1. OpenAI gets the banhammer too

The GPT-5.6 family (Sol, Terra, Luna) shipped on June 26, and only to about 20 government-cleared partners, through the API and Codex, at the request of the US government. The same week, Commerce partly reversed the Anthropic ban: Mythos 5 is back for 100-plus vetted US critical-infrastructure organizations (and, under the deemed-export rule, their foreign-national employees), while Fable 5 stays suspended worldwide for everyone else. Last issue this was a precedent to watch. It is now the operating rule for both leading labs.

  • The control point #107 flagged has hardened: who gets the best US models is now decided in Washington, not on a product roadmap, and it applies across the board, not to one lab.

2. The AI IPO window wobbles

OpenAI is reportedly weighing a delay to 2027 (per the New York Times), after SpaceX debuted, spiked past $225, then slid back near its $135 listing price within two weeks. Anthropic, which filed on June 1, is widely expected to feel the same pressure. The "imminent triple IPO" everyone priced in a month ago is now a question mark.

  • When the bellwether listing fades and a flagship model can be switched off by directive, the revenue story gets a risk discount, which is exactly what a public-market window reflects.

3. The AI tax hits your Xbox

Apple and Microsoft raised prices within five hours of each other on June 26, Macs, iPads, and Xbox consoles up $100-$150, both blaming AI-driven memory and storage shortages. The same week, Micron posted record ~85% margins off that exact shortage, with supply booked solid into 2027. AI-flation stopped being a forecast and showed up on a price tag.

  • Consumer electronics were one of the few categories where prices reliably fell. The data-center buildout is reversing that, and the pass-through has years to run.

4. China's IP war goes on record

Anthropic told the US Senate that Alibaba ran its largest known distillation attack, 28.8 million exchanges through roughly 25,000 fraudulent accounts between April and June (Alibaba denies it). It is the sourced version of a debate the community has had for weeks: if a competitor can distill your model and open-weight the result, you lose control of the asset. That is a large part of why the access gates above are closing: the IP these labs spend billions to build is leaking out faster than it can be protected.

  • Distillation turns frontier R&D into a one-way subsidy to fast-followers. For anyone licensing a closed model, the security posture of your vendor is now your security posture.

5. The tokenmaxxing party is over

Uber, Meta, and AT&T are capping internal AI spend and routing simpler jobs to cheaper or open models, after a year of treating raw token consumption as a proxy for productivity. The shift from "use as much as possible" to model routing and budget discipline is the demand-side reason cheap open weights like GLM-5.2 suddenly matter. State-of-the-art is rarely needed for most tasks, and finance teams have noticed.

  • The question is no longer "are we using AI," it is "are we using the right model for the job": routing, caching, and visibility, not a single frontier default.

More AI stories of the week