Chinese AI Models Are Winning US Enterprise Traffic

American companies are quietly rerouting a growing share of their AI workloads to models built in China. According to reporting by CNBC on July 7, 2026, the share of tokens that US companies run on Chinese AI models through model-routing gateways has sat above 30% every week since February 8, 2026, and has climbed as high as 46%. The average over the prior 12 months was just 11%, and it was as low as 4.5% in the first half of 2025. The shift has been fast enough that, by July 8, US lawmakers had opened an inquiry into the trend.

The story is not that Chinese labs suddenly out-engineered their American rivals on raw capability. It’s that they got close enough, opened the weights, and undercut everyone on price — and that developers noticed.

The numbers behind the shift

The data comes from usage on model-routing gateways, where developers send a prompt and the service picks or forwards it to a chosen model. That vantage point makes gateways a useful barometer of what teams are actually running in production, as opposed to what they say in surveys.

By that measure, Chinese models crossed 30% of US token volume in early February and have not dropped below it since, with individual weeks touching 46%. For a category that was a rounding error in early 2025, that is a structural change in a year and a half.

The driver, by nearly every account, is cost. Open-weight Chinese models can be 60% to 90% cheaper to run than the leading systems from OpenAI and Anthropic. “Price is doing the work here,” an analyst, Arora, told CNBC. “When a task doesn’t need the best model, teams are beginning to route it to the cheapest one that’s good enough, and the recent wave of models coming out of China is winning that trade.”

That “good enough” framing matters. Not every request needs a frontier reasoning model. Classification, extraction, summarization, routing, and first-draft generation are high-volume, low-stakes tasks where a capable open model at a tenth of the price is an easy call — especially once a workload is metered in billions of tokens a month.

Which models, and why now

The Chinese field is no longer a single name. DeepSeek remains the best-known, but the surge is broad. Z.ai’s GLM 5.2, released in June, saw what one gateway operator called the fastest adoption of any model it tracked in 2026: in its first full week, daily token volume grew roughly 27x and the number of customers using it grew about 80x. Alibaba’s Qwen, Moonshot AI’s Kimi, and MiniMax round out a deep bench of open-weight options.

Two things changed at once. Capability converged — the gap between the best open models and the best closed ones narrowed on many everyday tasks, a trend we covered in open-source AI models closing the gap. And distribution got frictionless: with open weights and gateway routing, swapping a model is often a one-line config change rather than a migration. For background on the specific systems fueling the move, see our explainers on GLM 5.2 and Z.ai.

The clearest sign of how sharp the trade can be: the AI startup Lindy said it moved 100% of its traffic from Claude to DeepSeek, a switch it expects to save millions of dollars. Not every company will go all-in, but the anecdote captures the calculus — when margins are thin and token bills are large, a cheaper model that clears the quality bar is hard to argue against.

The cost pressure on US labs

The flip side of cheap Chinese models is expensive American ones. As frontier labs raise prices to fund enormous training and inference budgets, the spread between “the best model” and “a good enough model” widens in dollar terms even as it narrows in capability. Teams that once defaulted to a single premium provider are now building routing layers that send each request to the cheapest model that will do the job.

Techniques that used to be optimizations are now table stakes. Aggressive prompt caching to cut LLM costs, smaller models for narrow tasks, and multi-provider routing all point the same direction: treating inference as a commodity to be sourced competitively rather than a single-vendor relationship. Chinese open-weight models slot neatly into that mindset because they can be self-hosted or rented from many providers, which keeps pricing competitive.

The complications

Cost is only half the story. Running workloads on Chinese models raises a cluster of concerns that boards and regulators are now weighing.

• Data governance. Where prompts and outputs are processed, logged, and retained becomes a compliance question — particularly for regulated industries and anything touching personal data. Self-hosting open weights sidesteps some of this; calling a Chinese-hosted API does not.
• Content and values. Models trained under one regulatory regime carry that regime’s constraints. Responses on politically sensitive topics can differ, and defaults around refusals and filtering may not match a US company’s expectations.
• Geopolitical exposure. The same week the adoption numbers surfaced, reports emerged that Beijing was weighing whether to restrict some AI model releases as US reliance on them came into focus. A supply that can be turned off by policy is a supply that carries risk, no matter how cheap.
• Scrutiny. With lawmakers now probing the trend, companies leaning on Chinese models may face disclosure requests, procurement rules, or outright restrictions — a regulatory overhang that closed US models don’t carry.

None of these is necessarily disqualifying, and open weights that run on infrastructure a company controls blunt several of them. But they turn what looks like a simple pricing decision into a governance decision.

What it means

The gateway data marks a real inflection: for a large and growing slice of everyday US enterprise AI work, the default is no longer an American frontier lab. That doesn’t threaten OpenAI or Anthropic at the top of the market — the hardest reasoning, agentic, and coding tasks still favor the best closed models, and those are the workloads customers pay premium prices for. But it does compress the profitable middle, where high-volume, low-complexity requests used to flow to premium providers by inertia.

Who wins: buyers, first and foremost, who now have real price leverage and a credible alternative for routine work. Chinese labs win distribution and mindshare inside the world’s most demanding software market. And the model-routing and gateway layer wins relevance, because sorting requests by cost and capability is now a core part of running AI economically.

Who’s squeezed: US labs that priced on the assumption of default loyalty, and any provider whose pitch is “one model for everything.” The competitive answer isn’t to match Chinese pricing on commodity tasks — it’s to keep the capability lead wide enough that the premium tier stays clearly worth it, the same dynamic shaping reasoning models.

What to watch: whether Washington moves from probing to restricting; whether Beijing actually curbs model releases; and whether the 46% peak becomes a ceiling or a floor. The number to track isn’t any single model’s benchmark score — it’s the share of production tokens, because that’s where the money, and now the politics, actually flow.