Robots sprinting, crashing and folding towels in Beijing tell us something much bigger about China, America, Europe and where AI goes once it leaves the screen.
A few days ago in Beijing, a humanoid robot called Tiangong Ultra ran 100 metres in 8.64 seconds.
That is faster than Usain Bolt’s 9.58-second human world record, although before anyone starts rewriting the history of athletics, the comparison is obviously imperfect. Robots do not have tendons, they do not start under exactly the same conditions and World Athletics has yet to announce an electrified division.
The number that interests me much more is 21.50 seconds. That was Tiangong’s winning time at the same competition last year.
In twelve months, the winning performance on essentially the same problem went from an awkward 21.5 seconds to 8.64. During this year’s competition another robot crossed the finish line, failed to stop properly, crashed and briefly caught fire. Others fell over, lost pieces of themselves or discovered that walking around a messy environment remains considerably more difficult than looking impressive in a promotional video.
This is why I find the whole thing fascinating.
The second World Humanoid Robot Games brought together 2,056 robots, 666 teams from 16 countries and 51 events, up from 280 teams and 26 events in the inaugural edition. There were the obvious crowd-pleasers, running, martial arts, football, dancing, but also.. 21 scenario-based competitions covering factories, hotels, logistics, retail, emergency response and other environments much closer to actual work.
It would be very easy to watch the viral clips and file the event under China doing strange futuristic things for television.
I think that would miss the important part. The Robot Olympics are not really about sport. They are about building a gigantic, highly visible machine for experimentation.
The spectacle is part of the engineering
From a conventional engineering perspective, organising an Olympic Games for machines that still occasionally run into walls seems unnecessary.
If you want to improve robotic locomotion, you can test it in a laboratory. If you want a machine to assemble something, you can put it in a factory. You certainly do not need an opening ceremony.
But look at what the event produces.
Hundreds of teams have a common deadline. Their machines are tested against comparable tasks. Failures become visible rather than disappearing inside a company demo. Hardware suppliers see what breaks. Engineers see what competitors solved differently. Universities, startups, manufacturers, investors and government officials end up looking at the same problems at the same time.
And the rules themselves are changing to force more autonomy. The important transition is away from a human remotely controlling an impressive piece of machinery and toward a system that has to perceive what is happening, decide what to do and execute the action itself.
That distinction is everything. A remotely controlled humanoid is an unusual machine. An autonomous humanoid is an AI system with a body.
China is explicitly trying to push the latter from demonstration into deployment. A programme launched this year by Chinese authorities calls for humanoids and other embodied-intelligence systems to be trained in real manufacturing, logistics, retail, medical, care, safety and emergency-response environments. The stated goal is to use those environments to improve algorithms, accumulate real-machine data, validate applications, identify more than 100 high-value scenarios and build the capability for deployment at the scale of tens of thousands of units.
That does not mean tens of thousands of useful humanoid workers are about to appear next Tuesday. Far from it.
Reuters recently examined more than 150 Chinese companies competing in humanoid robotics and found an industry still struggling with intelligence, dexterity, reliability, useful training data, cost and, rather importantly, genuine customer demand. Conventional industrial robots remain better at many factory jobs. Government support is already creating fears of overcapacity, falling prices and an eventual shakeout in which quite a few of today’s robot companies disappear.
That makes the experiment more interesting, not less.
China does not appear to know whether general-purpose humanoids will become a huge industry either.
It is simply arranging to find out quickly.
We have seen this pattern before
Humanoid robotics would be easier to dismiss as another industrial obsession if China had not already run variations of this playbook elsewhere.
Electric vehicles are probably the clearest example.
China did not merely conclude that people would eventually buy more electric cars. It understood that electrification changed where value sat inside the vehicle: batteries, power electronics, software, sensors, motors and increasingly autonomous systems mattered more, while some of the historical advantages of incumbent combustion-engine manufacturers mattered less.
Then it spent years building around that shift.
In 2025, China produced around three quarters of the world’s electric cars. It accounted for more than 80% of global battery-cell production, while Chinese manufacturers supplied almost 75% of batteries deployed in electric cars worldwide.
Renewable energy shows the same appetite for scale. China added almost 500 GW of renewable capacity in 2025 alone, accounting for more than 60% of global growth that year. Solar manufacturing has become extraordinarily concentrated there as well.
Industrial robotics is less glamorous but perhaps even more relevant to what comes next. China installed 295,000 industrial robots in 2024, 54% of all installations worldwide, and its operational stock passed two million machines. For the first time, Chinese manufacturers also captured a majority of their own domestic market.
Now add AI to that industrial base.
Then add something else that receives much less attention in European discussions about China: biotechnology.
Chinese companies were responsible for 32% of global clinical trials in 2025, according to IQVIA data cited by Reuters, compared with just 2% in 2009. Licensing of Chinese-developed drug assets has exploded as global pharmaceutical companies increasingly look there for molecules worth developing internationally.
Medical technology is being pushed strategically too. China’s drug and medical-device regulator has introduced measures specifically targeting medical robots, advanced imaging, AI-powered medical devices and new biomaterials, including accelerated review mechanisms and work on new standards for AI systems, surgical robots and brain-computer interfaces.
Cars. Batteries. Renewable energy. Industrial robots. AI. Biotech. Medical devices.
At some point, “they happened to get EVs right” stops being a very satisfying explanation.
There is a broader industrial idea underneath it: identify technologies likely to change the architecture of important industries, build capacity early, create domestic demand, allow multiple companies to compete, accept substantial waste and see what survives.
It can be brutal. It can create bubbles. It can produce terrible capital allocation.
It can also produce BYD.
This is not a “China great, Europe bad” story
There is a temptation whenever China comes up to pick one of two internet positions.
Either China has discovered the secret to economic civilisation and will dominate everything by 2032, or the entire thing is debt-fuelled theatre that will collapse as soon as somebody looks carefully at the balance sheet.
Neither is particularly useful.
China’s model has serious weaknesses. Local governments can chase national priorities with such enthusiasm that industries end up with absurd duplication and excess capacity. Political direction can move capital toward sectors before real demand exists. Price wars can destroy margins. State support can keep weak businesses alive considerably longer than a functioning market normally would.
Humanoid robotics may currently contain all of those ingredients.
And despite China’s progress across many technologies, it still has important dependencies in advanced semiconductor manufacturing equipment, aerospace and other high-end systems. A five-year plan is not a magic wand.
What interests me is not whether China’s system is superior.
It is the amount of uncertainty it is prepared to tolerate in pursuit of learning.
China seems willing to finance many wrong experiments if doing so increases the probability of discovering several very important right ones.
That sounds wasteful because it is. But learning is often wasteful.
Scientific research is wasteful. Venture capital is wasteful. Product development is wasteful. Most experiments are failures by definition; if you already knew the result, you would not need to run the experiment.
And this is where China and the United States, despite being almost comically different systems, start to look surprisingly similar.
America uses another fuel: money
The American mechanism is much less centrally coordinated.
Instead of a government deciding that humanoid robots deserve 150 competing companies, Silicon Valley has become extraordinarily good at ensuring that someone will fund almost anything carrying a sufficiently large imagined future.
In AI, the difference in capital is staggering. Private AI investment in the United States reached $285.9 billion in 2025, more than 23 times the $12.4 billion measured in China. Stanford also counted 1,953 newly funded American AI companies in a single year. Those figures understate China’s total commitment because state-guided funding is not captured well in private-investment statistics, but they illustrate the American mechanism perfectly.
One investor funds autonomous laboratories. Someone else funds humanoid workers. Another funds drones. Another funds an AI wearable. Several people fund things that will eventually produce excellent post-mortems.
Capital gets sprayed across possibilities, companies die, founders move, engineers carry what they learned somewhere else, and occasionally an entirely new category appears.
China and America therefore use very different machinery, but they share something important.
Both systems generate a lot of attempts.
And that brings me to Europe.
While robots were running in Beijing, a German notary was reading
As I was reading about the Robot Olympics, a completely different story appeared in my feed.
Patrick Collison, the co-founder of Stripe, wrote that he had recently met a German founder and asked whether stories about how difficult it was to build startups in Germany were exaggerated.
The founder apparently answered that they were understated.
According to Collison’s account, during an investment round he spent an entire day having a 90-page investment contract read aloud to him by a notary, who then charged €30,000. His next company was incorporated somewhere else.
I obviously cannot verify that particular invoice, so I would not construct an economic theory around one unhappy founder and one very well-paid afternoon.
The reading part, however, is real. Section 13 of Germany’s Notarisation Act states that the content of a notarial record must be read aloud to the parties in the presence of the notary before it is approved and signed.
There are reasons why that system exists. Notaries provide identity checks, legal certainty, reliable registers and safeguards against fraud. Those things matter.
But the story landed at an almost comically appropriate moment because Europe is currently debating EU Inc., the Commission’s attempt to create a common optional corporate form for companies that want to operate across the Union.
The Commission’s proposal would allow an EU Inc. to be incorporated fully digitally, within 48 hours, for a maximum of €100 and with no minimum share capital. It exists because what we continue to call the Single Market currently contains 27 national company-law systems and more than 60 corporate legal forms.
And Germany’s Federal Chamber of Notaries has formally pushed back against important elements of it. To be precise, the Bundesnotarkammer does not oppose European competitiveness or digitisation. It says it supports the objective but argues that parts of the proposal weaken reliable identification, legal certainty and preventive controls. It commissioned a legal opinion criticising the proposal and held a dedicated parliamentary event in Brussels to make the case.
Those concerns deserve to be heard.
But the contrast still captures something that bothers me about the older Western European core.
Faced with a potentially important new technology, China often starts by asking: where can we deploy it and see what happens?
America frequently asks: who is willing to fund it?
Large parts of Europe still begin with: what existing protections could be weakened if we change the system?
That is a perfectly sensible question.
It becomes dangerous when it is consistently the first one.
Estonia makes the European story much more interesting
The good news is that you do not have to cross an ocean or adopt Chinese industrial policy to see another model.
You can go to Estonia.
Once you have Estonian e-Residency and the relevant digital identity, a company can be registered entirely online. The government’s own guidance puts the registration process itself at roughly 15 minutes to one hour. The broader e-Residency system lets people authenticate themselves, sign documents and administer a company remotely.
Estonia is not an unregulated offshore experiment.
It is a member of the European Union.
It has simply spent years approaching public infrastructure almost like a good product team would: establish the outcome that needs to be achieved, build digital identity and data infrastructure around it, then keep asking which pieces of the old process are still necessary.
That gives us a much more useful distinction than “Europe regulates, America innovates.”
Europe is not one thing.
Northern and Baltic countries have been remarkably aggressive in digitising government. Some European research institutions are world class. Germany remains an industrial engineering powerhouse. Europe has major advantages in aerospace, pharmaceuticals, automation, precision manufacturing and medical technology.
Mario Draghi’s competitiveness report did not describe a continent without capabilities. It described one struggling to turn them into enough new growth. EU companies spent around €270 billion less on R&D than American companies in 2021, and only four of the world’s top 50 technology companies were European. The European Investment Bank separately found that by their tenth year, European scale-ups raise about 50% less capital than comparable San Francisco companies, while US venture investment runs six to eight times higher than in the EU.
That is a different problem.
We have ingredients.
We struggle with the cooking time.
Maybe the metric we should watch is experiments per year
The more I look at this, the less convinced I am that the most useful comparison is simply regulation, investment or even R&D spending.
Perhaps there is another metric hiding underneath them:
How many serious experiments can an economy run every year?
Not PowerPoint pilots. Not strategy documents. Not announcements of future innovation hubs.
Experiments where something actually touches reality and can fail.
How cheaply can somebody test an idea? How quickly can a university prototype enter a factory? Can a hospital test a new system under controlled conditions? Can a startup incorporate, hire, raise a round and sell across borders without learning administrative anthropology? Can public procurement create an early customer? Can a failed founder start again without spending the next three years discussing the previous failure with six institutions?
And, critically, what happens to the information generated when the experiment fails?
China’s Robot Olympics are interesting because they compress that loop.
Build something, test it, watch it fall over, collect data, change the hardware, change the model, come back next year.
A robot that ran 21.5 seconds twelve months ago now runs 8.64.
Again, running quickly does not mean it can competently stock a supermarket or help an elderly person get dressed. The post-race collision rather elegantly demonstrated the difference between optimising a benchmark and understanding the world.
But that is precisely why you keep testing it.
Europe sometimes appears to want evidence that an experiment will work before allowing enough experiments to produce the evidence.
There is a small logical problem in there.
Why am I writing about robots in a newsletter about AI?
Because the next phase of AI will increasingly happen outside the rectangle we have spent the last few years staring into.
Until now, most people have experienced artificial intelligence through a screen. You type something, upload something or ask something; a model processes information and gives you more information.
That is already changing.
Google DeepMind calls the next challenge embodied reasoning: enabling machines not only to understand language and images but to reason about physical environments and take actions inside them. Its latest Gemini Robotics work connects multimodal reasoning with navigation, manipulation, dexterity and whole-body control.
The European Innovation Council uses the broader term Physical AI, describing systems that combine intelligent perception, adaptive learning and autonomous decision-making in complex real-world settings.
Humanoid robots are simply the most photogenic expression of it.
A drone that can inspect an unfamiliar structure, understand what looks abnormal and decide where to fly next is Physical AI.
An autonomous vehicle dealing with something it has never encountered before is Physical AI.
A warehouse robot adapting its grip to an object that was not neatly positioned for it, an agricultural machine deciding which individual plants require treatment, a surgical system responding to what it sees, or an autonomous laboratory choosing and executing its next experiment are variations of the same transition.
AI is moving from producing answers to producing actions.
And once that happens, the industrial game changes.
Software intelligence needs compute, models, data and distribution. Physical intelligence needs all of those plus motors, batteries, sensors, cameras, actuators, materials, manufacturing, energy systems, simulation, safety engineering and enormous quantities of data generated by machines interacting with the physical world.
Suddenly China’s apparently separate industrial bets begin to connect.
The electric-vehicle industry built competence in batteries, motors, power electronics, computer vision and high-volume advanced manufacturing. Industrial automation built suppliers, integrators and robot engineers. Renewable-energy investment built power infrastructure and manufacturing scale. AI contributes perception and reasoning. Medical technology contributes precision, control and regulation.
A humanoid robot is not one new industry appearing from nowhere.
It is a convergence point.
And the same will be true for drones, autonomous machinery, medical robots and whatever categories we have not invented names for yet.
Europe does not need to become China
I do not think the lesson is that Europe should copy Chinese industrial policy, subsidise 200 humanoid startups and hold robot kickboxing competitions outside the European Parliament.
Although I would absolutely attend. Nor do I think we should copy the American model wholesale and assume that sufficiently large quantities of venture capital will eventually turn every unresolved social problem into a SaaS multiple.
Europe has strengths worth protecting: engineering quality, scientific depth, worker protections, consumer trust, healthcare systems, strong institutions and a tradition of demanding evidence before technologies are deployed into high-risk environments.
The mistake would be treating those strengths as arguments for moving slowly.
Safety and speed are not opposites.
A regulatory sandbox is regulation. A controlled clinical trial is regulation. A test track for autonomous cars is regulation. A factory environment in which robots can be deployed under defined conditions is regulation.
All of them also create room to experiment.
The question is whether the objective of regulation is to make experimentation safe or whether the practical result is to make experimentation scarce.
Estonia shows that the distinction exists even inside Europe.
China shows what can happen when enormous industrial ecosystems are allowed to iterate aggressively.
America shows what happens when capital is prepared to finance uncertainty at extraordinary scale.
None has solved the problem perfectly.
But the next technological cycle will probably reward whichever systems learn fastest.
The robots can keep falling over
I suspect most of the humanoid robots currently competing in China will not become products that you or I ever use.
Many of the companies building them may disappear. Some of the government money supporting them will be wasted. There will be warehouses containing machines whose principal contribution to economic history was performing an excellent dance at a trade show.
That does not make the exercise pointless.
The actuators remain. The control systems improve. Sensors get cheaper. Manufacturers learn how to assemble the machines. Models acquire more physical-world data. Engineers move to other companies. Supply chains deepen. Standards develop. Components originally designed for humanoids end up inside machines that look nothing like people.
Eventually, one of those adjacent applications may become enormous. That is what portfolios of experiments do. You do not need every experiment to work.
You need the learning from enough of them to compound.
This is why the image I cannot get out of my head is not really Tiangong crossing 100 metres in 8.64 seconds. It is all the robots that fell over before it did.
China appears remarkably comfortable putting those failures on a stage, learning from them and coming back twelve months later with something substantially better.
Europe should not envy everything about that system.
But as AI begins to drive cars, fly drones, manipulate objects, operate machinery and act autonomously in the physical world, we should probably become much more interested in the one thing China is producing in enormous quantities alongside all those machines: attempts.
Because there is a compounding effect to experimentation that is easy to underestimate. The first robot teaches you about balance. The hundredth teaches you about manufacturing. The thousandth teaches you about cost, reliability, suppliers, maintenance, edge cases and all the strange things humans do around machines that nobody thought to include in the laboratory test. Each failure creates data, engineers move between companies, components improve, prices fall and occasionally an application nobody originally considered turns out to be the useful one.
I do not think the important question is whether humanoid robots themselves will become a trillion-dollar industry. They might. They might also end up being a rather inefficient shape for many of the jobs we currently imagine giving them. A warehouse may prefer wheels to legs, a factory may prefer a specialised arm to a synthetic human, and I personally have yet to identify the economic necessity of a robot capable of performing kung fu.
But that almost misses the point.
China is building motors, actuators, batteries, sensors, vision systems, control software, physical-world datasets, manufacturing capacity and engineers who understand how AI behaves when it meets gravity. Those capabilities do not disappear if today’s preferred robot shape turns out to be wrong. They migrate into drones, autonomous vehicles, industrial machinery, medical devices, logistics, defence, agriculture and whatever category we have not named yet.
This is where the Robot Olympics stop being an amusing Chinese curiosity and become relevant to anyone paying attention to AI.
For the first phase of generative AI, the interface was mostly a rectangle. We typed something, a model processed it, and words, images or code came back through a screen. The next frontier is considerably messier. AI is acquiring sensors, motors and the ability to produce consequences outside the computer.
A model generating a mediocre paragraph is annoying. A model making a mediocre decision at 120 kilometres per hour requires a rather different engineering discipline.
That shift from digital AI to physical AI connects technologies that we often discuss as separate industries. Autonomous cars, drones, robots, intelligent medical equipment, automated laboratories and adaptive industrial machinery all require some combination of perception, reasoning, action and continuous learning from the physical world.
And suddenly China’s previous industrial bets begin to look less independent from one another.
Electric vehicles created enormous capabilities in batteries, electric motors, power electronics, sensors, computer vision and high-volume advanced manufacturing. Industrial automation created suppliers and customers for robotics. Renewable energy strengthened power electronics and manufacturing scale. AI provides increasingly capable perception and reasoning. Humanoid robotics sits somewhere near the intersection of all of them.
That does not guarantee China will win the next technological cycle. Nothing does.
The United States has advantages China still struggles to reproduce: extraordinary capital markets, leading frontier-model companies, deep research ecosystems and an almost pathological willingness to finance ideas whose business plans occasionally resemble science fiction with a spreadsheet attached.
Europe also has cards that should not be underestimated. In fact, physical AI ought to suit us unusually well. Europe remains exceptionally strong in industrial engineering, automation, aerospace, automotive technology, pharmaceuticals, medical technology, precision manufacturing and scientific research. Germany, Italy, France, the Nordics, the Netherlands and several Central European countries contain exactly the kind of industrial environments where physical AI can become useful rather than merely impressive.
And Estonia, along with some of the other smaller European economies, demonstrates something equally important: being European does not automatically require institutional arthritis.
This is why I am uncomfortable with the easy conclusion that China is dynamic, America innovative and Europe doomed. It is intellectually satisfying because it compresses a complicated problem into three stereotypes, but it does not help much.
The more useful distinction is between systems that make experimentation cheap enough to happen frequently and systems that make experimentation exceptional.
America does it primarily through capital. China combines capital, industrial policy, manufacturing infrastructure and enormous domestic test environments. Some European countries have shown that regulation, trust and speed can coexist perfectly well. The problem is that too much of the older Western European core still behaves as if avoiding visible mistakes is evidence of institutional quality.
Sometimes it is. Sometimes it simply means somebody else gets to make the discoveries.
The German notary debate around EU Inc. is a tiny example, but that is precisely why I find it useful. Nobody is going to lose the robotics race because incorporating a company requires a notary. Economies rarely decline because of one spectacularly stupid rule. They accumulate thousands of perfectly defensible frictions, each supported by a legitimate concern, an existing institution and someone explaining why changing it would create risk.
Then, twenty years later, everyone wonders why nothing new seems to scale.
Perhaps this is the part of China’s Robot Olympics that Europe should pay attention to. Not the medals. Not the dancing machines. Not even the 8.64-second sprint.
The willingness to put unfinished technology into the world, watch it embarrass itself, collect the evidence and come back twelve months later with something dramatically better.
There is plenty about China’s economic and political model that Europe should not copy, and plenty about Silicon Valley’s capital allocation that should probably come with adult supervision.
But there is one habit worth stealing from both: make it easier to try.
Because the next phase of AI will not be developed entirely inside data centres. It will increasingly be tested on roads, inside factories, in hospitals, warehouses, farms, laboratories and homes, where progress depends on discovering thousands of things that cannot be learned from a benchmark.
China has already built an Olympic Games for doing exactly that.
It looks slightly ridiculous.
That may be the most serious thing about it.