A Threshold Few Expected to See
The share of tokens that US companies route to Chinese open source AI models has climbed to as much as 46 percent on OpenRouter, a widely used model routing platform, and has stayed above 30 percent every week since early February. To appreciate how sudden this is, consider the baseline: the trailing twelve month average was roughly 11 percent, and in the first half of 2025 it sat as low as 4.5 percent. In the span of a year, models from DeepSeek, Alibaba's Qwen, Moonshot's Kimi, and Z.ai have gone from curiosity to a meaningful chunk of American enterprise inference.
We think this is one of the most consequential shifts in enterprise AI this year, and it has almost nothing to do with a breakthrough in raw capability. It is a procurement story. As the cost of frontier models from OpenAI and Anthropic keeps climbing, buyers are quietly discovering that a large fraction of their workloads never needed a frontier model in the first place. The result is a rerouting of real production traffic toward whatever is cheapest that still clears the bar, and increasingly that is Chinese.
Price Is Doing the Work
The economics are blunt. Open source Chinese models can run 60 to 90 percent cheaper than the leading systems from Anthropic and OpenAI, and for routine tasks the quality gap has narrowed to the point of irrelevance. Vercel, which watches this behavior across its developer base, put it plainly: when a task does not need the best model, teams are beginning to route it to the cheapest one that is good enough, and the recent wave of models coming out of China is winning that trade. Price, in Vercel's words, is doing the work.
The individual anecdotes are just as striking as the aggregate numbers. One developer described paying about ten dollars an hour on Claude versus under fifty cents on DeepSeek for comparable work. Z.ai's GLM 5.2, released to fanfare in June, saw the fastest adoption of any model Vercel tracked in 2026, with daily token volume growing roughly 27 times in its first full week. By some measures GLM 5.2 performs within a percentage point of Claude Opus 4.8 at about a fifth of the cost. That is the kind of ratio that changes budgets, not just benchmarks.
The Companies Already Making the Switch
This is not a theoretical migration. Lindy, a San Francisco AI startup, moved 100 percent of its traffic from Claude to DeepSeek, a switch the company says will save it millions. Founder Flo Crivello captured the mood with characteristic bluntness, arguing that you do not need God to write your email, and that if you can get lower tiers of intelligence for a tenth of the price, it would be foolish not to do it. That sentiment, unsentimental and cost driven, is spreading through engineering organizations that are under pressure to show AI returns.
Larger names are circling too. Microsoft has been exploring DeepSeek as a lower cost alternative for parts of Copilot, and Anysphere, the company behind the coding tool Cursor, has used open models such as Qwen and Kimi in building its infrastructure. Many enterprises hedge by accessing these models through American cloud providers rather than calling Chinese endpoints directly, which softens some concerns while keeping the cost advantage intact. The pattern is consistent: keep the expensive models for the hardest work, and offload the rest.
Washington Starts Asking Questions
The political system has noticed. Lawmakers have opened investigations into Airbnb and Anysphere after the companies disclosed using Chinese open models in their AI infrastructure, a sign that what began as an engineering optimization is now a policy flashpoint. The concerns cluster around data security, potential censorship or bias baked into model behavior, and the broader discomfort of American commercial software leaning on Chinese research at a moment of intense geopolitical rivalry.
We would caution against treating this as simple protectionism. The underlying question is legitimate: what does it mean for critical business logic and sensitive data to flow through models whose training and governance sit outside the reach of US oversight. But the scrutiny also collides with a market reality that regulation cannot easily wish away. When the price difference is an order of magnitude, prohibition tends to push usage into less visible channels rather than eliminate it, and blunt restrictions risk penalizing the very companies being transparent about what they run.
The Calculus for Regulated Buyers
For enterprises in regulated sectors, the decision is genuinely hard. A bank, hospital, or defense contractor cannot treat a 90 percent cost saving as a free lunch when data residency, auditability, and supply chain provenance are contractual and legal obligations. For these buyers, geopolitical and data security risk still deters direct adoption, and rightly so. The frontier US models retain a real advantage that has nothing to do with benchmarks: a governance and compliance story that a procurement committee can defend.
The pragmatic middle path is emerging around routing and segmentation. Organizations are learning to classify workloads by sensitivity and criticality, sending low risk, high volume tasks to cheap open models while reserving proprietary frontier systems for anything touching regulated data or core decision making. Analysts at Brookings estimate Chinese models trail US rivals by only six to nine months, which means this triage will only get more tempting over time. The skill enterprises need is less about picking a winner and more about building the plumbing to switch.
What This Means for the AI Vendor Hierarchy
The strategic takeaway is that the ground under the AI vendor hierarchy has shifted from capability to cost efficiency. For two years the pitch was that the best model wins, and buyers paid accordingly. That story assumed the best model was also the necessary one for most work, and the token data now says otherwise. Once agentic capability becomes a baseline expectation at every price tier, the differentiator moves to how cheaply and reliably a model can do the ordinary work, and that is a game commoditized open weights can play.
None of this means OpenAI and Anthropic are in trouble. Their frontier lead, enterprise trust, and compliance posture remain formidable, and the hardest, highest value workloads still flow to them. But the comfortable assumption that American labs would capture the entire enterprise inference market is gone. The winners over the next year will be the enterprises that treat models as interchangeable inputs to be routed on price and risk, and the vendors, wherever they are based, that make that routing painless.
