Anthropic and OpenAI Confront GPU Shortages and Interruptions

An Insatiable Demand That Exceeds Supply

The artificial intelligence industry is facing a major capacity crisis, fueled by a growing demand for AI agents, these autonomous tools capable of performing complex tasks without human intervention. According to the Wall Street Journal, this demand has led to a series of disruptions, including frequent outages at major providers, canceled or scaled-back products, and a significant increase in prices for essential chips like GPUs.

Anthropic: Development Hampered by Interruptions

Anthropic, the company behind the chatbot Claude and the coding application Claude Code, is particularly affected by this crisis. Since mid-February, outages have increased, prompting some clients to turn to other providers such as OpenAI. David Hsu, the founder of the software platform Retool, told the Wall Street Journal that he preferred Anthropic's Opus 4.6 model but had to migrate to OpenAI due to service interruptions. The report indicates that the availability rate of the Claude API over a 90-day period ending April 8 was 98.95%, well below the 99.99% standard typically maintained by established cloud providers.

OpenAI: Strategic Resource Reallocation

OpenAI is not spared from this growing pressure. The company recently announced the shutdown of its video generation application Sora, in order to free up computing resources for its coding and enterprise products, built on a new AI model named Spud. The web and app versions of Sora will be taken offline on April 26, with the API following in September. According to the Wall Street Journal, token usage on OpenAI's API has significantly increased, rising from 6 billion to 15 billion per minute between October and March. Sarah Friar, OpenAI's Chief Financial Officer, explained that she spends a lot of time researching short-term computing capabilities, and that the company must make tough decisions regarding which projects to set aside due to a lack of resources.

New Restrictions to Manage Demand

In order to manage this growing demand, providers have implemented new limitations. GitHub announced new limits for its Copilot tool on April 10, explicitly citing rapid growth and heavy usage as reasons. OpenAI also modified its Codex billing for businesses, shifting from a message-based pricing model to a token-based billing system in early April. Windsurf replaced its credit system in March with daily and weekly quotas, offering additional capacity at API prices. Anthropic adjusted its session limits at the end of March and temporarily offered double usage during off-peak hours to distribute the load more evenly.

The overall trend is clear: regular chat and agent work are increasingly being priced separately, with heavy workloads managed by dedicated pools, credits, and token-based surcharges.

A Soaring Spike in GPU Prices

Prices for GPUs, essential for processing AI tasks, have seen a dramatic increase. According to the Ornn Compute Price Index, one hour of usage of a Nvidia Blackwell chip now costs $4.08, a 48% increase compared to just two months ago. The Wall Street Journal reports that Coreweave, one of the largest publicly traded AI cloud providers, has raised its prices by over 20% towards the end of 2025 and now requires small clients to sign three-year contracts instead of one. Analysts at Bank of America expect demand to exceed supply until at least 2029.

J.J. Kardwell, CEO of Vultr, told the Wall Street Journal that the current capacity crisis is unprecedented in his more than five years of managing his cloud infrastructure company. He cited long hardware delivery times, slow data center builds, and the fact that available power until 2026 is already booked as major bottlenecks. Raising prices is one way to respond to the shortage. However, for major AI companies engaged in a fierce battle for users, this would represent a risky move.