The 90-Minute Ultimatum Trump Gave Anthropic — Fareed Zakaria Makes the Case for a 'Fed for AI'

Last Friday the US government did something extraordinary. It effectively forced Anthropic — the company behind the frontier model Mythos and its commercial cousin Fable — to pull its newest product from the market with roughly 90 minutes of warning. Easy to miss in the shadow of the Iran retreat, but as Fareed Zakaria argues, this may matter more in the long run.

The real fight isn't about Anthropic. It's the first visible battle over who governs AI: by rules and institutions, or by improvisation and raw power. AI is now plainly a general-purpose technology that flows into everything, and increasingly it's as powerful as people imagined. Senator Mark Warner relayed testimony from Gen. Joshua Rad — head of both NSA and Cyber Command — that Mythos could penetrate nearly all of the government's classified systems in hours, not days or weeks. The question now isn't whether to regulate AI, it's how.

The Trump administration's approach has been deeply confusing. Two and a half weeks earlier, an executive order set up a voluntary review process for frontier models and gave agencies 60 days to design a framework for identifying systems that needed heightened scrutiny — structured oversight, advanced testing, clear thresholds. Sensible direction.

Before that process could even be created, the Pentagon declared Anthropic a supply chain risk, and federal agencies were told to stop using its systems. There are real Anthropic missteps in the record — a Washington Post report noted Mythos access was expanded beyond what officials understood to be approved, and the company was slow responding to who-was-using-it concerns. That's exactly why process matters. Decisions of this consequence can't be made ad hoc by whichever faction wins a weekend bureaucratic skirmish. The 90-minute notice sends the world a clear signal: build on US AI infrastructure and Washington can flip the off-switch arbitrarily, without warning or explanation.

Zakaria's proposal is sharp. Build a 'Fed for AI' — an independent institution staffed with experts in technology, national security, and business. The Federal Reserve has worked because it combines public authority with private expertise, talks to markets constantly while staying independent, runs exams and stress tests, sets capital requirements, publishes guidance, and intervenes in graduated steps. That's roughly what advanced AI needs: pre-release access for evaluation, transparent thresholds for dangerous capabilities, a ladder of responses (warning, remediation, conditional deployment, restriction), and rules that apply to everyone equally. America builds first; Europe and Japan create parallel institutions; democracies coordinate standards the way central banks already do.

The same day, Anthropic published its own striking statement. AI is moving faster than expected; humans are needed less; the models are getting close to writing their own successors — recursive self-improvement, possibly arriving in years rather than decades. Anthropic co-founder Jack Clark used a brake-pedal metaphor on air: right now we have a gas pedal and no brake. The proposal isn't to slam the brakes, it's to make sure the option exists, coordinated across multiple national labs under shared conditions. Cold War nuclear arms-control precedents show two rival powers can stabilise a race when both want to. AI may need the same.