Azeem Azhar spent time inside Chinese AI labs and came back with a data point nobody expected: Chinese researchers are getting 4-7 times more intelligence per compute unit than their US counterparts. Not because they have better hardware - they're working with chips that are 2-3 years behind. Because export controls forced them to be ruthlessly efficient.
The result: Chinese open models now sit 6-8 months behind the US frontier while costing 11 times less to run.
The Hardware Disadvantage Became an Efficiency Advantage
US export restrictions cut Chinese labs off from the latest GPUs. The assumption was this would slow them down. Instead, it forced optimisation. When you can't throw more compute at a problem, you learn to use less compute better.
According to Azhar's research, Chinese labs developed techniques to extract more capability from older hardware. Model compression, quantisation, distillation - methods that were academic curiosities in the US became production necessities in China.
The irony: the constraint that was supposed to hold them back created a structural cost advantage. Chinese models are cheaper to train and cheaper to run. That changes the economics of deployment.
The Performance-Cost Trade-Off
Chinese open models are 6-8 months behind GPT-4 or Claude in raw capability. But they cost 11 times less to run. For most applications, that trade-off is compelling.
A chatbot doesn't need frontier performance. A content moderation system doesn't need GPT-4. A code autocomplete tool doesn't need Claude. They need "good enough" performance at a price that makes the unit economics work.
Chinese labs are building for deployment, not benchmarks. The models are smaller, faster, cheaper - and for 80% of use cases, that's better than bleeding-edge capability you can't afford to run at scale.
What Export Controls Actually Did
The goal of export controls was to slow Chinese AI development. The effect was to force Chinese labs to develop a different approach. They can't compete on raw compute, so they compete on efficiency.
US labs have access to the best hardware. Chinese labs have to make do with older chips. But "making do" forced innovation. When you can't scale up, you optimise down. When you can't buy the latest GPU, you learn to do more with less.
The result is a bifurcation: US labs optimising for capability, Chinese labs optimising for cost. Both strategies work. But for commercial deployment, cost often wins.
The Open Model Strategy
Chinese labs are releasing their models openly. Not out of altruism - out of strategic necessity. Open models build ecosystems. They create standardisation. They make the world depend on your architecture.
US labs keep their best models proprietary. Chinese labs give theirs away. The playbook is familiar: Android vs iOS, Linux vs Windows. Open systems trade margin for reach. They win by being everywhere.
If Chinese models become the default for cost-sensitive applications, that's a win. Not in revenue - in dependency. The infrastructure layer matters more than the application layer. Whoever owns the models that power 80% of use cases owns the future, even if they don't own the frontier.
What This Means for Builders
If you're building on LLMs, the cost structure just shifted. Chinese models offer 11x cheaper inference with performance that's close enough for most tasks. That changes the maths on what's economically viable.
Applications that didn't make sense at GPT-4 pricing suddenly work at Chinese model pricing. The use cases expand. The accessible market grows. The constraint isn't capability anymore - it's distribution.
For developers, this means: test Chinese models. Don't assume US models are the only option. If your application doesn't need frontier performance, you're paying for capability you're not using. The 11x cost difference is real money at scale.
The Bigger Pattern
Export controls were supposed to create a moat. Instead, they created parallel innovation paths. US labs have compute abundance. Chinese labs have efficiency constraints. Both are innovating - just in different directions.
The question isn't who's ahead. It's which innovation path matters more for the next decade of deployment. Capability or cost? Frontier performance or ubiquitous access?
Azhar's research suggests the answer might not be what the export controls assumed. Constraints don't always slow you down. Sometimes they force you to find a better path.