Open-Weight Models from China Are Capturing a Growing Share of AI Usage

Two of the most important data sources on real-world AI usage in production are currently painting the same picture: open AI models, whose weights are freely available, are taking a rapidly growing share of tokens processed worldwide — and most of these models come from China. That’s what both the latest Production Index from the Vercel AI Gateway and usage data from the model platform OpenRouter show. Both datasets, however, come with an important caveat: they only capture a slice of the market.

Vercel AI Gateway: 29 percent of tokens, but only 4 percent of spend

According to the Vercel AI Gateway’s Production Index, published in July and based on tens of trillions of tokens routed monthly between enterprise applications and AI providers, open-weight models already accounted for 29 percent of all tokens in June 2026 — up from 11 percent in April. That means the share has nearly tripled in just two months. The ratio between volume and cost is striking: open models handle nearly a third of tokens, yet account for less than 4 percent of spend.

The biggest driver is DeepSeek. The Chinese AI lab now accounts for 22.6 percent of token volume on the gateway, putting it in third place — less than two percentage points behind Google, whose share slipped to 24 percent. The newcomer GLM 5.2 from Chinese provider Z.ai, an MIT-licensed model aimed at agentic tasks priced at roughly a fifth of Anthropic’s Opus 4.8, broke into the platform’s highest-volume models within two weeks of release. Roughly one in eight of Vercel’s enterprise customers now runs an open-weight model in production.

Spend tells a very different story: the four major US frontier labs together captured 95 percent of gateway spend in June. Anthropic alone took 61 percent of spend on 32 percent of tokens — and more than 72 percent of spend in the highest-stakes use cases, such as coding agents, back-office agents, and app generation. The pattern that emerges: high-volume, low-stakes work moves to cheap open models, while high-risk work stays on the expensive frontier models.

OpenRouter: US models collapse from 70 to roughly 30 percent

The shift is even more pronounced on OpenRouter, the largest neutral model router, which lets developers flexibly distribute API calls across dozens of models. According to analyses of OpenRouter usage data, reported among others by Bloomberg based on figures from Exponential View, US models from Google, OpenAI, and Anthropic still held around 70 percent of token share on the platform in June 2025. A year later, in June 2026, that figure is down to roughly 30 percent.

DeepSeek is now the single largest provider on OpenRouter by token volume, at 16.3 percent — ahead of Google, Anthropic, and OpenAI. Chinese providers such as DeepSeek, Tencent, Xiaomi, MiniMax, and Qwen together account for roughly 44 percent of token volume among the top 10. On a weekly basis, Chinese models now process around 18 trillion tokens per week according to platform data, compared with about 5.5 trillion for US models — a ratio of more than three to one. 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.

A study published by OpenRouter together with venture firm Andreessen Horowitz, which analyzed more than 100 trillion tokens of anonymized usage data, confirms the trend: Chinese open-weight models grew their share from 1.2 percent to nearly 30 percent within a year. The main drivers are seen as price — DeepSeek’s current models are offered at a fraction of the cost of comparable US frontier models — and a shift in workloads toward coding, where Chinese models are considered particularly strong on price-performance. US policy also played a role: export controls that temporarily pulled Anthropic’s new Claude Fable 5 model from the market just days after its June launch, along with the initially restricted rollout of OpenAI’s GPT-5.6, pushed additional international demand toward freely available Chinese alternatives.

Why neither dataset shows the full picture

However telling the Vercel and OpenRouter numbers are for spotting trends, they capture only part of the AI market. Both platforms record exclusively the traffic that flows through their own routers. But a very large share of global AI usage happens directly through the APIs of Anthropic, OpenAI, or Google (Gemini) — without an intermediary, and therefore invisible to these statistics. On top of that comes the business of the cloud giants: enterprise customers procure AI models at scale through AWS (via Bedrock, for instance), Google Cloud (Vertex AI), or Microsoft Azure, where OpenAI’s models in particular run. The enormous volumes generated by consumer products like ChatGPT, Gemini, or Claude themselves don’t show up in router data either.

Router platforms like OpenRouter also attract a specific clientele: cost-conscious developers who can switch models quickly and compare price-performance on a daily basis. That segment reacts fastest to cheap open-weight alternatives — which likely inflates the apparent share of Chinese models in these statistics. Still, as an early indicator of where developer preferences move when there’s no vendor lock-in, the data remains valuable. And the direction is the same in both datasets: open models, particularly from China, are taking over the volume — while the US frontier labs, Anthropic foremost among them, keep collecting the money.

Putting it in perspective: router platforms are only a fraction of the overall market

Just how small a slice Vercel and OpenRouter capture becomes clear when looking at absolute token volumes. The Vercel AI Gateway itself speaks of “tens of trillions” of tokens per month. OpenRouter, by its own figures, most recently processed more than 20 trillion tokens per week, or roughly 80 to 90 trillion tokens per month — the study published with a16z covered more than 100 trillion tokens, though spread across roughly a year.

Compare that with the numbers from the major providers, to the extent they are known at all. Google cited the largest figure in the industry to date at I/O 2026 in May: over 3.2 quadrillion tokens per month across all of its own products — from Gemini to AI Overviews in Search to its cloud APIs — a sevenfold increase over the prior year’s 480 trillion. Google’s model APIs alone process around 19 billion tokens per minute according to the company, which extrapolates to more than 800 trillion tokens per month. OpenAI most recently cited more than 6 billion tokens per minute through its API alone at its DevDay in October 2025 — extrapolating to a good 260 trillion tokens per month, not even counting the enormous ChatGPT traffic, which now serves more than 800 million weekly users. Anthropic does not publish token figures; Barclays analysts, however, see the company clearly in the lead on enterprise inference tokens, and Anthropic has overtaken OpenAI in enterprise spending according to Ramp data.

The orders of magnitude speak for themselves: Google alone processes roughly 35 to 40 times what runs through OpenRouter each month — and a multiple of what flows through the Vercel gateway. Even OpenAI’s API volume from late 2025 exceeds both router platforms combined by several times over. The shift toward Chinese open-weight models on the neutral marketplaces is real and clearly measurable — but by far the largest share of global token traffic still flows through the infrastructure of the US giants, where these statistics simply can’t see.