Author | Hu Qingwen Editor | Xu Xiaofei
This time last year, Silicon Valley was still debating whether Chinese models could compete. This year, however, a set of data has completely overshadowed this question.
The latest data from OpenRouter shows that the share of Chinese models in the weekly token usage of U.S. enterprises surged from 4.5% in the first half of 2025 to consistently stabilizing above 30% week by week since early 2026, with a peak of 46%.
This is not a fleeting spike driven by a single hit model, but a growth curve that has not looked back since February 9.
What is even more intriguing is that the protagonists of this "defection" are not just Silicon Valley startups, but also major allies of OpenAI like Microsoft.
Pricing, billing, and a collective shift across the Atlantic are turning Chinese models from a purely technical topic into an inescapable arithmetic problem for foreign enterprises.
01From 4.5% to 46%: A Curve No One Expected
If the biggest consensus in the AI industry over the past two years was that U.S. models win on quality while Chinese models win on cost, then in 2026, this consensus is being rewritten as Chinese models are winning on usage volume.
"The biggest change this year is not that Chinese models have comprehensively surpassed U.S. ones, but that large language models are entering a phase of full commoditization," defined Zhu Yijun, head of AI products, characterizing this growth curve.
He believes that past debates focused on whose model was smarter and had stronger reasoning capabilities, but now the logic of the entire industry has shifted. How to accomplish the vast majority of tasks at a low cost has become the top priority for enterprises when adopting models.
According to the latest statistics from OpenRouter, the weekly token share of Chinese models invoked by U.S. enterprises was only 4.5% in the first half of 2025, with an average of just 11% over the past 12 months. However, starting from the week of February 8, 2026, this figure has never dropped below 30%, peaking at 46% mid-year.
Notably, this growth curve has emerged against the backdrop of the entire market expanding explosively.
The total weekly processing volume on the OpenRouter platform itself expanded from about 5 trillion tokens in April 2025 to over 20 trillion in April 2026. In other words, Chinese models are not fighting for share in a shrinking pie; rather, they are taking an increasingly larger slice of a rapidly expanding pie.
In this regard, Zhu Yijun also pointed out an easily overlooked transmission path. "OpenRouter is inherently a developer platform. From an industry perspective, developers have always migrated to large models with a price advantage first, followed by enterprises. In other words, the current curve is likely just a prelude to changes in the procurement structure of B2B enterprises."
Meanwhile, data from Vercel AI Gateway proves that the speed at which Chinese models are entering the global developer ecosystem is accelerating.
According to platform statistics, DeepSeek's token usage share rose from less than 1% to 17% within a single month in May 2026, making it the third-largest token source on the platform. Meanwhile, Zhipu's GLM-5.2, launched in June, saw its daily token volume grow by approximately 50 times in just two weeks, quickly climbing to the top of the Vercel AI Gateway model invocation rankings.
Even within the U.S. AI investment community, GLM-5.2 has sparked rare attention.
Marc Andreessen, co-founder of A16Z, posted on X: "Multiple AI industry practitioners judge that GLM-5.2 is the first Chinese AI model capable of matching, or even beating, publicly available models from top U.S. labs across multiple tasks, without making obvious compromises."
From Silicon Valley to Europe, numerous overseas enterprises are adopting Chinese AI models to break free from their sole reliance on model products from leading U.S. AI companies.
Reportedly, multiple enterprises including DoorDash, Siemens, and Airbnb have already deployed Chinese AI model products. Chinese large models are favored by overseas enterprises due to their lower prices, steadily improving performance, and support for on-premises deployment and operations.
German industrial giant Siemens stated that the company currently uses a mix of multiple AI model products internally, which includes large models from Chinese companies such as DeepSeek and Zhipu.
"The procurement logic itself is changing, with more and more enterprises leaning towards privatized or hybrid cloud deployments," Zhu Yijun noted. He believes that some European enterprises highly value data security and privacy protection and are also concerned about being locked into a single U.S. or European vendor, so they naturally prioritize Chinese AI.
In a recent report by the Financial Times, numerous European enterprises have stated that the previous export control policies introduced by the U.S. targeting Anthropic's large models prompted them to reduce their reliance on U.S. AI tools. "Even if the supply of Fable models is reopened, the concerns in the minds of enterprises will be hard to eliminate."
02The "Collapsing" Bills of U.S. Enterprises Are Benefiting Chinese AI Models
When discussing the ultimate reason for this collective defection, several interviewees pointed to the same factor: price. The answer lies hidden in the increasingly thick billing statements of U.S. AI model Agent workflows.
In April this year, institutions including the Stanford Digital Economy Lab published a paper titled "How Do AI Agents Spend Your Money?" The empirical data in the paper shows that for the same coding task, an AI Agent automated workflow consumes an average of 4.17 million tokens, while a standard code Q&A consumes only 3,390 tokens—a difference of over 1,000 times.
The reason is that Agents need to continuously read context, invoke tools, and plan iteratively. What is truly expensive is not answering the question, but keeping the AI working continuously.
As of July 2026, the input token price for DeepSeek V4 Flash is $0.14 per million, while the quote for OpenAI GPT-5.5 is $5, with the price gap fluctuating between 4 and 100 times. OpenRouter stated that the prices of open-source Chinese models are generally 60% to 90% cheaper than the flagship products of Anthropic and OpenAI.
When such a price gap is distributed across a task that requires hundreds of model invocations to complete, the magnitude of the bill is completely amplified.
Under such billing pressure, Microsoft confirmed to Axios that it is exploring the use of a fine-tuned version of DeepSeek V4 deployed on Azure as a low-cost model option for Copilot Cowork, to replace some workloads currently handled by OpenAI and Anthropic models, while also shifting the product from a fixed fee to a pay-as-you-go billing model.
Zhu Yijun interprets this series of moves as Microsoft entering the "multi-model strategy era," meaning the relationship between OpenAI and Microsoft is becoming non-exclusive.
"What Microsoft is selling now is not just GPT; Copilot, Microsoft 365, and GitHub are the core. The model is merely the underlying output capability. Especially since some Chinese large models are now good enough, stable enough, and highly cost-effective, they can all be invoked. Behind this is a shift in corporate mindset."
Shen Wei, an investor long focused on the tech sector, sees another side of the same trend from Microsoft's moves: "Future enterprise clients will definitely use a mix of multiple models, combining cheap and advanced ones." In his view, this "mixed usage" is transforming from an expedient measure for individual enterprises into a default configuration.
Voting with their feet even earlier than Microsoft is a batch of small and medium-sized companies.
San Francisco startup Lindy has migrated all its AI Agents from Anthropic to DeepSeek V4. The founder stated that this move saved the company millions of dollars and improved many core performance metrics.
Airbnb CEO Brian Chesky once publicly stated that the company relies heavily on Alibaba's Qwen model, primarily for large-scale production scenarios, because Qwen is not only effective enough but also faster and more cost-effective.
However, Shen Wei warned that the other side of this shift is a price war from which no one can emerge unscathed. "Global large models have entered a fiercely intensified and even deteriorating melee; the price war has markedly accelerated over the past few months."
He observed that this is not a unilateral price cut by Chinese models, but a collective involvement of top players. OpenAI leads, Anthropic follows, and both have recently introduced incremental benefits to heavy users, which is essentially a disguised price cut; Meta has also joined the battle for high volume at low prices. This is commercial competition: once one party cuts prices first, the rest of the manufacturers must follow suit.
Shen Wei's view is straightforward: "This brutal battle is intensifying irreversibly, and it is hard to say that any company's business model is healthy."
In other words, the "collapse" of U.S. model bills, rather than being a unilateral victory for Chinese models, is more like the entire industry being dragged together into this price competition. And Chinese manufacturers just happen to be standing on the more advantageous side.
03Going Global: Which Scenarios Have Chinese Models Conquered First?
"Although our models have caught up in terms of usage volume, an undeniable fact is that Chinese models have currently captured mainly the 'middle layer' rather than the 'tip of the pyramid'."
Sun Hao, an analyst who has long tracked the AI sector, told Leiphone that although the gap in coding Agent benchmarks between Chinese AI models and leading U.S. AI models has narrowed to the single-digit percentage level, the gap in frontier models has not yet been bridged in areas such as multimodal understanding, ultra-long context consistency, and highly sensitive scenarios involving strict regulation.
"What Chinese models have captured are high-traffic, low-sensitivity tasks, such as batch classification, content drafting, and code scaffolding, where the cost of data leakage is not catastrophic. This is a substitution of 'good enough' for 'the optimal,' and cannot be seen as a comprehensive frontal defeat."
And how far this path can go depends on a more pointed question: Can the market share exchanged for low prices be turned into hard cash profits?
Sun Hao believes that Chinese models are currently experiencing a subtle phase: they have won on token usage, but their bills have not grown proportionally. The low-price strategy has helped them quickly enter developer and enterprise workflows, but the commercial value behind the massive token invocation volume has not been fully released.
"Chinese models are winning on usage, but have not yet won on profits." Regarding the paradoxical situation of soaring token consumption without matching revenue, Zhu Yijun provided a specific "pathological analysis":
"First, model pricing is too low, which is essentially a multiplication problem; second, open-source dilutes inference revenue, and large model companies can only reap limited licensing revenue; third, there is a lack of upper-layer applications. Taking GPT as an example, the model is just the underlying foundation, and what truly generates revenue is still the upper-layer applications; fourth, enterprise software revenue is insufficient. Chinese large model companies mainly rely on API invocation fees, capturing only a meager share in the value distribution of the industry chain."
Investor Shen Wei places this issue in a larger coordinate system: "This is not a problem unique to Chinese models. The current business of large models relying on selling tokens has not yet run through a complete commercial closed loop, and the industry's business model is still under continuous exploration."
He pointed out that AI is more like electricity than the internet. Model API and token billing are the first layer of revenue, AI subscriptions are the second, enterprise customized software markups are the third, Agent automated workflows are the fourth, and vertical industry solutions are the fifth.
"There is no single answer to how AI ultimately makes money; rather, it is a layered profit combination."
At the same time, Shen Wei provided several core observation indicators to judge the trend of this industry battle: whether Agents can move from demo prototypes to real business implementation, whether enterprise clients can move from pilots to renewals and expansions, whether the token increment corresponds to real revenue growth, and the dynamic changes in inference costs and gross margins.
"Only tokens consumed in high-value business scenarios possess true commercial value."
A year ago, the core question facing Chinese models was whether the performance gap could be closed.
A year later, the core industry question has shifted to: when the performance gap continues to narrow and the price advantage is fully realized, who will lead in establishing a brand-new AI commercial order.
The watershed of the AI industry may not lie in who possesses the most tokens, but in who can create the highest industrial value from massive token consumption.