China blocks SoftBank’s $5.4bn takeover of ABB’s robotics division
China has told SoftBank it will not allow the $5.4bn acquisition of ABB’s robotics division to go ahead in a shock move that threatens to throw Masayoshi Son’s ambitions in physical AI into disarray. (…)
China is the biggest market for ABB’s robotics unit, with more than 90 per cent of its robots for Chinese customers manufactured locally. The Swiss-Swedish company employs about 12,700 people in China. (…)
The ABB division’s operations in China were a significant attraction for the deal, say people close to the Japanese group.
The NDRC review had included detailed scrutiny of ABB’s Chinese supply chains and a demand to retain some of the robotics unit’s manufacturing capabilities in China, according to people familiar with the matter.
The deal cleared China’s competition review in February. It then became subject to a foreign investment security review that highlighted Beijing’s concern over the transfer of assets developed in China to a Japanese company at a time of heightened tensions with Tokyo and strategic competition with Washington.
The NDRC had been concerned about robotics data being transferred overseas, where it could be used to train AI models, including those developed by OpenAI, in which SoftBank is a major investor, said people familiar with the matter.
ABB set up a joint venture with a Chinese partner in 1992 and its China operations, particularly its robotics business, have long benefited from government support. Shanghai’s government also helped ABB build its $150mn robotics mega-factory in Pudong, which began operating in 2022.
According to people briefed on their thinking, regulators had also examined keeping intellectual property relating to some of ABB’s robotics business in China, potentially requiring parts of the business to be spun off to Chinese owners for the deal to proceed.
Very significant, well beyond the impact on Softbank.
Physical AI is where AI interacts with the real world through robots, autonomous machines, vehicles and industrial equipment. Digital AI creates and processes intelligence; physical AI uses that intelligence to perform work in the real world.
Physical AI could become a market worth hundreds of $billions annually. McKinsey estimates that general-purpose robotics alone could generate a $370bn market by 2040 in its base scenario.
With ABB, SoftBank was attempting to position itself for the transition from AI that thinks to AI that acts. That transition could be one of the defining investment themes of the next 10–20 years.
That gets right into (against) China’s strategy.
China’s rejection is effectively asserting control over industrial/robotics AI data.
Industrial robots generate valuable information about:
-
How objects are manipulated.
-
How machines move and interact with their surroundings.
-
How manufacturing processes operate.
-
How robots respond to unexpected situations.
-
How physical tasks can be performed more efficiently.
The data generated by millions of industrial operations could become a strategic AI resource.
China is applying a national-security bend to that data, rather than treating it simply as the commercial property of a multinational company.
This has potentially profound implications for international AI development.
It suggests that access to industrial data could become as strategically sensitive as access to advanced semiconductors.
China understands the implications of a Japanese company, already a major investor in American OpenAI, acquiring a European company with substantial advanced robotics operations/knowledge/data in China.
A clear no-no for Beijing.
The wider message:
Multinational companies are discovering that owning a business globally does not mean they can freely transfer control of all its assets, intellectual property and data across national borders.
This could make future cross-border acquisitions in robotics, AI, semiconductors, autonomous vehicles and advanced manufacturing considerably more complicated.
The blocking of Meta’s acquisition of AI company Manus earlier in 2026 seemed relatively minor but it was the first indication that China is serious about AI leadership and corporate data.
It thus looks like Physical AI will also be split into 2 ecosystems.
Physical AI requires integration of several things:
-
Advanced AI models.
-
Computing power.
-
Sensors and machine vision.
-
Robotics hardware.
-
Industrial manufacturing expertise.
-
Real-world operational data.
Control over any one of these can create strategic advantages. China considers that the entire combination as strategically important.
The risk is that companies will find it increasingly difficult to develop globally integrated AI and robotics businesses.
The AI industry is moving from a competition over who builds the most capable models to a much broader competition over who controls the computing, data, machines and industrial ecosystems that make those models useful.
Our research trip to China last December included robotics.
From Bots On The Ground:
(…) On the battlefield for technological supremacy, robotics is now on the front line and China’s efforts are centered in Shanghai.
So far, robotics was essentially supplying fixed industrial robots. China already totally dominates this field.
The battle is now getting more serious: this is the “bots on the ground” era.
The Shanghai Humanoid Robot Kylin Training Ground, covering approximately 5,000 square meters (54,000 sq ft) in the Zhangjiang area of Shanghai, is China’s first national-level, public facility designed to mass-train humanoid robots from different companies in a shared environment. (…)
China clearly displayed its robotics objectives by inserting the word “kylin” above. “Kylin” is a standard brand name for homegrown Chinese “core technologies” intended to replace foreign dependencies.
The facility’s core purpose is to accelerate the entire Chinese robotics industry by centralizing data collection and standardizing how robots learn complex tasks.
It is a “heterogeneous” facility, meaning it trains robots from competing manufacturers (like AgiBot, Fourier, and Unitree) side-by-side rather than serving just one company.
It aims to collect 10 million high-quality physical data entries (robot movements, interactions) to build a shared “brain” for the industry. (…)
The facility also acts as a standard-setter. Just as cars need to pass crash tests, robots here must pass proficiency exams in these standardized environments to prove they are ready for factory work. (…)
The “US model” relies on private enterprise, venture capital, and a university system that fosters fundamental research and innovation.
The “China Model” offers several strategic advantages over the decentralized US approach.
Its primary superiority lies in treating AI training data as a public utility rather than a private asset, which theoretically allows the entire industry to advance faster than any single company could on its own.
In the US, valuable training data is trapped in “walled gardens.” Tesla does not share its Optimus data with Boston Dynamics, and Figure AI does not share with Agility Robotics. This means every US company must solve the same basic problems (e.g., how to open a door, how to walk on gravel) from scratch.
The Kylin center aggregates data from 100+ (1000+ in 2027) different robots into a single, massive dataset targeting 10 million entries. This creates a shared “industry brain.”
A startup’s robot can theoretically learn from the mistakes made by a competitor’s robot yesterday, accelerating the baseline intelligence for everyone. (…)
The US market typically waits for a clear winner to emerge before setting standards (e.g., waiting for Tesla’s NACS charging port to become the standard). This can lead to years of fragmentation where parts and software aren’t compatible.
The Kylin center acts as a central certification authority from day one. By forcing all robots to pass the same “exams” in the same physical facility, it enforces early standardization. This ensures that supply chains (batteries, joints, sensors) can be standardized faster, lowering component costs for the whole industry.
The Kylin center is “heterogeneous,” meaning it trains robots with different body shapes and mechanical structures side-by-side. This diversity is rare. Most US labs only contain one type of robot. Training AI on diverse hardware makes the algorithms more robust.
The China model (centralized infrastructure) and the US model (market competition) generate completely different economic benefits. While the US model maximizes corporate value and radical innovation, the Chinese model is designed to maximize manufacturing scale and supply chain dominance. (…)
US labs have text data; Chinese labs have physical data. This will create a new “data wall” that the US cannot easily climb because it lacks the manufacturing environments to deploy these robots at scale. (…)
Per the International Federation of Robotics:
China totally dominates the Industrial Robots field …
… and has boosted the share of its domestic suppliers from 32% to 55%, creating a huge domestic, efficient ecosystem.
As of September 2026, IFR has identified 174 humanoid robot manufacturers globally. More than 50% are in China, a proportion rising rapidly as a wave of new company creation coincides with breakthroughs in Gen AI.
But all is not lost for the US. Watch this Apple video: Perceptive Humanoid Motion from Simulation.
Here’s what the authors wrote (BTW, 5 Chinese and 3 Europeans):
Humanoid robots have the potential to automate a wide range of physical tasks.
Progress toward generalist humanoid autonomy, however, remains constrained by a data bottleneck: mapping pixels and free-form language to low-level joint commands requires paired supervision, and no corpus of such data exists at sufficiently large scale.
Some approaches rely on human teleoperation and wearable capture devices, which require at least an hour of human time to produce an hour of data.
We present HALO, a humanoid autonomy system powered by a pipeline that synthesizes vision-language-motion data across a thousand procedurally generated homes. The corpus scales with compute rather than human labor.
HALO acquires a repertoire of physical skills, including sitting on chairs, navigating to named objects, stepping sideways through half-open doors, crouching and crawling beneath tables, and pushing boxes. It orchestrates these skills to solve composite, long-horizon tasks.
Without ever seeing a real home during training, HALO transfers zero-shot to the physical world, achieving a 91% success rate across 360 trials in six unmapped real-world environments.
US software teams (including Apple with HALO) often must run their models on Chinese platforms like the Unitree Robotics G1 because equivalent domestic hardware is either scarce or cost-prohibitive.
The US seeks to overcome the data collection bottleneck using systems like Apple’s HALO, Nvidia Cosmos, and Google DeepMind Genie to generate hundreds of thousands of synthetic hours of photorealistic motion and perception across randomized 3D environments.
While simulation excels at visual navigation and high-level trajectory planning, physics engines struggle to accurately model complex micro-frictions, soft-body deformations, and tactile grasping. Analyses from robotics practitioners at Field Motion AI emphasize that contact-rich manipulation (such as tool use or precise mechanical assembly) still heavily relies on real-world demonstration data
Analysts from the Carnegie Endowment for International Peace note that Chinese state initiatives and factory deployments generate massive feedback loops. Deploying thousands of robots directly onto manufacturing floors yields corner-case physical failure data that synthetic models cannot anticipate.
And the US must convert whatever algorithmic lead it may get into commercial dominance, requiring overcoming reliance on foreign hardware manufacturing.
Trump: New Iran attacks “quite unlikely”
President Trump said Monday that a renewed military campaign against Iran is “quite unlikely” and “doesn’t seem to be necessary.”
Trump stopped short of saying the war was over, but his comments on Truth Social and at an impromptu press conference were the furthest he went in that direction since February.
The president in recent days has been weighing three options: launch new strikes, continue to strangle Iran’s regime economically, or cut a deal that includes significant concessions to the Iranians.
- Trump signaled on Monday that he’s going with the strangle strategy for now. That means keeping up the “status quo” — continuing the naval blockade, in addition to a new operation dubbed “Iron Gate” to cut off land routes that Iran uses to move advanced weaponry and other technology.
- Trump said the war is “effectively over” militarily. But he added that he doesn’t want to formally announce that before he makes sure U.S. allies in the region are safe.
Trump made his comments hours after meeting with Qatari Prime Minister Sheikh Mohammed bin Abdul Rahman Al Thani about the negotiations with Iran.
- “Qatar, Saudi Arabia, UAE, Bahrain, Kuwait, and others, would prefer the status quo” to resuming the war, Trump said.
- The Qatari prime minister also met separately with Vice President Vance, who’s been leading the Iran talks for the administration in recent weeks. Al Thani wrote on X that he discussed with Trump ways “to progress the U.S.-Iran talks.”
- A diplomat briefed on the details of the meeting said Trump made clear to Al Thani that he wants a deal and will give negotiations a chance. But he also noted that if Iranians don’t take negotiations seriously, he could reverse course, the diplomat said.
“Should there be any change with respect to developments in the Middle East, the United States of America is locked and loaded and ready to go. At this point, however, that seems to be quite unlikely,” Trump wrote on Truth Social.
- “We are seeing positive momentum toward an end to this Conflict, and will ensure that America’s interests are at all times protected. At the conclusion of this Military Engagement, the American People, and the World, will be far safer and more secure,” he added.
Trump said that Iran is “talking very politely” with the U.S. and noted that there is a “good chance that could get ended,” seemingly referring to the military conflict.
“Hope we are gonna be able to make a deal with them,” he added.
Iran’s Supreme National Security Council secretary, Mohsen Rezaei, wrote on X shortly after Trump made his comments that the war “has wrecked the U.S. government.”
He called Trump comments about negotiations with Iran “imaginary” and called on the U.S. “to SURRENDER” to Iran’s conditions for ending the war.
In effect, the GCCs told Trump to stop hurting them.