_________

Embodied AI is new and everyone is talking about it. A Chinese or even global leader in this emerging discipline of artificial intelligence has not yet been crowned, however. But there is now a first contender for the title.

Goldman Sachs believes Xiaomi could soon be classified as a "leader in physical AI". If the company continues to deliver strong results, the market will come to see it that way, analysts led by Timothy Zhao wrote.

What prompted this assessment happened within a single week in March. Xiaomi unveiled three AI initiatives in rapid succession. First, it surprised the market with a trio of proprietary large AI models. A few hours later came a new electric vehicle with an AI-powered driver assistance system built on one of those models. And third, founder Lei Jun announced plans to invest more than 60 billion yuan (roughly US$8.3 billion) in artificial intelligence over the next three years.

Embodied AI is widely regarded as the next great frontier in AI research. Here, algorithms and models do not merely compute on screens. They are expected to prove themselves in the physical world.
Xiaomi has built an ecosystem that combines the development of AI models with the ability to integrate them into millions of its own household appliances and smart electric vehicles.
This combination is rare, represents an entirely new kind of challenge for European industrial companies, and is the subject of this week's Must Read:

_________

Beijing-based Botsing Technology has released what it calls the welding industry’s first industrial-grade embodied AI agent. The system combines a rail-free crawling robot with a proprietary welding large model called “Jiluan”, built on millions of welding data points.

Industrial validation shows it can shorten welding development cycles by 60 per cent and cut production costs by more than 40 per cent. It is already deployed in oil and gas, shipbuilding and nuclear power equipment.

Welding automation in China stands at roughly 30 per cent, yet welding is fundamental to shipbuilding, energy and rail. With a shortage of 5 million welders and fewer than 7 per cent of young people willing to enter the profession, embodied AI agents are becoming an industrial necessity.

Zhongce Rubber Group, one of China’s largest tyre manufacturers, has deployed 18 AI agents across its entire value chain since June 2025. An AI testing agent can run 300 virtual durability tests per second at 120 km/h, compressing a process that once took six months into days.

AI-driven factory scheduling has reduced a 10,000-worker production line to 2,000, increasing labour efficiency fivefold while cutting pollution and energy consumption by 80 per cent.

Zhongce’s vice president says the company “will no longer be just a manufacturing company, but a big data-driven AI company.” When a traditional tyre maker redefines itself through AI, it signals how deeply agentic AI is penetrating Chinese manufacturing.

Shanghai-based UniVista has launched UDA 2.0, which it describes as China’s first fully agent-based EDA tool built on a domestic architecture. The platform can autonomously complete RTL design, verification, debugging and optimisation after receiving guidance from engineers.

EDA software is a critical bottleneck in China’s semiconductor industry. Synopsys, Cadence and Siemens EDA hold more than 75 per cent of the global market. Washington briefly banned all three from supplying China in May 2025 before reversing course in July. UniVista’s push shows that domestic substitution efforts are accelerating regardless.

China Southern Power Grid has launched the country’s first large-scale AI data analytics platform for new-type energy storage in Guangzhou. It connects eight pilot stations across three provinces, collecting real-time data from 2.38 million data points across 300 energy storage systems.

After one year of trial operation, fault rates at connected stations fell by 34 per cent and renewable energy consumption rose by about 30 per cent. China’s “new-type” installed energy storage capacity now exceeds 130 GW. Managing this growing and diverse fleet requires AI.

AstraZeneca and Tsinghua University have jointly established a research centre for AI-driven drug discovery. The collaboration focuses on AI-powered drug screening, translational medicine and clinical development.

Tsinghua’s DrugCLIP model increases the speed of traditional drug screening by one million times, opening access to a chemical space estimated at 10^60 possible drug-like molecules.

All of the world’s top 20 pharmaceutical companies have now established AI partnerships. Roche is deploying Nvidia-powered “AI factories”, Eli Lilly has launched “LillyPod.” China is becoming a key node in this global network, with AstraZeneca’s head of R&D China calling the country a “driving force” in pharmaceutical innovation.

For details and more analysis, please click here to read the full, free version of this week’s Highlights:

_________

_________

The 15th Five-Year Plan elevates China’ AI Plus initiative as a top national priority. AI is mentioned 52 times throughout the 141-page document.

Source: South China Morning Post

China’s AI-related industries will be valued at more than 10 trillion yuan (US$1.45 trillion) by the end of the 15th Five-Year Plan period (2026–2030).

Source: Xinhua

Weekly token consumption by Chinese AI models on the platform OpenRouter from 16 to 22 March, overtaking US models for the third consecutive week.

Source: Science and Technology Daily (Kechuangban Ribao)

The AI Plus initiative includes a target of integrating AI into 90% of China’s economy by 2030.

Source: China’s 15th Five Year Plan

_________

The new standard, released on March 26, establishes a unified benchmarking and testing framework for embodied AI systems. For the first time, it establishes how embodied AI systems should be evaluated in both simulated and real-world environments.

The new embodied intelligence standard will come into effect on June 1, 2026. It was released by the China Academy of Information and Communications Technology (CAICT), together with more than 40 other organisations.

The standard defines system architectures and capability requirements for embodied AI. It also standardises benchmarking frameworks, methods and metrics for both simulation and real-world environments. The evaluation system covers basic capabilities, cognitive reasoning and full end-to-end closed-loop performance across four testing approaches.

More than 10,000 test tasks have already been created, covering 300 task categories across manufacturing, households, retail and logistics. Supporting tools for data collection, simulation task generation and automated metric calculation have been developed alongside the standard.

A regulatory architecture for humanoid robots and embodied AI was unveiled at the annual meeting of Humanoid Robots and Embodied Intelligence Standardization (HEIS) in Beijing on 28 February 2026. It is the first national standard system covering the entire industrial chain and lifecycle of these technologies, the State Council Information Office reported.

The framework comprises six key components: basic commonality, brain-like and intelligent computing, limbs and components, complete machines and systems, application, and safety and ethics. It was developed collaboratively by over 120 research institutions, enterprises and industry users under the Ministry of Industry and Information Technology’s technical committee for HEIS.

2025 is considered China’s first year of humanoid robot mass production, with over 140 domestic manufacturers releasing more than 330 different models, according to MIIT.

OpenClaw by Peter Steinberger has been met with an enthusiastic response in China, with citizens queuing in long lines to have the software installed on their laptops for free by tech companies using the moment for marketing.

Local governments all over the country, including district governments in Shenzhen, Wuxi and Shanghai have rushed to ride this AI agent boom, promising subsidies of up to five million yuan (US$690,000) for “OPCs” or one person companies built around agentic AI.

Longgang district in Shenzhen was the first to release a dedicated policy framework specifically for OpenClaw and “one-person companies” (OPCs) on 7 March. Others were quick to follow suit, while the central government issued warnings about data security and asked government officials to stay clear of OpenClaw at the office.

More than anything else the nationwide hype about “raising lobsters”, as the Chinese media have dubbed the OpenClaw wave after the lobster claw in the developer’s logo, shows how enthusiastically the Chinese people are embracing artificial intelligence.

While serious AI developers were making fun of retirees and housewives queuing in Shenzhen and elsewhere, and while analysts predicted that the boom might be short-lived, there was also a clear upside. With „AI literacy“ on the to-do list of just about every company on the planet, a huge part of the Chinese public that until recently couldn’t really say what a LLM was for was now tinkering with one on their own hard disc.

The 15th Five-Year Plan mentions AI more than 50 times and includes a sweeping “AI+ action plan.” Special treasury bonds worth 200 billion yuan (roughly US$28 billion) are earmarked for upgrading traditional industries.

The government work report, delivered by Premier Li Qiang on 5 March during the National People’s Congress (NPC), calls on the nation to leverage AI for two main objectives at the same time: the upgrading of traditional industries, mainly manufacturing, plus simultaneously the development of future industries like robotics, drones or AI-assisted drug discovery.

Upgrading traditional industries is one of the top priorities, supported by 200 billion yuan (US$28 billion) in special treasury bonds for equipment renewal. Minister of Industry and Information Technology Li Lecheng said China would “vigorously promote AI Plus manufacturing” and accelerate breakthroughs in next-generation AI products.

The nationwide transformation propagated by the central government is well underway. The country already has more than 5,300 AI companies, including 71 unicorn startups.

The central government in Beijing is also committed to building “hyper-scale” computing clusters supported by cheap electricity, including direct “point-to-point” links between solar- and windfarms and AI data centres, and to making open-source AI a competitive advantage against the United States, according to official government websites and reports in Chinese state media.

Two co-founders of the AI startup Manus have been barred from leaving the country while regulators review Meta’s acquisition of the firm. Meta had valued Manus at US$2–3 billion.

The case highlights Beijing’s growing concern about the loss of valuable AI technology to geopolitical rivals. Manus had been hailed as a Chinese AI success story, reaching US$100 million in annualised revenue within months. For foreign companies considering acquisitions of Chinese AI firms, the Manus case is a warning that regulatory scrutiny of outbound technology transfers is intensifying.

_________

Chinese AI models such as DeepSeek and Kimi have become highly popular with software developers from California to Bavaria because they perform well and cost far less than their American competitors.

For the first time, token consumption by Chinese LLMs has overtaken that of GPT, Claude and their peers. Four of the five most popular models on the API platform OpenRouter were Chinese.

When a student in San Francisco types a prompt for DeepSeek, tokens race through a fibre-optic cable under the Pacific into a Chinese data centre and back again. That is an indirect export of Chinese energy, because between 46 and 60 per cent of the operating costs of AI data centres consist of expenses for electricity, according to the International Data Corporation (IDC).

The term 'tokenomics' was coined in the context of Bitcoin. Now there is, in effect, 'Tokenomics v2.0', because tokens for AI inference have become an internationally tradable commodity. China holds a strong position in this particular segment of the global AI race, because it already has vast wind and solar capacity and thus more cheap energy than the United States and most other countries.

For the full story behind China's energy-based value-added exports, we recommend our Deep Dive of the week.