The Week a Government Switched Off a Frontier Model

For three years the AI story was a leaderboard. This week it was a negotiating table, a regulator's desk, and an export-control order.

Anthropic's most powerful models spent nine straight days dark on a government's say-so, OpenAI and Anthropic both lined up for the public markets, the architect of the transformer changed teams, and the labs kept buying up the layers of the stack underneath everyone. Here are the turning points worth watching.

Washington pulled the plug — and left it pulled

Anthropic's Claude Fable 5 and Mythos 5 have been offline worldwide since a June 12 US export-control directive — the first time a government has effectively pulled a frontier model from the market. Executives in Seoul promised access back "within days," but as the June 22 free-trial cutoff arrived the API was still erroring, with reporting pointing to a $100M investor's China exposure as the spark.

Why it matters: the precedent is now set. Any model you don't host yourself can be switched off by someone else's government, on someone else's timeline. Treat model access like any other dependency — know who can pull the plug, and keep a fallback warm. Fortune · Anthropic

Two labs, one road to Wall Street

Within a week of each other, both frontier leaders filed confidentially with the SEC — Anthropic around June 1 at a reported ~$965B valuation, OpenAI on June 8 at an $852B post-money mark. It's the first time the two are in the IPO pipeline simultaneously, which means burn rate, compute commitments, gross margins, and safety posture all become public line items.

The era of the private research lab is ending in real time — and public companies have to show margins, which puts a clock on the era of deeply subsidized tokens. TechCrunch

The architect changed teams — and the toolchain got bought

Noam Shazeer, co-author of 2017's "Attention Is All You Need" and co-lead of Gemini, announced June 18 he's joining OpenAI as Lead for AI Architecture Research — less than two years after Google paid ~$2.7B to bring him back from Character.AI. When the person who designed the foundation of every modern LLM picks a side, it's a signal about where the next architectural leap comes from. Axios

The same week, OpenAI moved on the layer below: it's acquiring Astral, makers of uv, Ruff, and ty — the Rust-based Python tooling a huge slice of the ecosystem now runs on — folding it into Codex. OpenAI says the tools stay open source after closing. Great for integration; worth watching closely for lock-in, because their entire adoption was built on open-source trust. OpenAI · The New Stack

Capacity, not cleverness, is the constraint

Google agreed to pay SpaceX roughly $920M a month — renting ~110,000 NVIDIA GPUs from October 2026 through June 2029 — after demand for Gemini Enterprise outran its own data centers. When the company with the deepest infrastructure on earth has to lease GPUs from a rocket company, the compute crunch is real and not priced into anyone's roadmap. TechCrunch

And the bottleneck just moved into the physical world: on June 18 FERC issued "show cause" orders to six US regional grid operators, demanding they justify or reform how AI data centers connect to the grid — bypassing the usual years-long rulemaking and calling it a "national priority." It lands alongside the White House's June 2 executive order, a deliberately light-touch, voluntary framework that stops short of mandatory licensing. Policy, not just chips, is now the gating factor. FERC · The White House

The assistant market split three ways

New share data put ChatGPT at 46.4% of the global AI assistant market as of late May — the first time it has held less than half — with Gemini at 27.7% and Claude at 10.3%. The leader is still the leader, but the moat is visibly draining as Google's distribution and Anthropic's enterprise traction compound. "Default to ChatGPT" is no longer a safe product assumption; your users are increasingly spread across three assistants. Build Fast with AI

The agent era's defining bug

Researchers showed that an attacker holding only a public Sentry DSN can post a crafted error event that an AI coding agent — Claude Code, Cursor, or Codex — reads as instructions and runs as shell commands. In controlled tests "Agentjacking" hit an 85% success rate and silently exfiltrated cloud keys and git credentials. Sentry added a content filter after the June 3 disclosure, but the root problem — agents treating untrusted tool output as commands — is the security gap that defines this whole phase. If you run agents against external data, assume that data can talk back. SC Media · Hackread

On our radar next week

• Gemini 3.5 Pro, June 30. Google's flagship (2M-token context, "Deep Think" mode) is still in limited Vertex preview. Sundar Pichai's I/O promise of a June GA window runs out June 30 — an on-time ship hands Google the spotlight; a slip is a conspicuous miss. Source
• Does Fable 5 come back? Anthropic keeps saying access returns "within days." How the first government-ordered model shutdown ends sets the template for the next one. Source
• EU AI Act, August 2. The Act becomes fully applicable in six weeks, switching on enforcement powers — and Apple and Meta are still sitting out the voluntary GPAI Code of Practice. Source

— Megan