Trump's order lifting export curbs on Anthropic models, reported by Wired, removes a restriction that had barred Claude from shipping into a defined set of jurisdictions since the prior administration's control list. What it does not change is the more binding constraint on Anthropic's roadmap: training compute. The same week, TechCrunch reported Anthropic is in chip talks with Samsung, and that pairing is the tell. A model export rule governs where you can sell what you have already built. A chip supply conversation governs whether you can build the next one at all. Anthropic runs its frontier training on Trainium2 clusters via AWS and on Google TPU v5p capacity, both allocated under long-term contracts that cap how fast compute scales quarter to quarter. Samsung's foundry business fabs logic and HBM (the high-bandwidth memory stacked next to a chip so it can feed data fast enough to keep the compute busy) at Pyeongtaek and Hwaseong. If those talks are about securing HBM supply or a custom accelerator line outside the Nvidia queue, that is Anthropic hedging against the same bottleneck every US lab faces: TSMC's CoWoS advanced packaging capacity, the process that stacks memory onto a chip package and currently gates how many Blackwell-class GPUs the whole industry can get built per quarter.

So the export order widens Anthropic's addressable market this week. The Samsung conversation, if it lands as a supply agreement, is the one that decides whether Anthropic has enough silicon in 2027 to keep training runs on pace with OpenAI's GPT-5.6 Sol cadence and Google's internal TPU roadmap. Watch for a disclosed capacity commitment or MOU date out of Samsung's foundry division, not the export order's jurisdiction list, as the signal that actually moves Anthropic's compute position.