A Wall Street Journal investigation says the AI compute crunch is now showing up in pricing, uptime, and customer churn. Spot rental rates for Blackwell GPUs rose from $2.75 per hour to $4.08 in two months, a 48% increase. At the same time, Anthropic’s Claude API posted just 98.95% normal execution uptime over the 90 days through April 8, well below the 99.99% standard commonly expected in enterprise settings, pushing some customers toward OpenAI.
Compute prices rise as power becomes the limiting factor
The report says Ornn’s compute price index, OCPI, was recently added to the Bloomberg Terminal, giving institutional investors a way to monitor spot GPU rental prices in real time. Demand from agentic AI is a major driver. These systems do more than answer prompts on a web page; they run longer and consume far more sustained compute capacity.
Vultr CEO J.J. Kardwell said this is the worst compute shortage he has seen in more than five years running the company. He added that data center build-out takes so long that available power for 2026 has already been fully reserved. That shifts attention beyond chip supply. Power access and data center infrastructure are now central constraints.
Claude API instability pushes Retool to switch
Anthropic appears to be one of the companies under the most pressure. According to the report, Claude API’s normal execution rate was 98.95% over the 90 days ending April 8. The gap from 99.99% may look small on paper, but in enterprise production use it translates into nearly eight extra hours of downtime per month.
Retool founder and CEO David Hsu said Opus 4.6 was, in his view, the best enterprise model, but his company still moved to OpenAI because Anthropic kept going down. For companies embedding models into core workflows, reliability is not secondary. The shortage is starting to affect contract decisions directly.
Anthropic began rate limiting in late March, restricting token usage from 5 a.m. to 11 a.m. Pacific Time on weekdays. Earlier in March, it had also promoted a plan that doubled usage during off-peak periods, an apparent attempt to shift traffic away from peak hours and free up capacity.
OpenAI reallocates capacity while CoreWeave tightens terms
The strain is not limited to Anthropic. The report says OpenAI’s API token processing volume climbed from 6 billion per minute in October 2025 to 15 billion per minute by late March this year. CFO Sarah Friar said the company has been hunting for every last bit of available compute and making painful trade-offs, with some projects dropped because capacity was not there. The report says OpenAI’s decision to take down the video generation app Sora was partly tied to redirecting chip resources toward coding tools and enterprise products.
On the supply side, CoreWeave is also reshaping who gets access. The report says the company raised rental prices by more than 20% late last year and began requiring small and mid-sized customers to sign contracts of at least three years, up from one year before. That is a tougher bar for startups and mid-market firms that need flexibility. At the same time, CoreWeave announced on April 10 a multi-year agreement with Anthropic, which committed up to 1 GW of compute capacity using Nvidia Grace Blackwell and next-generation Vera Rubin hardware.