Compute is tight
Demand for large AI agents is pushing GPU rental prices and causing rationing, with reported hourly rent for Nvidia Blackwell GPUs jumping to $4.08 from $2.75 over two months. Spot-market strain is producing outages and access limits that hit startups and teams without captive capacity first. ( )
Renting the newest Nvidia artificial-intelligence chips now costs sharply more by the hour, and cloud providers are starting to ration access. (wsj.com) The Wall Street Journal reported on April 13 that spot rental for one Nvidia Blackwell graphics processing unit reached $4.08 an hour, up from $2.75 two months earlier, citing the Ornn Compute Price Index. Ornn says its index tracks live traded spot prices across hardware including Nvidia B200 Blackwell chips. (wsj.com, ornn.com) The squeeze is showing up in products. The Decoder, citing the Journal, said Anthropic’s Claude application programming interface posted 98.95 percent uptime over the 90 days ending April 8, below the 99.99 percent benchmark common in mature cloud services. (the-decoder.com) Anthropic’s own status page shows a stream of incidents in late February, March, and April 2026, including elevated errors on Claude Opus 4.6 and Sonnet 4.6 and a Claude.ai outage on April 13. (status.claude.com) A graphics processing unit is the workhorse chip that trains and runs large artificial-intelligence models, and Nvidia’s Blackwell line is built for the biggest jobs. CoreWeave says its GB200 systems are aimed at training 500 billion-parameter models and “massive-scale inference,” the step where a model generates answers for users. (docs.coreweave.com, hai.stanford.edu) The pressure comes from a new kind of workload. The Journal and The Decoder tied the jump to “agentic” artificial intelligence systems, which do not just answer one prompt but carry out multi-step tasks, consuming far more tokens and compute than a single chat reply. (wsj.com, the-decoder.com) Stanford’s 2026 Artificial Intelligence Index says generative artificial intelligence reached nearly 53 percent population-level adoption within three years, a pace it says exceeded the personal computer and the internet. More users and heavier agent workloads are landing on the same finite pool of chips, power, and data-center space. (hai.stanford.edu) The winners are the companies that already control capacity. CoreWeave’s documentation shows Blackwell GB200 instances are live in multiple United States availability zones, while teams without long-term contracts are left shopping the spot market where prices move fastest. (docs.coreweave.com, ornn.com) The immediate effect is uneven access, not a total shutdown. Big model makers can redirect workloads and sign multi-year infrastructure deals, while startups, coding tools, and video generators are more likely to hit limits first when compute gets scarce. (the-decoder.com, wsj.com) For now, the price of a useful artificial-intelligence answer is being set as much by physical supply as by software. The chips are faster, but the bottleneck has moved to who can actually get time on them. (ornn.com, wsj.com)
