A lot of China’s military secrets have, in effect, been loaded into American AI. Not every war plan walked out the door. The pipe did. Users thought they were talking to Kimi or DeepSeek. The backend was Claude. Surveillance video, internal code, live credentials, police matching systems—and scenes Anthropic flagged as likely military-linked—landed in a U.S. company’s logs. Secrets are not mainly lost to hacking. They are lost when your own people paste them into a “domestic” interface. People ask: China is not short of talent. Tsinghua, USTC, returnees—why can’t it break through the way the United States did? The shortage is not brains. It is soil. What grew OpenAI, Anthropic, and Google was three things stacked: capital willing to burn for a decade, labs allowed to fail, and a space where firms both cooperate with the state and fight it. A frontier model is not a quarter’s slide deck. It is compute, data, engineering, safety constraints, and years of wasted runs. Nobody gets a guarantee they can IPO next season. China has the people. The soil is thin. Bosses, as a rule, have no patience. They want a demo now, a raise now, a leaderboard win now. Distillation, wrappers, relay stations, fake accounts in Singapore—same instinct: someone else already paid the training bill; we just move the answers and stamp them “self-reliant.” Fast, yes. It also moved users and sensitive work into the other side’s server room. Shortcuts are not a tactic. They are a culture: “catching up” means looking the part; “breakthrough” means the spreadsheet looks good. The CCP’s deep grip on private firms salts what soil is left. Chips restricted. Data export restricted. Model filings restricted. Talking points restricted. At the same time the demand is: ideologically correct, technically sovereign, results immediately. Companies are crushed between three impossibilities: they cannot legally plug into U.S. models, they cannot take years to train from scratch, and they must still ship something that can fight under the right slogans. The rational short-term move is gray-market proxies, fake accounts, and pass-through. The tighter the controls, the sweeter the shortcut. The sweeter the shortcut, the easier secrets leak. So it is not that Chinese people are stupid. The system will not let anyone finish the stupid, slow work. Talent splits. Some leave. Some join big firms and grind apps. Some learn to sell a wrapper as innovation. People willing to burn five years on their own reasoning chain get squeezed out by quarterly reviews and inspection packets. America is impatient too, and full of foam. It still leaves a layer that can protect a lab from the trading desk. That layer is what China lacks. Then the loop closes: The louder the talk of self-reliance, the more the product depends on stolen outputs.
The louder the talk of secrecy, the more internal questions leave the country through a patriotic app.
The louder the talk of concentrating power to do great things, the more those things become free labeling for Anthropic. Military secrets in American AI are not mainly the work of traitors in the field. They are what you get when “fast” is the only KPI and the soil has been replaced by approvals and slogans. The talent is still there. There is nowhere to turn talent into a breakthrough—only into freight. Freight can win a quarter of headlines. It cannot win a real model war. The keys are already in the other pocket. The fix is not another post that says “we did not distill.” It is patience, room to work, and legal compute and data so the slow work can actually get done. Without those three, the next dump will only be larger.