Washington's attempt to cage the frontier labs might just hand the keys to everyone else.
The Summary
- New White House export controls on OpenAI and Anthropic are designed to slow China's AI progress but may backfire by accelerating global open-source development
- Perplexity CEO warns the restrictions could inadvertently boost China's domestic AI innovation, potentially reshaping who leads the global AI race
- Chinese AI already matches Anthropic's capabilities in cybersecurity, suggesting export controls arrived too late to maintain the gap Washington hoped to preserve
The Signal
The White House is rolling out export controls targeting America's frontier AI labs. The theory: restrict access to cutting-edge models from OpenAI and Anthropic, and you slow down adversaries. The reality shaping up: you might just incentivize the world to route around you.
Perplexity's CEO laid it out plainly. When you cut off access to closed models, you don't stop AI development. You redirect it. China doesn't throw up its hands and quit. China doubles down on open-source alternatives and domestic model development. Every restriction is a forcing function for independence.
"Export controls may inadvertently boost China's AI innovation, potentially reshaping global AI leadership."
The evidence is already mounting. Chinese AI systems now match Anthropic's performance in cybersecurity tasks, a domain where capability gaps matter most. This isn't theoretical future risk. This is present-day parity in one of the most sensitive application areas. If the goal was to maintain American advantage through access control, that ship has already left port.
Here's what happens next if this pattern holds:
- Global AI development fragments into closed US models and increasingly capable open alternatives
- Countries and companies locked out of frontier models invest heavily in open-source AI infrastructure
- The clampdown strengthens global open-source AI ecosystems by creating massive demand for alternatives
The geopolitical calculation gets messy fast. Export controls work when you control a scarce resource with no substitutes. But AI isn't uranium. The knowledge is already distributed. The architectures are published. The compute is expensive but not impossible to acquire. When you restrict access to GPT-5 or Claude 4, you don't stop AI progress. You create a market incentive worth billions for anyone who can build something close enough.
The Implication
If you're building in the agent economy, watch where open-source models go next. The combination of export pressure and commercial incentive could accelerate open model capabilities faster than the slow-and-careful approach of frontier labs. Models you can run on your own infrastructure, audit completely, and modify freely start looking a lot more attractive when geopolitics makes API access unreliable.
For policymakers, this is the hard lesson: in software, walls rarely work as intended. They just change who builds what, and where. The question isn't whether to regulate AI. It's whether regulation designed for hardware export control can possibly work for software that wants to be free.