AI BOARD
AI Board
AI tools, policy moves, and public decisions are tracked in plain language.
The Six Axes of China's AI Battlefield: In-Depth Analysis
The Six Axes of China's AI Battlefield: In-Depth Analysis
Thursday, March 12, 2026 | KIMKJ.COM - POLITICAL ARCHIVE
As of March 2026, China's AI industry has entered a decisive phase of shifting "from experimentation to proof of revenue." With AI mentioned more than 50 timesin the 15th Five-Year Plan (2026-2030) and elevated to a core national strategy, daily token consumption exploded by 300 times(from 100 billion to 30 trillion+ tokens) in 18 months. Huawei Ascend chips rose to about 40%of China's AI semiconductor market, Chinese-made humanoid robots accounted for 80~90%of global sales, and domestic models such as DeepSeek, Qwen, and Doubao took 61%of OpenRouter's global token volume, rapidly giving concrete shape to an independent "Chinese-style AI" ecosystem.
1 From Tools to ROI: The Great Shift Toward Practical AI and Industrial Application
The year 2026 is a watershed in which China's AI industry shifts the paradigm "from Chat to Agent". At Tsinghua University's AGI-Next Summit, an industry consensus formed that "the era of conversational AI is over, and now agents that execute work are the core of competition."
Real-use indicators support this. Daily token calls for the Doubao model reached 50 trillion(December 2025), a tenfold increase year over year, and more than 100 enterprise customers each consumed a cumulative 1 trillion tokensor more. Average daily token consumption for the MiniMax M2 series surged by 6 timesin February 2026 compared with December 2025. According to an Avnet Insights survey, 85%of Chinese companies said they had already put AI products into mass production and delivery, far above the global average of 56%.
The tone of analysis from securities and economic media is changing too. HSBC projected that "2026 will be the first year in which Chinese companies gradually shift from model and computing investment to investment in AI applications." iiMedia Research analyzed that "a turning point will emerge in 2026 when profitability becomes the key metric in valuing AI companies." Bank of America estimated China's AI investment in 2026 at 700 billion yuan (about $99 billion), a 48% increase from 2025.
AI adoption by central state-owned enterprises (SOEs) is also accelerating. State Grid used AI to lower transmission loss rates by 1.2%p, and Sinopec shortened oil and gas field discovery cycles by 18 months. Across 16 core industries, more than 800AI application scenarios are running at state-owned enterprises. The manufacturing AI adoption rate rose from 19.9% in 2024 to 25.9%in 2025, concentrated in customer service and operations (40%+) and R&D (30-40%).
2 From Digital to Physical: Acceleration of the Embodied Intelligence Revolution
A. Overwhelming Quantitative Advantage
China's expansion in physical AI is a reality confirmed by numbers. Of the roughly 13,000 to 18,000 humanoid robots shipped globally in 2025, China accounted for 87~90%, and Morgan Stanley doubled its 2026 forecast for Chinese humanoid robot sales to 28,000 units.
B. Trends by Major Company
Unitreeranked first globally, shipping about 5,500 units in 2025, and its 2026 target is 10,000 to 20,000 units. At the 2026 Spring Festival Gala (CCTV Chunwan), its G1 robots performed autonomous group martial arts. The company is valued at about $3 billionand is preparing for a $7 billion IPO. AgiBotshipped 5,168 units in 2025, ranking second globally, and unveiled an online store and a Robot-as-a-Service (RaaS) rental platform at MWC 2026.
UBTechdeployed Walker S1/S2 at Zeekr, NIO, Audi China, and BYD factories, demonstrating multi-robot collaboration on real manufacturing lines for the first time. BYD established its own embodied intelligence laboratory, aiming to deploy 1,500 units in 2025 and 20,000 units in 2026. Galbot raised more than $300 millionand reached a valuation of $3 billion, while Noetix launched Bumi, a home companion robot priced at $1,400, attempting to break consumer pricing.
C. Autonomous Driving and Unmanned Delivery
Baidu Apollo Go has completed a cumulative more than 17 millionfully driverless rides, and weekly orders have exceeded 250,000. Pony.ai is the only company operating fully autonomous robotaxis in all four of China's first-tier cities, and it reduced the BOM cost of its seventh-generation vehicles by 70%, achieving unit-economics breakeven in Guangzhou.
A sharp price decline created a turning point in unmanned delivery. The price of unmanned delivery vehicles plunged from more than 1 million yuan in 2018 to less than 20,000 yuanin 2025. China Post placed the world's largest procurement order, for more than 7,000unmanned vehicles, while ZTO Express operates more than 2,000 unmanned vehicles in more than 200 cities, delivering more than 200,000parcels per day. Revenue for autonomous delivery suppliers is expected to grow more than tenfold, from 35.2 billion yuanin 2026 to 373.4 billion yuanin 2030.
D. Investment and Policy
Investment in embodied intelligence has also surged. In 2025, there were 325 related investment deals totaling 39.8 billion yuan, up 216% and 326% year over year, respectively. IDC projected that China's spending on embodied intelligence robots would jump from $1.4 billionin 2025 to $77 billionin 2030, a 94% compound annual growth rate. The 15th Five-Year Plan specified "a leap across the full embodied intelligence chain", and the Ministry of Industry and Information Technology (MIIT) announced that China currently has more than 140domestic humanoid manufacturers and about 330models made public.
3 Signs of Maturity in an AI Market Where Overheating and Cooling Coexist
A. The Spring Festival Battle and Subsidy Frenzy
During the 2026 Spring Festival, BAT collectively spent 4.5 billion yuan (about $625 million)on AI app promotions. Alibaba's Qwen launched a 3 billion yuancampaign, and its MAU exploded in one month from 31 million to 203 million, while Tencent Yuanbao distributed 1 billion yuanin digital hongbao. Doubao's MAU remained first at 315 million.
B. The Retention Problem After Subsidies
Tencent Yuanbao experienced meaningful user churn after the Spring Festival holiday. Pan Helin, a member of the MIIT expert committee, warned: "Holiday hongbao and free offers can induce a temporary surge, but the real test comes after the incentives disappear. Only AI products that solve real problems and integrate into daily life will survive." According to a RevenueCat report, subscription churn for AI apps is 30% fasterthan for non-AI apps, and refund rates are 20% higher.
C. The End of the Hundred Models War and Vertical Specialization
The "Hundred Models War" is over. According to a report by CAICT (China Academy of Information and Communications Technology), the number of foundation models is converging, and surviving companies are shifting to vertical specialization. Baichuan pivoted to medical AI, and 01.AI pivoted to customized enterprise deployment. The basis of competition has moved from "model benchmark performance" to "scenario penetration, ecosystem building, and extraction of application value".
D. Bubble Risk Assessment
On China's AI bubble risk, UBS assessed it as structurally lowerthan that of the United States. There are three reasons. Leading Chinese model companies fund themselves not through circular financing but through parent-company cash flow; China's major-company AI capital expenditure in 2025 was about 400 billion yuan ($55 billion), around one-tenthof the U.S. level while achieving similar model capabilities; and actual data-center utilization is high. The stock market, however, shows signs of overheating. Cambricon's share price surpassed Kweichow Moutai to become the "highest-priced stock" on the A-share market, and Moore Threads' STAR Board IPO 425%surged, while MiniMax's Hong Kong IPO was oversubscribed 1,848 times and rose 109% on its first day.
4 Acceleration of AI Industrialization Through a Three-Level Central-Local-Private Combination
A. Large-Scale Funding at the Central Level
Funding for China's AI industry is expanding at an unprecedented scale. The National AI Industry Investment Fund of 60.06 billion yuan ($8.2 billion), the National Venture Capital Guidance Fund of about 1 trillion yuan ($138 billion, operated over 20 years), and Bank of China's AI industry-chain support of 1 trillion yuan (over five years)have been created.
B. Active Investment by Local Governments
Shenzhen has a 10 billion yuan (about $1.39 billion)fund for AI and robotics, while Shanghai will invest more than 70 billion+ yuan in about 50 major technology projectsin 2026. Hangzhou provides computing-power vouchers of up to 8 million yuan (about $1.1 million)per company, one of the highest levels of support nationwide. According to Stanford data, 71%of AI companies that received both government and private VC funding received government investment first, showing that government acts as a traffic light for private VC. From January to March 2026, $5.21 billionwas raised across 72 equity investment rounds, a 143% increasefrom the same period a year earlier.
C. Major Startup Funding and IPO Status
•Zhipu AI - Listed on the Hong Kong stock market on January 8, 2026, becoming the first listed Chinese AI LLM company. Raised HK$4.35 billion (about $560 million) through the IPO, with a market capitalization of about $13 billion
•MiniMax - Listed in Hong Kong on January 9, 2026, surged 109% on the first day, and was oversubscribed 1,240 times. Listed only four years after founding, setting the shortest-time-to-listing record in the AI industry. Market capitalization about $14 billion to $15.2 billion
•Moonshot AI (Kimi) - Pursuing a new $700 million+ round in February 2026 at a valuation of about $10 billion to $12 billion. Revenue exceeded all of 2025 only 20 days after K2.5 launched. China's fastest decacorn
•DeepSeek - Zero external VC investment. Self-funded with founder Liang Wenfeng's High-Flyer hedge fund money ($8 billion AUM). An unusual model
•Galbot - Raised $300 million+, valuation $3 billion. One of the largest investments in the humanoid robot field
•AI² Robotics - Raised 1 billion+ yuan ($145 million) in a February 2026 Series B, valuation $1.5 billion+
Across the industry, total Chinese AI investment in 2025 reached 890 billion yuan (about $125 billion), up 18% year over year and accounting for about 38%of global AI investment. By region, Beijing (28%), Shenzhen (24%), and Shanghai (19%) accounted for 71%of the total, cementing a three-city structure.
5 AI Penetrating Consumption, Distribution, and Everyday Services
A. The Consumer Penetration Turning Point Shown by the Spring Festival
In the Alibaba Qwen app, it became possible to "order milk tea with one sentence." A complete closed loop from AI voice command to payment to fulfillment ran through Taobao Flash Sale, Feikou, Hema, and Alipay, processing about 200 million shopping ordersduring the Spring Festival holiday. McDonald's x NIO launched an in-car AI voice ordering system, where if a driver says, "I want McDonald's," the system automatically recommends the best store and menu.
B. Deepening AI Integration in E-Commerce
In 2025, 34%of Tmall Supermarket orders came from AI recommendation channels. Douyin e-commerce GMV surpassed 4 trillion yuan, and AI digital-human livestream commerce spread rapidly. During the 618 period, more than 50%of all livestreamers on Taobao and Kuaishou were AI. With Baidu's digital-human technology, Luo Yonghao's digital avatar achieved 55 million yuanin GMV in a single broadcast, exceeding the real person's one-hour sales performance in 26 minutes.
C. Rural and Campus Penetration, and the Scale of Generative AI Users
Rural penetration of unmanned delivery also stands out. In one county in Hunan Province, unmanned vehicles were deployed at a shared delivery center and handle county-to-township parcels. In Pingwu Town, Zhejiang Province, a customer operating more than 30 unmanned vehicles achieved about 30% savingsin transport costs and recouped the investment within three years. Cainiao operates more than 700 Xiaomanlv delivery robots across 15 universities and about 400 campuses in 11 cities. The number of generative AI users reached 602 millionas of December 2025 (CNNIC). The State Council's AI+ guidelines target AI agent penetration at 70%by 2027 and 90%or more by 2030.
6 Concrete Shape of an Independent "Chinese-Style AI" Path
A. Semiconductor Self-Sufficiency: Huawei Ascend and the Four Little Dragons
Huawei Ascend 910C has reached about 80%of H100 performance, and about 600,000 unitsare planned for production in 2026, twice the previous year. Bernstein projected that Huawei's share of China's AI chip market would reach about 50%in 2026, while NVIDIA would fall to about 8%. The next-generation Ascend 950DT (Q4 2026) will include self-developed HBM (HiZQ 2.0), providing 144GB of memory and 4TB/s bandwidth. Cambricon aims to ship 500,000AI accelerators in 2026, three times the previous year. The "GPU Four Little Dragons," Moore Threads, MetaX, BirenTech, and Enflame, listed one after another, and their combined market capitalization surpassed 1.3 trillion yuan (about $186 billion).
B. System-Level Competition Strategy
SMIC developed N+3 (5nm-class) using DUV multipatterning, but yields remain around 20%, about one-third of TSMC's level, and costs are 40-50% higher. With access to EUV equipment blocked, China chose a "system-level competition" strategy. Huawei's Atlas 950 SuperCluster connects more than 500,000NPUs to implement 524 EFLOPSFP8 operations, taking an approach of "responding with rack-scale solutions by clustering thousands of lower-performance chips, rather than relying on individual chips."
C. Open-Source Strategy and the Global Ecosystem
Chinese models' share of Hugging Face downloads is about 30%of the global total, nearly twice the U.S. share of 15.7%, and this is a deliberate strategic choice. DeepSeek, Qwen, and ERNIE have all been released under MIT or Apache 2.0 licenses, building a global developer ecosystem and moving first to occupy technical standards. Morgan Stanley analyzed that "China focuses more on bringing AI to marketthan on building the strongest AI capabilities."
D. U.S. Export Controls and the Response
The Trump administration eased H200 chip export licenses to China to a "case-by-case review" in January 2026, but NVIDIA's CFO said in the February 2026 earnings call that "H200 China revenue is still zero." It was reported that Chinese authorities blocked H200 imports through tariffs and warned companies to refrain from buying U.S. AI technology unless unavoidable. DeepSeek's R1 model was trained for $5.6 million, achieving costs 10 to 40 times lowerthan the hundreds of millions of dollars spent by Western competitors. With mHC architecture and Sparse Attention, it trains larger models with fewer computing resources.
Major Company and Policy Trends (March 2026)
Key Announcements by Big Tech Companies
•DeepSeek - V4 release is imminent but unreleased as of March 12. Expected features include about 1 trillion parameters, MoE, a 1-million-token context window, and native multimodality
•Baidu - Released ERNIE 5.0 on January 22 (2.4 trillion parameters, MoE). Ranked first in China and eighth globally on the LMArena text leaderboard. ERNIE Bot surpassed 200 million MAU
•Alibaba - Three key Qwen team members departed in succession in early March. CEO Eddie Wu held an emergency all-hands meeting and created a "foundation model task force." Plans to invest 380 billion yuan (about $52 billion) in AI infrastructure over three years
•ByteDance - Released Doubao 2.0 (Seed 2.0) on February 14, claiming performance at the level of GPT 5.2 and Gemini 3 Pro. Ranked first among Chinese AI apps with 155 million weekly active users
•Huawei - Unveiled Atlas 950 SuperPod (8,192 Ascend 950DT chips, 16 EFLOPS FP4) and Atlas 950 SuperCluster (500,000+ chips, 524 EFLOPS FP8) at MWC 2026
•Tencent - Globally released Hunyuan 3D 3.0, with 3 million+ Hugging Face downloads. Yuanbao AI chatbot surpassed 50 million DAU
•Moonshot AI (Kimi) - Open-sourced K2.5 under MIT on January 27. 384-expert MoE. Agent Swarm technology with up to 100 subagents and 1,500 parallel tool calls
•SenseTime - Subsidiary Dashao Robotics open-sourced ACE-Brain-0, a "general model for spatial intelligence," on March 9. Added to the MSCI China Index on February 11
Core Government Policies
•15th Five-Year Plan (2026-2030) - Announced at the National People's Congress on March 5. AI mentioned more than 50 times, including the "AI+ Action Plan." Semiconductor subsidies of up to 500 billion yuan (about 60 billion euros)
•NPC Standing Committee - On March 9, Chairman Zhao Leji declared AI legislative research the top priority for 2026
•MIIT "AI+ Manufacturing" Action Plan - Jointly announced by eight ministries. Target of spreading 500 representative application scenarios by 2027
•Cybersecurity Law Amendment - Took effect January 1, 2026. New Article 20 explicitly states comprehensive support for AI R&D
•Cross-Border Data Standard - Took effect March 1, 2026. China's first safety standard for cross-border personal information processing
Conclusion: Five Key Implications
First, 2026 is the "year of proof."
China's AI industry has shifted from competition over model performance to competition over profitability and application value. A 300-fold explosion in token consumption, 85% of companies entering mass production, and the turn toward agentic AI support this.
Second, physical AI is not the "next battlefield" but the "current battlefield."
With an overwhelming quantitative advantage of 87-90% of global humanoid sales, deployment on real manufacturing lines at BYD, NIO, Audi, and others, and unmanned delivery vehicles entering the 20,000-yuan price range, physical AI is already a commercial reality.
Third, "Chinese-style AI" has established its own model of efficiency, openness, and system integration.
DeepSeek's 10-40x cost efficiency, a 30% global open-source share, and Huawei's system-level clustering systematically compensate for disadvantages at the chip level. This is a fundamentally different path from the U.S. model of concentration, closure, and platforms.
Fourth, the three-level structure linking central government, local governments, and the private sector is the engine of AI industry funding.
The national AI fund of $8.2 billion, the national VC guidance fund of $138 billion, local-government computing vouchers, and private big-tech startup investments are systemically interlocking, while government investment acts as a leading indicator for private VC (71%).
Fifth, the coexistence of overheating and cooling is a healthy sign of maturity.
The simultaneous progress of stock-market speculative overheating and real-use validation, the shift to vertical specialization after the end of the "Hundred Models War," and 30% faster AI app subscription churn show that the market is filtering from "AI for everyone" to "AI that creates real value."
This report is a draft automatically collected and written by AI. Please check the original articles for exact details.
KIMKJ.COM — POLITICAL ARCHIVE
Search Keyword Hashtags
#ChinaAI #ChinaAIPolicy #EmbodiedIntelligence #EmbodiedAI #PhysicalAI #HumanoidRobots #Unitree #AgiBot #DeepSeek #HuaweiAscend #Cambricon #MooreThreads #MiniMax #ZhipuAI #MoonshotAI #Kimi #BaiduERNIE #Doubao #Qwen #15thFiveYearPlan #AIIndustrialization #UnmannedDelivery #Robotaxi #AISemiconductors #GPUFourLittleDragons #AIBubble #HundredModelsWar #AIAgent #AttorneyKimKyungjin #KIMKJ