China's Agent Era Begins: National Policy Framework, AI Terminal Standards, and the Trillion-Dollar Global AI Race

Something fundamental changed in China's AI industry during the first week of May 2026.

It wasn't a single model release. It wasn't a funding round. It was the simultaneous convergence of policy, standards, infrastructure, and global capital that signaled a structural shift: the age of AI agents has officially begun in China.

On May 8, three ministries — the Cyberspace Administration of China (CAC), the National Development and Reform Commission (NDRC), and the Ministry of Industry and Information Technology (MIIT) — jointly issued the "Implementation Opinions on the Standardized Application and Innovative Development of AI Agents" (《智能体规范应用与创新发展实施意见》). For the first time at the national level, China defined what an "AI agent" is, laid out 19 prioritized application scenarios, and established a governance framework that balances innovation with safety.

The same day, the GB/Z 177—2026 national standard series on "AI Terminal Intelligence Grading" was released, creating a unified L1–L4 classification system for smart devices from phones and TVs to glasses and headphones.

The same day, news broke that Anthropic is preparing a $50 billion fundraising round at a $900 billion pre-money valuation — potentially surpassing OpenAI's $85.2 billion to become the world's most valuable AI startup.

And the same day, China Mobile announced it would open a trillion-token service experience package to accelerate AI adoption across industries.

This wasn't a normal week. This was the week the agent economy became real.

Executive Summary: The Week Agents Went Mainstream

Source: Xinhua, CAC official website, Financial Times, 36Kr, People's Daily, China Mobile, MIIT announcements

What unifies these eight events is a single narrative: 2026 is the year AI transitions from chatbot to agent, from experiment to infrastructure, and from hype to revenue. China's policy framework gives the industry guardrails. The terminal standards give hardware makers a roadmap. And the global capital surge — Anthropic's potential trillion-dollar valuation — proves that investors now believe agent-based AI is the next computing platform.

1. The Three-Ministry Policy: China's First National Agent Framework

1.1 What the Policy Actually Says

The "Implementation Opinions on the Standardized Application and Innovative Development of AI Agents" is not a vague white paper. It is a 38-point operational framework with specific mandates, timelines, and accountability structures.

For the first time, a Chinese government document formally defines an AI agent:

"An intelligent agent is an intelligent system with autonomous perception, memory, decision-making, interaction, and execution capabilities, and is an important form of AI products and services."

— CAC/NDRC/MIIT Joint Document, May 8, 2026

This definition matters because it distinguishes agents from simple chatbots. An agent doesn't just respond to prompts — it perceives its environment, maintains memory across sessions, makes decisions autonomously, interacts with other systems, and executes actions in the real world.

Source: CAC official website, Xinhua, China News Service

1.2 The 19 Application Scenarios

The policy identifies 19 specific application scenarios where agents should be deployed first. This isn't aspirational — it's a procurement roadmap for government and enterprise buyers.

Source: CAC Policy Document, Section IV

The explicit inclusion of content creation, digital culture, and short video/dynamic manga as priority scenarios is notable. Chinese AI companies have already demonstrated 10x production efficiency gains and 60% cost reductions in these areas. The policy effectively sanctions and encourages scaling these proven use cases.

1.3 Governance Innovation: Classified Risk Management

Perhaps the most sophisticated element of the policy is its classified governance framework:

Source: CAC Policy Document, Section III.3

This tiered approach is designed to avoid regulatory overreach that would stifle innovation in low-risk consumer applications while maintaining strict control over domains where agent failures could cause physical or financial harm. It's a pragmatic balance that reflects lessons learned from the early, chaotic years of AI governance.

1.4 The "Smart Internet" Vision

The policy also introduces a forward-looking concept: the "Smart Internet" (智能互联网) — an evolution of today's internet architecture designed specifically for agent-to-agent communication.

Key components include:

- Agent registration platforms with digital identity management

- Agent Interconnect Protocol (AIP) as a national standard

- IPv6-based end-to-end agent communication

- Multi-agent collaboration frameworks with conflict resolution mechanisms

This is China's answer to a question the rest of the world hasn't fully asked yet: How do autonomous AI agents securely discover, authenticate, and transact with each other at scale?

2. AI Terminal Intelligence Standards: The L1–L4 Grading System

2.1 What GB/Z 177—2026 Establishes

On the same day as the agent policy, China's standardization authorities released the GB/Z 177—2026 series: "AI Terminal Intelligence Grading." This is a unified national standard that classifies all AI-powered consumer devices into four intelligence levels.

The standard uses a "2+N" architecture:

- "2" = Foundational framework (Part 1: Reference Architecture; Part 2: General Requirements)

- "N" = Device-specific standards for each product category

Source: MIIT Announcement, May 8, 2026

Note: L4 specifications are explicitly flagged for future revision as technology evolves — an acknowledgment that fully collaborative agent intelligence is still emerging.

2.2 The Seven Device Categories

The initial standard release covers 7 product categories, with more planned:

Source: MIIT Standard Release Document

2.3 Why This Standard Matters Globally

This standardization effort has three global implications:

First, it creates a de facto export requirement. Any AI device manufacturer wanting to sell in China — the world's largest consumer electronics market — will need to certify against these grades. This mirrors how China's early 5G standards became global standards through market scale.

Second, it harmonizes the fragmented AI device landscape. Today, every manufacturer uses proprietary terminology: "AI-powered," "smart assistant," "intelligent companion." The L1–L4 system gives consumers a clear, comparable framework.

Third, it accelerates the agent-hardware feedback loop. By explicitly defining L3 and L4 capabilities, the standard tells hardware makers exactly what compute, memory, and connectivity specs they need to build for agent-native devices.

3. The Geopolitical Layer: China-US AI Dialogue

3.1 The May Summit Signal

On May 8, China's Foreign Ministry spokesperson Lin Jian responded to Reuters reporting that China and the US are considering launching official AI dialogue talks, potentially during the mid-May Beijing summit between the two countries' leaders. US Treasury Secretary Scott Bessent would reportedly lead the American side.

Source: Foreign Ministry Press Conference, May 8; Reuters; World AI Conference 2025

3.2 Why AI Dialogue Matters Now

The timing is not coincidental. Both sides face structural pressures that make unilateral approaches increasingly costly:

For China: US export controls on advanced chips have accelerated domestic alternatives (Huawei Ascend, Cambricon), but the transition costs billions and slows frontier research. A dialogue channel could create predictability.

For the US: Chinese open-source models now dominate global API usage (60%+ of OpenRouter traffic). Attempting to contain China's AI industry through hardware controls alone is becoming strategically ineffective — software spreads regardless of silicon restrictions.

Analyst George Chen of The Asia Group framed the dynamic well: "Two camps are gradually taking shape. China clearly insists on advancing AI through multilateralism, while the US wants to build its own camp primarily targeting China's rise in AI."

4. The Global Capital War: Anthropic's Trillion-Dollar Bid

4.1 The Numbers Are Staggering

While China was releasing policy frameworks, Silicon Valley was rewriting valuation records.

Source: Financial Times, 36Kr, The Information, company disclosures

Anthropic's projected $1 trillion valuation would make it the most valuable AI company in the world, surpassing OpenAI. Its revenue run rate has grown 5x in five months — from $90B to a projected $450B annually — driven by Claude Code (developer tool) and Cowork (general productivity agent).

4.2 What This Means for China's AI Industry

The global capital surge has three direct implications for China:

First, talent competition intensifies. Anthropic, OpenAI, and Google can now offer compensation packages that rival Wall Street. Chinese companies must compete with equity structures, research autonomy, and domestic market scale rather than cash alone.

Second, the compute arms race escalates. Anthropic's $50B raise is explicitly for "massive compute expansion." Google's reported commitment of up to $40B (including potential $30B performance-linked tranches) means Anthropic will have access to unprecedented TPU capacity. China's domestic chip ecosystem must scale faster.

Third, agent-based revenue models are validated. Anthropic's revenue growth comes from products that are essentially agents — autonomous systems that execute tasks rather than just answer questions. This validates the Chinese policy focus on agent development. The market is voting with its wallets.

5. Infrastructure Layer: China Mobile's Trillion-Token Bet

5.1 The Announcement

At the China Mobile 2026 Mobile Cloud Conference on May 8, Chairman Chen Zhongyue unveiled a bold initiative: a trillion-token service experience package opened to industry partners, designed to lower the barrier to enterprise AI adoption.

Source: China Mobile 2026 Mobile Cloud Conference; Xinhua Economic Observer

5.2 Token as the New Metric

Chen Zhongyue framed the strategic shift eloquently: "Tokens have become the new measure of artificial intelligence."

National data confirms this. According to the National Data Administration's "National Data Resources Survey Report (2025)":

Source: National Data Administration, OpenRouter, industry statistics

China Mobile's trillion-token package is a demand-side stimulus. By making tokens effectively free for pilot deployments, the company is betting that enterprises will discover AI value quickly and transition to paid consumption at scale.

6. Model + Embodied Intelligence: DeepSeek and Zhiyuan Robotics

6.1 DeepSeek V4: The Final Gap Closes

DeepSeek's multimodal gray-scale testing (launched May 1–7) represents the completion of its capability matrix. The company that once defined itself through text-only reasoning excellence now sees.

Source: Company announcements, technical papers

The simultaneous announcement that DeepSeek-V4 runs entirely on Huawei Ascend chips for both training and inference is the technical story beneath the product story. It proves that China's domestic AI silicon ecosystem is ready for production workloads at scale.

6.2 Zhiyuan Robotics: The 10,000th Unit Milestone

Zhiyuan Robotics' Expedition A3 humanoid robot produced its 10,000th unit in early May 2026 — a milestone that marks the transition from prototype to product.

Source: 21st Century Business Herald, industry announcements

The 10,000-unit milestone is significant because software-only AI has no physical impact on the world. Agents that can manipulate objects, navigate spaces, and perform physical tasks represent the ultimate expression of autonomous intelligence. China's simultaneous progress in model intelligence (DeepSeek) and physical embodiment (Zhiyuan) suggests the two streams will converge faster than in Western markets, where they are developed by different companies with less coordination.

7. The Patent Advantage: 60% Global AI IP

7.1 The Data Point

On May 3, People's Daily reported that China has surpassed the US to become the world's largest holder of AI patents, with Chinese entities holding approximately 60% of global AI patent stock.

Source: People's Daily, May 3, 2026; Patent analytics

7.2 Why Patents Matter in the Agent Era

In the chatbot era, open-source model weights were the primary competitive asset. In the agent era, patented techniques for task planning, tool use, multi-agent coordination, and long-term memory become critical differentiators.

China's patent dominance in these exact domains — particularly NLP and machine learning — means Chinese companies have stronger defensive IP positions as agents become commercial products rather than research demos.

Additionally, the report noted that China's 6G R&D has completed first-phase technical trials with 300+ key technology patents filed. The convergence of 6G connectivity and agent intelligence will enable real-time, low-latency agent coordination across distributed devices — a capability that requires both AI and telecommunications IP.

8. Synthesis: The Agent Economy Takes Shape

8.1 Four Layers, One Ecosystem

The events of May 8–10, 2026 fit together into a coherent picture of China's emerging agent economy:

8.2 The 2027 Target

The State Council's "AI+" Action Plan (issued April 2026) set a specific target: by 2027, AI agent application penetration rate exceeds 70%. The May 8 policy framework is the operationalization of that goal.

What does 70% penetration mean in practice? It means:

- 70% of smartphones will have L3+ agent capabilities

- 70% of enterprises will deploy at least one agent in operations

- 70% of government services will offer agent-assisted interfaces

- 70% of content production will involve agent assistance

This is not a distant vision. It's an 18-month target with government backing, standardized frameworks, and proven technology.

9. Global Context: The Parallel Race

9.1 China vs. US: Different Paths to the Same Destination

Source: Comparative policy analysis

Both approaches have merits. China's standardized, top-down framework enables rapid deployment at scale. America's decentralized, capital-intensive approach pushes the frontier faster but creates fragmentation.

The critical question for the next 18 months: Which approach produces better agent outcomes — measured by utility, safety, and adoption?

Social Media Reactions: What People Are Saying

"智能体终于有国家定义了！以前各说各话，现在至少知道什么叫L3级智能终端。我的手机明年能不能自动帮我回微信？"

*"Finally, a national definition for agents! Before everyone used their own terminology. Now at least we know what an L3 smart terminal means. Can my phone auto-reply to WeChat messages next year?"*

— Weibo user @科技爱好者小王, 12K likes

"Anthropic估值要破万亿？Claude Code确实好用，但这不是泡沫是什么？中国开源模型占了60%的API调用，他们凭什么值1万亿？"

*"Anthropic breaking $1T valuation? Claude Code is good, but isn't this a bubble? Chinese open-source models account for 60% of API calls — what justifies $1T?"*

— Zhihu user, 890 upvotes

"中国移动开放万亿Token，这是在抄AWS的免费套餐策略。先用免费养成习惯，再收费收割。但不得不说，对企业来说这是好事。"

*"China Mobile opening a trillion tokens — they're copying AWS's free tier strategy. Hook users with free, then charge. But honestly, this is great for enterprises."*

— X (Twitter) user @cloudarch_cn, 2.4K retweets

"三部门文件最震撼的是那19个应用场景。内容创作、短剧、数字文化全在里面。这意味着AI生成内容即将获得国家层面的认可和鼓励。"

*"The most striking part of the three-ministry document is the 19 application scenarios. Content creation, short videos, digital culture — all included. This means AI-generated content is about to receive national-level endorsement and encouragement."*

— WeChat official account "AI Product Notes", 100K+ views

"L4协同级标准还没细化，说明真正的自主智能体还在路上。现在的L3手机助手，离真正的agent差得远。但方向是对的。"

*"The L4 collaboration level hasn't been detailed yet, showing true autonomous agents are still on the way. Today's L3 phone assistants are far from real agents. But the direction is correct."*

— Bilibili tech commentator, 15K views

"DeepSeek跑在华为昇腾上，这才是真正的去美国化。不是喊口号，是产品级落地。如果多模态版本也能全链路国产化，那CUDA垄断就真的结束了。"

*"DeepSeek running on Huawei Ascend — this is real decoupling, not slogans. Product-level deployment. If the multimodal version can also fully domesticate the chain, then CUDA's monopoly truly ends."*

— Hacker News translation thread, 340 comments

Related Articles

Continue exploring China's AI landscape:

- China's AI "War of the Gods": DeepSeek Hits $45B, Kimi Raises $2B

- China's AI Infrastructure Awakens: Wuwenxinqiou's $100M Bet

- The Great AI Compute Crunch: 8 Trillion Tokens and Counting

- The End of Free AI: Doubao Paywalls, Unitree Mass Production, and Regulation

*This article was originally published on May 10, 2026. For corrections or updates, contact editor@ainchina.com.*