)