Chinese AI Models Now Handle Up to 46% of US Enterprise AI Traffic
Quick Answer
A CNBC investigation published July 7 found that Chinese open-source AI models now handle 30% to 46% of enterprise AI token traffic in the US, up from just 4.5% in early 2025. The reason isn’t politics — it’s price. These models run 60% to 90% cheaper and perform within a point or two of the top US models on real work. Big companies are proving, with real budgets, that “good enough and cheap” is beating “best and expensive.”
What Actually Happened
On July 7, CNBC published an investigation into usage data from OpenRouter, one of the biggest AI model marketplaces developers use to route requests. The number that stood out: Chinese-built open-source models have handled between 30% and 46% of the AI tokens routed by US companies every single week since February 8, 2026. A year earlier, that share averaged just 11%. In the first half of 2025, it was only 4.5%.
The clearest example of why is Z.ai’s GLM-5.2, released in June. It’s now the fastest-adopted model Vercel has tracked all year. In its first full week live, daily token volume on the platform grew about 27 times over, and the number of customers using it grew roughly 80 times over. Vercel’s head of agentic infrastructure, Harpreet Arora, gave CNBC the reason in four words: “Price is doing the work here.”
That price gap is the whole story. OpenRouter’s data team says open-source Chinese models run 60% to 90% cheaper than the leading models from Anthropic and OpenAI, while landing within a point or two of them on tough coding and agent benchmarks. For a company routing millions of tokens a day, that’s not a rounding error. It’s the difference between a manageable AI bill and a runaway one.
46%
Peak weekly share of US enterprise AI tokens now going to Chinese models
90%
Cheaper than leading US models, at the high end of the gap
27x
Jump in GLM-5.2’s daily token volume in its first week live
Why This Isn’t Just an Enterprise Story
You’re probably not running a fleet of AI agents across your whole operation, or migrating an entire team off one model the way some companies reportedly have. But you’re making a smaller version of the exact same call every time you open a new AI tool. Do you jump to whatever just launched with the loudest hype? Or do you stick with something that’s already good enough and put your energy into actually getting results from it?
Companies with the biggest budgets and the best data teams on earth just answered that question with real dollars. They didn’t pick the flashiest model. They picked the one that was good enough, at a fraction of the cost — and they’re not switching back. That’s not a small signal. It’s the market’s most well-resourced buyers telling you the newer, pricier option isn’t automatically worth the jump.
What It Means For You
We’ve made the case before that most builders lose more to switching tools than they ever lose to picking the “wrong” one — and this week, a CNBC investigation just proved it at enterprise scale. The companies gaining the most right now aren’t the ones chasing every new release. They’re the ones who found something good enough, then stopped restarting the clock every time a shinier option showed up.
The lesson here isn’t “go find the cheapest AI tool.” It’s simpler than that. The tool you’re using probably doesn’t need to be replaced — it needs to be used better. That’s the whole argument, now playing out with enterprise budgets instead of hype cycles.
Go deeper:
Frequently Asked Questions
Why are enterprises suddenly using more Chinese AI models?
Mostly price. Open-source Chinese models run 60% to 90% cheaper than the top US models and now land within a point or two of them on real benchmarks, so companies are routing everyday work to the cheapest model that’s still good enough.
Does this mean Chinese AI models are now better than the top US models?
Not necessarily. US labs still lead on the hardest, most demanding tasks, and Chinese models reportedly trail by several months on frontier benchmarks. What’s changed is that “good enough” now costs a fraction of the price, so the bulk of everyday work is shifting.
What does this mean if I’m not running an enterprise AI stack?
The same logic applies at any size. Constantly switching to the newest tool rarely pays off. The bigger win is picking one tool that’s good enough and getting deep with it, instead of resetting every time something new launches.