Where we dig into one subject across several stories — the one piece to read closely today
Two hours by train from Beijing, roughly 100 data centers have sprung up in Ulanqab, Inner Mongolia. Cheap electricity, cold winters, and surplus wind power have drawn China's AI companies to this plateau — but underneath sit two unresolved problems: water scarcity and coal. One city tells you how China's AI infrastructure math actually works.
By 김현국
Two hours by train from Beijing, roughly 100 data centers have sprung up in Ulanqab, Inner Mongolia. Cheap electricity, cold winters, and surplus wind power have drawn China's AI companies to this plateau — but underneath sit two unresolved problems: water scarcity and coal. One city tells you how China's AI infrastructure math actually works.
Z.ai's GLM-5.3 was welcomed as a tool for defenders, while TamperBench — built by the University of Waterloo and FAR.AI — showed that the safeguards on 21 widely used open-weight models can be stripped away with a single round of fine-tuning. Put the two stories side by side, and the real issue around open-weight models comes into focus.
Z.ai's gong from its IPO, an ordering system rebuilt with Kimi K3, free DeepSeek access, and a 300-person party at WAIC — founders, developers, and investors keep converging on this one spot, and it's turning into lived-in infrastructure for China's AI ecosystem.
NATO Special Operations School researcher outlines an AI-driven expertise collapse scenario in "Tragedy of the Cognitive Commons"
That's the last story.
