Chinese artificial intelligence models have accounted for more than 30% of weekly token usage by US companies since February 8, peaking at 46%, according to data from OpenRouter — a platform that lets developers route queries to dozens of AI models. The surge marks a dramatic reversal: over the prior 12 months the equivalent figure averaged just 11%, and as recently as the first half of 2025 it sat as low as 4.5%.
What Happened
OpenRouter data, tracked by CNBC and AI Weekly, shows that Chinese models — led by DeepSeek, Alibaba’s Qwen, and offerings from Baidu and Zhipu AI — crossed 30% of US-based weekly token consumption on February 8 and have stayed above that threshold every week since. Globally, Chinese models briefly overtook their US counterparts in total weekly token usage during the first two weeks of March, occupying the top three spots in overall consumption. The trend was confirmed separately by Rest of World, which documented US developer teams actively routing lower-stakes workloads to Chinese models to manage inference costs.
Cost is the primary driver. Open-source Chinese models can be deployed at prices “60% to 90% cheaper” than leading US alternatives, according to developers quoted by CNBC. As US AI pricing has remained elevated — with Claude Fable 5 now priced at $10 per million input tokens and $50 per million output tokens — cost-sensitive engineering teams are routing workloads to whichever model clears a quality threshold at the lowest price per token.
Why It Matters
The shift carries both economic and geopolitical dimensions. Economically, it demonstrates that the quality gap between frontier US models and Chinese alternatives has narrowed to the point where many enterprise workloads — summarisation, classification, code completion, customer support — can be served adequately by Chinese models at a fraction of the cost. This creates direct competitive pressure on US AI companies whose growth strategies depend on enterprises paying premium prices for premium capability. Anthropic recently crossed $30 billion in annualised revenue, but sustaining that trajectory becomes harder if a significant share of mid-tier inference volume migrates to cheaper open-source alternatives from China.
Geopolitically, the trend illustrates the limits of US export controls on shaping AI competition. Despite restrictions on Nvidia chip exports to China, Chinese labs have continued producing competitive models — including Tencent’s recently released Hy3, a 295-billion-parameter model launched under the open Apache 2.0 licence. The pace of Chinese releases has not slowed, and Meituan’s LongCat-2.0, a 1.6-trillion-parameter model built entirely on domestically produced Chinese chips, offers further evidence that China’s AI capabilities are advancing on its own semiconductor base.
Background and Context
The shift began accelerating after DeepSeek released its R1 reasoning model in January 2025, demonstrating performance competitive with OpenAI’s o1 at dramatically lower training cost. The release triggered a widespread reassessment in developer communities of the assumption that US models would maintain an unclosable quality advantage. Since then, Chinese labs have maintained an aggressive release cadence, with Alibaba’s Qwen 3 series, Baidu’s ERNIE 5, and multiple open-weight releases from smaller Chinese research groups expanding the available options continuously.
US AI companies have generally framed the competitive pressure as an incentive for faster capability development and investment. But the OpenRouter data suggests that the response on cost — where Chinese models hold a structural advantage due to lower training expenditure and open-weight deployment models — has not yet closed the gap. The result is a bifurcating market: US models for frontier tasks demanding the highest capability and lowest latency; Chinese models for commodity inference workloads where cost per token is the dominant variable.
What Comes Next
The question for US AI companies is whether they can defend market share in mid-tier inference, or whether that segment becomes effectively commoditised by Chinese open-source alternatives. Some companies are experimenting with tiered pricing strategies; others are emphasising data privacy, regulatory compliance, and enterprise support as differentiating factors that Chinese models cannot easily match. The trajectory, if sustained, would represent a meaningful structural challenge to the assumption that AI inference revenue will remain concentrated among a small number of US providers.
For enterprises, the shift offers a meaningful opportunity to reduce AI inference costs substantially. But it raises legitimate questions around data sovereignty, supply chain risk, and the auditability of safety practices at Chinese AI labs. How companies navigate those trade-offs — and whether governments introduce new frameworks to guide the decision — will be a defining thread in enterprise AI strategy through the remainder of the decade.
