Chinese Open-Weight Models Triple Their Enterprise Token Share in Two Months
The share of enterprise AI workloads running on open-weight models, the majority of which originate from Chinese companies, has nearly tripled in two months, according to new data from Vercel’s AI Gateway Production Index. In June 2026, open-weight models processed 29% of all tokens routed through Vercel’s production gateway, up from 11% in April. Despite handling close to a third of all traffic, these models accounted for less than 4% of total spending, a ratio that captures the core commercial logic driving their adoption: they are available at roughly one-tenth of the average token price charged by US frontier model providers.
DeepSeek Reaches Third Place on Vercel Gateway; GLM-5.2 Breaks In Within Two Weeks
The data, based on tens of trillions of tokens flowing monthly between enterprise applications and AI model providers, offers one of the clearest pictures yet of how Chinese AI models are penetrating production workloads in Western companies. DeepSeek emerged as the single biggest beneficiary of the shift, accounting for 22.6% of token volume on the Vercel gateway in June, placing it third overall, just 1.8 percentage points behind Google, whose share slipped to 24%.
Z.ai’s GLM-5.2, an MIT-licensed model designed for agentic tasks and priced at roughly one-fifth the cost of Anthropic’s Opus 4.8, broke into the platform’s highest-volume models within two weeks of its release. Approximately one in eight of Vercel’s enterprise customers now runs an open-weight model in production.
Harpreet Arora, Vercel’s head of agentic infrastructure, described the dynamic in straightforward terms: “Price is doing the work here. 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.” He added that data residency and privacy protections remain key considerations for customers evaluating open-weight models for production deployment, a caveat that has so far slowed adoption in regulated industries but has not reversed the overall trend.
US Frontier Labs Still Capture 95% of Spending Despite Volume Losses
The revenue picture tells a different story from the volume numbers. The four leading US frontier AI companies, Anthropic, Google, OpenAI, and one other, captured 95% of total spending through the Vercel gateway in June. Anthropic alone accounted for 61% of spending despite processing only 32% of tokens, and more than 72% of spending in the highest-stakes use cases: coding agents, back-office automation, and application generation. The pattern that emerges is a bifurcated market in which high-volume, low-stakes work migrates to cheap Chinese open-weight models while high-risk, high-value work remains on expensive US frontier models.
OpenRouter Data: Chinese Models Now Process Three Times More Tokens Than US Models
A parallel dataset from OpenRouter, the largest neutral model router, shows an even more pronounced shift. US models from Google, OpenAI, and Anthropic held approximately 70% of token share on OpenRouter in June 2025. A year later, that figure has fallen to roughly 30%. DeepSeek is now the single largest provider on OpenRouter by token volume at 16.3%, ahead of all US competitors. Chinese providers collectively, including DeepSeek, Tencent, Xiaomi, MiniMax, and Qwen, account for approximately 44% of the top ten providers’ token volume.
On a weekly basis, Chinese models process around 18 trillion tokens, compared with roughly 5.5 trillion for US models, a ratio of more than 3:1. The tipping point reportedly came in the week of February 9–15, 2026, when Chinese systems processed more tokens than American ones for the first time.
Video Generation: ByteDance Seedance Dominates Premium Spending
In video generation, the Chinese advantage is even more concentrated. ByteDance’s Seedance model accounted for nearly half of all video-related spending on the Vercel gateway despite generating approximately one-third of videos, a premium positioning that reflects the model’s quality reputation relative to its cost. ByteDance’s broader AI strategy has consistently prioritized deployment scale over frontier model competition, and the Seedance data suggests that approach is translating into a durable revenue share in specialized AI categories.
The Vercel and OpenRouter figures come with an important caveat: both platforms capture only the traffic that flows through their own routers, which represents a fraction of global AI usage. The overwhelming majority of enterprise AI inference happens directly through the APIs of Anthropic, OpenAI, and Google, or through cloud platforms such as AWS Bedrock, Google Vertex AI, and Microsoft Azure, where US models dominate.
Google alone processes volumes roughly 35 to 40 times what flows through OpenRouter each month. The shift toward Chinese open-weight models on neutral marketplaces is real and measurable, but the US frontier labs retain a commanding position in the overall market, particularly in the enterprise segments where accuracy, compliance, and vendor accountability matter most.