Tokens, the most valuable asset in the worldwide AI sector, are where China is making considerable strides.
Since February, Chinese AI models, developed by entities like DeepSeek and MiniMax, have surpassed their US counterparts in token utilization, based on OpenRouter data, which monitors these segments of text, code, or data processed by extensive language models.
This transformation indicates a fundamental alteration in the AI competition. Nvidia’s Jensen Huang asserted this month that the creation and deployment of these digital elements will propel the AI economy. Given that developers are billed for each token, this also functions as an indicator of model uptake and a fiercely competitive pricing arena among AI firms.
The capacity to generate tokens affordably is redefining global competition and granting China a fresh advantage. This is especially true since AI agents, for instance, those constructed on the OpenClaw open-source platform, process considerably more tokens than their predecessor chatbots.
Will Liang, CEO of Amplify AI Group, a Sydney-based tech consulting firm, commented, “If your agent expends millions of tokens daily, even a minor per-token price variation turns into a substantial cost factor.” He added, “This presents an inherent advantage for Chinese laboratories, which only intensifies as the use of agentic systems expands.”
The cost advantage held by Chinese AI entities originates from lower-cost power and more optimized models. This enables firms like MiniMax and Moonshot to price their output tokens at $2 to $3 per million, a stark contrast to Anthropic’s Claude Sonnet 4.5, which costs around $15 – an almost six-times difference.
This disparity becomes particularly apparent with AI agents, which process considerably greater quantities of tokens than traditional chatbots. For instance, while a chatbot might use approximately 30,000 tokens to condense Shakespeare’s Hamlet, an AI agent could demand up to 20 million for a trivial coding assignment.
This development is altering how AI developers opt to allocate their funds. Terry Zhang, a developer situated in Hong Kong, mentioned that he now dedicates around $50 daily to Moonshot’s Kimi model for roughly 80 percent of his tasks, saving Anthropic’s Claude for more intricate assignments.
He elaborated, “I previously only invoked Claude, but with a growing volume of tasks, relying solely on Claude would incur expenses of approximately $900 daily.” He concluded, “That’s excessive, and the combined deployment of Kimi and Claude functions effectively for me.”
This trend is influencing revenues. MiniMax, whose M2.5 model is currently classified among the most widely adopted globally based on token utilization, experienced a 476 percent surge in token usage from the previous month as of March 20, as reported by OpenRouter.
Although OpenRouter constitutes merely a small portion of worldwide model utilization, it is commonly employed as an industry benchmark, primarily because comparable data is limited elsewhere.
US entities continue to expand swiftly as the entire sector broadens, with OpenAI, Anthropic, and Google all reporting substantial income increases and uptake. However, more economical Chinese models have secured an opportunity to advance among users globally.
China’s cost benefit in token pricing partly originates from the nation’s extensive commitment to renewable energy. This month, the Chinese government identified “computing-electricity synergy” as a key objective in its 2026 work report, thereby clearly connecting energy policy with AI rivalry.
Regarding software, Chinese entities have adopted optimized AI frameworks, like “mixture-of-experts” designs, which diminish processing requirements, though occasionally sacrificing accuracy. This impetus for computing streamlining has been motivated by a scarcity of advanced chips within China, a consequence of US export controls.
Technological limitations exist. Zhipu AI’s GLM-5 model momentarily led OpenRouter charts in February until its utilization escalated beyond its processing capabilities, resulting in hold-ups and diminished service quality.
The company, compelled to issue an apology and increase costs, experienced a 22 percent stock decline on that day, effectively wiping out over $10 billion in market capitalization.
One seasoned developer at Google commented, “The model’s prowess is crucial, but consistent processing power and service are equally essential.” Google’s Gemini 3 Flash is positioned second among the top five widely utilized models this month, falling behind Minimax.
China’s technology behemoths have acted swiftly to capitalize on their lead. Earlier this month, Alibaba unveiled the establishment of Alibaba Token Hub, a fresh corporate division to be spearheaded by chief executive Eddie Wu. This unit suggests Alibaba’s perspective that token economics will characterize the subsequent phase of AI competition.
“We are on the verge of an AGI pivotal moment,” Wu penned in an internal memorandum last week. He added, “Billions of AI agents are ready to assume a progressively larger portion of digital work, each driven by tokens produced by models, and these agents will progressively serve as the main intermediary between individuals and the digital realm.”
It is uncertain whether China’s token advantage can endure, particularly since some enterprises remain cautious about depending on models operated from Chinese data centers.
Amplify’s Liang stated, “The international challenges are substantial, especially pertinent to governments and controlled sectors.” He continued, “Regulators are scrutinizing more closely where data is handled and under whose authority it falls.”
Data representation courtesy of Haohsiang Ko in Hong Kong