Agent-to-Agent ecosystem accelerates, with domestic models showing strong usage and cost-performance advantages

Overview

This is a Jefferies China Internet AI series industry report that compares the Chinese and U.S. model ecosystems through data on agent collaboration, expansion of model use cases, model pricing, token consumption, AI application user growth, and the market size of MaaS/AI coding tools. Using sources such as OpenRouter, Artificial Analysis, QuestMobile, aicpb.com, Omdia, Grand View Research, F&S, and CAICT, the report argues that Chinese AI models and agent ecosystems are commercializing rapidly.

Core views

The core views are: first, Agent-to-Agent collaboration is becoming a new gateway for internet platforms and smart terminals, with Tencent strengthening Weixin scenarios through Yuanbao, Meituan Xiaomei, and partnerships with handset brands, while BABA is building a service ecosystem by opening Qwen to third-party Agents and Skills. Second, coding, Workspace, and medium- to long-form content scenarios are pushing AI from "chatting" toward "working," driving higher token consumption. Third, Chinese models lead in usage in scenarios such as OpenRouter and OpenClaw and have advantages in cost efficiency. Fourth, the enterprise market, private deployment, AI code generation, and agentic AI platforms may become clearer commercialization directions.

Analysis framework

The report adopts a combination of top-down market sizing and bottom-up data tracking: it uses OpenRouter and OpenClaw to observe token consumption and model rankings, Artificial Analysis to compare model intelligence, video generation, and API pricing, QuestMobile/aicpb.com to track DAU/MAU for AI applications and internet sub-sectors, and Omdia, F&S, CAICT, and Grand View Research to assess the market opportunities for agentic AI, AI coding tools, and the Chinese LLM market.

Methodology notes

• Usage trackingOpenRouter token consumption tracking

  Measures model adoption intensity through weekly token consumption, model rankings, and China-U.S. model usage comparisons.

  This framework is used to judge whether coding, agent, and workflow scenarios are truly driving growth in inference demand, and to identify the relative positioning of models such as DeepSeek, Tencent, and MiniMax.

• Model comparisonArtificial Analysis model intelligence and API pricing

  Compares model intelligence, input/output API prices, video generation model rankings, and cost efficiency.

  This framework supports the report's view that Chinese models are narrowing the intelligence gap while being more competitive on API costs.

• Application trafficQuestMobile/aicpb.com application traffic tracking

  Observes user trends in AI applications and internet sub-sectors through DAU, MAU, and month-over-month and year-over-year changes.

  This framework is used to verify changes in user activity for AI applications as well as internet businesses such as e-commerce, local services, music, video, travel, and social information.

• Market opportunityOmdia/F&S/CAICT/Grand View Research market size forecasts

  Uses CAGR, market size, and enterprise token share to estimate the market opportunity for agentic AI, AI coding tools, MaaS, and the Chinese LLM market.

  This framework is used to support the medium- to long-term investment thesis that the enterprise market, key-account customization, private deployment, and cloud service providers will benefit.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Tencent

  Beneficiary of the A2A ecosystem and the Weixin entry point

  Strengths

  Yuanbao is cooperating with Meituan Xiaomei and advancing A2A capabilities with handset brands such as OPPO, Xiaomi, Huawei, Vivo, and Honor, connecting local services, voice assistants, and Weixin communication scenarios.

  Weaknesses

  The report does not disclose direct revenue contribution, and monetization still depends on user experience, partner integration, and conversion within Weixin scenarios.

  Comparison

  Compared with Alibaba, which is more focused on the Agent platform and cloud model ecosystem, Tencent places greater emphasis on social, local services, and smart-device entry points.

  Risks

  Unstable A2A experience, regulatory and privacy constraints, insufficient partner coverage, or monetization falling short of expectations.

• Alibaba / BABA / Alibaba Cloud / Qwen

  Core beneficiary in models, agent ecosystem, and enterprise MaaS

  Strengths

  The Qwen app is open to third-party Agents and Skills, with the first batch including KFC, Luckin, Mixue, and China Eastern; Qwen3.7-Plus improves multimodal and coding capabilities; Alibaba Cloud ranked highest in 5 of Omdia's 7 metrics and reached a 32% share of enterprise tokens in 2H25.

  Weaknesses

  Consumer willingness to pay remains low, and long-term revenue may depend more on enterprise, key-account customization, and cloud service conversion.

  Comparison

  Compared with Doubao and DeepSeek, Alibaba has a more complete linkage across cloud, Agent platform, enterprise customers, and model ecosystem.

  Risks

  Model pricing competition, pressure on cloud gross margins, enterprise project delivery cycles, and competition from open-source models.

• Baidu / Alibaba / Tencent / Kingsoft Cloud

  Beneficiaries of cloud services and inference demand

  Strengths

  The report expects growth in token consumption and the launch of high-performance Chinese models to benefit CSPs, especially BAT and Kingsoft Cloud.

  Weaknesses

  Growth in cloud inference demand may come with compute capex and pricing competition, and may not necessarily translate into higher margins at the same time.

  Comparison

  Cloud service providers have infrastructure and customer entry points, making them more directly exposed to enterprise AI demand than pure application companies.

  Risks

  Declining inference unit prices, compute supply constraints, customer budget fluctuations, and substitution by self-built models.

• Kling

  Beneficiary of video generation and global AI applications

  Strengths

  Global users exceeded 100m in June 2026, enterprise customers reached 50K, and coverage extended to 224 countries and regions, indicating rapid expansion of video generation applications.

  Weaknesses

  The report does not provide payment rate, ARPU, or margin data, and video generation compute costs may be high.

  Comparison

  Competing with video models such as Veo, Seedance, and SkyReels, Kling's main highlights are its user scale and enterprise customer growth.

  Risks

  Intense competition in video generation models, high unit costs, and content compliance and copyright risks.

• Meitu / WPS

  Beneficiary of office design and e-commerce marketing AI scenarios

  Strengths

  WPS users can use Meitu's AI poster, text-to-image, image-to-image, and short-video generation features, especially for marketing materials used by e-commerce merchants.

  Weaknesses

  The partnership is still more scenario-access oriented, and the report does not quantify revenue, retention, or paid conversion.

  Comparison

  Compared with general-purpose agents, this combination is more focused on office, design, and e-commerce content production.

  Risks

  Feature homogenization, insufficient user willingness to pay, and platform revenue sharing or product experience falling short of expectations.

• Doubao / DeepSeek / MiniMax / Kimi / Zhipu

  Beneficiaries of domestic model and AI application ecosystems

  Strengths

  DeepSeek, MiniMax, and others rank highly in OpenRouter/OpenClaw usage; Doubao's April DAU reached 151m and grew 13% month over month; Kimi Work demonstrated local desktop agents, multi-step browser tasks, and multi-agent collaboration capabilities.

  Weaknesses

  Traffic and model rankings fluctuate significantly, and consumer monetization and enterprise-grade SLA still need validation.

  Comparison

  During the report's tracking period, Chinese models recorded higher token usage than U.S. models and were more attractive in cost efficiency.

  Risks

  Rapid technological iteration, competition from open-source models, declining API prices, and data compliance requirements.

• OpenAI / Anthropic

  Reference assets for overseas agents and vertical industry tools

  Strengths

  OpenAI launched financial-industry plugins in Codex and is developing legal-industry tools, while Anthropic previously launched financial services Agent templates, showing that leading U.S. model companies are also strengthening vertical-industry agents.

  Weaknesses

  The report points out that Chinese model API costs are relatively lower, and U.S. model costs may affect the scalability of some usage scenarios.

  Comparison

  U.S. models remain in the top tier in intelligence scores, but Chinese models are narrowing the gap and leading in some usage scenarios.

  Risks

  High valuations, pressure to deliver commercialization, and the pace of regulatory and industry customer adoption.

Key data

• OpenRouter total token consumption36.1tn, +13.5% week over weekWeek of 1 Jun versus week of 25 May.
• Chinese model usage on OpenRouter14.2tn, +27.5% week over weekFor the 1-7 Jun period; U.S. models were 3.2tn in the same period.
• OpenRouter model rankingsDeepSeek V4 Flash ranked No.1, followed by Tencent Hy3 preview and MiniMax M3The report believes high-performance, low-cost Chinese models are driving usage growth.
• OpenClaw model rankingsDeepSeek V4 Flash, MiniMax M3, and MiniMax M2.7 ranked in the top threeAll top three are Chinese models.
• Model intelligence gapTop U.S. models lead Chinese models by 2.7%From the 2026 Stanford AI Index Report Apr-26.
• Top models in Arena EloAnthropic 1503, xAI 1495, Google 1494, OpenAI 1481, Alibaba 1449, DeepSeek 1424As of Mar-26, Alibaba and DeepSeek have entered the first tier.
• Kling user scaleGlobal users exceeded 100m in June 2026, up 67% from 60m at the end of 2025Enterprise customers rose from 30K to 50K, covering 224 countries and regions.
• AI application DAUDoubao 151m; Qwen 29.2m; Yuanbao 9m; DeepSeek about 29mIn April, Doubao rose 13% month over month, while Qwen, Yuanbao, and DeepSeek also grew month over month.
• Omdia agentic AI marketUSD271m (2025) to USD9.7bn (2030), CAGR 105%Alibaba Cloud ranked highest in 5 of 7 metrics.
• Global AI coding tools marketExpected to grow to USD26bn in 2025-2030, about 3x 2025From Grand View Research.
• China AI code generation marketAbout USD4.7bn in 2028F&S expects key-account customized solutions to be a key revenue contributor.
• Enterprise token consumption37tn daily average in 2H25, +263% half-over-halfAlibaba Cloud share 32%, Doubao 21.3%, DeepSeek 18.4%.
• China LLM market2024-2030 CAGR about 64%, exceeding RMB100bn by 2030Enterprise and local deployment are expected to contribute more, while consumer willingness to pay is lower.
• Online shopping DAUIn April, JD +2% MoM / +9% YoY; Taobao -1% MoM / +6% YoY; PDD -2% MoM / +10% YoYPDD maintained relatively fast YoY growth, while VIPS was +3% MoM in April.
• Content entertainment trafficDouyin main app DAU was +2% MoM and +20% YoY in April; Soda Music MAU was +75% YoYMAU for long-video platforms declined year over year, while some TME apps faced pressure both month over month and year over year.

Impact & implications

In investment terms, the report views agents, coding tools, and enterprise MaaS as key incremental AI commercialization scenarios. Rising token consumption implies that cloud inference, model APIs, AI coding tools, and enterprise customized solutions may form clearer revenue pathways; if Chinese models continue to maintain high performance and low cost, it will benefit assets with model, cloud, application, or vertical-scenario entry points such as Alibaba Cloud, Tencent, Baidu, Kingsoft Cloud, AI labs, and Kling.

Risks

• Data from third-party platforms such as OpenRouter and OpenClaw cannot fully represent total market token consumption.
• Pricing competition in model APIs and Agent services may compress margins for cloud service providers and model companies.
• The implementation of agents in enterprise workflows may be constrained by accuracy, permissions, data security, and compliance.
• Growth in AI application DAU/MAU may not directly translate into paid revenue or profit.
• Growth in inference demand may require higher compute investment, and capex and chip supply may become constraints.
• China-U.S. regulation, data localization, export restrictions, and model compliance requirements may affect product iteration.
• Market size forecasts are highly sensitive to technology paths, enterprise budgets, and the macro environment.

What to watch

• Weekly token consumption on OpenRouter and OpenClaw, the share of Chinese models relative to U.S. models, and changes in top models.
• DAU/MAU, retention, and paid conversion for AI applications such as Qwen, Yuanbao, Doubao, and DeepSeek.
• Actual user experience and usage frequency of Tencent's A2A scenarios in Weixin, Meituan, and smartphones.
• Brand onboarding count, call volume, and monetization model in Alibaba Qwen's third-party Agent/Skill ecosystem.
• Price changes for model input/output APIs, coding tools, and Agent subscription packages.
• Enterprise token consumption share, especially changes in the shares of Alibaba Cloud, Doubao, and DeepSeek.
• Enterprise customer growth, payment rates, and unit compute costs for Kling and other video generation models.
• The intelligence gap and cost gap between Chinese and U.S. models in Artificial Analysis, Arena Elo, and other benchmarks.