Mythos Export Controls Lift, PRC Compute Gap Stays
The Chip Was Never It
Start with what actually happened. The Trump administration lifted the Mythos export controls this week. That was a licensing regime that had blocked a specific tier of Nvidia training silicon from being sold to buyers in mainland China since 2023. In plain terms: any Chinese company that wanted that chip simply could not legally buy it. Read the headlines and you would think Beijing just got the missing piece it needed to catch up. It did not. The missing piece was never the chip. Alibaba, Tencent, and DeepSeek have spent two years training their models on Huawei's Ascend 910B, a domestic chip that trails Nvidia's H100 on raw processing speed. But that speed gap was never the real reason PRC labs lagged on deployment, meaning actually getting a model into paying customers' hands. The real reason is simpler, and no export license can fix it: none of the three PRC labs owns a commercial cloud platform with the enterprise customer base that Microsoft Azure gives OpenAI, or that Amazon Web Services gives Anthropic. Washington lifted the licensing regime this week. The ten thousand signed enterprise contracts that gap requires are not part of the deal.
Korea Bets On Distribution
Seoul committed $550 billion to its chip sector this week, on top of a separately reported $1 trillion program spanning chips and robotics. Officials are framing this around what they call the chip crunch. That is not a shortage of factories that make chips. It is a shortage of high-bandwidth memory, the specialized memory chips that feed data into AI training clusters fast enough to keep them running. SK Hynix and Samsung already supply that memory to Nvidia. So Korea's bet is essentially this: keep being the supplier everyone needs, rather than trying to build the winning model yourself. Compare that to Anthropic's move this week to seek redress from Alibaba over what it calls model cloning, the allegation that a competitor trained its own model on outputs generated by Claude rather than on original data of its own. Both stories share the pattern set by Mythos. The fight has moved from who has the best model to who controls the pipeline that model runs on, whether that pipeline is a licensed enterprise API, a signed cloud contract, or a legally defensible claim to your own training data. Korea is not betting on winning a model race. It is betting on owning one link in somebody else's supply chain. The $550 billion committed to chips this week, plus the separately reported $1 trillion program, buys a seat at Nvidia's and Samsung's table. It does not buy a seat at OpenAI's.
So if the chip ban was not the wall, what was Washington actually holding back with it for two and a half years? Time, not capability. Two and a half years for US labs to build the enterprise contract base PRC labs still lack. The export controls lifted this week. The contract gap did not. That gap, not the chip, is what the next round of licensing decisions will actually be about.