The artificial intelligence gold rush is rapidly depleting an indispensable resource for AI developers: computing power. This severe computing power shortage has caused anxiety among heavy users, forced companies to abandon some products, and caused service stability issues. These problems have sounded the alarm for the AI craze—just as a large number of users are beginning to rely on advanced AI tools to improve productivity, computing power bottlenecks may limit the actual effectiveness of these powerful new tools.
In the past few months, the demand for "intelligent AI" has exploded. These autonomous AI tools can independently complete a variety of tasks, from writing software code to scheduling house tours for real estate agents. Enterprises are scrambling to seize computing resources to serve an expanding user base with significantly increased usage intensity.
“Everyone is talking about oil, but I think the real global shortage is tokens,” said Ben Pouladian, a Los Angeles engineer and technology investor. Token is a unit of measurement used in AI to measure the computing power consumed by tasks. "Today's AI is no longer just a chatbot that stands in front of the refrigerator and asks for recipes. It is coordinating tasks and becoming more and more intelligent."
All this points to a recurring classic problem in the history of technological prosperity: from the expansion of railways in the 19th century to the explosion of telecommunications and the Internet in the early 21st century, demand growth has always far exceeded the speed at which companies can obtain resources and build infrastructure. Historically, price increases have often been the only means to address supply shortages, but such a move carries huge risks for cutting-edge AI companies, which are competing fiercely for users.
Hourly rental prices for graphics processing units (GPUs) used to train and run AI models have surged since last fall. Anthropic, the company that develops the popular chatbot Claude and its popular programming tool Claude Code, has experienced frequent service outages recently. The company has begun to impose token usage limits on users during peak hours, but this measure has caused a large number of users to complain, saying that they quickly reached the usage limit.
According to reports, OpenAI has discontinued its video generation application Sora, partly to free up computing power to support new AI model-driven programming and enterprise products code-named Spud.
OpenAI’s API platform token usage for enterprise users soared from 6 billion per minute in October last year to 15 billion in late March.
"I did spend a lot of time looking for all the computing power that could be temporarily called upon." Sarah Fryer, OpenAI's chief financial officer, said in a recent public video interview with investors. "Due to insufficient computing power, we are currently being forced to abandon some projects and make very difficult choices."
At the end of last year, CoreWeave, one of the largest listed AI cloud computing companies, raised prices by more than 20% and required smaller customers to sign service contracts for at least three years, compared with the previous one-year period. Analysts at Bank of America restored their rating on the company late last month and gave it a "buy" rating, saying demand for its services will continue to exceed supply until at least 2029.
Data from New York data provider Ornn shows that spot rental prices for NVIDIA's full range of GPUs in cloud data centers have increased significantly in recent months. The Ornn Compute price index shows that the single-hour rental rate for Nvidia's most advanced Blackwell series chips has reached US$4.08, an increase of 48% from US$2.75 two months ago.
Since mid-February, Anthropic's entire system has experienced frequent downtime, and some enterprise customers have switched to using models from other AI vendors.
David Hsu, founder and CEO of software development platform Retool, said that he originally preferred to use Anthropic's Opus 4.6 model to drive the company's AI agent tool because he believed this was the model most suitable for enterprise scenarios. But he has recently switched to OpenAI’s model. "Anthropic's service has been experiencing frequent outages."
The reliability of Internet core services is usually measured by "several nines". The four nines are an online rate of 99.99%, which is also a common promise made by software companies to customers. As of April 8, the online rate of Anthropic's Claude API in the past 90 days was only 98.95%.
“This is abnormal,” said Amir Haghighat, co-founder and chief technology officer of Baseten, an AI inference startup. “Think about AWS, databases, RDS, or Stripe. They all require extremely high stability and online rates. But that is not the case in the AI field right now. This is not the quality of service you want to get from a company that provides intelligent capabilities for applications.”
Anthropic is experiencing explosive growth amid frequent service outages. At the end of 2025, the company's annual revenue reached US$9 billion; by February this year, this number soared to US$14 billion; two months later, it doubled again to US$30 billion.
In late March, Anthropic suddenly announced that it would limit users’ token consumption during the peak hours of 5 a.m. to 11 a.m. Pacific time on weekdays. Users complained on social platforms. One user wrote on
"We are doing our best to meet Claude's growing needs." Boris Cherny, head of Claude Code and developer, said on