Beijing's answer to OpenAI just made the strongest case yet that the West's AI export controls are backfiring spectacularly.
The Summary
- Zhipu AI founder Zhang Peng argues frontier AI should remain open and accessible, rejecting calls for concentrated control by select entities
- This directly contradicts the Western push toward closed, export-restricted AI development led by US policy and frontier labs
- China's largest AI companies are now positioning themselves as champions of open access while US competitors lock down their models
The Signal
Zhang Peng, founder of Zhipu AI, one of China's most prominent AI research labs, just threw down a gauntlet that Silicon Valley probably didn't expect. His argument for keeping frontier AI broadly accessible isn't just ideological posturing. It's a strategic wedge aimed directly at the contradictions in Western AI policy.
Zhipu built China's GLM series of models, which power applications across the country's tech stack. They're not a scrappy startup. They're a well-funded operation with deep ties to Tsinghua University and backing from major Chinese tech firms. When Zhang speaks about AI access, he's speaking from a position of strength, not scarcity.
"Frontier AI should remain broadly accessible rather than controlled by select individuals."
The timing matters. US export controls on advanced chips and AI models have created a bifurcated development landscape. American labs like OpenAI, Anthropic, and Google are increasingly secretive about model architectures, training techniques, and capability thresholds. The stated reason is safety. The actual effect is a moat. Meanwhile, Chinese labs are positioning themselves as the open alternative, the place where researchers worldwide can actually see what's under the hood.
This creates a fascinating inversion:
- Meta releases Llama models openly, but faces pressure to restrict access to cutting-edge versions
- Chinese labs release competitive models with fewer restrictions on who can use them
- Researchers in neutral countries now face a choice: closed Western models with export controls, or open Chinese models with fewer strings attached
Zhang's argument taps into a real tension in the AI research community. Many researchers, especially outside the US and China, resent being treated as security risks. Export controls don't just block Chinese access. They block Indian developers, Brazilian startups, Nigerian researchers. Anyone deemed insufficiently aligned with US interests gets frozen out.
Key strategic implications:
- China gains soft power by positioning as the "open" alternative to Western restrictions
- Talent flows toward whoever provides the best tools, regardless of origin
- The global AI commons fragments along geopolitical lines, but not the way Washington expected
The counterargument, that open frontier AI poses catastrophic risks, loses credibility when "open" actually means "open to US allies only." Real openness would include safety research, red-teaming, and global participation in governance. What we have now is selective closure masquerading as caution.
Zhipu isn't altruistic here. They benefit enormously from open development norms. They can study Western models, incorporate techniques, and leapfrog without doing all the foundational research. But their incentives align with a large chunk of the global research community who want to actually build things without navigating a maze of export compliance.
The Implication
The US is losing the narrative war on AI governance. When Chinese labs can credibly position themselves as champions of open access while American labs sound increasingly like they're hoarding capability for competitive advantage, something has gone badly wrong with the messaging.
Watch who builds the next generation of developer ecosystems. If Chinese models become the default for researchers in the Global South, that's not just a commercial win. That's a structural shift in where AI capability gets built and who controls the standards. The Fourth Web might speak Mandarin if the West keeps treating openness as a liability instead of a strategy.