AI's Computing Power Demand Exhausts Global Supply, Constraining Growth

Electricity, water...and now computing power? The AI boom was projected to grow at infinity times 100 gazillion (a slight exaggeration...). That is, until the reality of how much existing data, electricity, water and now, computing power could be resourced in addition to new capacity to support the Silicon Valley hype machine growth fantasy. 

Yes, AI will be impactful and it will change many things about the way we work, make money and live, just as new technology has always done. But also just like transformational technologies of the past, there will be periods of delayed or slowed growth because the existing infrastructure cannot yet support it. Concerns about electricity have been extant for some time, based in part on simple calculations about the world's capacity to apply the components, building materials and people to construct the data centers required. Latterly, there has been a growing awareness of the fact that many of the data centers are being built in places with favorable tax and regulatory regimes - but which were already facing existential water shortages. Now, computing capacity availability is being cited as a constraint. The overall implication is that, in the real world, some growth takes time and that, in turn, affects the ability of an industry like AI to meet enthusiastic but, perhaps, overly optimistic, expectations. JL

Angel Au-Yeung and Robbie Whelan report in the Wall Street Journal:

The AI gold rush is rapidly drying up the supply of computing power. capacity crunch has forced companies to scuttle products and led to reliability problems. The issues are a warning sign for AI, as they limit the utility of new AI tools just as users have begun to rely on them. Companies have been scrambling to secure computing to serve a growing base of customers who are significantly increasing AI use. This is a classic problem in tech booms throughout history, from 19th-century railroad expansion to the internet of the early 2000s. Demand is growing far faster than companies are able to access resources and build out infrastructure. Token use in OpenAI’s API rose from six billion a minute in October to 15 billion a minute in late March. "Why don’t we just deploy more gear? Data center build times are long, the power that’s available through 2026 is already all spoken for.”

The artificial intelligence gold rush is rapidly drying up the supply of the one resource that AI developers can’t do without: computing power. 

The sharp capacity crunch has caused consternation among power users, forced companies to scuttle products and led to reliability problems. The issues are a warning sign for the AI boom, as they may limit the utility of powerful new AI tools just as massive amounts of users have begun to rely on them to boost productivity. 

Over the past few months, demand has exploded for “agentic” AI, autonomous tools that use the technology to independently perform tasks, from writing software code to scheduling house tours for real-estate brokers. Companies have been scrambling to secure the availability of computing capacity needed to serve a growing base of customers who are also significantly increasing their AI use.

“Everyone’s talking about oil, but I think what the world is mainly short of is tokens,” said Ben Pouladian, an engineer and tech investor based in Los Angeles. A token is a unit of measurement in AI to track how much computing resources are being used for a task. “AI is at this point no longer just some chatbot that we ask for a recipe while we stand in front of the fridge. It’s orchestrating tasks, it’s getting smarter,” Pouladian said.

All of it points to a classic problem that has popped up in technology booms throughout history, from the 19th-century railroad expansion to the telecom and internet explosion of the early 2000s. Demand is growing far faster than companies are able to access resources and build out infrastructure. Historically, price increases have been among the only ways to address a supply crunch, but such a move could be perilous for frontier AI companies, who are in a ferocious competition to gain users.

Hourly rental prices for GPUs, the microchips used to train and run AI models, have surged since the fall. Anthropic, the maker of popular chatbot Claude and viral coding app Claude Code, has been plagued recently by frequent outages. The company has begun metering computing supply to users during peak hours, but the rollout has been marred by customers who have complained that they are reaching the limit far too quickly.

OpenAI scrapped its Sora video-generation app in part to free up computing resources to power coding and enterprise products that would work on a new AI model, code-named Spud, The Wall Street Journal reported. 

Token use in OpenAI’s API—a platform where mostly enterprise users access its software—rose from six billion a minute in October to 15 billion a minute in late March.

“I do spend a lot of time trying to find any last-minute compute available,” Sarah Friar, OpenAI’s chief financial officer, said in a recent public video interview with an investor. “We’re making some very tough trades at the moment on things we’re not pursuing because we don’t have enough compute.”

Toward the end of last year, CoreWeave, one of the largest publicly traded AI cloud companies, raised prices by more than 20% and started asking smaller customers to sign contracts committing them to use the company’s services for at least three years, up from one year before. Bank of America analysts reinstated coverage of the company with a “Buy” rating late last month, saying demand for its services is likely to outstrip supply through at least 2029.  

Spot-market prices to access Nvidia’s GPUs, or graphics processing units, in data-center clouds have risen sharply in recent months across the company’s entire product line, according to Ornn, a New York-based data provider that publishes market data and structures financial products around GPU pricing. Renting one of Nvidia’s most-advanced Blackwell generation of chips for one hour costs $4.08, up 48% from the $2.75 it cost two months ago, according to the Ornn Compute Price Index. 

“There’s a massive capacity crunch that’s unlike anything I’ve seen in the more than five years I’ve been running this business,” said J.J. Kardwell, chief executive of Vultr, a cloud infrastructure company. “The question is, why don’t we just deploy more gear? The lead times are too long. Data center build times are long, the power that’s available through 2026 is already all spoken for.”

Since mid-February, outages for systems across Anthropic have become so common that some of its enterprise clients are switching to other AI model players. 

David Hsu, founder and CEO of software development platform Retool, said he prefers to use Anthropic’s Opus 4.6 model to power his company’s AI agent tool because he believes it is the best model for enterprise. He recently changed to OpenAI’s model to power his company’s agent. “Anthropic has just been going down all the time,” he said.

The reliability of core services on the internet is often measured in nines. Four nines means 99.99% of uptime—a typical percentage that a software company commits to customers. As of April 8, Anthropic’s Claude API had a 98.95% uptime rate in the last 90 days. 

“That is not normal,” said Amir Haghighat, co-founder and chief technology officer at Baseten, an AI inference startup. “Think about AWS, databases, RDS or Stripe—these need to be very resilient with a very high uptime. But that is not the world we live in when it comes to AI. That’s not the quality of service that you want to be getting from the company that’s providing intelligence for your application.”

Claude API outages, in hours

Monthly

uptime

Opus 4.5 released

0

1

2

3+

None

Oct.

99.82%

Nov.

99.73%

99.81%

Dec.

Jan.

99.53%

Feb.

99.24%

March

98.32%

April

99.35%

1st

10th

20th

30th

Note: Data as of noon
Source: the company
James Benedict/WSJ

The frequent outages at Anthropic are happening as the AI lab is experiencing explosive growth. At the end of 2025, the company hit $9 billion in annual run rate, which means the company was on track to make that amount of revenue in the next 12 months. By February, that figure ballooned to $14 billion. Two months later, it doubled to $30 billion.  

In late March, Anthropic suddenly announced it would limit the amount of tokens that users could burn through during peak hours from 5 a.m. to 11 a.m. Pacific Time on weekdays. Customers have taken to social media to complain about the change. “I haven’t hit my Claude Code terminal limit in weeks but this week I hit it in like 45 minutes,” wrote one user on X.

“We’ve been working hard to meet the increase in demand for Claude,” wrote Boris Cherny, creator and head of Claude Code, on X. “Capacity is a resource we manage thoughtfully and we are prioritizing our customers using our products and API.”