China’s AI labs have essentially solved the model problem and still can’t solve the money problem, and that second one is harder.
That’s the whole story in one line, but it deserves unpacking, because both halves of it get mangled in the usual coverage. One camp insists Chinese models are distilled knockoffs coasting on American research. The other insists the US lead evaporated sometime last year. The numbers support neither.
The capability gap is now a rounding error
As of March 2026, US and Chinese models sat roughly level on Arena, the leaderboard where users compare two anonymous responses and vote for the better one. That matters more than benchmark scores because nobody can game their own preference. You don’t know which model you’re voting for. You just pick the answer that helped.
Stanford put the performance gap between the top American and top Chinese models at 2.7 percent by March 2026. For anyone who has actually run these things side by side in production, 2.7 percent is noise. It’s smaller than the variance you get from a prompt rewrite. It is not a gap you can build a business strategy on, and it is definitely not a gap you can build a trillion-dollar valuation on.
The cost side is where it stops being close and starts being uncomfortable. Chinese models are delivering 90 percent or more of frontier capability at 5 to 10 percent of the cost. I’ve tested enough tools to know what that does to buying decisions. If you’re building an agent that makes thousands of calls per task, a tenfold price difference isn’t a preference, it’s the entire margin. The market is voting with its wallet, and the wallet does not care about national pride.
Where the money actually sits
Now the other half. The US dominates AI spending and computing power, and that dominance isn’t subtle. Anthropic is valued around $965 billion. OpenAI around $852 billion. The combined valuation of the major Chinese AI labs is roughly $455 billion, meaning the two biggest American labs together are worth about four times the entire Chinese field.
Interesting part: China is ahead on research output while behind on capital. That combination is unusual and it tells you something about where each side’s constraint actually is. American labs are buying their way forward. Chinese labs are engineering their way forward because they have less to spend.
Efficiency under constraint is a real competence. It is also not a substitute for compute when the next jump requires compute. You can out-engineer a funding gap for a while. You cannot out-engineer it forever.
The distillation argument, briefly
American companies have accused Chinese competitors of model distillation, the technique where outputs from a stronger model train a weaker one. Worth addressing because it comes up every time someone posts a Chinese model’s benchmark results.
Two things are true at once. Distillation would explain fast capability gains at low cost. It would not explain sustained research leadership, and it doesn’t make the cheaper model less useful to the person paying the bill. If your API spend drops 90 percent, the provenance of the training data is an interesting legal question and an irrelevant operational one. Harsh, but that’s how procurement works.
Kimi K3 and the pressure it creates
Moonshot AI showed Kimi K3 at the 2026 World AI Conference, and it did exactly what DeepSeek and Qwen have been doing for a while now, which is make US spending levels look like a choice rather than a requirement. That’s the actual threat to American labs. Not being beaten. Being made to look expensive.
Those enormous valuations rest on an assumption that frontier capability stays scarce and defensible. A 2.7 percent gap at a tenth of the price is not scarcity. It’s a commodity market forming in real time, and commodity markets are brutal to anyone who priced in a monopoly.
What I’d actually tell you to do
- Stop treating model choice as a geopolitical statement. Test both, measure your own task, pick what wins on your data.
- If your workload is high-volume and latency-tolerant, the cost math is hard to argue with.
- If you need the absolute top of the capability curve, that 2.7 percent might matter. For most people it doesn’t.
- Price in regulatory and availability risk on both sides. That’s a business risk, not a quality one, and it belongs in a different column of your spreadsheet.
The capability race is close to settled as a practical matter. The capital race is not, and that asymmetry is going to define the next couple of years more than any leaderboard will. American labs have the money and need to justify it. Chinese labs have the efficiency and need to fund the next jump. Both problems are real. Only one of them can be fixed by writing better code.
🕒 Published: