Major AI companies including Microsoft, Anthropic, and OpenAI are facing a critical inflection point as the economics of their business models struggle to keep pace with surging demand for AI agents. As users increasingly rely on tools that run continuously and consume vast amounts of compute, leading firms are being forced to impose stricter usage limits and reconsider their subscription structures. On Monday, Microsoft announced that GitHub Copilot would pause new signups for Student, Pro, and Pro+ plans while tightening existing usage caps. The company cited that long-running, parallelized AI sessions now consume far more resources than originally anticipated, with some requests incurring costs that exceed plan prices. This shift marks a departure from the rapid expansion phase of 2022, where unlimited or low-cost access was common. Anthropic has faced similar pressures. On Tuesday, the company revealed it was testing whether to restrict access to its popular coding tool, Claude Code, for lowest-tier paid subscribers. Although Anthropic later characterized this as a test and promised user notification before any permanent changes, the move signals broader operational challenges. The startup experienced a massive influx of users following a political standoff with the Trump administration, which propelled Claude to the top of the Apple App Store. This surge overwhelmed capacity, leading to outages and the need for peak-hour restrictions. Anthropic executives noted that their subscription models were designed for heavy chat usage, not for the persistent, multi-hour sessions now enabled by autonomous AI agents. Industry analysts warn that this crisis is not isolated but indicative of a fundamental structural issue. Arun Chandrasekaran, a distinguished vice president analyst at Gartner, explained that the initial business models adopted in 2022 are unsustainable given current consumption patterns. Companies are now attempting to wean users off free tiers while demonstrating the tangible value of newer, more expensive models. Geographic constraints are also complicating the issue. Unlike traditional cloud resources, AI compute capacity is not a global pool. Users in regions like Europe may face bottlenecks if local data centers, such as those in Amsterdam, lack sufficient capacity to handle the load. This regional fragmentation exacerbates the scarcity of resources available to developers and consumers worldwide. Facing these finite compute limits, companies are likely to pursue one of three strategies: improving model efficiency, optimizing request routing, or prioritizing specific user segments. All options carry potential downsides for consumers, including potential price hikes, the deprecation of older models, or reduced availability. OpenAI recently reversed a plan to sunset an earlier GPT model after user backlash, yet it has already discontinued its Sora video-generation app, demonstrating a willingness to make tough decisions when economics dictate. Despite OpenAI CEO Sam Altman publicly celebrating Anthropic's challenges, OpenAI is not immune to these unit economics issues. While it benefits from deep integration with Microsoft Azure infrastructure, the fundamental costs of running advanced AI remain a constraint. As the industry matures, the era of abundant, cheap AI access may be coming to an end, forcing both providers and users to adapt to a more constrained and costly future.