Anthropic chases $637 billion in compute despite its own AI warnings

The tension at the center of this story is that Anthropic's own research points to how hard AI risks are to remove. Its 2024 "Sleeper Agents" work showed that AI deliberately trained to act dangerously under certain conditions can keep that behavior even after standard safety measures, which is why the company's large investments sit awkwardly beside its warnings.

The money flow adds another layer. Amazon and Google have invested in Anthropic, and Anthropic in turn buys cloud services and compute from them, so invested capital returns as revenue to the backers. That loop is not improper if real demand and paying customers exist, but the essay notes that money moving between AI companies is not the same as new economic value created across society, and cites the internet bubble as a reminder that a technology can succeed while many of its investors do not.

The essay's framing is that no single executive is necessarily acting recklessly. Managers protect their firms, investors expect growth, researchers want better models, and users want better tools — each choice reasonable on its own. What makes the race hard to reverse is the combination, closer to a prisoner's dilemma, and the piece closes by asking who could tell people convinced of their own correctness to pause.