Making sense of the panic over Chinese AI

The pattern is now so familiar it could be scripted. A Chinese lab releases a model. Benchmarks flash green. Twitter erupts. Venture capitalists quote TAM projections like scripture. Frontier lab executives subtweet about national security. Three days later the model still cannot reliably run a bash script, and the panic evaporates until the next release. Moonshot AI's Kimi just ran this play for the third time in a year. DeepSeek did it before. The script does not change because the incentives do not change.

What makes this cycle dangerous is not the hype. Hype is cheap. What makes it dangerous is the quiet lobbying campaign running in parallel. OpenAI and Anthropic have reportedly pressed regulators in Washington to restrict open-weight Chinese models. They frame it as competitiveness. It looks like moat-building. When the companies most threatened by open distribution ask the government to choke off that distribution, the press should treat their patriotism with the same skepticism it reserves for telecom mergers.

Kirsten Korosec put the question cleanly on the Equity podcast: are we accelerating American victory in the AI race, or are we ensuring that certain frontier labs do better than others? The answer determines whether policy serves the public or the cap table. Restricting Chinese open models does not make American models better. It makes American models richer. It hands pricing power back to the labs that spent two years arguing that open weights were unsafe — right up until their own open-weight competitors started winning benchmarks.

Sean O'Kane called the industry "jumpy." That is generous. The industry is performative. Watch the same accounts that spent Saturday declaratively stating that Kimi had replicated macOS in thirty minutes. They posted screenshots of a pixel-perfect Aqua interface. They did not post the kernel. They did not post the driver stack. They did not post a single line of working POSIX compliance. A graphical skin is not an operating system. A benchmark score is not a product. The industry knows this. It pretends not to on weekends because outrage engages better than nuance.

The "touch grass" line landed because it described the actual pathology. Grown engineers spent seventy-two hours trading barbs on X about a demo they could have debunked in ten minutes by asking for the repo. They did not ask. They did not want to debunk it. They wanted the panic. Panic justifies budget. Panic justifies headcount. Panic justifies the next round at a higher valuation. The frontier labs have monetized anxiety so effectively that they now outsource the anxiety generation to their own executives.

Chinese labs are not innocent actors. They optimize for benchmark visibility because benchmark visibility attracts Western capital and Western talent. They release open weights because open weights are the only distribution channel that bypasses the App Store gatekeepers and the cloud provider lock-ins. The openness is strategic, not ideological. But strategic openness still produces public goods. Developers in Lagos, São Paulo, and Warsaw can fine-tune Kimi today. They cannot fine-tune GPT-4o tomorrow. That asymmetry matters more than any single benchmark.

The lobbyists in Washington know this. They know that restricting Chinese open models will not stop Chinese research. It will only stop Chinese distribution. The models will still train. The weights will still exist. They will simply move to mirrors in Singapore, Dubai, and Frankfurt — jurisdictions that have no interest in enforcing American export controls on mathematics. The frontier labs know this too. They are not lobbying to stop the technology. They are lobbying to delay the commoditization of their own moats.

Commoditization is the real threat. Every open release pushes the price of intelligence toward zero. Every benchmark parity announcement erodes the narrative that only three companies can build frontier models. The panic is not about Chinese superiority. It is about the evaporation of pricing power. The lobbyists are buying time. They are asking Congress to subsidize their margins with national security rhetoric.

Congress should say no. Not because Chinese models are harmless — they are not. Not because open weights carry no risk — they do. Congress should say no because the remedy proposed by the lobbyists cures the wrong disease. The disease is not Chinese openness. The disease is American concentration. Three labs control the frontier. They want four. They want five. They want the number to stay countable on one hand. Restricting Chinese open models keeps the number small. It does not make the remaining models safer. It makes them more expensive.

The alternative is boring. Invest in domestic open research. Fund the compute grants that let universities train competitive open models. Build the evaluation infrastructure that lets enterprises audit any model, Chinese or American, without trusting the vendor's benchmark blog post. Treat AI like the boring infrastructure it is becoming — electricity, not Excalibur. That work does not generate weekend Twitter wars. It does not juice valuations. It does not make for good podcast clips. It simply builds the thing the lobbyists say they want: American leadership that does not depend on keeping everyone else in the dark.

The panic will return. Another Chinese lab will release another model. Another benchmark will flash green. Another weekend will burn. The industry will jump again. The question is whether Washington jumps with it, or whether Washington looks at who is holding the leash.