The export-control reversal is the specific event here, so the architecture opens on that.

The Bureau of Industry and Security's October 2023 rule set the wall: no advanced AI accelerator, and by extension no model trained past a certain compute threshold, could ship to a list of destinations without a license. On July 3 the Trump administration lifted that wall for a named set of Anthropic's models, Mythos and Fable, clearing them for release outside the US for the first time. Anthropic did not get this for free. Wired reported the company added a new security control, likely usage monitoring or geofencing tied to the export terms, specifically to satisfy the administration after an earlier safety-testing episode had stalled the deal. The Verge separately confirmed Fable 5, which had been sitting on the shelf since the export question was unresolved, is now cleared to launch. Read together, the sequence is a licensing negotiation, not a policy reversal: Washington traded market access for a monitoring hook it can audit.

The next fact to watch sits at Samsung, not at BIS. TechCrunch reported Anthropic is in talks with Samsung for a custom AI chip, the same move OpenAI made with Broadcom and Google made with its own TPU line: once a lab's enterprise API revenue is large enough, the next constraint is not export policy but fab allocation, and Samsung's Hwaseong and Pyeongtaek lines are the capacity Anthropic would be bidding against Nvidia's own order book for. Meanwhile OpenAI is running the inverse trade in Washington, offering the US government a 5% equity stake to defuse political pressure over its structure, the same currency Anthropic just spent on a security feature. Both labs are paying Washington in different coin for the same thing: room to ship. The chip talks, if they produce a term sheet, are the next dated event to track, since a Samsung fab commitment would be the first sign Anthropic is hedging its compute supply away from a single vendor before the next BIS licensing cycle.