1. DC switches off Anthropic's best model

The most powerful model ever shipped, and a week later the government switched it off. On Tuesday, Anthropic released Fable 5, the first publicly available model above Opus, a safeguarded version of Mythos 5 (its most capable model, which stays limited to vetted "Project Glasswing" partners). By Wednesday Fable was the most controversial launch in AI history: it silently degraded answers for AI researchers, blocked basic biology questions, and forced 30-day data retention even on zero-retention enterprises, enough for Microsoft to bar internal use. Anthropic reversed the silent- degradation policy in a day ("we made the wrong trade-off"). Then on Friday at 5:21pm ET, the Commerce Department ordered both models cut off for any foreign national on national-security grounds, and to comply Anthropic disabled them for every customer. The trigger, per the Wall Street Journal, was Amazon (Anthropic's largest investor and chip supplier), whose CEO Andy Jassy told officials his researchers had coaxed cyberattack-useful information out of Fable 5. By the administration's account (via AI adviser David Sacks), officials then asked Anthropic to fix the flaw or pull the model, and it refused. Anthropic, which had notified the government of the launch in advance without objection, counters that the flaws are basic, found in other public models including GPT-5.5, and not a true jailbreak. As of the weekend the models remain offline while the two sides negotiate.

  • The optics are remarkable: the flag came from Anthropic's own biggest backer, which also uses Claude to hunt software flaws. For everyone else the lesson is plainer: if your most powerful tool can vanish overnight on one company's tip, that is a dependency, not a strategy.
  • The reach stunned allies as much as adversaries: the order bars every foreign national, anywhere. A Brit, a Singaporean, an ally-country team, even Anthropic's own foreign-born researchers, now locked out of the models they helped build. For most of the world, frontier access became a question of passport overnight.

2. Anthropic asked for the brake

Two days before that shutdown, Anthropic argued for exactly that power. On June 10, Dario Amodei published a sweeping policy essay, "Policy on the AI Exponential," calling for FAA-style mandatory third-party testing of frontier models, with the government able to block or reverse any release that fails on cybersecurity, bioweapons, loss-of-control or automated-R&D risk, paired with $350M for the labor fallout. Then on Friday the government blocked Anthropic's own model, and the company objected that this was not the transparent, fair process it had called for.

  • With Washington now willing to pull a deployed model, "who gets to decide what ships" just went from hypothetical to live, and the essay itself is read as either principled risk management or a moat that locks in the few labs able to clear the bar.
  • It is also the latest round of a deeper standoff. Anthropic has refused to let its models be used for autonomous weapons or mass surveillance, the Pentagon has designated it a security risk (which it is challenging in court), and, the same week, OpenAI finalized a Defense Department deal. Values, not just safety, are shaping who gets the government's business.

3. SpaceX lands, OpenAI files

SpaceX debuted Friday at $135, opened at $150, closed up roughly 19%, and passed Tesla to become one of the largest US companies. Days earlier OpenAI filed confidentially, following Anthropic on June 1. All three frontier names are now in listing territory. But the pop deserves an asterisk: most holders were locked up, so a float of only about 4% of shares met demand that was twice oversubscribed, with banks managing the open.

  • Read it as choreography, not pure price discovery. The pop rode a sliver of stock and 2x-oversubscribed demand, not fundamentals. The mechanical pressure is still ahead: fast-tracked index inclusion is set to pull passive funds into roughly 30% of that float within about 15 trading days, and the real test comes when lockups expire.

4. The Chinese AI wall goes up

Both capitals spent the week walling off frontier AI. The US foreign-access ban (story 1) drew a line around the top American model. On the China side, Meta completed its split from Manus, firewalling all data sharing after Beijing ordered the $2B deal unwound on national-security grounds, and Beijing is now pushing other Chinese AI startups to abandon their offshore structures and reincorporate at home. Underneath the walls, the open-weight camp gained, with Google shipping DiffusionGemma and Gemma 4 alongside Mistral.

  • Where your model lives, and whose government can reach it, is now strategic. Open-weight is the emerging hedge against a switch-off, but Chinese open weights carry their own provenance and jurisdiction questions that belong in the diligence, not the marketing.

5. OpenAI's $500B compute moonshot

OpenAI is in talks to lease a 10-gigawatt data center on a former uranium-enrichment site in Pike County, Ohio, at up to $500B over 20 years, with the first phase in 2028. The structure is the story: the federal government would own the 9.2GW power plant (built with Japanese trade-deal money and operated by SoftBank's SB Energy), while OpenAI leases the campus and Nvidia backstops both OpenAI's rent and SB Energy's financing, a chip supplier underwriting its own customer, on government-owned power. The same week, Apollo and Blackstone arranged $35B for Anthropic's compute.

  • Compute is now financed like national infrastructure, not a startup. Watch the circularity as chipmaker, lab, and lender become one money loop, even as the social license cracks (moratoriums spreading: Seattle 9-0, New York statewide, 70-plus localities).

More AI stories of the week