Frontier AI makers face compute crisis as agentic tools 'max out' resources

The "all-you-can-eat" era of frontier AI appears to be hitting a hard ceiling as major providers struggle to reconcile fixed subscription prices with the massive compute costs of agentic AI.

This week, Microsoft’s GitHub Copilot took the drastic step of pausing new signups for its Student, Pro, and Pro+ plans, citing usage that regularly exceeds the revenue generated by the plan prices.

Joe Binder, GitHub's VP of Product, noted that long-running, parallelised sessions—often powered by tools like OpenClaw—now consume resources in ways that the original business models never anticipated.

Anthropic is facing similar growing pains. Following a surge in demand that saw its app reach number one in the Apple App Store last month, the company has begun testing usage restrictions for its popular Claude Code tool.

Anthropic's Head of Growth, Amol Avasare, explained that their "Max" plan was originally designed for heavy chat usage, not for autonomous agents that run for hours.

These hardware constraints have been exacerbated by regional data centre hurdles; as Gartner analyst Arun Chandrasekaran noted, providers cannot always rely on a global pool of compute, leaving users in specific regions like Belgium or the Netherlands more vulnerable to throttling.

While OpenAI CEO Sam Altman has publicly mocked Anthropic’s struggles, his company is not immune to the shifting economics.

Although OpenAI recently rolled out "ChatGPT Images 2.0" and reset rate limits for its Codex tool, it also permanently sunsetted its popular 4o model in February 2026 and discontinued the Sora video app last month.

As the industry matures, analysts expect a shift toward three survival strategies: increasing model efficiency, more aggressive request routing, or significantly higher pricing for priority access to high-compute "reasoning" models.