Trump's Rollback Reopens the Mythos Compute Map
The Rule That Moved
Start with what actually changed this week. The Mythos export controls were a Biden-era rule. They restricted which AI chip designs could ship to a list of countries outside China, specifically Nvidia's high-bandwidth interconnect chips, the ones built for training runs above a certain compute threshold. Trump's rollback this week lifts that ceiling. In plain terms: this was never really a China rule. It controlled where US allies and neutral buyers, think Gulf states and Southeast Asian data center operators, could route large orders. Lifting it means Nvidia can now sell higher-end training clusters into markets that were previously capped. That timing matters, because those same markets are where the next wave of sovereign AI buildouts is trying to lock in supply. Korea's newly announced $1 trillion chip and robotics program is one of them. With the cap gone, whoever places their order first gets first claim on chips everyone else is now also allowed to buy. Nvidia can book orders from Gulf and Southeast Asian buyers the same week Seoul commits $550 billion to its own chip fund. The rule that moved this week decides who gets to place that order first.
What Alibaba Actually Copied
Anthropic's complaint against Alibaba, reported this week, says a version of Alibaba's Qwen model reproduced outputs closely enough to Claude's to suggest it was trained on Claude's outputs rather than on independent data. Put simply: the complaint alleges Qwen learned by copying Claude's homework instead of doing its own. Set the legal merits aside for a moment and look at what the complaint reveals about the shape of the competition. If Alibaba's fastest path to a competitive model is distilling a rival's outputs, that tells you the weights, the trained parameters that make up the model itself, are not the moat. They are not what actually protects a lab's advantage. Anthropic is signaling the same read with Sonnet 5's launch this week: the company's own messaging leans on workflow integration, the tools and enterprise pipelines built around the model, over headline benchmark gains. Here is why that distinction matters. A cloned model can match Claude's scores on a benchmark, a standardized test used to compare AI models. What it cannot do is walk into the bank that already wired Claude into its compliance review and signed a multi-year contract for that integration.
Korea's $1 trillion is a chip and robotics bet, not a model bet. Seoul did not announce a sovereign foundation model this week. Neither did Riyadh or Jakarta. The fabs and the racks are getting funded; the model layer is still a three-lab-per-bloc game. Next quarter's filings will show whether any of those buyers put money behind a model instead of a rack.