The Department of Defense is poised to move classified artificial-intelligence workloads off Anthropic by October, according to reporting this week. The break is not about performance. It is about the company’s usage policy, and about a lab that has been willing to lose a contract rather than loosen a rule.
For two years, Washington treated frontier models as a strategic resource and frontier labs as vendors. Anthropic’s stand — limits on certain military and intelligence uses — always sat awkwardly next to that story. The awkwardness is now a deadline.
Amodei’s Saturday essay makes the split stranger. The same chief executive who wants governments to force a slowdown is losing the government customer that most cares about control. One reading is integrity. Another is that national-security buyers will go where the terms are quietest, and that the quietest terms are not written in San Francisco.
OpenAI, Google, and a line of defense startups will be happy to take the work. The evaluators Amodei wants to embed at Anthropic will not follow the models to their next home unless someone makes them. That is the part of “pace the frontier” no essay can do alone.
