Why AI Tools Are Hitting Usage Limits: The Compute Crunch Explained

In early 2026, power users of Anthropic’s Claude began burning through five-hour usage limits in just 20 minutes. Simultaneously, OpenAI shuttered Sora as developer demand for Codex surged to 4 million users per week. These aren’t isolated incidents — they signal what experts are calling the “compute crunch,” a growing gap between AI demand and the physical infrastructure required to sustain it.

AI policy expert Lennart Heim, formerly of the RAND Center on AI, Security, and Technology and co-founder of Epoch AI, explains the fundamental issue in a Scientific American interview: “If 10 times more people use AI 10 times more heavily, you need close to 100 times more compute.”

Why Flat-Rate Subscriptions Break Down

Unlike traditional internet services where a power user barely costs more than a light user, AI usage is directly tied to compute cost. Using AI ten times more intensively costs the provider roughly ten times as much. This makes flat-rate $20/month plans economically unsustainable at scale — and rate limits are the natural result. Companies like Anthropic are already defaulting users to lighter models to manage costs.

The Supply Chain Can’t Keep Up

• The International Energy Agency projects global data center electricity use will double by 2030; Anthropic’s projections call for 50 gigawatts of U.S. AI capacity by 2028 — roughly the output of 50 nuclear reactors
• TSMC, which fabricates most of the world’s advanced AI chips, committed $56 billion in 2026 alone to expand capacity — and customers are still asking for more
• Bottlenecks cascade across gas turbines, memory chips, and the specialized manufacturing facilities needed to produce them — industries that had been flat for years before AI demand surged

If AI becomes the interface for coding, medicine, customer service, and office work, access to compute becomes access to economic speed. The limits showing up in consumer products today are an early signal of a much larger constraint.

🔗 Read the full article on Scientific American