Global large model competition is entering a new phase, with usage data pointing to a shift toward AI agents and a rising “token economy” reshaping pricing, infrastructure demand, and cloud business models.
Data for late March 2026 shows Chinese large models have led global usage rankings for a full month, highlighting a shift from model capability to application-scale demand.
China models lead token surge
OpenRouter data shows global usage reached 22.7 trillion tokens in March 23-29, up 11.2% week-over-week. Chinese models generated 9.857 trillion tokens, rising 33.94%, versus 3.007 trillion tokens for US models, up 1.79%. Chinese models have led for four consecutive weeks.
Six of the top nine global models were Chinese, generating 9.82 trillion tokens, up 33.4%. Models led by Anthropic and Google totaled 2.99 trillion tokens and declined.
The top four models were all Chinese: Xiaomi MiMo-V2-Pro at 3.96 trillion tokens, StepFun Step 3.5 Flash at 1.49 trillion, MiniMax M2.7 at 1.29 trillion, and DeepSeek V3.2 at 1.24 trillion. Token consumption has expanded since February, indicating accelerating demand.
Analysts cite OpenClaw (Lobster) as a key driver, contributing about 20% of token consumption at its peak and marking a shift from conversational AI to task-executing agents.
Z.ai CEO Zhang Peng said at the Zhongguancun Forum that moving from “chatting” to “doing” requires long-task planning, context compression, and multimodal processing, increasing token use per task by several to dozens of times and raising compute costs.
Agent boom reshapes cloud pricing
Rising demand is reshaping pricing. Z.ai, Alibaba Cloud (Aliyun), and Tencent Cloud have raised fees. Tokens are emerging as the core pricing unit, reflecting compute intensity and usage scale, with pricing shifting from discounting to cost-based premiums.
China’s lead reflects cost advantages and application scale. DeepSeek and MiniMax have lowered API prices to attract developers, while WeChat, Ding Talk, and Lark expand reach and drive long-tail demand. Open-source strategies further accelerate adoption versus closed overseas ecosystems.
According to Yicai, the OpenClaw-driven agent boom has triggered rapid product launches across Tencent, Alibaba, Baidu, JD.com, Youdao, Xiaomi, and Qihoo 360, expanding offerings to dozens within weeks.
Deployment speed has accelerated. Tencent WorkBuddy was prototyped in January and launched within days, later adding remote control and integration with WeChat Work and QQ. Product head Wang Shengjie said Tencent’s CodeBuddy infrastructure enabled rapid rollout.
Cloud vendors are expanding ecosystems. Tencent Cloud outlined a full agent stack spanning infrastructure, models, applications, and security, while opening Skills and SkillsHub with enterprise integrations.
The immediate impact is a surge in token consumption. Tencent said higher token use is driving compute demand and will support revenue growth. JD.com reported a 455% week-on-week increase in token calls, while IDC expects agent usage to grow tenfold by 2027, with token and API calls rising 1,000-fold.
OpenClaw is lowering barriers from conversational to execution-layer AI, accelerating enterprise adoption. AI-to-B is emerging as the largest growth driver for cloud vendors.
Agent applications are expanding from code generation to contract review, bid evaluation, video analysis, auditing, customer service, and drone inspection, replacing manual workflows.
Token economy drives costs, risks
Rising agent adoption is forcing cloud vendors to rethink monetization. Earlier price competition is giving way to recalibrated pricing as demand surges.
Alibaba has set up a Token Hub unit led by CEO Eddie Wu, targeting token generation and MaaS growth as a core revenue driver. ByteDance’s Volcengine has seen daily calls rise over 60% in under two months and raised its MaaS revenue target above CNY10 billion (approx. US$1.4 billion).
Tencent Cloud has upgraded its MaaS platform to TokenHub, signaling a shift toward token-centric management alongside SaaS and infrastructure services.
Executives warn that focusing solely on token consumption without improving efficiency could raise costs and drive churn. Token revenue lacks stickiness, making low-price strategies unsustainable.
Structural challenges are emerging. Nonlinear token demand is straining compute supply and raising costs for GPUs, memory, and data centers, while business models remain in flux.
OpenClaw is accelerating agent deployment and shaping the token economy. China’s lead reflects a shift from model competition to an agent-driven ecosystem where compute, cost control, and ecosystem integration will define the next phase.
Article edited by Jack Wu
