Anthropic recently disclosed an annual revenue run rate of $30 billion, enough to pass OpenAI. That is one company, one year, one number. Now hold that next to a smaller Chinese model called Qwen3-30B-A3B, which has posted benchmark results better than models considerably larger than itself, and which you can download without paying anyone a cent.
Those two facts sitting side by side explain the whole story of AI right now. The capability gap between China and the U.S. has narrowed to something close to a rounding error. The money gap has not.
What “matching the models” actually means
If you follow this stuff casually, you have probably heard that China is “catching up” in AI for about three years running. The difference now is the specificity. Alibaba shipped Qwen3 as an open-source answer to U.S. models, and parts of that family beat bigger competitors on benchmarks. DeepSeek says its newest models nearly match what OpenAI and Anthropic are putting out. Baidu’s Ernie 4.0 is described as being on par with GPT-4.
For a non-technical reader, here is the translation: the underlying brains are becoming interchangeable. If you are building an AI agent that reads your email, drafts replies, pulls data from a spreadsheet, and files a report, the quality difference between a Chinese open-source model and an American commercial one is no longer the thing that decides whether your agent works.
That is genuinely new. For most of the past few years, the answer to “which model should my agent use?” was short: the best American one you can afford. Now the honest answer is “it depends on what you are optimizing for,” which is a far more interesting conversation.
The part money still buys
So if the models are comparable, what does $30 billion in revenue get you? Quite a lot, as it turns out, and almost none of it is intelligence.
- Capacity. Running a model for millions of simultaneous users is an infrastructure problem, not a research problem. Serving is expensive, continuously.
- Distribution. American labs are wired into the tools people already use at work. That placement is bought and negotiated, not invented.
- Staying power. Revenue at that scale funds the next training run, and the one after that, without needing anyone’s permission.
- Trust infrastructure. Compliance teams, security reviews, enterprise contracts, support staff. Boring, costly, and the reason big companies sign.
Chinese labs are not standing still on the hardware side either. ByteDance has designed two chips with TSMC that it reportedly plans to mass produce by 2026, which is an attempt to solve the capacity problem directly rather than out-spend it. But designing silicon is a long game, and the revenue scoreboard today is not close.
Why open source flips the math
The strange twist is that being behind on money has pushed Chinese labs toward a strategy that is working. China’s open-source models are gaining ground precisely as U.S. models face new rollout restrictions. When the best American system is hard to get, slow to approve, or unavailable in your region, a free model you can download and run yourself stops being a compromise and starts being the practical choice.
Some U.S. startups have already made that switch, moving off Western models toward Chinese open-source ones. Not out of ideology. Because the numbers work.
If you are building agents, open weights change your options in concrete ways. You can run the model on your own servers, so sensitive data never leaves your building. You can fine-tune it on your own documents. Your costs become predictable because you are paying for compute rather than per-token pricing someone else sets. And nobody can deprecate the model out from under you next quarter.
What this means if you are not an engineer
You do not need to pick a side in a geopolitical race to make a sensible decision. A few things are worth keeping in mind.
Stop treating model choice as a status symbol. The frontier name is not automatically the right fit for a summarizer, a classifier, or a scheduling agent. Smaller models beating larger ones on benchmarks is the clearest signal that size and brand have stopped being reliable proxies for usefulness.
Design so you can swap. If your agent is built around one specific API in a way that would take months to unwind, you have handed away your negotiating power. Keep the model layer replaceable.
And do read the licenses and data terms rather than assuming. Open weights and open-ended permission are different things, and where a model runs matters for what you are allowed to put into it.
The capability race is turning into a draw. The business race is not. For anyone actually building, that gap is where the useful decisions live.
🕒 Published: