Previous Columns

Beijing tightened its own AI model exports while easing Nvidia chip import caps, but the real constraint on China's AI buildout sits in Korean memory supply, not chip licenses.

OpenAI shipped GPT-5.6 and killed Atlas, lost its AGI chief and safety chief, and got sued by Apple, all in the same ten days. That is a company folding back to the one thing it still does better than anyone.

South Korea's trillion-dollar sovereign AI build depends on Samsung fabs whose most advanced line may end up making chips for a US lab that just got its models banned and unbanned by Washington in three weeks flat.

Korea just pledged $550 billion to fix a chip shortage that is really a power shortage, and that distinction determines who actually gets to train next-generation models.

Trump's rollback of the Mythos export controls and Korea's chip pledges are reopening the compute map for US labs while Alibaba's Qwen clone dispute exposes what China's labs still cannot buy: a deployment surface.

Washington lifted the Mythos export controls this week. Don't expect that to hand Beijing compute parity, because the chip ban was never the real constraint. The real constraint was who owns the deployment surface the chip feeds.

Anthropic's push to punish Alibaba for cloning Claude reveals that the harder Chinese labs compete on capability, the more the US advantage has quietly moved to enterprise deployment contracts, not model weights.

The Alibaba distillation attack and the Fable 5 ban arrived seven days apart and together expose a regulatory architecture being assembled around a gap it cannot close.

The directive from the Bureau of Industry and Security (BIS, Commerce's export enforcement arm) that pulled Fable 5 and Mythos 5 globally on June 13 delivered APAC governments a more compelling sovereign-AI argument than any policy paper has.

Fable 5 is suspended because Anthropic disclosed the jailbreak the government cited; the identical jailbreak is documented in GPT-5.5, which faces no equivalent order.

Anthropic's IPO prices its enterprise API business at a $47B run rate and Apple's concurrent decision to power Siri with Gemini defines exactly which market that business cannot enter.

Google's agent era declaration is backed by B200 NVL72 inference clusters that BIS October 2023 controls have already placed beyond PRC labs' procurement reach.

Greg Brockman's return to OpenAI product and ChatGPT's push into bank-account integration have together expanded OpenAI's liability surface past any US enforcement framework built to contain it.

OpenAI is pressing ChatGPT into banking while its CEO's credibility is formally contested in federal court; US tort law, not pending statute, is the constraint on the deployment surface.

Anthropic's $5B Colossus contract and Nvidia's $40B equity portfolio mark training-compute's shift from procurement to partnership, a structure PRC labs cannot access before 2028.

Pentagon procurement cuts and drone strikes on data centers, arriving together, mark the moment when AI's next constraint became physical rather than algorithmic.