China Embodied AI Push Pairs Humanoid Funding With Robot Brains

A bet on robots as the route to general intelligence

Most of the AI race is measured in chatbot benchmarks and data-center spending. China is placing a second, very large bet next to that one. Its claim is that AI will reach general intelligence by acting in the physical world, and that robots are how it gets there. In recent weeks, record funding rounds, new robot foundation models, state-run data facilities and factory deployments have all pointed the same way. Embodied AI is no longer a side project in Beijing's technology plans. It is becoming one of the main pillars.

The policy basis is clear. Beijing lists embodied AI among its "industries of the future," next to 6G and quantum technology, and makes it a priority in the 15th Five-Year Plan, which runs to 2030.1 The plan moves national robotics strategy away from traditional industrial automation and toward embodied intelligence and humanoids. It builds on the factory base and electric-vehicle supply chains that earlier plans created.59 Researchers at Carnegie's China AI Initiative argue that Washington and Silicon Valley could fall behind if they keep focusing on software alone.1

The argument goes further than economics. Leading Chinese scientists say embodiment may be an overlooked path to artificial general intelligence. Zhang Bo, founding dean of Tsinghua University's Institute for Artificial Intelligence, says intelligence depends on learning from and making decisions in the physical world, not just on cognition.1 Zhou Bowen of Shanghai AI Lab makes a similar case. He sees China's huge manufacturing base as a training ground where real-world data flows back into AI models.1 This is not a retreat from language models. Chinese labs such as DeepSeek and Z.ai are scaling those as hard as their American rivals.1

The money: record rounds, and investors getting pickier

Capital is flowing at a pace that would have been hard to picture two years ago. XPeng's robotics unit raised more than $900 million at a post-money valuation above $6.3 billion. The company calls it the largest single private round in China's embodied AI sector so far.1518 IDG Capital led the round, and Tencent and Alibaba joined as strategic investors. The money is meant to fund physical AI foundation models, data generation and mass-production facilities for the next-generation IRON humanoid.18 AGIBOT is reportedly close to closing a round of RMB 1 billion to 1.5 billion at a valuation of about RMB 50 billion. People familiar with the deal say the timing could push its planned Hong Kong IPO into 2027.20

The infrastructure layer is drawing money too. D-Robotics, which makes the Sunrise chips used in many robots, closed a $400 million Series C, bringing its disclosed funding to about $770 million.14 The company says it has shipped more than 8 million Sunrise chips and that revenue grew several times over in the first half of 2026.19 Unitree's Shanghai listing in August raised close to $1 billion, and its shares jumped more than 450 percent on the first day.1

Cumulative figures show how one-sided the funding race has become. One tracker counts $7.77 billion raised by Chinese humanoid startups over the past decade, compared with $3.11 billion in the United States.12 Another estimates that Chinese startups now account for 88 percent of humanoid funding rounds.11

The headline totals hide a change in what investors want. Chinese venture firms are moving away from the grand story of a general-purpose humanoid. They now look harder at commercial traction, data flywheels and proven use in specific industries.2 The founders in highest demand are veterans of autonomous driving. Kunlunxing Robotics, co-founded by an early leader of Li Auto's driving team, passed a $1 billion valuation after three rounds in under 90 days. Moqi Intelligence, founded by a former head of Huawei's autonomous-driving AI division, raised more than CNY 1 billion in its angel round before it had unveiled a product.2 The same reporting also notes that early-stage investment has started to cool in the second half of the year. Uncertainty over exits and regulators warning against "reckless growth" are part of the reason.2

There is a tension here, but not a contradiction. The biggest checks are going to companies with factories, chips or a corporate parent. Pure-concept startups now face tougher questions. This looks like a market sorting itself out more than a bubble starting to burst.

Robot foundation models and the race to build a robot "brain"

The AGI argument rises or falls on software that can generalize, and Chinese labs have released a wave of it. Light Origins launched Light-O1, which learns from human motion extracted from internet video before being adapted to specific robots. The company trained six 4-billion-parameter versions on up to 120 billion multimodal tokens. It reports that prediction errors fell in a power-law pattern as pretraining grew.23 The company is careful about the claim, though. Its scaling tests measured how well the model predicts the next action, not whether robots succeed at real tasks.23

Other releases follow the same pattern. The Chinese Academy of Sciences' Maxwell model scored 91.9 on the Meta-World simulation benchmark, the highest score recorded, and handled more than 200 tasks without fine-tuning.27 Unitree, long seen as a hardware company, open-sourced a 6-billion-parameter model that can run more than 60 household and tabletop tasks. It also plans to spend nearly half its IPO proceeds on model research.27 Ant Group's Robbyant released LingBot-VLA as open weights. It was trained on about 20,000 hours of real two-arm robot data, and the company says performance had not yet stopped improving as data grew.22 Xirang Kaiwu, founded by former Huawei AI leaders, is building what it calls a Large Physics Model and has raised money at a reported valuation of about $500 million.30

The timelines are getting more aggressive. Spirit AI expects a "GPT-3 moment" for robots by mid-2027, meaning robots that follow natural-language instructions through a series of physical steps.10 Unitree founder Wang Xingxing shortened his forecast for an embodied AI tipping point to one to two years after testing GPT-6 on Unitree robots.24 Independent tests point to limits, however. In one set of trials, GPT-6 Astra succeeded on 19 of 20 simple block-in-bowl tasks but only 2 of 20 precision insertions.24 Analysts describe a wide gap between lab demonstrations and reliable deployment. Vision-language-action models still struggle with the variety found in real homes and factories.3

My reading is that benchmark wins like Maxwell's are real but narrow. China's lasting advantage is less likely to come from any single model than from the data supply behind all of them.

Data as national infrastructure

That data supply is now being built on purpose. As of the end of June 2026, China had opened more than 70 embodied AI "training grounds," with over 40 more under construction or planned. At these sites, operators in motion-capture suits teach robots to fold laundry and scan parcels.4 Not everyone counts the same way. Another analysis puts the number of such facilities at about 40 by mid-2026, which probably reflects a narrower definition.5 In June, authorities launched a plan requiring each province to identify at least 20 priority scenarios for real-world robot training. A national standard for the quality of embodied AI datasets takes effect on November 1.4 Beijing's humanoid innovation center has set a long-term goal of one million hours of high-quality interaction data.7

The main tension between Chinese and Western views shows up here. Chinese officials frame the training grounds as filling a gap. The internet holds plenty of text but almost no systematic record of physical interaction.4 Western analysts see the same facilities as building a flywheel: robots deployed at scale produce data that rivals without similar deployment cannot match.135

Warehouses and factories come first

Reporting agrees that logistics is where embodied AI will prove itself first. Chinese makers produced about 97 percent of the roughly 19,100 humanoids shipped worldwide in the first half of 2026. About 60,000 are expected for the full year.38 Researchers at Smart Analytics Global say the targets are manufacturing, logistics and warehousing. In those settings tasks are well defined and productivity is easy to measure.38 Counts for 2025 vary a little, from 14,400 to roughly 15,000 Chinese units, but every source puts China's share above three-quarters.45

The deployments are real but small. SF Express worked with Robot Era to put humanoids in more than 10 logistics centers, reporting 85 percent of human efficiency.17 UBTECH's new factory in Liuzhou can finish an industrial humanoid every 10 minutes. Inside the plant, its Cruzr robots do material handling rather than assembly.33 UBTECH and FAW-Volkswagen signed an agreement on October 8 to test humanoids in factory logistics. No order, task list or timetable has been announced.36 UBTECH sold 1,079 full-size humanoids in 2025. By contrast, many units Chinese makers ship go to research, education and data collection rather than paid work.32

What it means

Geopolitics is turning this into a contest between two blocs. In July, the US Federal Communications Commission added humanoid robots to its Covered List, which blocks approval of new Chinese models.33 Beijing is answering with its own push for self-reliance. It recently unveiled a robot it describes as running on a fully domestic electronic architecture.26

The overall reading is that China's AGI-through-robots thesis is still unproven. The strategy behind it, though, is coherent and well funded. Even if embodiment never leads to general intelligence, China is building the most important input for physical AI: large amounts of real-world interaction data, gathered in warehouses and factories that the state is opening to robots. If embodiment does turn out to matter for general intelligence, that head start in data will count for more than any single model release.