A shortage of computing power is emerging as the most challenging bottleneck in the era of the AI boom. Over the past few months, with the explosive growth in demand for "agentic" AI, the supply of computational resources has fallen far behind the rate of consumption. Cracks are spreading at the most vulnerable point of this technological surge, evident from frequent service outages to soaring prices.

As Los Angeles engineer and tech investor Ben Pouladian stated, everyone is talking about oil, but the real scarcity facing the world might be tokens.

Recently, Anthropic's popular coding tool, Claude Code, has experienced frequent service interruptions. The company has begun limiting user token usage during peak weekday hours, but the implementation has sparked numerous user complaints. Simultaneously, OpenAI chose to shut down its video generation application, Sora, to free up computing power for its next-generation AI models. Cloud computing provider CoreWeave has increased its service prices by over 20% and is requiring smaller clients to sign long-term contracts of at least three years.

The impact of this computing power crisis on the AI industry is already clear. Just as millions of business users are deeply integrating AI tools into their workflows, the reliability and availability of these tools are declining. The resulting erosion of trust could become a heavy burden for cutting-edge AI companies in the fierce battle for users.

Token, the New Scarce Resource A token is the fundamental unit in AI for measuring computational consumption, directly corresponding to the computing cost behind each model inference. On OpenAI's API platform, the token call volume surged from 6 billion per minute last October to 15 billion per minute by the end of March this year, a 150% increase in less than half a year.

The core driver of this growth is the rapid adoption of "agentic" AI. These tools can autonomously perform tasks, from writing software code to scheduling property viewings for real estate agents, deeply replacing workflows previously done by humans. Ben Pouladian pointed out that AI is no longer just a chatbot we query for a recipe in front of the refrigerator; it is orchestrating tasks and getting smarter.

Outages and Throttling: Anthropic Bears the Brunt Since mid-February this year, service outages at Anthropic have become increasingly frequent, leading some enterprise clients to migrate their workloads to other AI providers. As of April 8th, the uptime for Anthropic's Claude API over the past 90 days was only 98.95%. This figure falls far short of the internet industry's common standard, which typically requires software service providers to promise 99.99% availability.

Amir Haghighat, co-founder and CTO of AI inference startup Baseten, stated directly that AWS, databases, and payment platforms all require extremely high availability. However, the reality in the AI field is different, and this falls short of the service quality one expects from a company providing intelligence for applications.

David Hsu, founder and CEO of software development platform Retool, mentioned that while he initially preferred using Anthropic's Opus 4.6 model to power his company's AI agent tool, considering it performed best in enterprise scenarios, he has recently switched to OpenAI's model. He said, "Anthropic has been down constantly."

Under pressure, Anthropic announced at the end of March that it would impose token usage limits during peak hours on weekdays. This move quickly sparked a wave of user complaints on social media. One user wrote on platform X that they hadn't hit their Claude Code limit for weeks, but this week they reached it in about 45 minutes. Ironically, this backdrop of frequent outages coincides with Anthropic's hyper-growth. The company's annualized revenue leaped from $9 billion at the end of 2025 to $14 billion in just two months, subsequently doubling again to $30 billion.

Soaring Prices: GPU Rental Costs Surge 48% in Two Months Tight supply is simultaneously driving up computing costs. According to the computing power price index released by New York data service provider Ornn, the hourly cost to rent Nvidia's latest-generation Blackwell chips is currently $4.08, a 48% increase from $2.75 two months ago.

J.J. Kardwell, CEO of cloud infrastructure company Vultr, described the current situation as a "large-scale compute shortage unlike anything seen in over five years." He explained that the problem isn't simply deploying more equipment, as lead times are too long, data center construction cycles are extensive, and available power capacity is fully booked until 2026.

Analysts at Bank of America recently resumed coverage of CoreWeave with a "buy" rating, anticipating that demand for the company's services will continue to outstrip supply until at least 2029. OpenAI's Chief Financial Officer, Sarah Friar, also admitted in a recent public investor video interview that she spends a significant amount of time trying to find any temporarily available computing power. She stated that the company is making very difficult choices on projects it would like to advance, simply due to a lack of compute.

Historical Cycle: Infrastructure Always Lags Demand This predicament is not unique to the AI industry but is a classic theme throughout technological history. From the railway expansion of the 19th century to the telecom and internet bubbles of the early 21st century, every wave of technological fervor has been accompanied by lagging infrastructure development, where demand growth consistently outpaces the ability to acquire resources and expand capacity.

Historical experience shows that price increases are often one of the few effective means to alleviate supply shortages. However, for cutting-edge AI companies currently locked in intense competition for users, this option is particularly tricky. Raising prices could directly harm user growth, which is a core pillar of their valuation logic. Finding the balance between maintaining service quality, controlling computing costs, and preserving competitiveness will be the central challenge the AI industry must confront in its next phase.