Perplexity CEO’s China AI Warning: Did U.S. Controls Accidentally Make Open-Source China Stronger?
Chinese open models like DeepSeek and GLM-5.2 are now forcing American developers to ask a painful question: what if containment made China’s AI ecosystem more resilient?
The China AI debate has moved from fear to embarrassment. The old assumption was simple: restrict chips, protect U.S. model companies, slow Beijing, preserve American dominance. But the result may be more complicated. China did not stop. It adapted.
A quote attributed to Perplexity’s CEO has gone viral because it captures the mood: whatever the U.S. did to stop China from catching up did not really matter. China caught up anyway, and the most dangerous part is not simply that Chinese models are good. It is that the best Chinese models are open enough for American developers to build on.
DeepSeek made that argument impossible to ignore. It offered strong performance at low cost and forced Silicon Valley to confront the possibility that frontier-style capability did not require the same enormous spending model. Now Z.ai’s GLM-5.2 is being promoted as another leap: open weights, long context, strong coding benchmarks and performance close enough to leading proprietary models to matter.
The exact benchmark gap will be debated. Every model launch comes with marketing. Terminal-Bench, coding tests, long-context tasks and agent evaluations can be gamed, misunderstood or cherry-picked. But the strategic trend is clear: Chinese labs are not only building models. They are building ecosystems.
Open models change the race because they diffuse. A closed American model may be more powerful, but it sits behind pricing, policy and access controls. An open Chinese model can be downloaded, fine-tuned, hosted locally, modified and integrated into products by teams that do not want to depend on U.S. vendors. That includes developers in India, the Middle East, Latin America, Africa, Europe and even America.
This is where U.S. policy faces a paradox. Export controls may slow China’s access to the best chips. But they also push Chinese companies toward efficiency, model compression, open deployment and domestic supply chains. If containment makes your rival more self-reliant, the strategy has a hidden cost.
Washington still has enormous advantages: frontier labs, capital markets, cloud infrastructure, top universities, chip design, global talent and military funding. But China has scale, manufacturing, state coordination, open-source momentum and a domestic market large enough to absorb experimentation.
The question is not only who has the best model this month. It is who controls the defaults. Developers are practical. They follow price, performance, licenses and community support. If a Chinese model is good enough and cheap enough, ideology may not stop adoption.
That creates a security problem. Governments worry about hidden bias, censorship, data leakage, backdoors, dependency and influence. If Chinese-origin models shape codebases, customer-service systems, education tools, medical triage or national platforms, the geopolitical layer becomes unavoidable.
But the same concern applies to U.S. closed models. Countries outside the West ask why they should trust American AI infrastructure more than Chinese AI infrastructure. Both are shaped by national laws. Both can censor. Both can collect data. Both can be weaponized. The difference is that open weights at least allow inspection and local hosting, while closed APIs require trust.
That is why Europe is watching nervously. It does not want Chinese dependency, but it also does not want total American dependency. The open-source race gives Europe a chance to build on available models while creating its own infrastructure. The danger is that Europe becomes a regulator of other people’s technology rather than a producer of its own.
For American developers, the issue is immediate. If DeepSeek, Qwen or GLM models are cheaper and strong enough, should they use them? If a startup cannot afford premium closed models, open Chinese models may be the difference between existing and dying. National security arguments rarely pay a startup’s inference bill.
The headline says China caught up. The more careful version is that China has built enough open AI capacity to disrupt the assumption of permanent U.S. control. That alone changes the game.
The next AI superpower may not be the country with the single best closed model. It may be the country whose models become the most widely copied, modified and embedded. By that measure, open source is not a side issue. It is the battlefield.