The AI gold rush is finding that there is not enough compute to go round, and the shortage is rattling power users, killing products and turning “reliability” into a bit of a laugh.
According to the Financial Times, the capacity crunch is forcing firms to scuttle launches and cope with outages, right as more users lean on AI tools to get work done.
Demand has spiked for “agentic” AI in recent months, with autonomous tools performing tasks such as writing code and booking house tours for estate agents. Companies are scrambling to lock in the computing capacity needed to support a growing customer base that is burning through more AI.
Los Angeles-based engineer and tech investor Ben Pouladian said, “Everyone’s talking about oil, but I think what the world is mainly short of is tokens. AI is no longer just a chatbot we ask for a recipe from while standing in front of the fridge. It’s orchestrating tasks, it’s getting smarter.”
It has the familiar stink of past tech booms, from 19th-century railways to the telecom and internet frenzy of the early 2000s. Demand is racing ahead of infrastructure, and price rises are usually the next step.
Hourly rental prices for GPUs, the chips used to train and run models, have jumped since autumn.
Anthropic, maker of chatbot Claude and the viral coding app Claude Code, has been hit by frequent outages. It has started metering supply during peak hours, annoying customers by having limits that arrive far too quickly.
OpenAI scrapped its Sora video-generation app partly to free up compute for coding and enterprise products tied to a new model codenamed Spud.
Token use in OpenAI’s API, where mostly enterprise customers plug into the software, jumped from six billion a minute in October to 15 billion a minute in late March.
OpenAI’s chief financial officer, Sarah Friar, said, “I do spend a lot of time trying to find any last-minute compute available. We’re making some very tough trades at the moment on things we’re not pursuing because we don’t have enough compute.”
CoreWeave, one of the largest publicly traded AI cloud firms, raised prices by more than 20 per cent toward the end of last year. It has been pushing smaller customers into at least three-year contracts, up from one year previously.
Bank of America analysts reinstated coverage late last month with a “Buy” rating, arguing demand will outstrip supply through at least 2029. Spot prices for access to Nvidia GPUs in data-centre clouds have risen sharply across the product line, according to Ornn, a New York data provider that tracks the market and structures financial products around GPU pricing.
Ornn says renting an advanced Nvidia Blackwell chip for an hour now costs $4.08, up 48 per cent from $2.75 two months earlier, based on the Ornn Compute Price Index. Those are not gentle moves for something meant to be “on demand”.
Claude Code, creator and head, Boris Cherny wrote on X, “We’ve been working hard to meet the increase in demand for Claude. Capacity is a resource we manage thoughtfully, and we are prioritising our customers using our products and API.”
