AI系列报告#33：代码生成与智能体驱动下的中美同行ARR增长

China (PRC) | Internet AI Series #33: Coding and Agents Drive ARR inChina and US Peers Highlights (1)OpenRouter: (week of 30 Mar), overall token consumptionincreased by 19 % vs prior week; (2) Fast ARR growth in China and US peerson surging tokens demands on coding and agents, as seen from latest ARRfrom Anthropic and AI Labs; (3)Volcano Engine: daily token consumptions In this report, we provide Investors with over 50 data points on AI models, OpenClaw, user trendsby sub-sector, and other industry data for analysis. OpenRouter: Token consumption was up by 19% vs prior week.Based on the analysis ofthe week of 30 Mar (vs the week of 23 Mar) (Exhibit 1), number of weekly tokens consumedincreased by 19% to 27tn. For China models, number of tokens increased by 31.5% to 12.96tn,surpassing that of US models at 3tn (up 0.76%). In terms of ranking, BABA's Qwen3.6 Plus (free)ranked no. 1 (new in rankings, 4.6tn tokens), followed by Xiaomi's MiMo V2 Pro (ranking down Fast growth is seen in ARR in China and US peers on surging tokens demands on codingand agents.On 6 Apr, Anthropic announced ARR surpassed USD30bn, up from about USD9bnend of 2025. This is driven by acceleration in Claude customers which increases from over 500(each spending annualized over USD1m) in Feb to over 1000 customers in less than 2 months.In China market, (1)MiniMax: (a) ARR increased by 50% from USD100m in Dec-25 to USD150min Feb-26; (b) the average daily token consumption of M2 series text models grew to over 6x inFeb-26 vs Dec-25, among which token consumption from Coding Plan increased by over 10x; Source: Visible Alpha, Jefferies Volcano Engine: Daily token consumption surpassed 120tn in Mar-26.According to media(e.g. The Paper), it represents the doubling of 63tn tokens in Dec-25. Number of enterpriseswith cumulative tokens (over 1tn) reaches 140, up from 100 end of 2025. This is mainly drivenby AI video generation (e.g. Seedance 2.0) and Agents (e.g. ArkClaw) adoption. (Exhibit 73) Qwen3.6-Plus makes further breakthrough and advances coding and agentic capabilities.BABA released Qwen3.6-Plus last week which has top tier performance in coding andreasoning as well as 1m contexts for repository level tasks. It has set a new frontier in agenticAI. Highlights (1) In agentic coding, Qwen3.6-Plus ranked no. 2 after Claude 4.5 Opus in SWE-bench Pro / Verified / Multilingual; (2) Qwen3.6 Plus (free) Ranked no. 1 in OpenRouter in termstoken consumption and is the first one to break 1tn tokens in a single day; (3) input/output Continued overleaf... Thomas Chong * | Equity Analyst852 3743 8016 | thomas.chong@jefferies.com Zoey Zong * | Equity Analyst852 3743 8163 | zoey.zong@jefferies.com BAT, KC, and AI Labs are Beneficiaries.We expect surge in token consumption on rising demandin AI models and agents to benefit CSPs likeBaidu, Alibaba, Tencent (BAT), Kingsoft Cloud (KC),and AI Labs.According to analysis from Artificial Analysis, (1) there is a narrowing intelligence gapbetween US and China models; (2) API cost is just a fraction vs US benefiting from MoE, linear China vs US comparison AI Models According to OpenRouter, rising adoption ofOpenClaw and release of high-performanceand cost-efficient China models drive token According to OpenRouter, token consumption went up significantly in Feb-26 (a)surge in adoptionof open-source OpenClaw, which allows people using AI from "to chat" to "to work". It connects withlarge models which can be deployed locally or on servers which allows you to automate workflowthrough natural language chat; (b)release a number of high performance and cost-efficient model According to OpenRouter (between 30 Mar and 5 Apr), China models token consumption increasedby 31.5% compared to prior week reaching 12.96tn. This compares to US models at 3tn token China models surpass US models again interms of token consumption the week of 30 On token consumptions by company, Qwen (20.8%) comes first, followed by Xiaomi (16.6%), Google(10.9%), Anthropic (10.4%) and MiniMax (7.8%). On the other hand, top 5 weekly token consumptionby models include Qwen3.6 Plus (free) (4.6tn), followed by MiMo-V2-Pro (3.08tn) and Qwen3.5Flash (free) (1.64tn).Qwen3.6 Plus (free) ranked first in terms oftoken consumption the week of 30 Mar Based on Artificial Analysis, China models DeepSeek, Qwen, Kimi, MiniMax and Zhipu are closingthe gap with US models in terms of intelligence. China models are highly cost-efficient on Mixtureof Experts (MoE) architecture, linear attention, and MLU (Model Flog Utilization). China models are narrowing the gap with USmodels in terms of intelligence. On Input/Output API, China large models are highly In terms of Input/Output API price, China models are just a fraction vs US models. OpenClaw Pricing Plans for Different OpenClaw Offerings byTencent, Alibaba, Baidu, and MiniMax PricingPlansfordifferentOpenClawOfferings by Tencent, Alibaba, Baidu, and According to OpenRouter, top 3 models for OpenClaw b