Demand for agentic AI that autonomously handles specific tasks is surging, deepening a situation in which the supply of computing power needed to run AI is failing to keep up with demand.

This appears to be leading not only to rising AI costs but also to outcomes in which AI itself cannot be used properly.

According to a recent Wall Street Journal report, hourly rental fees for GPUs, a core AI infrastructure, have jumped sharply since last fall. Anthropic has faced customer backlash after limiting computing resource usage during peak hours.

OpenAI, which is known to have poured far more money into AI infrastructure, is also not immune to the issue of depleted computing resources. OpenAI recently halted its video-generation AI app, Sora. Analysts say the decision was also influenced by a judgment that more computing resources needed to be allocated to coding and enterprise products.

Token use via OpenAI's API rose to 15 billion tokens per minute by the end of March from 6 billion tokens per minute in October last year, the WSJ reported. Tokens are the basic unit of text processed by AI models and are also seen as an indicator of how much computing power AI uses.

GPU fees are also rising sharply. CoreWeave, a cloud computing service provider specialised in AI, raised prices by more than 20 percent late last year. It has also demanded that smaller customers extend contract terms from 1 year to 3 years. According to data provider Ornn, spot prices for Nvidia GPU products across cloud services have risen sharply in recent months. Hourly rental fees for Nvidia's latest Blackwell chips are up 48 percent from 2 months ago.

A shortage of computing resources also appears to be directly affecting service stability at AI companies. Anthropic, which is experiencing explosive growth, has suffered frequent outages since mid-February. That prompted some customers to switch to other AI models.

The computing resource shortage that has hit the AI sector does not appear likely to be resolved quickly.

According to the WSJ, JJ Cardwell, CEO of cloud infrastructure company Vultr, said, "We are experiencing severe capacity shortages at a level I have never seen in more than 5 years of operating the business." He added, "People may ask why we don't deploy more equipment, but the problem is the lead time is too long. Data centre construction also takes a long time, and power that can be used through 2026 has already been preempted."

Ben Poladian, an engineer and tech investor, said, "Everyone talks about oil, but what is most scarce in the world right now is tokens." He added, "At this point, AI is no longer a simple chatbot that asks for a recipe in front of the refrigerator. AI is directing tasks and becoming smarter."