A new piece of data that deserves more attention than it has received: in the week ending April 26, China's top nine AI models processed 4.37 trillion tokens on OpenRouter, against 4.98 trillion for US models. That gap is close. Earlier this month it briefly inverted: in the week ending April 5, Chinese models processed 12.96 trillion tokens to US models' 3.03 trillion, a record that was driven partly by a surge in usage following DeepSeek's breakthrough earlier in the year. The figures come from Jefferies, the brokerage, which has been tracking AI usage volumes as a proxy for the underlying capability and commercial momentum of each country's models.
Token throughput is not a perfect measure of anything. Different models charge different prices per token, handle different use cases, and attract different user populations. A trillion tokens of basic code generation is not the same as a trillion tokens of complex reasoning. But volume is a reasonable first approximation of real-world adoption, and the convergence of Chinese and US volumes on the most widely used AI routing platform is not the picture most Western observers have been working from.
The price story is equally significant, and points in the opposite direction from "Chinese models are just cheap substitutes." A year ago, Chinese models charged roughly 5 percent of what US models charged per token. That figure has climbed to at least 17 percent, according to Morgan Stanley analysts, who note the price rise as a consequence of growing user demand. Chinese AI providers are, in other words, beginning to discover pricing power. The race-to-zero model that initially drove adoption is giving way to something more sustainable, which in turn makes these businesses more attractive to investors.
That investor interest is moving from private to public markets. Knowledge Atlas Technology, the company behind Zhipu AI, and MiniMax both listed in Hong Kong in January and have seen their shares surge significantly. Both are expected to join the Hang Seng Tech Index on June 8, an event Morgan Stanley analysts predict will generate between $1.25 billion and $1.75 billion in passive inflows as index funds rebalance. The analysts raised price targets on both stocks substantially and predict that each frontier Chinese AI model can achieve at least $1 billion in annual revenue this year, with more than double that in 2026.
The products themselves are differentiated in ways that matter. Zhipu's models have become known for coding capabilities; MiniMax has built a broader range of abilities spanning text and audio generation, and has been adopted by users of OpenClaw, an AI agent platform, partly because Chinese models remain cheaper than US alternatives even as prices rise. Moonshot, which operates the Kimi model, and StepFun remain privately held and are not yet publicly traded, suggesting the IPO wave has further to run.
The backdrop to all of this is China's longer-term trajectory in digital services exports: the country is the world's sixth-largest exporter of digitally-delivered services, with exports growing at 15.2 percent annually over the past two decades to reach $245 billion. AI models are a natural extension of that export capacity. Unlike physical goods, AI inference can cross borders without touching a port or a tariff. That characteristic makes AI-driven digital exports both harder to restrict and harder to measure, which is part of why token volume data on neutral routing platforms like OpenRouter has become a meaningful data source.
The narrative that has dominated coverage of China's AI capabilities since DeepSeek's debut has focused on efficiency: doing more with constrained compute. What this week's data suggests is something simpler: Chinese models are now being used, at scale, by real users, generating real revenue, and attracting mainstream institutional capital. The capability gap may be narrowing; the commercial gap, for a moment, has essentially closed.