Chinese AI Models Now Run Nearly Half of US Enterprise AI Traffic — And US Tech Stocks Are Feeling It

The clearest sign that Chinese AI labs have moved from "competitive" to "structurally disruptive" isn't a benchmark score. It's a usage number. Chinese-origin AI models hit a weekly peak of 46 percent of US enterprise token usage on OpenRouter this year, according to data tracked by IT Home and cited across multiple outlets — up from an 11 percent average over the prior twelve months, and just 4.5 percent in the first half of 2025. Since February 8, Chinese models have held at least 30 percent of enterprise token volume every single week. This isn't a spike. It's a sustained migration of real, paying enterprise workloads away from US frontier labs.

The driver isn't mysterious, and it isn't really about capability parity — it's about price, at a scale that makes the decision close to automatic for cost-sensitive enterprise deployments. Open-source Chinese models are running 60 to 90 percent cheaper than comparable offerings from Anthropic and OpenAI. The specific numbers are stark: DeepSeek's V4 Flash model costs roughly $0.14 per million input tokens, against $5.00 per million for OpenAI's GPT-5.5 — a more than 35-to-1 price gap for workloads where the cheaper model is good enough for the task at hand. For any enterprise running AI at real production volume, where token costs compound quickly across millions of daily requests, that gap isn't a rounding error in a budget meeting. It's the difference between a viable AI deployment and an unaffordable one.

Wall Street has been repricing US AI exposure in real time as this data becomes harder to wave away. US tech stocks fell again this month after China's latest AI model announcements at the World AI Conference in Shanghai, part of a now-familiar pattern of Chinese model releases triggering US selloffs — Alphabet shares dropped 4 percent on reports of a delayed flagship AI model launch, then slipped a further 2 percent the next session, while Nvidia shares fell more than 2 percent as the broader chip-sector selloff accelerated. At least one prominent tech CEO has gone further, warning publicly that the AI spending bubble in the US may be "about to pop" as cheaper, genuinely competitive alternatives erode the pricing power that current valuations assume US labs will keep indefinitely.

What makes this moment different from earlier scares — including the original DeepSeek shock in January 2025 — is the shift from potential to measured, ongoing enterprise behavior. A benchmark win is a claim about capability. A sustained 30-to-46 percent share of actual paid enterprise token volume is a claim about what businesses are actually doing with their production AI budgets, week after week, for months running. That's a much harder data point for US labs and their investors to argue away, and it directly complicates the case for the scale of capital expenditure US AI companies have committed to on the assumption that being the best model matters more than being the affordable one.

None of this means US frontier labs are losing the capability race outright — the broadest aggregate benchmarks, as this outlet has covered, still generally favor Fable 5 and GPT-5.6 Sol over their Chinese counterparts. But capability leadership and market share are turning out to be two different competitions, and the enterprise token data suggests American AI companies may be winning the first while steadily losing ground in the second. For an industry whose valuations are built substantially on assumptions about durable pricing power, that's the more consequential number to be watching.