Signals Matter More Than Products

Viewed separately, these are simply two corporate milestones.

Musk appears at the Optimus production line in Fremont.

Unitree moves into the IPO registration process.

But viewed together, they appear to be sending the same message to the market.

For founders, signals influence fundraising.

For investors, signals shape expectations.

For an emerging industry, signals often reshape the market long before products do.

Startups Raise Confidence Before They Raise Capital

The biggest challenge for a startup is rarely technology.

It is the shortage of believers.

Investors wonder whether the market is real.

Suppliers question whether demand will last.

Customers worry about long-term support.

Banks worry about cash flow.

Top talent wonders whether the company will survive the next funding round.

Everyone is asking the same question:

Who else believes this will work?

The best startups do more than build products.

They build confidence in the future.

Musk Is Sending a Demand Signal

The most important part of Musk’s factory visit is not the photo.

It is the location.

A laboratory represents research.

A factory represents production.

Production represents procurement.

Procurement represents years of future orders.

To suppliers, an active production line is worth far more than a hundred product launches.

Because suppliers invest not in today’s revenue, but in tomorrow’s cash flow.

What Musk is signaling is not that humanoid robots have already succeeded.

He is signaling that the industry is now worth investing in.

Unitree Is Sending a Supply Signal

Unitree’s IPO carries a different message.

Many founders think of an IPO primarily as a financing event.

For manufacturing businesses, however, its greater value lies in reducing uncertainty across the entire commercial ecosystem.

Suppliers become more willing to expand capacity.

Customers become more comfortable signing long-term agreements.

Banks become more willing to extend credit.

Industrial investors become more willing to commit capital.

An IPO does more than raise money.

It tells the market that the company intends to be around for the long haul.

That certainty becomes a business asset in its own right.

Together, They Are Educating the Market

From a competitive perspective, it is easy to conclude that Tesla and Unitree are racing toward the same destination.

From an investor’s perspective, something else is happening.

Every new industry needs more than one successful company.

Without multiple market leaders, there is no stable supply chain.

Without a stable supply chain, there is no large-scale manufacturing.

Without scale, costs remain high.

And without lower costs, markets struggle to expand.

This was true for electric vehicles.

It was true for cloud computing.

It was true for smartphones.

Humanoid robotics is unlikely to be any different.

Together, They Reduce Supply Chain Risk

For suppliers, this is not an either-or decision.

Tesla scaling production means growing global demand.

Unitree going public means more stable demand from China.

Whether orders come from America or China is not the deciding factor.

The real question is whether future demand is becoming more predictable.

When both companies send long-term commitment signals at the same time, suppliers gain confidence that humanoid robotics is no longer a temporary trend.

It is becoming a sustainable industry.

That confidence encourages new factories.

New equipment.

New R&D investment.

Those investments reduce manufacturing costs across the entire ecosystem.

In the end, the beneficiaries are not only Tesla or Unitree.

Every company entering the industry stands to gain.

What Changes Is Market Expectation

Investors rarely invest in today’s business.

They invest in the speed at which tomorrow’s market will emerge.

An industry does not truly take off because one company succeeds.

It takes off because participants begin believing that others will succeed as well.

Economists describe this as coordinated expectations.

Once that shared expectation takes hold, the independent decisions of thousands of companies reinforce one another, pushing the industry toward a new equilibrium.

That is why the valuation of the same company can increase tenfold within a few years.

The technology may remain largely unchanged.

What changes is the market’s expectation.

One Validates Demand. The Other Validates Supply.

The most interesting aspect of these two events is not who moved first.

It is how each reinforces the other.

Musk’s factory visit validates demand.

Unitree’s IPO validates supply.

One strengthens confidence that a market exists.

The other strengthens confidence that the industry can support it.

Together, they reduce uncertainty across the entire robotics ecosystem.

Win-Win Is How Industries Mature

For founders, this makes fundraising easier.

For investors, it lowers perceived risk.

For suppliers, it justifies expanding capacity.

For the robotics industry, it marks the crossing of a far more important threshold.

Not a technological threshold.

A confidence threshold.

Industries rarely explode because of a single breakthrough product.

They grow because enough people begin believing the market will exist for decades.

Capital follows.

Supply chains expand.

Costs decline.

Demand accelerates.

A self-reinforcing cycle begins.

Years from now, few people will remember which production line Musk toured or the exact date Unitree filed for its IPO.

What they may remember instead is that this was the moment when capital, manufacturing, and supply chains all reached the same conclusion:

Humanoid robotics was no longer an experiment.

It had become a long-term business.

For two decades, real estate was the economy’s largest container for capital. Land, mortgages, construction, home appliances, and local government finance formed a self-reinforcing cycle that absorbed enormous amounts of liquidity. Later, the internet briefly assumed a similar role. Instead of concrete and steel, it monetized traffic, advertising, payments, and consumer spending, transforming capital into ever-higher valuations.

Today, neither engine is sufficient.

Property is deleveraging. Consumer internet has entered a zero-sum phase. Software-as-a-Service has become increasingly commoditized. Price competition dominates many consumer sectors. The result is an economy facing an abundance of capital but a shortage of investable assets capable of generating both attractive returns and broad economic spillovers.

This is where embodied AI enters the story.

In 2025, China’s embodied AI and robotics sector attracted RMB 73.5 billion in financing across 744 investment rounds. Momentum accelerated further in the first half of 2026, with disclosed funding exceeding RMB 46 billion across 288 transactions.

This is more than another venture capital craze.

It represents China’s search for a new physical container for capital.

Unlike large language models, which primarily create value inside digital interfaces, embodied AI reconnects artificial intelligence with the physical economy. A chatbot requires cloud computing; a humanoid robot requires motors, reducers, sensors, batteries, semiconductors, production lines, testing facilities, field engineers, and enormous quantities of real-world data.

One consumes computation.

The other consumes—and revitalizes—an industrial ecosystem.

That distinction matters.

When capital flows into an embodied AI company, it rarely remains on the company’s balance sheet. It finances research campuses, testing facilities, manufacturing plants, and production equipment. It becomes purchase orders for suppliers of motors, force sensors, cameras, precision bearings, actuators, PCBs, and industrial software. It pays salaries for mechanical engineers, AI researchers, robotics technicians, data operators, and manufacturing workers.

In other words, embodied AI is not merely another technology sector.

It is a mechanism for redistributing capital throughout the real economy.

Its greatest attraction lies not in today’s robot sales, but in its ability to channel investment across land, labor, manufacturing, software, computing infrastructure, and ultimately national strategic capability.

That is precisely why the sector fits China’s economic structure.

Over the past decade, China has built the world’s most comprehensive electric vehicle supply chain. Batteries, electric motors, lightweight materials, autonomous driving software, precision manufacturing, and large-scale production have created industrial capabilities unmatched in scale.

Humanoid robots inherit much of that foundation.

Robot joints borrow from electric drivetrain technology. Perception systems leverage advances originally developed for autonomous vehicles. Manufacturing processes resemble automotive supply chains. Eventually, the robots themselves can be deployed inside automotive factories, logistics centers, power grids, warehouses, and industrial facilities.

Embodied AI therefore represents more than a new industry.

It is arguably the second major outlet for China’s advanced manufacturing capacity after electric vehicles.

If automobiles transformed how people move, robots may transform how physical work is performed.

This helps explain why industrial groups, local government funds, and state-backed investors have entered the market so aggressively.

They are not simply financing robot companies.

They are investing in the next industrial platform.

For local governments, embodied AI promises employment, industrial clusters, tax revenues, and regional competitiveness. For state-owned enterprises, it offers automation for dangerous, repetitive, and labor-intensive operations. For automotive manufacturers, it provides technologies that may ultimately improve their own factories.

Embodied AI sits at the intersection of three strategic priorities.

Artificial intelligence needs access to the physical world.

Manufacturing requires new sources of high-value growth.

National security increasingly depends on technological self-sufficiency.

Robotics is where all three converge.

This also explains why sectors such as power transmission, energy infrastructure, industrial inspection, emergency response, and logistics are emerging as the first commercially viable deployment scenarios.

These industries share several characteristics: relatively structured environments, high labor costs, significant safety risks, and measurable economic returns. Robots do not need to replace every human task. They simply need to perform specific high-frequency, high-risk operations more safely, more consistently, and more economically than people.

Yet beneath today’s enthusiasm lies an uncomfortable reality.

Capital formation is advancing far faster than industrial validation.

Many robotics companies have already achieved multi-billion-dollar valuations, while a substantial portion of their revenue still originates from research institutions, demonstration projects, government pilots, and laboratory procurement. Too many robots remain demonstration platforms rather than productive assets.

Capital markets celebrate the phrase “the year of mass production.”

The more important question is different:

Are these robots working on factory floors—or performing on exhibition stages?

The distinction is critical.

Selling robots to universities and research laboratories does not constitute industrial adoption. Genuine commercialization begins only when robots generate measurable returns in automotive assembly, power grid maintenance, warehouse automation, chemical processing, healthcare, and elderly care.

Today, many humanoid robots still face significant technical constraints in precision, reliability, operating endurance, maintenance costs, and general-purpose task execution. Industrial customers require millimeter-level accuracy, continuous operation, predictable maintenance, and verifiable safety—not impressive demonstration videos.

Private markets can price stories.

Public markets price cash flow.

That is where the valuation cliff begins.

As the current generation of embodied AI companies eventually enters public markets, investors will shift their attention away from technological narratives toward revenue quality, gross margins, customer retention, operating efficiency, and free cash flow. Companies whose businesses depend primarily on research grants, government subsidies, or promotional demonstrations may discover that valuation premiums disappear surprisingly quickly.

None of this implies that embodied AI is a bubble.

Quite the opposite.

Industries that reshape economies almost always generate speculative excess in their early years. Railways, electricity, the internet, and electric vehicles all experienced periods of overinvestment, duplicated infrastructure, and inflated valuations.

Bubbles are not evidence of failure.

They are often the price capital pays for financing infrastructure before demand fully materializes.

The real question is not whether speculation exists.

It is whether, after speculation subsides, the industry leaves behind productive assets, resilient supply chains, engineering talent, proprietary data, and commercially viable applications.

Viewed through this lens, China’s embodied AI financing boom represents a form of macroeconomic substitution.

As legacy growth engines slow, capital searches for a new destination.

As digital AI reaches the limits of software monetization, investment seeks a physical outlet.

As manufacturing searches for its next productivity revolution, robotics offers a new industrial narrative.

As technological competition increasingly defines geopolitical competition, embodied AI becomes the physical expression of digital sovereignty.

Money flowing into embodied AI is therefore purchasing far more than equity in robotics startups.

It is purchasing an option on whether China can evolve from manufacturing products to manufacturing labor itself.

That option is extraordinarily expensive.

It is also extraordinarily risky.

If humanoid robots fail to leave the laboratory, fail to achieve industrial-grade precision and reliability, or fail to deliver measurable returns in real production environments, today’s financing boom will ultimately produce little more than expensive hardware, inflated valuations, and exhausted investor patience.

But if they succeed, China will gain something far more valuable than a handful of robotics unicorns.

It will acquire a new economic flywheel.

Capital finances robots.

Robots strengthen manufacturing.

Manufacturing generates data.

Data improves AI models.

Better AI creates better robots.

Better robots expand into new industries.

Capital begins to circulate once again.

That, more than any technological breakthrough, explains why embodied AI has become China’s hottest investment theme.

It is not merely another industry.

It is the country’s next physical container for capital.

Some people spend their entire lives searching for love, yet the moment genuine kindness finally reaches them, they push it away.

You support them, and they question your motives.

You trust them, and they assume it’s temporary.

You show patience, and they begin testing your limits.

Eventually, the people around them become exhausted and walk away. Ironically, every departure only reinforces the belief they already held:

“See? Nobody ever stays.”

This isn’t simply a personality flaw. It’s what psychologists call an Internal Working Model—the unconscious blueprint through which we interpret relationships.

How we see the world is often decided long before we enter it as adults.

After World War II, psychologist John Bowlby studied children who had been separated from their parents for extended periods.

His conclusion changed modern psychology.

The defining factor in a child’s emotional development wasn’t material comfort. It was whether they had experienced a stable, reliable attachment figure.

Children who consistently received warmth and emotional security developed an unconscious assumption:

The world is generally safe. People can be trusted.

Children who grew up with neglect, inconsistency, or constant criticism often formed the opposite belief:

People eventually leave. Trust is dangerous.

These assumptions don’t disappear at adulthood.

They quietly follow us into every friendship, every marriage, every company we build.

What many people casually describe as “someone who lacks love” is, in reality, someone whose sense of emotional security never had the chance to develop.

Another famous experiment illustrates this perfectly.

Psychologist Harry Harlow raised infant monkeys with two artificial mothers.

One was made of wire and provided milk.

The other was covered with soft cloth but offered no food.

According to conventional wisdom at the time, the monkeys should have preferred the one providing nourishment.

Instead, they spent almost all their time clinging to the soft cloth mother, running to the wire mother only when hungry before immediately returning.

When frightened by unfamiliar environments, they instinctively sought comfort from the cloth mother.

The experiment fundamentally changed developmental psychology.

Humans—and many animals—need far more than food or material support.

They need consistent emotional safety.

This pattern doesn’t disappear in the workplace.

Every experienced manager has met someone like this.

A routine piece of feedback feels like personal rejection.

A delayed reply from a colleague becomes evidence of exclusion.

A project proposal isn’t selected, and suddenly the entire organization is perceived as being against them.

Instead of solving problems, they spend enormous amounts of energy interpreting other people’s intentions.

Economics teaches us that every resource is scarce.

Attention is no exception.

Every hour spent defending imagined threats is an hour not spent creating value.

The result is predictable:

Communication slows.

Collaboration becomes expensive.

Decision-making deteriorates.

Trust erodes.

This is precisely why leading companies increasingly invest in psychological safety.

Organizations built on fear rarely outperform organizations built on trust.

The same phenomenon appears repeatedly in startups.

Some founders insist on reviewing every line of code, approving every product decision, and participating in every conversation.

It isn’t always because their teams lack competence.

Often, they simply cannot believe that anyone else will do the job well enough.

As the company grows, so does the founder’s workload.

Eventually, the bottleneck is no longer funding, product, or market demand.

It becomes the founder’s inability to trust.

Many people describe this as a control problem.

Psychologically, it is often an anxiety-management strategy.

Control temporarily reduces uncertainty.

But scaling a company requires giving up control.

Some of the greatest founders don’t evolve because they master better management frameworks.

They evolve because they gradually learn to trust other people.

The same dynamic quietly destroys personal relationships.

Some people repeatedly ask,

“Do you really love me?”

They’re not actually searching for an answer.

They’re searching for certainty.

But certainty cannot be permanently borrowed from another person.

Today’s reassurance fades by tomorrow.

The cycle repeats.

Reassurance.

Temporary relief.

Doubt.

More reassurance.

Eventually, the other person becomes emotionally exhausted.

Psychologists call this an inability to internalize security.

The problem isn’t that love isn’t being offered.

The problem is that it cannot be retained.

Fortunately, psychology offers hopeful news.

Early experiences influence us, but they do not imprison us.

Many people who grew up without emotional security eventually build healthy relationships through self-awareness, therapy, supportive partners, close friendships, and years of intentional growth.

Psychologists call this earned secure attachment.

Our childhood shapes our starting point.

It does not determine our destination.

From an economic perspective, relationships resemble long-term capital investments.

Trust is capital.

Cooperation is compound interest.

People who cannot trust others struggle to build enduring partnerships.

People who cannot receive kindness often fail to sustain it.

Whether in friendship, marriage, or entrepreneurship, the most valuable asset is rarely intelligence, capital, or technical ability.

It is possessing a psychological structure stable enough to trust others, respond to goodwill, and extend it in return.

When that foundation exists, the transaction costs of human cooperation fall dramatically.

Teams move faster.

Organizations become more resilient.

Businesses scale more naturally.

Conversely, those who interpret every act of kindness as manipulation, every offer of help as control, and every relationship as a future betrayal pay an invisible tax throughout their lives.

They don’t merely lose a friend, a partner, or an employee.

They lose repeated opportunities to accumulate the one form of capital that compounds faster than money itself:

Trust.

Not because hatred creates value, but because rivalry changes incentives.

In economics, the biggest challenge for a new strategic industry is not only technology. It is coordination. Capital, customers, governments, suppliers, and talent all move according to different incentives. A visible rival helps align them.

This is the economic logic behind the Thucydides Trap.

When a rising power challenges an established power, fear increases on both sides. That fear can be dangerous politically. But economically, it also creates mobilization.

Airbus is a classic example.

Boeing was founded in 1916. Airbus was founded in 1970. When Airbus appeared, Boeing had already spent more than half a century building dominance in commercial aviation.

Europe did not create Airbus merely to build another aircraft company. It needed a rival narrative. Boeing became the visible enemy. That narrative helped European governments, airlines, suppliers, engineers, and investors justify long-term coordination.

The same pattern appeared in semiconductors.

In the 1970s and 1980s, Japan’s semiconductor industry rose quickly and challenged American leadership. For the United States, Japan became the visible rival. That rivalry helped turn semiconductors from a business issue into an industrial strategy issue.

This is the useful side of the Thucydides logic.

A rising competitor makes the established player feel pressure.

An established competitor gives the rising player a target.

Both sides become easier to mobilize.

This is why American robotics companies and Chinese robotics companies can both benefit from a strong rival narrative.

For American robotics companies, China can be framed as the rival in physical AI, manufacturing automation, and humanoid robotics.

For Chinese robotics companies, the United States can serve the same role.

The point is not to create hostility.

The point is to create urgency.

A rival narrative raises the perceived cost of inaction. It makes investors more willing to fund, customers more willing to adopt, governments more willing to support, and ecosystems more willing to organize.

Industrial history rarely moves only because someone says:

“We are building a better product.”

It moves faster when people believe:

“If we do not move now, we will fall behind.”

Airbus had Boeing.

American semiconductors had Japan.

Japan had American industrial leadership.

Robotics may follow the same rule.

The company that defines the rival does not merely define its competitor. It defines the economic reason for everyone else to join.

Someone once asked me what embodied AI product managers struggle with most every day.

After thinking about it, I realized it’s probably not requirements, roadmaps, or even technology.

In an early-stage startup—especially one fueled by venture capital—the question that keeps coming back is surprisingly simple:

Will what I’m doing today increase the company’s valuation?

It sounds a little cynical, but that’s often the reality.

Sometimes a team spends weeks refining a feature. The engineering is solid, the product works, and everyone feels good about the result. Yet nothing really changes for the company.

On the other hand, some initiatives may not look impressive from a pure product perspective, but they can completely change how investors view the business.

Landing a flagship customer. Deploying the first meaningful batch of robots. Proving that proprietary data can be collected at scale. Discovering a use case that can be replicated across thousands of sites.

None of these things necessarily make the robot smarter overnight. But they can make the company’s future look dramatically different.

That’s why many embodied AI product managers aren’t really doing traditional product management.

They spend much of their time deciding where limited resources should go:

What is worth building?

What can become a defensible advantage?

What moves the company one step closer to the next stage?

Because in this industry, technology changes constantly. Models change. Hardware changes. Markets change.

The only thing that cannot stop is the company’s ability to prove that its future is becoming more valuable.

And more often than not, the product manager is standing at the front of that process.

The more time I spend in robotics, the more I think this isn’t really a camera question.

It’s a startup-stage question.

Early-stage startups are rarely buying hardware.

They’re buying time.

A basic camera can absolutely capture video. But once you start building a robot, the camera itself is only a small part of the problem.

You still need drivers, calibration, synchronization, ROS integration, data pipelines, debugging tools, and long-term maintenance.

None of these problems are particularly difficult on their own, but together they can consume weeks or months of engineering effort.

For a startup trying to validate a product, a few months of engineering time are usually far more expensive than a few hundred dollars of hardware.

That’s why RealSense became so common.

You plug it in and immediately get RGB images, depth information, calibration tools, SDK support, and a huge community that has already solved many of the problems you’re about to encounter.

Instead of becoming camera experts, teams can focus on collecting data, training models, and understanding customer needs.

What’s interesting is that recent progress in imitation learning and VLA models has also shown that many robotics tasks don’t require the visual quality people once assumed.

For tasks like grasping objects, opening drawers, organizing items, or basic manipulation, stable data often matters more than ultra-high-resolution images.

In that sense, RealSense is often a very reasonable choice for the validation stage.

But that doesn’t mean it should stay forever.

One mistake I occasionally see is treating a development tool as a product component.

When a company begins productization, the conversation changes completely.

The question is no longer:

“Can this work?”

The questions become:

* Can we manufacture it at scale?

* Is the supply chain reliable?

* Is the power consumption acceptable?

* Does it fit the industrial design?

* Is the cost appropriate for our target market?

* Can it be serviced and maintained efficiently?

These are product questions, not research questions.

At that point, the vision system should be re-evaluated from the ground up.

The best camera for development is not necessarily the best camera for production.

Some products may only need low-cost RGB cameras. Others may require custom stereo systems. Some may use multiple cameras distributed across the robot. The right answer depends entirely on the product and its use case.

Personally, I think the transition should happen much earlier than many teams expect.

Not at the thousandth robot.

Not after mass production starts.

The moment a company commits to a product direction and enters productization, it should begin designing the vision system around manufacturing, cost, power consumption, industrial design, and supply-chain realities.

RealSense is an excellent development tool.

A product, however, should be designed for customers, manufacturing, and scale.

Those are very different goals.

In robotics startups, it’s easy to optimize too early for hardware cost and miss the bigger picture.

The real objective in the beginning is not building the perfect camera system.

It’s finding something customers actually want.

Once you’ve found that, then it’s time to build the camera system your product truly needs.

Twenty-seven slides.

Market size. Competitive analysis. Five-year projections. A beautifully designed revenue model that stretched all the way to Series C.

He walked into the investor meeting confident. The investor flipped through the deck in less than three minutes.

“Looks good,” he said.

The founder smiled.

Then the investor closed the laptop.

“Tell me about the time your company almost died.”

The founder paused.

He talked about the customer who refused to pay.

The engineer who quit before launch.

The night he drove six hours to fix a machine himself because nobody else could.

The month he went without salary.

The deal he walked away from because the product wasn’t ready.

For the next hour, nobody mentioned the deck.

They talked about customers, failures, mistakes, and decisions. A week later, the company received its first investment.

Years afterward, the founder asked the investor why he invested.

The investor laughed. “It wasn’t the plan.” “What was it then?” “The person who could survive when the plan stopped working.”

The founder looked at the old deck stored somewhere in the cloud.

Almost every number was wrong.

The market had changed.

The product had changed.

The business model had changed.

Even the company mission had changed.

Only one thing remained the same:

The person who kept adapting after every version of the plan failed.

That, it turned out, was what the first check had really funded.

Because it clearly exposes a reality that has never changed: every country acts according to what it believes best serves its own interests. In the AI era, many people had not fully appreciated this fact. Now, they are starting to wake up to it.

Imagine a future where Fable has been widely deployed in humanoid robots, national infrastructure, hospitals, factories, and even ordinary homes. Then, with a single policy decision or remote command, an entire country’s critical systems could be disrupted or disabled. That is a frightening scenario.

After leaving Qianxun last year, I designed a conceptual eldercare robot called AroOne. One of its core AI design principles was straightforward: local model deployment and on-device inference.

A truly reliable humanoid robot—the kind people can trust and depend on—must be able to operate locally. This is especially true for companion robots. They should perform AI inference on-device, function completely offline when necessary, and remain independent of external cloud services.

The discussion around Fable is not really about one model or one company. It is a reminder that autonomy, resilience, and trust are not optional features in embodied AI—they are foundational requirements.

Provide a caption (optional)

Back then, people were amazed by the breakthrough of a robot mastering a single, specific task.

Fast forward to 2026. Embodied AI companies are raising billions of dollars, while compute, models, and datasets have improved by several orders of magnitude. Yet towel folding remains one of the industry’s favorite demonstration tasks.

From an entrepreneurship and investment perspective, technological progress is not always linear. Often, a decade is spent not answering “Can it be done?” but rather “Can it be done reliably, affordably, and at scale?”

I don’t chase every wave, nor do I mock every bubble.

What truly matters is not whether a robot can fold a towel today. What matters is who, ten years from now, can turn towel folding into a business, technology into value, and a demo into an industry.

Yong’s Field Memo | June 2026For hard tech founders, GPs/LPs, and anyone still debugging a line at 2 a.m.

In March I posted a timeline on X of everything I’ve shipped over twenty years — a China robotics blog, BMW/Tesla line systems, JIEQI, OmniEdge, DeepFashion, product work at Spirit AI, the elder-care robot concept AroOne https://77.tech

, then OmniNervous and https://OpenKoi.ai. Long list. Most people stop scrolling halfway through.

When I re-read it myself, I skip two entries too:

• RoboticsFAQ (2012–2016) — a non-profit robotics Q&A community

• Schulung / 诩阆 (2015–2017) — a free engineer training MOOC

No ARR. No valuation. No Demo Day screenshots. On qianyong.me they sit in the Non-Profit Projects section — just a domain and a date range.

But if you read backward like an investor — or like a founder still asking where does the first production order actually come from — these two projects carry a signal that’s still worth money today:

In hard tech, the earliest PMF often doesn’t show up in a contract. It shows up when peers keep coming back with the same field question.

2012 vs. 2026: Same industry, different universe

In 2026, more than 150 entities in China claim to be building embodied AI or humanoids. Funding activity in the first half alone is back near historical peaks. Every few weeks a new demo drops — backflips, tea pouring, warehouse runs.

2012 was a different density. No one said “embodied AI.” No one debated VLAs. Capital wasn’t really in the room. The scarce asset wasn’t an algorithm or a demo — it was an engineer who could keep a production line running.

Those years I was bouncing between factory floors globally. Days: coordinating process, sensors, cameras, PLC–robot handshakes. Nights: hotel room, forums, the same threads resurfacing every few weeks:

• Critical paragraphs in English manuals no one on the integrator side could parse;

• Alarm codes senior techs could work around, but no searchable write-up existed;

• Junior engineers who wanted structured training — courses were either expensive or useless on a real line.

My read at the time was blunt: what blocked people wasn’t usually a missing technology. It was not knowing anyone who’d seen that exact failure mode before. Information asymmetry doesn’t stay abstract — it becomes schedule slip, rework, and hidden labor burn. Founders who’ve done integration and investors who’ve run DD both know that math.

That’s how RoboticsFAQ started. Q&A, papers, application videos — ask, answer, archive, search.

RoboticsFAQ: A pre-revenue demand database

Stack RoboticsFAQ against today’s paid communities, knowledge planets, or startup “developer ecosystems” and it looks primitive. No recommendation engine. No live commerce. No fundraising narrative.

What it had — and what a lot of hard tech still lacks — is this: questions from the field, not from the deck.

Active threads were narrow and ugly:

• Intermittent e-stop on a six-axis — which log patterns usually point to which root causes?

• Vision calibration drifting under line vibration — mechanical first or exposure first?

• Multi-vendor PLC talking to a robot — which gateway is cheaper to maintain long-term?

No single right answer. But reproducible debug paths. Leave the path in the thread; the next person picks up where you left off.

If I were an LP diligencing early hard tech, I’d separate two kinds of “community”:

Q&A is horizontal, not vertical. The asker might have two years on the job; the responder might have ten on the floor. Bad answers get corrected in-thread. That error-correction loop is closer to how plants actually work than any one-way training program.

From 2012 to 2016 I never tried to monetize it. Servers were cheap. I covered dev costs myself. Content came from peers. No sales org. Charging would have filtered out exactly the users I knew best — small integrators, new engineers, maintenance crews in second- and third-tier cities.

Schulung: When point fixes aren’t enough

RoboticsFAQ handled spot failures. By ~2015 I kept seeing a different gap: finding one alarm-code fix wasn’t enough. People didn’t know what to learn next.

A junior could follow an existing program. They were still a long way from designing a full robot system. Integrator owners could hire — but couldn’t tell whether someone was missing PLC depth, vision, robot dynamics, or the ability to own the whole stack.

So I built Schulung (诩阆):

• Chinese name: 诩阆

• English name: Schulung

• Domain:

  xulung.com

• Format: engineer training MOOC

• Timeline: 2015–2017 (I launched operations in 2016)

Also free.

Crude analogy: RoboticsFAQ was a field triage manual. Schulung was an onboarding syllabus.

MOOCs weren’t new. The hard part in industrial automation was the gap between lab curriculum and line reality — grease, takt time, heterogeneous equipment. Schulung tried to fill that gap: how to read safety interlocks, offline programming vs. teach pendant workflow, PLC handshake timing, when to call the OEM vs. when to grep the logs yourself.

Not sexy. Very usable for integrator teams outside tier-one cities — closer to shipment than a keynote on “the next decade of humanoids.”

For founders the through-line is obvious: community tells you where users get stuck; curriculum tells you where talent is thin. Stack both and you’re approaching scalable delivery — not just a moving demo.

One timeline, not two careers

2004 China robotics blog
2006 Weibo robotics channel (于仁颇黎@机器人)
2012 RoboticsFAQ (Q&A community)
2015 Schulung (training MOOC)
2008+ Industrial robot systems (wafer handling, laser cutting, BMW/Tesla auto lines)
2018 JIEQI (industrial IoT)
2021 OmniEdge (enterprise mesh VPN — 7,000+ users across 26 countries)
2023 DeepFashion (fashion design AIGC)
2025 Spirit AI product work / AroOne / OmniNervous / OpenKoi

Community work and commercial products overlapped. Not chapter one charity, chapter two startup.

RoboticsFAQ and Schulung kept me inside real engineer questions for years — which issues resurfaced every few weeks, which topics paid courses skipped, which traps were documented in English but never translated for the Chinese field context. Later, building JIEQI and OmniEdge, the same needs kept reappearing in the roadmap: remote troubleshooting, connect gear across geographies, reduce friction to get the right answer fast. OmniNervous rewrote the P2P layer in Rust for AI and industrial fleets — different stack, same obsession I picked up in those two communities:

Field knowledge shouldn’t be hostage to geography.

In one line:

RoboticsFAQ / Schulung → JIEQI → OmniEdgeTransport went from forum threads and screen-recorded lessons to mesh VPN. The problem didn’t: machine in city A, expert in city B, line can’t wait.

For investors, that arc matters more than “founder did philanthropy.” It shows product instinct can run a user loop at zero revenue, then migrate to billable infrastructure. In hard tech DD, that’s a more believable pattern than “raise first, build community later.”

If I were an LP: three signals I’d still underwrite

RoboticsFAQ and Schulung never hit ARR or a cap table. But they left fingerprints I still use in diligence — not slide-deck methodology, field observations:

Signal 1: User insight before monetization.

What got searched and re-asked often became feature priority later — not from a competitive matrix, but from questions already asked a hundred times in the forum.

Founder read-through: if a hard tech team’s principals have never lived inside their own user community — GitHub issues, Discord, integrator WeChat groups — I’d question whether they actually know where deployment breaks. Demos can be outsourced. Debug paths can’t.

Signal 2: “Can ship” ≠ “can teach.”

Schulung forced me to modularize a decade of field work. Doing the job and diagnosing where someone else stalls are different skills. The second one sets hiring velocity, support cost, and retention.

Investor read-through: I’ve seen teams with gorgeous algorithm metrics in the deck and empty training docs, delivery SOPs, and integrator enablement. Their ramp usually runs 1–2× slower than the financial model assumes.

Signal 3: Free community ROI isn’t ARR — it’s trust inventory and talent pipe.

Servers, dev, course production, moderation, spam removal — all real cost. The return: peers remember you showed up when it wasn’t a product pitch. Later, when you ship OmniEdge, publish a field memo, or make an investment call, they listen past the first sentence.

Founder read-through: early hard tech is rarely short on Twitter followers. It’s short on the first integrators and field engineers willing to test when you’re still unreliable. Community is one of the cheapest ways to recruit that cohort — not the only way, one of the most underpriced.

2026: Capital is loud. The help thread is still the same.

Robotics today has orders of magnitude more capital than 2012. Humanoid demos reset expectations every few weeks. Government guidance funds are everywhere.

I still think about RoboticsFAQ’s plain thread titles and Schulung’s unglamorous slides.

Back then there was no embodied AI label, no VLA discourse, no humanoid valuation comps. One scene on repeat: machine down, line lead waiting, engineer posting, hoping someone who’s seen that floor before replies.

One thing hasn’t changed — and it’s still one of the most expensive mistakes in the category:

The gap between lab Sim and line Real isn’t just algorithms. It’s who has read the English manual and who has cleared that e-stop pattern before.

If you’re underwriting embodied AI in 2026 and you only track demo reels and round sizes — not deployment friction, integrator enablement, or how field knowledge compounds and transfers — you’re probably missing the layer RoboticsFAQ and Schulung lived in every day.

Those two projects were my answer to that layer when I was younger. Free not because they weren’t valuable — because the person on the line usually had no time and no budget for a ticket in.

About the author

Yong (涌) — 20-year industrial automation and robotics operator; three exits across IIoT, AIGC, and OSS in the US and China. Former Head of Products at Spirit AI.

I didn’t fully appreciate how hard it was until I built and delivered my first AIoT hardware device to a customer seven years ago.

Getting a hardware product to market isn’t just about designing a circuit board. It’s everything that comes with it: PCBAs, enclosure tooling and fixtures, wiring, antennas, manufacturing, testing, and quality control.

Then there’s the software stack: device drivers, embedded software, cloud services, SaaS platforms, mobile apps, deployment, and ongoing maintenance.

Every component has dependencies. Every small change can create unexpected problems somewhere else. Success requires all of these pieces to work together as a single system.

That first product taught me a lesson I still carry today: building hardware is difficult, but building hardware that works reliably in the real world is even harder.

You learn by doing.

I’m excited to share that my books are now available on Amazon:

📘 How to Build Fashion AI SaaS From Scratch
A practical guide to building an AI-powered fashion platform using Stable Diffusion, modern cloud infrastructure, and SaaS business principles. Designed for founders, engineers, and AI builders interested in turning generative AI into real products.

📗 Building Your Enterprise Peer-to-Peer VPN SaaS From Scratch
A deep dive into building a production-ready networking platform based on WireGuard, P2P networking, edge computing, cloud orchestration, and SaaS architecture. Based on lessons learned while building OmniEdge.

📙 The Complete Guide to Former Founder Employment Success
For founders transitioning back into the job market. Covers founder bias, positioning your experience, interviewing effectively, and turning entrepreneurial experience into a competitive advantage.

These books are not theory. They are collections of lessons learned through years of building products, solving technical problems, leading teams, raising capital, serving customers, and occasionally making expensive mistakes.

Whether you’re a startup founder, software engineer, product leader, robotics professional, or someone considering a new chapter in your career, I hope these books help you move a little faster and avoid a few of the pitfalls I encountered.

Author page:

Amazon Author Page – Yong Qian

https://www.amazon.com/author/yongqian

Thank you to everyone who has supported my journey as an engineer, founder, builder, and lifelong learner. More books on robotics, embodied AI, industrial systems, and deep-tech entrepreneurship are coming soon.

Building a startup isn’t about looking more successful than everyone else. It’s about solving problems consistently and staying in the game.

Funding rounds, media coverage, a bigger office, a growing headcount, and social media buzz can all create the illusion that a company is doing well.

The praise from people around you, the admiration of employees, the attention from partners, and the interest from the outside world can be intoxicating. Before you realize it, you’re floating comfortably above reality.

First-time founders are especially vulnerable to it.

Over time, I’ve learned that the things that truly matter are far less glamorous: whether customers keep paying, whether people continue using the product, and how long the cash in the bank will last.

Many companies don’t run into trouble because of competitors. They run into trouble because they become too attached to external validation. They expand to protect their image, chase valuations to support a narrative, and drift away from customers in pursuit of growth.

After years of building companies, I’ve come to believe that real progress isn’t about receiving more applause. It’s about seeing problems sooner and having the discipline to face them.

As the project and fundraising efforts progressed, it became clear that we held different views regarding the company’s future direction. As a result, I decided to conclude my involvement and end the collaboration in late May.

Although no non-disclosure agreement was signed between me and the project, I will refrain from sharing further details out of respect for all parties involved and in keeping with a principle of avoiding unnecessary harm. This statement is intended simply to inform friends, clients, and business contacts who were already aware of my participation.

Thank you for your understanding.

Yong

Unitree Robotics recently filed for its STAR Market IPO, disclosing hundreds of pages of application materials. This is not just a regulatory financial filing, but a highly practical playbook of hard-tech survival. Founder Wang Xingxing spent 10 years scaling a garage operation started with ¥100,000 to an expected IPO valuation of ¥42 billion (~$6.2 billion) and the world’s leading humanoid robot shipment volume. In August 2025, I wrote an article Unitree’s Trajectory Towards Dominance in Robotics, noting: Unitree is following in DJI’s footsteps, positioning itself as a hardware-led company that leverages China’s manufacturing advantages to capture the market. It has been proven that this positioning successfully enabled Unitree to become the first—and currently the world’s only—profitable humanoid robot company. Post-IPO, Unitree will venture deeper into industrial scenarios and utilize crowd-sourced data to build better AI model ecosystem products.

Zooming past the high-level AI hype, the most compelling part of the prospectus is its micro-level corporate architecture. The real story is written in these operational details. Unitree has built a world-class hardware, software, and developer ecosystem, with CEO Wang Xingxing concurrently serving as Chief Product Officer (CPO). A hard-tech startup must have a dedicated CPO—either the founder themselves, or someone with the actual budget and authority to drive product execution.

Unitree is a hard-tech benchmark that succeeded under extreme uncertainty. Behind its success lies the reality of hundreds of similar robotics startups that failed due to wrong technology roadmaps, cash flow collapses, or supply chain traps. Unitree’s breakout is not only due to its brilliant micro-level corporate design but also a favorable macroeconomic window. Nonetheless, its practices in equity defense, in-house cost reduction, dual-track product portfolios, and ESOP design remain a highly replicable hard-tech survival template.

10 Key Observations and Data Points from Unitree’s Journey

1. Commercializing prototypes over pitch-deck ideas

   • Evidence: Wang built the XDog prototype at Shanghai University and founded Unitree directly to commercialize it, bypassing the PPT concept phase.

2. Retaining manufacturing and design of core parts

   • Evidence: Self-developed integrated joints, torque motors, algorithms, and sensors rather than outsourcing them.

3. Remaining patient during the foundational period

   • Evidence: Spent four years (2016-2019) in pure R&D, keeping commercial revenues to a minimum to focus on component iteration.

4. Deploying a dual-track product strategy

   • Evidence: Go series quadrupeds lock in supply chain scale, while enterprise B/G/H series units secure high-margin enterprise sales.

5. Going global from Day 1

   • Evidence: Export revenue exceeded 55% of total sales for several years, de-risking domestic price wars.

6. Sequencing financing stages

   • Evidence: Credibility-focused early VCs (Sequoia, Matrix), followed by strategic buyers (Meituan), and pre-IPO backing from national giants (China Mobile, Tencent).

7. Accessing liquidity concurrently with primary financing

   • Evidence: Wang sold ¥50M in secondary shares at the primary round price during the Series D financing to relieve personal debt stress.

8. Deploying dual-class shares early to protect control

   • Evidence: Set up special voting rights when shareholding diluted to 23%, retaining 65% voting control post-IPO.

9. Filing for IPO only after proving self-sustainability

   • Evidence: Delivered ¥672 million in net operating cash flow and ¥600 million in non-recurring net income in 2025.

10. Funneling proceeds into future technology

    • Evidence: 74.5% of the target ¥4.2B IPO proceeds will go directly to AI model and next-gen hardware R&D.

I. The Story Arc: From ¥100K to ¥42B Expected IPO Valuation

II. Technical and Product Strategy: Beating the Market with Hardware Realism

1. Conviction Over Hype: The 4-Year Foundation

Wang Xingxing built his quadruped prototype (XDog) at Shanghai University using a pure hard-tech roadmap. He didn’t start with pitch decks to raise seed money; he immediately commercialized the hardware.

Technically, Unitree’s joint design inherits the “Quasi-Direct Drive” (QDD) philosophy pioneered by the MIT Biomimetic Robotics Lab—combining a high-torque-density motor with a low-reduction planetary gearbox to enable high-bandwidth torque control. Wang referenced open academic literature on these design principles during XDog’s early development (2013-2015). However, when MIT open-sourced its Mini Cheetah scheme in 2019, Wang noted that their motor and electronic configuration was highly similar to the design he had independently developed and published for XDog three years prior. Unitree’s core contribution was not inventing this QDD academic concept from scratch, but rather executing extreme cost reduction, engineering reliability, and mass manufacturing to transition it from a laboratory prototype into a commercial product.

In my observation of the hardware ecosystem, Unitree spent its first four years (2016-2020) in the workshop with almost no revenue growth. They developed their own motors, reducers, and control algorithms from scratch. While many hype-driven teams dismissed this approach as too slow, this early hardware experience determined the design parameters for their future humanoid systems.

Unitree’s Development Path vs. Conventional Robotics Teams

• Entry Point:

  • Unitree: Developed the XDog prototype independently during grad school, transitioning directly to production upon graduation.

  • Others: Follow academic paper concepts → assemble pitch decks → seek VC funding to validate ideas.

• First Product:

  • Unitree: Launched Laikago (2017) — a commercialized quadruped with a real BOM (bill of materials) and immediate shipping capacity.

  • Others: 3D renders, acrylic-housed laboratory demos, or conceptual MVPs.

• R&D Depth:

  • Unitree: Full-stack R&D across core hardware (motors, gearboxes) and low-level control systems.

  • Others: Outsource hardware assembly, focusing only on high-level application software.

• Foundational Patience:

  • Unitree: Weathered a 4-year (2016-2020) pure R&D phase with minimal commercial revenue.

  • Others: Seek to complete product cycles and scaling within a year of funding.

The joint modules of the H1 and G1 humanoid systems are fundamentally a combination of high-density frameless motors and precision reducers, which shares the same physical foundation as quadruped drive units. This development timeline cannot be bypassed. If they hadn’t spent years addressing CAN bus packet drops, gearbox tooth wear, and driver board overcurrent issues on quadruped platforms, the system complexity of launching a humanoid robot would have been unmanageable.

2. Full-Stack Self-Development as a Cost Control Lever

Unitree’s prospectus repeatedly highlights “全栈自研” (full-stack self-development) as its core moat. In deep tech, full-stack design isn’t about manufacturing every bolt; it is about strategically owning the components that determine the final BOM (bill of materials) cost and product performance, while keeping supply chain integration flexible:

• Core Actuation & Joints: Unitree owns the core power module—integrated joint assemblies, frameless torque motors, and control driver boards. For planetary and harmonic reducers, while they maintain internal R&D and manufacturing capacity, they also integrate local suppliers (e.g., Zhongdali De and Green Harmonic) to share production volume.

• Perception Modules: A hybrid strategy of “in-house design + off-the-shelf procurement” is utilized. For instance, Unitree developed its L1 4D LiDAR, but integrates industrial DJI (Livox) LiDARs on premium enterprise configurations to balance engineering speed and cost.

• Control Loops & System Software: The motion control stack extends down to FOC magnetic field control, thermal dynamics simulation, and power management driver layers, avoiding black-box proprietary solutions.

• Brain Models: Open-sourced UnifoLM (Unified Large Model) models (such as the WMA-0 world model for simulation, the VLA-0 vision-language model for natural language, and the WBT teleoperation dataset) cultivate the developer ecosystem, while the industrial UnifoLM-X1-0 model is tested directly on their joint assembly lines.

In the early phases, outsourcing speeds up time-to-market, but at scale, component markup eats all margins. Self-developing the critical motors and motion loops allows a startup to control pricing rather than leaving margin on the table for component suppliers.

3. The Dual-Track Product Matrix: Volume and Margins

Unitree avoided the trap of choosing between high-end research and low-cost manufacturing. Instead, they built a dual-track portfolio:

The Dual-Track Product Portfolio

• Laikago (2017) | Academic Research

  • Pricing/Margin: Premium pricing / Academic exploration

  • Strategic Purpose: Proved quadruped dynamics and closed the control loop.

• A1 (2020) | STEAM Education

  • Pricing/Margin: Mass-market entry

  • Strategic Purpose: Scaled up early batch assembly lines and tested consumer demand.

• Go1 (2021) | Mass Consumer

  • Pricing/Margin: ¥16,000 (disrupted industry pricing)

  • Strategic Purpose: Drove down entry costs to quickly scale supply chain volume.

• B1 (2022) | Industrial Inspection

  • Pricing/Margin: High premium / IP67 dustproof & waterproof

  • Strategic Purpose: Captured high-margin enterprise inspection clients.

• Go2 (2023) | Consumer

  • Pricing/Margin: ¥9,997 (sub-¥10K retail)

  • Strategic Purpose: Minimal BOM cost; established global consumer volume monopoly.

• B2 (2023) | Specialized Industry

  • Pricing/Margin: High premium / Explosion-proof and disaster response

  • Strategic Purpose: Replaced expensive foreign imports for power grid/utility clients.

• H1 (2023) | Labs & AI Research

  • Pricing/Margin: Premium full-size humanoid

  • Strategic Purpose: Established early positioning and brand mindshare in the humanoid space.

• G1 (2024) | Humanoid Research & Mass Market

  • Pricing/Margin: Sub-¥100K (¥99K starting price)

  • Strategic Purpose: Capitalized on supply chain advantages to capture global developer market share.

This dual strategy is highly effective: consumer-grade units (Go-series, G1) drive volume, establish ecosystem scale, and amortize fixed tooling/purchasing costs; enterprise-grade units (B-series, H-series) capture high ASPs and secure fat margins.

III. Financials and Product Economics: The Path to Profitability

1. Revenue Trajectory

• 2022: Revenue: ¥122.9M | Gross Margin: 44.18% | Net Profit: -¥22.1M | Adj. Net Profit: -¥8.1M

• 2023: Revenue: ¥159.1M (+30% YoY) | Gross Margin: 44.22% | Net Profit: -¥11.1M | Adj. Net Profit: -¥18.0M

• 2024: Revenue: ¥392.4M (+147% YoY) | Gross Margin: 56.41% | Net Profit: ¥94.5M | Adj. Net Profit: ¥77.5M

• 2025 (Q1-Q3): Revenue: ¥1,167.5M | Gross Margin: 59.45% | Net Profit: ¥105.3M | Adj. Net Profit: ¥430.6M

• 2025 (Full Year): Revenue: ¥1,708.2M (+335% YoY) | Gross Margin: 60.27% | Net Profit: ¥287.6M | Adj. Net Profit: ¥600.1M

(Note: The 2025 net profit was affected by a one-time share-based payment of ¥349M. The adjusted net income represents its true operational profitability.)

Unitree ran moderate losses in 2022 and 2023. They achieved profitability in 2024 (¥77.5M adjusted net profit) and reached ¥600M in adjusted net profit in 2025. They filed for IPO only after proving their cash-generation engine.

2. Component Scaling and Shifting Product Mix

Improving gross margins from 44% to over 60% in three years illustrates the economic levers of hardware scaling:

1. Amortization of self-developed parts: Internal component costs (motors, drives) drop rapidly as manufacturing volume increases.

2. Shift to high-ASP units: The high-margin humanoid business grew from nothing to 51.53% of revenue (¥595M) in Q1-Q3 2025, while quadrupeds dropped to 42.25%.

The financial logic here centers on using the cash flows and supply chain scale of a mature product (quadrupeds) to fund the development of the next-generation platform (humanoids). Because the two lines share core motor and control systems, reusing these technologies lowers technical risk and expands net margins.

3. Global Revenue as a De-Risking and Margin Engine

• 2022: Overseas Revenue: ¥69.4M | 57.2% of total revenue

• 2023: Overseas Revenue: ¥87.6M | 55.6% of total revenue

• 2024: Overseas Revenue: ¥215.7M | 55.7% of total revenue

• 2025 (Q1-Q3): Overseas Revenue: ¥452.8M | 39.2% of total revenue

Relying on a single domestic market limits a startup’s options. Unitree maintained overseas revenue above 55% for years. Even with the domestic humanoid spike in 2025, international sales grew to ¥452.8M. Overseas research and enterprise buyers are far less price-sensitive, providing higher ASPs and buffer margins that fund the domestic cost-reduction roadmap.

IV. Capital Efficiency: Sequencing High-Quality Funding

1. Financing Round History

• Pre-2022 (Seed/Angel & Pre-A):

  • Investors: Sequoia China, Astrend IV (Source Code), Matrix Partners, Vertex Ventures, Decent Capital

  • Capital Raised: Undisclosed

  • Valuation: Estimated at ~¥200M to ¥500M post-money

• January 2022 (Series A):

  • Investors: Matrix Partners, Hexagon, ESOP platform

  • Capital Raised: ¥91.5M

  • Valuation: Implied post-money of ~¥900M

• March 2022 (Series A+):

  • Investors: Local state funds, Astrend IV (Source Code) follow-on

  • Capital Raised: ¥108.4M

  • Valuation: Implied post-money of ~¥1.1B

• June 2022 (Capitalization):

  • Action: Conversion of capital reserve to share capital (8x increase in shares)

  • Valuation: Bookkeeping restructure for subsequent joint-stock conversion

• June 2024 (Series B):

  • Investors: Meituan (Hanhai), CITIC Goldstone, Source Code Capital, Xinjiang Shenzhen Capital

  • Capital Raised: Undisclosed

  • Valuation: Valued at ~¥7B to ¥8B post-money

• August 2024 (Series C):

  • Investors: Beijing Robot Fund, Zhongguancun Science Park, Sequoia Capital, CITIC Securities

  • Capital Raised: ¥284.7M

  • Valuation: Valued at ~¥8B to ¥10B post-money

• June 2025 (Series D):

  • Investors: China Mobile (Zhongyi Hechuang), Tencent, Wuxi Jinqiu, and 23 existing investors

  • Capital Raised: ¥694.5M (plus secondary shares)

  • Valuation: ¥12.0B pre-money / ¥12.7B post-money

• 2026 (STAR Market IPO - Planned):

  • Investors: Public offering and strategic placement

  • Target Raising: ~¥4.2B

  • Valuation: Expected IPO listing valuation of ¥42B ($6.2B)

2. Sequencing the Investor Base

Looking closely at Unitree’s cap table shows a logical sequencing of capital partners:

1. Early Stage (2016-2021) — Credibility: Top-tier financial VCs (Sequoia, Matrix, Vertex) established the brand’s reputation, opening doors to university and research institute procurement.

2. Growth Stage (2022-2024) — Order Volume: Commercial and industrial strategic partners (like Meituan) entered to validate real-world scenarios such as deliveries and logistics.

3. Late Stage (2024-2025) — Ecosystem Scale: Big tech and state-backed funds (China Mobile, Tencent, municipal robot funds) laid the groundwork for large-scale industrial deployment and major integration project bids.

Every round was structured to prepare for the next stage of commercial rollout.

3. Clean and Concurrent Founder Liquidity

During the June 2025 Series D round, Wang Xingxing sold 11,378 shares to Source Code Capital for ¥50M.

In the Series D round with a valuation exceeding ¥10B, a ¥50M secondary sale is standard practice. Ten years of hardware development is a long grind, and a founder’s personal finances are under high stress. Relieving this pressure with a transparent secondary transaction allows the team to maintain strategic focus, preventing short-term decisions after listing. However, such liquidity events require the founder to maintain absolute voting control. This is a crucial lesson for all founders: without early defenses, dilution will leave you vulnerable during growth rounds, potentially resulting in a loss of corporate control and hindering future development.

V. STAR Market Listing and Corporate Architecture

1. Safeguarding Control via Dual-Class Voting

Unitree opted for the SSE STAR Market, utilizing a dual-class share structure:

• Listing Criteria: Minimum expected market cap of ¥10B (met via ¥12.7B Series D valuation, with expected IPO valuation at ~¥42B).

• Voting Power: Wang Xingxing holds direct A-class shares carrying higher votes (approx. 2.67 votes per share).

• Control Line: Wang controls 68.78% of the voting power pre-IPO with 23.82% economic equity. Post-IPO, his voting power remains above 65.3%.

This voting control shields the engineering roadmap from being derailed by short-term market pressures, even when economic ownership drops below 25%.

2. Launching the IPO at the Intersection of Metrics and Brand

Unitree timed its filing when several key lines crossed:

1. Financial Strength: ¥600M in adjusted net profit in 2025, demonstrating solid self-funding capability.

2. Narrative Peak: Embodied AI and humanoid commercialization entered the late-stage testing phase globally.

3. Volume Leadership: Shipping 5,500+ humanoid units in 2025, establishing an undeniable volume lead.

3. Allocating IPO Proceeds to R&D

Unitree’s target ¥4.2B IPO proceeds are allocated as follows:

• Intelligent Robot Model R&D: ¥2.02B (48.1%) — Embodied AI models, reinforcement learning, and cloud systems.

• Robot Body R&D: ¥1.11B (26.4%) — Next-generation hardware dynamics.

• New Product Development: ¥0.45B (10.6%) — Expanding the catalog.

• Manufacturing Site: ¥0.62B (14.9%) — Expanding output.

Directing 74.5% of IPO proceeds to R&D signals to the public market that Unitree is an AI and robotics engineering platform, rather than an assembly-only business.

VI. Equity and ESOP Abstractions

1. Mitigating Dilution with Structured Voting Rights

Wang Xingxing’s ownership journey:

• Founding (2016): Economic Ownership: 100% | Voting Power: 100%

• Pre-2022 (Early VCs): Economic Ownership: ~41% | Voting Power: ~41%

• Post-2022 Rounds: Economic Ownership: ~36% | Voting Power: ~36%

• Post-2024 Rounds: Economic Ownership: ~27% | Voting Power: ~27%

• Post-2025 Series D: Economic Ownership: 23.82% | Voting Power: 68.78% (established dual-class voting)

• Post-IPO (Planned): Economic Ownership: 21.44% | Voting Power: 65.31%

In control defence, waiting until ownership drops below 30% to set up mechanisms is usually too late. Unitree incorporated AB dual-class voting in May 2025 during its transition to a joint-stock company. Resolving this structure early in the listing process avoids regulatory hurdles closer to IPO.

2. Avoiding Equity Proxy Pitfalls: The LP ESOP Structure

The prospectus details the cleanup of historical “equity proxy/nomination” arrangements.

Between 2017 and 2021, Wang held shares on behalf of 11 employees for privacy reasons. In listing audits, such proxy holdings are major red flags due to potential ownership disputes. In 2025, Unitree consolidated employee shares into the Shanghai Yuyi partnership platform.

From a governance perspective, proxy arrangements introduce severe compliance risks. A cleaner approach is using a limited partnership (LP) platform where the founder acts as the GP (general partner) and employees act as LPs (limited partners). The GP retains all voting power, while employees receive the economic returns. This rewards talent without fragmenting company control.

VII. Headcount and Organization

At the time of filing, Unitree had only 480 employees, generating approximately ¥3.56M ($490K USD) per employee. This is an exceptionally high efficiency rate for hardware manufacturing:

• Revenue per employee: Approx. ¥3.56 million, putting the company at the top tier of structural hardware manufacturing.

• R&D Makeup: The R&D team consists of only 124 people (37.8% of headcount). This lean engineering core developed and maintained multiple quadruped and humanoid lines.

• Leadership Stability: The core technical leadership team (Yang Zhiyu, Chen Li, Zhang Yangguang) has been stable for 7 to 9 years, holding substantial stakes in the early ESOP.

In my experience, a flat and highly technical R&D structure avoids administrative overhead. Keeping the team flat and highly technical minimizes alignment friction and lets the founder stay close to the physical details.

VIII. Key Risk Exposures We Must Face

Unitree’s risk disclosures detail their real points of vulnerability, which are standard for the robotics sector:

1. Key Brain Dependency: Heavy reliance on Wang Xingxing and a small group of core technical experts.

2. Inventory Obsolescence: Holding ¥278M in inventory by Q3 2025. In a fast-moving engineering sector, component obsolescence can rapidly write down margins.

3. FX Exposure: With over 35% of revenue from international markets, currency fluctuations introduce cash flow volatility.

About the Author

涌 Yong — 20-year industrial automation and robotics veteran, former product lead at Spirit AI (valuation ¥20B+), three-time founder and exit in both the US and China (IIoT / AIGC / OSS). Currently focused on cross-border venture research in embodied intelligence and deep tech.

• Personal Website: qianyong.me

• Substack: yongqianme.substack.com

• LinkedIn: linkedin.com/in/yongqian

• X/Twitter: @yongqianme

1. Have enough capital to support not only the company’s day-to-day operations, but also the lives of the people building it. A business needs time to grow — decisions shouldn’t be driven solely by short-term cash flow pressure.

2. A team needs more than aligned skills; it also needs goodwill toward one another. Over the long run, outcomes are shaped not just by professional ability, but by whether people choose understanding, trust, and support when things get difficult.

3. Introducing aggressive investor protections or performance-based financing terms too early is rarely easy. Capital may not intend to harm a startup, but the pressure created by certain terms can quietly drain a team’s energy, distort pace, and narrow future options before the company has matured.

4. Think about the consequences of failure from the beginning. A company can shut down, restart, or exit with dignity. Often harder than failure itself is getting stuck — unable to raise, under business pressure, and trapped in a state where neither moving forward nor stepping away feels possible.

More important than the growth curve is preserving enough room for the company to keep moving forward — or to change direction with clarity and dignity.

One thing I’ve learned is that 0-to-1 execution is rarely about high-level theory. It’s about solving concrete, messy technical and career challenges from scratch.

To help builders, systems engineers, and founders navigate these hurdles, I’ve packaged my boots-on-the-ground experience into four playbooks:

1. How to Build Fashion AI SaaS From Scratch

A practical guide drawing from my journey building and selling DeepFashion. It covers everything from training Stable Diffusion models on rented A100 GPUs and managing cloud GPU costs to leveraging AI to write 80% of the codebase.

Link: https://www.amazon.com/build-Fashion-Scratch-step-step-ebook/dp/B0F8WBNSPP

2. Building Your Enterprise Peer-to-Peer VPN SaaS from Scratch

A hands-on systems engineering guide based on my experience scaling OmniEdge VPN. It walks through WireGuard integrations, zero-trust mesh network overlays, cloud orchestration, and cross-platform native client deployments.

Link: https://www.amazon.com/Building-Enterprise-Peer-Peer-Scratch-ebook/dp/B0F8YSG1WW

3. The Complete Guide to Former Founder Employment Success

Transitioning from Founder/CEO back to an employee is one of the hardest career pivots. This book is a practical framework to help former founders counter recruiter bias, reframe their resumes from “doing everything” to “delivering key metrics,” and secure senior operator roles.

• English Edition: https://www.amazon.com/dp/B0FNRZP9DD

  

• 中文版本: https://www.amazon.com/dp/B0H2JPTLJL

  

Whether you’re building next-gen AI infrastructure, designing IIoT systems, or plotting your next career move, I hope these guides save you some hard-learned lessons.

Which of these topics are you currently tackling? Let’s discuss in the comments.

#AI #SaaS #SystemsEngineering #Robotics #IoT #FounderTransition #Books

As we navigate the commercialization wave in 2026, the humanoid robotics and physical AI sectors are undergoing a massive geopolitical split. On one side of the Pacific, the United States is running an AI-first, capital-heavy “foundation brain” playbook. On the other, China is leveraging its absolute dominance in electromechanical supply chains to run a deployment-led, cost-optimized “hardware scaling” race.

Below is my raw, data-backed teardown of how the top 20 startups in both regions stack up when paper valuations collide with the cold realities of the assembly line.

Comparative Structural Matrix: US vs. China Humanoid Ecosystems

China Humanoid Robotics & Embodied AI Top 20 Cohort

Data link

1. Valuation Dynamics: Silicon Valley’s Oligopoly vs. China’s Flat Long-Tail

The way capital is distributed between these two ecosystems reveals a profound difference in regional geopolitical and financial strategy.¹¹

Transatlantic Humanoid Valuation Concentration (Top 3 as % of Cohort):

1. United States (US): Figure AI ($39B) + Skild AI ($14B) + Physical Intelligence ($5.6B) = $58.6B (Approx. 80% of Top 20)

2. China (CN): Galaxy General ($3.0B) + Linkerbot ($3.0B) + Spirit AI ($2.8B) = $8.8B (Approx. 27.5% of Top 20)

The United States: A Billion-Dollar Oligopoly

The US physical AI ecosystem is exceptionally top-heavy.¹¹ While the cumulative valuation of the top 20 US startups hovers around $80 billion, a staggering 80% of that book value is concentrated in just three platform-scale entities ¹¹:

• Figure AI commands a private valuation of $39.0 billion, driven by its Helix vision-language-action (VLA) model and its massive factory pilot at BMW’s Spartanburg facility.

• Skild AI has reached a $14.0 billion valuation after SoftBank led a massive $1.4 billion Series C in January 2026 to fund its platform-agnostic “Skild Brain”.²⁰

• Physical Intelligence is valued at $5.6 billion, with active negotiations for a $1 billion round led by Founders Fund and Lightspeed targeting an $11 billion+ valuation.

The elite tier of US venture capital (Sequoia, General Catalyst, Thrive, Founders Fund) is not interested in funding a dozen different hardware assemblers. They are making a high-conviction bet to crown a select few software-first monopolies ²⁰ that will license the “brains” of tomorrow to whoever builds the physical shell.

China: A Hyper-Competitive Net of Survivors

In contrast, China’s $31.95 billion to $34.3 billion startup cohort is incredibly flat.¹⁰ Its top three unlisted players ¹⁰:

• Galaxy General (银河通用), building wheeled-bipedal systems, is valued at $3.0 billion.

• Linkerbot (灵心巧手), the global dominant force in high-DoF dexterous hands, is valued at $3.0 billion (and is actively seeking a $6 billion valuation for its next round).

• Spirit AI (千寻智能), focusing on next-gen universal embodied brains, is valued at $2.8 billion.

These three unlisted giants represent only 25.6% to 27.5% of the Chinese cohort’s total value.¹⁰ Even if you add the publicly traded UBTECH Robotics (market cap ~$7.0B) ¹⁰, the top-three concentration barely scratches 37% to 40%.¹⁰ China’s primary market supports a vast “long-tail” of over 150 active humanoid companies. Rather than crowning a single king early, Chinese investors let a massive network of component suppliers, bionic developers, and assembly specialists duke it out on the factory floor.

2. Geopolitics of Capital: State Guidance Funds vs. Big Tech Balance Sheets

Who funds your research dictates the parameters of your commercial execution.

China: The Rise of “Sovereign-Directed” Order Pipelines

In China, humanoid robotics is a state-level industrial mandate. More than 75% of China’s top 20 humanoid startups have direct state-owned capital, local government guidance funds, or national strategic platforms on their cap tables.

China’s Local Government Robotics War Chests:

1. Beijing Robot Industry Development Fund: ¥10.0 Billion
2. Hubei Humanoid Robot Parent Fund: ¥10.0 Billion
3. Shenzhen Embodied AI Fund Group: ¥3.0 Billion

Beyond the central government’s $138 billion state venture capital guidance fund, local municipals are highly aggressive in utilizing state funds to force localization:

• In March 2026, the National AI Industry Fund (Big Fund Phase III) directly led a massive ¥2.5 billion strategic round for Galaxy General.

• Heavyweight manufacturing and battery titans like CATL, BYD, and Luxshare Precision are strategically funding startups (e.g., CATL co-leading Galaxy General/GALBOT’s rounds, Luxshare co-leading EngineAI’s $200 million Series B).

This capital is not free money; it is a highly coordinated, order-driven pipeline. Startups get capital in exchange for deploying humanoid fleets directly on BYD car assembly lines, automating CATL’s battery production cells, or managing high-volume logistics hubs for JD.com and Shunfeng.

United States: Big Tech Hyperscalers and the Military-Industrial Shield

The US ecosystem is fueled by private tech balance sheets and the defense complex.

US private AI investment reached $109 billion—representing a 12-fold gap in private capital intensity compared to China’s $9.3 billion. Startups like Agility Robotics, Figure AI, and Apptronik are funded by the deep pockets of Amazon’s Industrial Innovation Fund, Microsoft, Google, and Nvidia.

Simultaneously, the US has a high-margin defense corridor that Chinese startups are completely barred from:

• San Francisco-based Foundation Future Industries (FFI) secured $24 million in Pentagon research contracts to test its 176-pound Phantom MK-1 humanoid.

• These tactical units were deployed directly in Ukraine for frontline logistics and reconnaissance testing—representing the first documented real-world combat deployment of humanoid hardware.

3. The Profitability Myth: Unitree’s Hardware Margin Moat vs. the Burn Rate

In my decades in the field, I’ve seen countless hardware companies go bankrupt because they designed beautiful products with zero path to positive gross margins.

2025 Financial Performance: Unitree vs. Competitors:

1. Unitree (Unlisted): ¥1.71 Billion Revenue | ¥600 Million Net Profit | 60.27% Gross Margin
2. UBTECH (HKG: 9880): ~¥2.0 Billion Revenue | ~¥790 Million Net Loss
3. Boston Dynamics: ₩3.28 Trillion Capital Injected | ₩440.5 Billion (~$320M) FY2024 Operating Loss

Unitree Robotics: The Cost-Down Outlier

According to the IPO prospectus submitted to the Shanghai STAR Market, Unitree Robotics has achieved something no other unlisted humanoid startup has managed globally: scale-level, self-sustaining profitability:

• It pulled in ¥1.71 billion in 2025 revenue, netting an adjusted net profit of ¥600 million.²¹

• Its gross margin escalated to 60.27%.

Unitree achieved this by refusing to outsource its bill of materials.²¹ They self-develop and vertically manufacture their planetary gearboxes, high-torque joint actuators, and electric motors in-house. This allows them to price their Unitree G1 humanoid at an aggressive $16,000 and still maintain high-margin unit economics. Moreover, over 73% of their humanoid revenues came from the high-margin, low-overhead academic research and developer markets, avoiding the costly custom-engineering drag of industrial deployment.

The Cash-Burn Reality of B-End Industrial Assembly

Outside of Unitree’s unique playbook, the humanoid landscape is a heavy cash-burn environment.

Even publicly listed UBTECH Robotics, which shipped a globally leading 1,079 full-size humanoids in 2025, ran an estimated net loss of ¥790 million on ¥2.0 billion in revenue. The immense overhead of customizing Walker S robots for specific automotive assembly lines eats away at software-style margins.

In the West, Boston Dynamics (majority-owned by Hyundai) continues to run deep losses, reporting a ₩440.5 billion (~$320 million) operating loss in FY2024 despite cumulative parent capital injections of ₩3.28 trillion. US venture-backed darlings like Figure AI, Apptronik, and 1X remain entirely pre-profit, utilizing their multi-billion-dollar cash runways to fund extreme R&D and high-talent acquisitions.

4. Supply Chain Gravity: The $131,000 “Optimus Penalty”

McKinsey’s April 2026 report laid bare the inescapable gravity of the Chinese manufacturing cluster:

• Attempting to build Tesla’s Optimus Gen 2 entirely outside of China’s electromechanical supplier network results in a 3x cost penalty, with the Bill of Materials (BOM) surging from $46,000 to $131,000.

• A typical humanoid robot assembled in China has an average BOM of $35,000 in 2026, on an aggressive trajectory to fall to $13,000 by 2035.

China’s Dominance in Key Humanoid Hardware Inputs:

1. Heavy Rare-Earth Processing (Motors): 99%
2. Permanent Magnet Market (Actuators): 93%
3. Electromechanical Component Supply: 80% ~ 90%

China controls 80% to 90% of key hardware components globally, 93% of the permanent magnet market essential for actuators, and 99% of the heavy rare-earth element processing required for high-torque robotic motors. This allows Chinese startups like 松延动力 (Noetix Robotics) to release its compact home humanoid “Dora/小布米” at a starting price under $1,400 (under 10,000 yuan), bypassing traditional, low-volume CNC machining bottlenecks.

5. Algorithmic Bifurcation: The “Zero-Shot Brain” vs. the “Data Pyramid”

When it comes to the “robotic brain” (cognitive model) and the “robotic hand” (dexterous manipulation), the US and China are executing completely different technological plays.

The US Pure-Play: Foundation Worlds and Semantic Generalization

The US strategy is built around high-compute simulation engines and native zero-shot generalization.

• Physical Intelligence’s universal model, (released in April 2026), can operate highly complex, completely unfamiliar home appliances (like a specific model of an air fryer) after seeing it only twice in its entire training set.

• US players leverage massive synthetic physics engines and world models. In Q1 2026 alone, $6 billion in venture capital flowed into dedicated world model developers, allowing robots to learn generalized motor skills within photorealistic simulations (such as NVIDIA Isaac Sim and AI Habitat) running at over 10,000 FPS on massive Blackwell compute clusters.

The Chinese Playbook: Pragmatic Data Acquisition and Hard-Target Tuning

China’s developers bypass high-compute simulation bottlenecks by industrializing physical data collection.

• 千寻智能 (Spirit AI) has bypassed traditional $20,000 teleoperation rigs by developing custom, 5th-generation wearable data-capture gloves that cut 数采 (data acquisition) costs to 1/10 of traditional methods.

• Spirit AI has deployed a 1,000-person team to physically record over 200,000 hours of real-world physical interaction data (targeting 1 million hours in late 2026).

• Chinese startups do not chase absolute, open-ended general intelligence. They focus on getting a robot to handle specialized, high-repetition industrial tasks—such as picking parts, sorting bins, and packing pallets—utilizing RAG (Retrieval-Augmented Generation) and custom physical training farms (such as the Hubei Humanoid Innovation Center).

6. Founders and Elite Pedigrees: Academic Giants vs. Serial Hustlers

The humanoid battle comes down to a high-wire tug-of-war between the elite minds of both nations.

The Chinese “Tsinghua-Berkeley” Academic Nexus:

1. Gao Yang (Spirit AI Co-founder) - Tsinghua CS, Berkeley BAIR PhD (Darrell / Abbeel Labs)
2. Chen Jianyu (RobotEra Founder) - Tsinghua, Berkeley PhD
3. Xu Huazhe (Galaxea AI Chief Scientist) - Tsinghua, Berkeley PhD

China’s Academic-Industrial Dream Teams

The core of China’s humanoid software brain layer is run by the highly concentrated “Berkeley Quartet” (伯克利归国四子)—a group of elite Chinese scholars who trained at the Berkeley Artificial Intelligence Research (BAIR) lab before returning to Tsinghua University’s Institute for Interdisciplinary Information Sciences.

These academic minds are systematically paired with veteran, battle-tested hardware operators:

• Spirit AI pairs Chief Scientist Gao Yang (Tsinghua CS, Berkeley PhD under Trevor Darrell and post-doc under Pieter Abbeel) with CEO Han Fengtao, the former CTO of Luoshi Robotics (ROKAE) who personally led the delivery of over 20,000 industrial robotic arms across China.

The US Hall of Fame: Deep Learning Pioneers and Billion-Dollar Founders

The US founding teams represent the absolute pioneers of deep reinforcement learning and legendary Silicon Valley serial entrepreneurs:

• Physical Intelligence was co-founded by Sergey Levine (UC Berkeley Professor, the most-cited academic in deep robotic learning) and Chelsea Finn (Stanford Professor, pioneer in meta-learning and MAML).²²

• Covariant was founded by Pieter Abbeel (Director of the Berkeley Robot Learning Lab).

• Skild AI was founded by Carnegie Mellon University Professors Deepak Pathak and Abhinav Gupta.

These academic titans are joined by legendary Silicon Valley product builders, including Kyle Vogt (co-founder of Twitch and former CEO of GM’s self-driving division Cruise, now leading consumer humanoid play The Bot Company) ²⁴, and Figure AI’s Brett Adcock (who previously founded Archer Aviation and Vettery).²⁵

7. Real Capital Efficiency: What Does Your Dollar Buy?

The ultimate metric of any physical tech business is capital-to-physical-output efficiency. How much real, functioning industrial capability did your startup construct per dollar of capital raised?

• AgiBot (上海智元) raised approximately $85 million in cumulative funding, yet successfully manufactured, shipped, and deployed over 10,000 physical humanoid units (including the general-purpose Expedition A3) as of late March 2026. This represents an incredibly lean capital-to-output yield of $8,500 in funding per deployed robot.

• Unitree Robotics raised $237 million ¹² and shipped 5,500 full-size humanoids in 2025 alone.

• Figure AI has raised over $1.9 billion, but has manufactured and deployed only around 150 pilot-stage units to strategic partners like BMW. This implies an astronomical capital-to-output yield of $12.66 million in funding per shipped unit.²⁷

• Apptronik has raised nearly $1 billion ($976M) while shipping only early double-digit pilot units of its Apollo platform.

In China, every dollar raised is immediately converted into steel, copper, physical testing fields, and direct, on-site industrial testing. In the US, the vast majority of your venture dollar is burned in the virtual world: renting massive H100/Blackwell GPU clusters, hiring AI researchers at $1 million+ salaries, and training high-fidelity simulation world models that still face a massive sim-to-real transfer gap.

8. Yong’s Product Truths: Cut the Demo Bullshit, Get to the Factory Floor

Having spent 20 years in the automation trenches, I have zero tolerance for investor-facing PR stunts. Here is my unfiltered product advice on how to look at this industry before 2030:

1. “Dark Factories” (完全无人化工厂) are a VC hallucination.
   If a factory manager buys a $150,000 humanoid robot that only has active battery runtimes of 2 to 4 hours per charge, losing precision when the factory lighting shifts or dust settles on its depth cameras, they will fire you on day one. Industrial buyers are brutally pragmatic. Standardized high-speed assembly tasks belong to 6-axis collaborative robotic arms. Humanoids only make economic sense in unstructured, low-payload, highly irregular environments where traditional rigid automation fails.

2. The battle is won or lost at the fingertips (Dexterous Hands).
   As Elon Musk noted, over 50% of the entire engineering and control workload of a humanoid robot is concentrated specifically on the hands. This is why 灵心巧手 (Linkerbot) commands a $3 billion valuation despite only building hands. If you cannot handle the complex force-feedback and micro-slip adjustments required to tighten a non-threaded screw or handle a flexible textile wire harness, your $100,000 robot is just a very expensive, slow-moving forklift.

3. The Strategic End-State (2026–2030):

• China is on track to win the physical hardware race. By dominating the electromechanical supply chains, cutting BOM costs to $13,000, and delivering thousands of units to academic, logistics, and retail spaces, China will commoditize the hardware shell of the humanoid robot.

• The United States is on track to win the high-margin software brain race. If generalist VLA models (like Skild Brain or Physical Intelligence’s ) achieve true spatial-temporal reasoning and zero-shot generalization, they will establish a dominant “Robotics Android OS”. They will license this software globally, capturing 80%+ of the industry’s economic value while leaving Chinese hardware manufacturers fighting over low-margin assembly subcontracts.

The ultimate winner of this transatlantic hard-tech duel will be decided by a single question: Will the robotic “ChatGPT moment” be triggered by making the hardware incredibly cheap and physically robust, or by making the software brain completely generalist and physically aware? ²⁷

Sourcing & Citations

• SNE Research & IFR Industry Brief (2025/2026): Confirmed China’s control of 99% of heavy rare-earth processing, 93% of actuator magnetic materials, and over 75% state (state-owned capital) Cap Table representation.

• Unitree Robotics (宇树科技) Shanghai STAR Market IPO Prospectus (March 2026): Verified FY2025 financial figures: ¥1.71B revenue, ¥600M net profit, 60.27% gross margin, and 5,500 humanoid shipments.

• AgiBot (智元机器人) Partner Conference Disclosures (April 2026): Verified cumulative shipment milestone of 10,000 units by Q1 2026, and its ¥2.0B “Yuansheng” industrial ecosystem fund.

• McKinsey & Co. Global Humanoid Supply Chain Analysis (April 2026): Detailed the 3x cost penalty for building Optimus outside China ($131,000 vs $46,000) and the BofA projection of $13,000 BOM cost by 2035.

• ¹ Reuters Linkerbot (灵心巧手) Executive Interview (May 2026): Verified its 80% global market share in high-DoF dexterous hands, $3B Series B+ valuation, and its target $6B valuation round.

• US Department of Defense (Pentagon) Procurement Portal (2026): Confirmed FFI’s $24M research contract for the Phantom MK-1, Eric Trump’s role as Chief Strategy Adviser, and its deployment for active logistics testing on the Ukrainian front lines.

• Hyundai Motor Group SEC filings (FY2024/FY2025): Confirmed Boston Dynamics’ financial performance, including a ₩440.5B (~$320M) operating loss in 2024.

• ⁴ SEC & TechCrunch Unicorn Board (2025/2026): Verified the $39B private valuation of Figure AI, the $14B valuation of Skild AI after its $1.4B round, and Physical Intelligence’s $11B+ active raise parameters.

Works cited

1. After AI, China unicorns are eyeing global domination in robotic humanoids, accessed May 17, 2026, https://m.economictimes.com/news/international/us/after-ai-china-unicorns-are-eyeing-global-domination-in-robotic-humanoids/articleshow/130742840.cms

2. Chinese Startup Linkerbot Targets a $6B Valuation Amid the AI Robotics Boom - incrypted, accessed May 17, 2026, https://incrypted.com/en/chinese-startup-linkerbot-targets-a-6b-valuation-amid-the-ai-robotics-boom/

3. Chinese humanoid robot maker Linkerbot seeks $6B valuation | Semafor, accessed May 17, 2026, https://www.semafor.com/article/05/04/2026/chinese-humanoid-robot-maker-linkerbot-seeks-6b-valuation

4. Skild AI Robotics Manufacturing Foundation Model Raised $1.4B — and It’s a Threat to Every Automation Vendor - Medium, accessed May 18, 2026, https://medium.com/@creed_1732/skild-ai-robotics-manufacturing-foundation-model-raised-1-4b-bffbaad0e70e

5. Skild AI Stock: $14B Valuation — Is It a Buy? | TSG Invest, accessed May 18, 2026, https://tsginvest.com/skild-ai/

6. Skild AI Valuation Soars to $14B in Stunning Funding Round Led by SoftBank and Nvidia, accessed May 18, 2026, https://cryptorank.io/news/feed/edb5f-skild-ai-14-billion-valuation-robotics

7. Skild AI funding, news & analysis | Sacra, accessed May 18, 2026, https://sacra.com/c/skild-ai/

8. Physical Intelligence in Talks to Raise $1 Billion at $11 Billion Valuation - MLQ.ai | AI for investors, accessed May 18, 2026, https://mlq.ai/news/physical-intelligence-in-talks-to-raise-1-billion-at-11-billion-valuation/

9. Physical Intelligence eyes $1B raise at $11B valuation, Founders Fund and Lightspeed in talks, accessed May 18, 2026, https://techfundingnews.com/physical-intelligence-1b-raise-11b-valuation-founders-fund-lightspeed/

10. 清华博士辍学造人形机器人：销售额已过亿，不想做短期套利, accessed May 17, 2026, https://m.mp.oeeee.com/a/BAAFRD0000202505301091115.html

11. Top Humanoid Robotics Startups by Valuation (2026) - New Market Pitch, accessed May 18, 2026, https://newmarketpitch.com/blogs/news/humanoid-robotics-top-startups-valuation

12. 2026年4月，国内具身智能机器人企业融资情况汇总 - 维科号, accessed May 17, 2026, https://mp.ofweek.com/robot/a256714022697

13. Humanoid Robotics In 2026: The Race From Pilot To Platform - KraneShares, accessed May 18, 2026, https://kraneshares.com/humanoid-robotics-in-2026-the-race-from-pilot-to-platform/

14. Latest Progress of Seven Chinese Humanoid Robot Unicorns ..., accessed May 17, 2026, https://cnmra.com/latest-progress-of-seven-chinese-humanoid-robot-unicorns-valued-at-over-10-billion-in-2025/

15. Top 10 Chinese Humanoid Robots of 2026, accessed May 17, 2026, https://humanoidroboticstechnology.com/articles/top-10-chinese-humanoid-robots-of-2026/

16. 未来科学城这个机器人“小孩哥”，又双叒叕火了！, accessed May 17, 2026, https://www.ncsti.gov.cn/kjdt/scyq/wlkxc/wldt/202504/t20250423_202596.html

17. Company Introduction-Noetix Robotics (Beijing) Technology Co., Ltd., accessed May 17, 2026, https://noetixrobotics.com/en/about-us

18. Dora - Humanoid robot guide, accessed May 17, 2026, https://humanoid.guide/product/dora/

19. China’s humanoid robot boom faces reality check as 150 companies chase a market where only 23% of buyers are satisfied - TNW, accessed May 17, 2026, https://thenextweb.com/news/china-humanoid-robot-boom-commercialisation-reality-check

20. 估值破百亿！星动纪元完成10亿元战略融资—— - 国际科技创新中心, accessed May 17, 2026, https://www.ncsti.gov.cn/kjdt/scyq/zgckxc/zgcdt/202603/t20260310_240429.html

21. Will investors embrace China’s humanoid robot champion?, accessed May 18, 2026, https://www.ft.com/content/721e3bed-285b-46d4-8151-8cf28cb5ef50?syn-25a6b1a6=1

22. Physical Intelligence - 2026 Company Profile, Team, Funding & Competitors - Tracxn, accessed May 18, 2026, https://tracxn.com/d/companies/physicalintelligence/__1JIN2UI-QvpeC93FXAryqyydTalOe6C4uxMhHvwMhTw

23. Physical Intelligence Seeks $1 Billion as Robotics Interest Grows | PYMNTS.com, accessed May 18, 2026, https://www.pymnts.com/artificial-intelligence-2/2026/physical-intelligence-seeks-1-billion-as-robotics-interest-grows/

24. The Bot Company: Funding, Team & Investors | Startup Intros, accessed May 18, 2026, https://startupintros.com/orgs/the-bot-company

25. Figure AI Stock: $39B Valuation — Is It a Buy? | TSG Invest, accessed May 18, 2026, https://tsginvest.com/figure-ai/

26. Figure AI - Wikipedia, accessed May 18, 2026, https://en.wikipedia.org/wiki/Figure_AI

27. PUDU D7 | Semi-Humanoid wheeled AI-powered Robot for Industrial & Hospitality sectors, accessed May 17, 2026, https://www.pudurobotics.com/news/917

I’ve spent my career bridging worlds that rarely meet: the disciplined, century-old precision of Stäubli in Switzerland (founded in 1892, a global leader in industrial and mechatronic solutions) and the engineering excellence of Frimo in Germany (over 50 years building tools and equipment for plastic components). Then, in just five years, I built and exited two companies with many projects in China and the United States. After that, I joined a humanoid robotics startup that’s barely one year old.

This path isn’t a resume flex. It’s a front-row seat to what actually separates enduring organizations from those that flame out.

When you look across legacy industrial companies and early-stage startups, a pattern becomes clear. The difference isn’t just speed or scale—it’s philosophy.

Legacy companies teach you how to build things that last. Startups teach you how to build things that matter right now. The founders and executives who thrive are those who learn to carry the first discipline into the second environment—without letting it slow them down.

What Century-Old Companies Get Right (and What They Pay For It)

At Stäubli and Frimo, processes weren’t optional. They were the product.

Quality wasn’t a slogan—it was embedded in every drawing, every tolerance, every supplier audit. Decisions moved slowly because mistakes could echo for years. In industrial markets, reputation compounds quietly—and once lost, it doesn’t come back.

People stayed for decades. Knowledge didn’t sit in slide decks or individual heads; it accumulated across the organization.

This creates incredible resilience. These companies survive wars, recessions, technological shifts, and leadership changes. They know how to scale manufacturing, manage global supply chains, and protect intellectual property across borders.

But the same strengths become liabilities in fast-moving markets. Consensus takes time. Risk aversion is cultural. Innovation often means incremental improvement on proven platforms rather than questioning the platform itself.

When I moved from that world into building my own companies, the contrast hit immediately.

One thing they do exceptionally well is treating people as a long-term asset, not a short-term resource. When employees align with the company over time, motivation and stability follow. That’s what enables multi-decade thinking—what I’d call true “horizon strategy.”

But that kind of thinking is hard to maintain when the market is moving in months, not decades.

The Startup Crucible: Agility, Risk, and the Five-Year Sprint

Startups operate on a completely different clock.

In robotics, AI, and industrial tech, you’re often compressing what used to take decades into five-year cycles—build, scale, exit.

In that environment, leadership looks different. You’re constantly bridging hardware, AI, and software. You’re not just building products—you’re integrating systems that actually work in the real world.

From building OmniEdge (IIoT/VPN) and DeepFashion (AIGC), one thing became clear:

The value is not the hardware or software.
It’s the integration.

Most founders underestimate this. They focus on “cool tech,” but the real challenge is making everything work together reliably—and doing it before you run out of time or cash.

The hardest parts aren’t technical. They’re operational and existential: bottlenecks, cash flow, hiring mistakes, wrong bets.

Staying alive is the strategy.

The Brutal Speed of Building and Exiting in China and the US

In China, execution velocity is unmatched.

You can hire a team, iterate a product, and get customer feedback in weeks—not quarters. The market rewards speed and adaptability above almost everything else. But it also punishes sloppy foundations: weak IP protection, unclear equity structures, or teams without skin in the game can destroy value overnight.

The US offers a different advantage. Capital is more accessible. Talent density is higher. Customers are more sophisticated. You can test ideas with buyers who actually understand what they’re buying.

But there’s a tradeoff. The pressure for hyper-growth—and the “move fast and break things” mindset—can erode the discipline needed to build something that lasts.

I built and exited both companies in five years. The exits taught me something humbling: speed without durability is just expensive motion.

The things that actually mattered came from my time in Switzerland and Germany:

• Process discipline prevented expensive mistakes

• Respect for accumulated knowledge saved time

• Long-term thinking made the companies more valuable to buyers

At the same time, startups forced me to let go of what didn’t work:

• Too many approval layers

• Over-engineering before validation

• Fear of shutting things down

The Real Insight for Founders: Borrow Durability, Don’t Import Bureaucracy

If you’re a founder (especially one coming from big industry or corporate life), here’s what I wish someone had drilled into me earlier:

1. Process is a feature, not a bug—until it becomes the product.

Use the hard-won systems from legacy companies for things that compound: quality systems, financial controls, IP management, supplier governance. But ruthlessly question every other process. In a one-year-old humanoid startup, “how we did it at Stäubli” is only useful if it directly accelerates our ability to ship working robots that solve real problems today.

2. Build for both the sprint and the marathon.

Your first product needs to work in six months, not six years. But the company you’re creating should be capable of still being relevant in at least 10. That means hiring people who can move fast and think in decades. It means equity structures and culture that reward long-term ownership, not just quick flips.

3. Speed without judgment is chaos; judgment without speed is irrelevance.

Legacy experience gives you excellent judgment. The startup environment teaches you speed. The magic is combining them. I’ve watched founders from big tech or industry fail because they imported too much process and moved too slowly. I’ve also watched pure “zero-to-one” founders crash because they had no respect for durability—burning cash, talent, and trust until nothing was left.

4. Culture scales faster than technology.

In old companies, culture is often unspoken and inherited. In startups, you design it explicitly every day.

Take the best of both:

• Craftsmanship and professionalism from industrial companies

• Ownership and urgency from startup environments

For Managers in Big Companies: Your Next Move Might Be Smaller, Not Safer

If you’re sitting in a large organization wondering whether to make the leap, understand this: the skills that make you valuable in a 100+ year company—deep domain expertise, risk management, cross-functional coordination—are incredibly powerful in startups. But only if you’re willing to trade comfort for ownership.

The startups I built, the humanoid startup I used to work almost had no history. Every decision feels existential. There’s no “we’ve always done it this way” because there is no “always.” That’s terrifying and liberating.

What I carry from my past is perspective: how systems scale, where things break, how to build something that doesn’t collapse under pressure.

What I’m still learning is speed.

The Bottom Line

Legacy isn’t a burden or a luxury. It’s raw material. The best founders and forward-thinking executives don’t reject their corporate or industrial heritage—they refine it. They keep the parts that create lasting value (discipline, craftsmanship, long-term thinking) and shed the parts that create friction (bureaucracy, risk aversion, slow decision-making).

Whether you’re building the next century-old industrial giant or trying to ship the first truly useful humanoid robots this decade, the game is the same: create something that works today and can still work tomorrow.

The companies that endure aren’t the ones that never change. They’re the ones that change deliberately—carrying forward the hard lessons of history while staying hungry enough to write new ones.

What world are you building from? What will you leave behind to make it real? And are you ready to live with what the choice you make?

The global industrial landscape is witnessing a structural shift as artificial intelligence transitions from purely cognitive, screen-based applications to embodied physical systems. This transformation, often termed “Physical AI,” represents the convergence of high-performance electromechanical hardware with foundational vision-language-action (VLA) models, creating a new class of general-purpose humanoid robots capable of operating within human-centric environments.¹ As of early 2026, the sector has moved beyond the “peak hype” of laboratory demonstrations into a pivotal phase of commercial deployment and industrial scaling.³ The acceleration of this industry is driven by a unique synergy between the maturation of the electric vehicle (EV) supply chain, breakthroughs in end-to-end neural network training, and a critical global labor shortage in manufacturing and logistics sectors.³ For investors and policy makers, the humanoid robotics market is no longer a speculative future but a core component of national productivity and sovereign technological autonomy.

The Global Macro-Economic Framework for Humanoid Adoption

The humanoid robotics market is projected to expand from a valuation of approximately million in 2024 to over billion by 2033, representing a compound annual growth rate (CAGR) of .⁶ Other estimates from financial institutions such as Goldman Sachs and Morgan Stanley suggest even more aggressive trajectories, with the potential for the market to reach between billion and billion by 2035, and a long-term total addressable market (TAM) that could eventually be measured in trillions as robots integrate into every facet of the global economy.⁴ This growth is not merely a result of technological progress but is fundamentally anchored in demographic realities. Falling fertility rates and aging populations across the G7 and East Asia are creating structurally smaller workforces, leading to persistent upward pressure on labor costs and a corresponding shift in corporate strategy from operational expenditure (OpEx) on wages to capital expenditure (CapEx) on automated labor.⁵

The current distribution of the market reveals a stark geopolitical divide. As of late 2025, Chinese manufacturers have achieved a dominant position, accounting for over of all humanoid installations worldwide.⁴ This lead is not due to a superiority in cognitive AI—where the United States remains the leader—but rather a superior ability to iterate hardware and leverage a vertically integrated electronics and EV supply chain.⁹ However, startups in the United States and Europe are carving out high-value niches focused on complex reasoning, safe human-robot interaction (HRI), and integration into heavily regulated industrial and domestic sectors.⁴

The Chinese Benchmark: Acceleration through Supply Chain “Involution”

The Chinese humanoid ecosystem is characterized by what local observers call “involution”—an intense, hyper-competitive environment where thousands of startups and established tech giants compete for dominance, leading to rapid technological refinement and cost reduction.¹⁰ Companies like AgiBot (Zhiyuan Robotics) and Unitree Robotics have set the global pace for production, with AgiBot reaching a milestone of units produced by early 2026.¹¹ The pace of this transition is unprecedented: it took AgiBot nearly two years to reach its first units, but the jump from to was completed in just three months.¹¹

The Adjacency of the EV Supply Chain

The structural advantage of the Chinese robotics sector lies in its “structural identity” with the electric vehicle industry. Components such as high-torque servo motors, lithium-ion battery modules, power electronics, and environmental sensors (LiDAR and RGB-D cameras) are essentially the same for both cars and humanoids.⁷ China currently controls approximately of the key companies in the global supply chain for humanoid components and roughly of the world’s battery cell capacity.⁹ This density allows a Chinese startup to complete six hardware iterations in the time it takes a Western competitor to complete two.⁹

Strategic Industrial Clusters: T-Chain and H-Chain

The Chinese market is organized around two primary technological clusters: the “Tesla Chain” (T-Chain) and the “Huawei Chain” (H-Chain).¹² The T-Chain focuses on high-performance actuators and thermal management, with suppliers like Sanhua Intelligent and Tuopu Group providing the mechanical foundations for platforms like Tesla’s Optimus.¹² Conversely, the H-Chain emphasizes the “HarmonyOS” ecosystem and AI integration, utilizing iFLYTEK for voice interaction and Leisai Intelligent for servo systems.¹² These clusters ensure that even small startups have access to world-class components at a fraction of the cost available to Western firms.⁹

US Humanoid Startups: Compute Moats and Venture Scale

The United States approach to humanoid robotics is defined by “Venture Scale”—massive capital infusions aimed at building the most advanced cognitive “brains” for the robots.⁵ American startups are less concerned with producing the cheapest hardware and more focused on achieving “General Intelligence” in physical form.¹

Tesla Optimus: The Vertical Integration Paradigm

Tesla is the most high-profile player in the US market, leveraging its expertise in AI inference, manufacturing automation, and energy storage.¹³ Elon Musk has positioned Optimus as potentially more valuable than Tesla’s automotive business, aiming for a retail price of under and mass production of million units annually.¹³ Optimus Gen 2 and Gen 3 have demonstrated significant breakthroughs in end-to-end learning, performing tasks such as cooking, cleaning, and industrial assembly by watching human demonstrations rather than through explicit coding.¹⁶ The strategy is to turn the “car” (a robot with four wheels) into a “humanoid” (a robot with two legs), using the same Full Self-Driving (FSD) neural networks and the Dojo supercomputer for training.¹³

Figure AI: The Collaborative Hyperscale Model

Figure AI represents the fastest-growing humanoid valuation in history, reaching billion by late 2025.¹³ Figure’s success is built on a “consortium” model, with backing from Microsoft, NVIDIA, OpenAI, Jeff Bezos, and Intel.⁴ Their Figure 02 and Figure 03 models utilize a proprietary “Helix” VLA model, which employs a dual-system architecture: System 1 operates at Hz for low-level motor control and reflexes, while System 2 operates at – Hz for high-level reasoning and task planning.¹³ This allows the robot to handle delicate tasks, such as manipulating objects with tactile sensors capable of detecting forces as small as grams.¹³

Pragmatic Industrialists: Agility Robotics and Apptronik

While Tesla and Figure chase general-purpose utility, Agility Robotics and Apptronik have focused on immediate commercial viability in the logistics sector.¹⁴ Agility’s Digit robot is currently the leader in real-world warehouse deployment, with active pilots at Amazon and GXO Logistics.⁴ Digit is designed to fit into existing workflows, performing “tote-to-person” tasks and handling material in environments where traditional automation is too rigid.⁴ Apptronik’s Apollo robot emphasizes “field serviceability” and modularity, using swappable -hour batteries to ensure near-continuous operation on factory floors.¹³

European Humanoid Ecosystem: Precision Engineering and Regulatory Moats

European humanoid startups are building a competitive advantage through high-precision engineering and a focus on “Cognitive Robotics”—machines that can safely interact with humans in unstructured domestic and healthcare settings.⁶ The European strategy is heavily influenced by the EU AI Act, which mandates transparency and safety, potentially creating the world’s most trusted platforms for human-centric applications.²¹

Neura Robotics: The Cognitive Pioneer

Metzingen-based Neura Robotics has emerged as Europe’s leader by adopting a “one-device” philosophy, viewing their 4NE1 humanoid as a “smartphone with arms and legs”.²⁰ Neura’s core innovation is the “Neuraverse” ecosystem, a platform that allows developers to create apps for robots using an open development environment and a security/operating system.²⁰ Their MAiRA and 4NE1 platforms are equipped with human-like sensing—vision, spatial audio processing, and -DOF force-torque sensors—allowing them to work alongside humans without protective cages.²⁰

1X Technologies: The Home Assistant Vision

Norway’s 1X Technologies (formerly Halodi Robotics) is betting on the consumer home market.¹⁷ Backed by the OpenAI Startup Fund, 1X is developing the NEO robot, which is designed with “muscle-like” actuators for safe interaction in domestic environments.¹ 1X is the first company to accept consumer pre-orders for a humanoid at a price point, targeting 2026 for initial deliveries.¹⁷ Their focus on “cloning human thought and behavior” intofoundation models for safety distinguishes them from purely industrial players.¹⁷

PAL Robotics and Shadow Robot: Niche Excellence

Spain’s PAL Robotics and the UK’s Shadow Robot Company represent the “deep tech” end of the European market. PAL’s KANGAROO platform is designed for dynamic locomotion and embodied AI research, while Shadow Robot’s “Shadow Hand” remains the industry standard for dexterous manipulation, utilizing optical D sensors and magnetic Hall effect arrays to provide human-level tactile feedback.²³

Similarity Mapping: Cross-Regional Convergence in Humanoid Strategy

Despite differences in funding and geographic origin, a “Transatlantic Consensus” is emerging on the technical and commercial requirements for humanoid success. US and European startups are converging on several key architectural and business strategies.

Technical Convergence: The Physical AI Stack

All leading startups have transitioned from traditional, pre-programmed control systems to end-to-end neural networks.¹ The standard “Physical AI” stack now consists of:

1. Foundational Models: Large Vision-Language-Action (VLA) models that allow robots to understand natural language commands and map them to physical motor outputs.²

2. Sim-to-Real Transfer: Utilizing high-fidelity digital twins and simulation environments (like NVIDIA Isaac Lab) to train robots on millions of variations of a task before they ever touch a real-world object.¹³

3. Multimodal Sensing: The integration of LiDAR, stereo cameras, and tactile skins to provide “contextual understanding” of the environment.⁴

Business Model Convergence: Robotics-as-a-Service (RaaS)

To overcome high upfront capital costs, startups on both sides of the Atlantic are adopting RaaS models.¹³ By charging an hourly rate—typically around to per hour—or a monthly subscription (e.g., for home units), these companies align their incentives with the user’s need for uptime and reliability.² This model effectively turns the robot into a “liquid labor” asset that can be scaled up or down based on seasonal demand.²⁵

Supply Chain Deep Dive: The Actuator and Component Moat

For investors, the most “investable” part of the humanoid revolution in 2026 is often the component layer rather than the robot OEMs themselves.⁵ The “Three Bs” of the humanoid supply chain—Brains (compute), Bodies (actuators), and Batteries—represent the critical bottlenecks for the industry.⁵

Brains: The Compute Infrastructure

Humanoids require massive on-board compute for real-time inference. NVIDIA’s Jetson Thor platform, released in 2025, has become the de facto standard, providing the performance of prior generations.² The software “brain” is increasingly moving toward decentralized identities (DIDs) and programmable wallets for robots, allowing them to autonomously pay for charging or software updates via blockchain protocols.²⁷

Bodies: The Actuator Market

The actuator is the single most expensive component of the humanoid, often accounting for to of the bill of materials (BOM).³ While Chinese suppliers like AgiBot have developed their “PowerFlow” integrated joints for peak torque density, Western suppliers like Maxon (Switzerland), Harmonic Drive (Japan/USA), and Kollmorgen (USA) maintain a lead in precision and reliability.²⁸

• Maxon Motor (CH): Specialized in high-efficiency joints (HEJ) that integrate motors, gearheads, sensors, and electronics into a single IP67-rated housing, capable of Nm peak torque.²⁹

• Kollmorgen (USA): Their TBM2G series provides frameless motors designed specifically for harmonic gearing, optimizing for torque density and weight in human-scale joints.³⁰

• Harmonic Drive (Global): Remains the monopoly holder for high-precision strain wave gearheads, though Chinese competitors like Green Harmonics are rapidly gaining domestic market share ( in China).¹²

Batteries: The Energy Density Challenge

Humanoids are significantly less energy-efficient than humans, requiring high-density lithium-ion packs to survive an -hour shift.⁵ While battery costs have fallen in ten years, the challenge remains thermal management during high-torque events.⁵ Tesla and Chinese firms like BYD have a strategic advantage here, as they can leverage their existing automotive battery production lines to reduce costs by compared to standalone robotics startups.⁹

Geopolitical Implications: The American Security Robotics Act and Beyond

Policy makers are increasingly viewing humanoid robotics through the lens of national security and economic sovereignty.¹⁰ The “robotics race” with China is now being equated to the semiconductor race in terms of strategic importance.²⁵

US Legislative Response

In March 2026, the US Congress introduced the American Security Robotics Act.³³ Proposed by Senators Tom Cotton and Chuck Schumer, the bill would:

1. Ban Federal Procurement: Prohibit the US government from purchasing or operating humanoid robots made by Chinese firms.³³

2. Mitigate Data Risks: Cite concerns that Chinese humanoids could serve as surveillance tools or be remotely controlled by foreign adversaries.³³

3. Protect Domestic Industry: Guard American research from “market flooding” by subsidized Chinese competitors.³³

Additionally, the United States has expanded Section 301 tariffs to include Chinese-made robotics and semiconductors, with initial rates set to increase significantly by mid-2027 to discourage reliance on the Chinese “T-Chain”.³⁴

The EU Regulatory Stance

Europe is pursuing a “Third Way,” focusing on the EU AI Act and the Digital Omnibus on AI Regulation.²¹ While the US focuses on defense and competition, the EU is prioritizing ethical frameworks:

• Prohibited Practices: The AI Act bans “harmful manipulation” and “emotion recognition” in workplaces, which could limit the deployment of certain Chinese or US bots that use facial analysis for productivity tracking.²¹

• Safety Certification: The revised ISO 10218-1/2:2025 and ANSI/A3 R15.06-2025 standards have shifted from certifying the “robot” to certifying the “deployment,” placing more responsibility on the factory owner to ensure a safe collaborative environment.³⁶

Horizon Europe and Sovereign Funding

To counter the massive subsidies of the Chinese government ( billion over 20 years), the European Commission has allocated over million in late 2025 for “Cognitive Computing” and “Soft Robotics” research.³⁷ The goal is to build “tech sovereignty” in AI, data, and robotics, ensuring that European industry remains competitive in Wave 2 (consumer/developer) and Wave 3 (medical/elder care) applications.³

The Future of Work: The National Autonomous Work Index (NAWI)

A novel economic framework is emerging for policy makers to measure national strength: the National Autonomous Work Index (NAWI), also known as the “Autonomous Workforce of the Nation” (AWN).²⁵ This metric measures the total annual hours of productive work performed by robots and automated systems.²⁵

The National Time Dividend

Policy makers are exploring the concept of a “National Time Dividend”—where the deflationary pressure created by humanoid productivity (e.g., reduction in the cost of domestic appliances) is distributed to the public through lower costs of living rather than direct taxation.²⁵ This approach aims to stabilize the economy as repetitive manual jobs shrink by to globally.¹

DARPA and Defense Innovation

The US Department of Defense, through the DARPA Triage Challenge, is driving humanoid innovation into high-risk medical and disaster-response scenarios.⁴⁰ By November 2026, the final event of this challenge will test whether autonomous systems can perform medical triage in mass-casualty incidents, providing a “dual-use” pathway for humanoid technology to move from the battlefield to civilian emergency services.⁴⁰

Conclusion: Actionable Insights for Investors and Policy Makers

The humanoid robotics sector has reached a genuine growth inflection point. The convergence of Physical AI software with mature hardware supply chains has made mass adoption a 2027–2030 reality.²

For Investors:

• Pivot to the Component Layer: The most defensive “moats” are held by precision actuator manufacturers and specialized AI chip designers (the “Three Bs”).⁵

• Prioritize Data-Rich OEMs: Firms like Figure AI, Tesla, and Neura Robotics that are actively amassing “real-world spatial data” through enterprise pilots will hold the long-term software advantage.¹⁰

• Sectoral Timing: Invest in “Wave 1” automotive and logistics specialists now, but keep a close watch on “Wave 2” consumer and “Wave 3” healthcare platforms for the 2028 horizon.³

For Policy Makers:

• Subsidize the Supply Chain, not just R&D: To counter China’s lead, the US and EU must incentivize domestic “Actuator Hubs” and battery capacity specifically for mobile robotics.⁹

• Streamline Regulatory Compliance: The EU must ensure the AI Act does not become a barrier to adoption, while the US must coordinate state-level AI laws to provide a “minimally burdensome” national framework for startups.²²

• Measure “Sovereign Work-Hours”: Adopt metrics like the NAWI to track national productivity and ensure that the “Robotics Boom” translates into a structural reduction in the cost of living for citizens.²⁵

The humanoid revolution is the final stage of the digital transformation. By embedding intelligence into physical form, we are unlocking the largest total addressable market in history.⁸ Success will depend on the ability to balance rapid iteration with uncompromising safety and national security.

Reference:

1. The Dawn of Humanoid Robots and Physical AI - Leo Wealth, accessed March 31, 2026, https://leowealth.com/insights/the-dawn-of-humanoid-robots-and-physical-ai/

2. Robotics: AI Moves into the Physical Economy - Global X ETFs, accessed March 31, 2026, https://www.globalxetfs.com/articles/robotics-ai-moves-into-the-physical-economy

3. The Global Humanoid Robots Market 2026-2036 - Advanced and Emerging Technology Market Research, accessed March 31, 2026, https://www.futuremarketsinc.com/the-global-humanoid-robots-market-2026-2036-3/

4. Humanoid Robot Market Size: $38B by 2035 — Growth Data & Forecasts - Blog, accessed March 31, 2026, https://blog.robozaps.com/b/market-size-for-humanoid-robots

5. AI gets physical: Innovation meets opportunity - Barclays Investment Bank, accessed March 31, 2026, https://www.ib.barclays/our-insights/series/impact-series/ai-gets-physical-innovation-meets-opportunity.html

6. Humanoid Robot Market Trends | Market Intelligence 2033, accessed March 31, 2026, https://www.skyquestt.com/report/humanoid-robot-market

7. A new era: physical AI and humanoids enter the stage - Betashares, accessed March 31, 2026, https://www.betashares.com.au/insights/a-new-era-robots/

8. Musk has a plan to make human labor obsolete. Billionaires are joining in., accessed March 31, 2026, https://www.washingtonpost.com/technology/2026/03/27/musk-optimus-robot-physical-ai/

9. China’s humanoid robot cluster – 80 percent global market share ..., accessed March 31, 2026, https://xpert.digital/en/chinas-humanoid-robot-cluster/

10. Is China Leading the Robotics Revolution? | ChinaPower Project, accessed March 31, 2026, https://chinapower.csis.org/china-industrial-robots/

11. AGIBOT Reaches 10,000 Units as Real-World Demand for Robots Accelerates, accessed March 31, 2026, https://www.prnewswire.com/news-releases/agibot-reaches-10-000-units-as-real-world-demand-for-robots-accelerates-302728295.html

12. Three core supply chains of humanoid robots (Tesla Chain, Huawei Chain, Unitree Chain), accessed March 31, 2026, https://www.yicaiglobal.com/star50news/2025_03_056800594033476370453

13. 30+ Humanoid Robot Companies Ranked [2026] - Blog - Robozaps, accessed March 31, 2026, https://blog.robozaps.com/b/humanoid-robot-companies

14. Top 12 humanoid robotics companies to watch in 2026 - Standard Bots, accessed March 31, 2026, https://standardbots.com/blog/humanoid-robotics-companies

15. Robozaps Blog - Humanoid Robotics News & Insights, accessed March 31, 2026, https://blog.robozaps.com/

16. Humanoids on the move: How 2025 became the breakthrough year ..., accessed March 31, 2026, https://techequity-ai.org/humanoids-on-the-move-how-2025-became-the-breakthrough-year-for-ai-driven-robotics/

17. 12 Fastest Growing Robotics Companies and Startups - Landbase, accessed March 31, 2026, https://www.landbase.com/blog/fastest-growing-robotics-companies

18. Figure vs Apptronik vs Agility Robotics - Sacra, accessed March 31, 2026, https://sacra.com/research/figure-vs-apptronik-vs-agility-robotics/

19. 31 Best Humanoid Robots [2026 Ranked] - Blog, accessed March 31, 2026, https://blog.robozaps.com/b/best-humanoid-robots

20. NEURA Robotics x Made for Germany, accessed March 31, 2026, https://neura-robotics.com/neura-robotics-made-for-germany-initiative/

21. EU AI Act - Updates, Compliance, Training, accessed March 31, 2026,

    https://www.artificial-intelligence-act.com/

22. The EU AI Act: A Story without a Plot? A Lack of Vision in the Digital AI Omnibus Proposal threatens constitutional AI Governance - Jean Monnet Saar, accessed March 31, 2026, https://jean-monnet-saar.eu/?p=322956

23. PAL Robotics | Home, accessed March 31, 2026, https://pal-robotics.com/

24. Humanoid Robotics Component Datasheet — Live Supplier, Pricing & Certification Updates, accessed March 31, 2026, https://humanoid.press/humanoid-press/datasheet/

25. The Autonomous Workforce of the Nation: A National Time-Bank Strategy - Liquid Labor, accessed March 31, 2026, https://www.liquid-labor.com/

26. Is robotics the next frontier for AI? | J.P. Morgan Asset Management, accessed March 31, 2026, https://am.jpmorgan.com/sg/en/asset-management/per/insights/market-insights/market-updates/on-the-minds-of-investors/is-robotics-the-next-frontier-for-ai/

27. physicalai Community Insights & Market Sentiment | Binance Square, accessed March 31, 2026, https://www.binance.com/en/square/hashtag/PhysicalAI

28. Robotic Actuator and Servo Motor Selection Guide - CubeMars, accessed March 31, 2026, https://www.cubemars.com/robotic-actuator-and-servo-motor-selection-guide.html

29. Powerful robot drives: Perfect solutions for modern robotics - Maxon Motor, accessed March 31, 2026, https://www.maxongroup.com/en-us/market-solutions/mobility-solutions/robotics

30. Humanoid Robots - Kollmorgen, accessed March 31, 2026, https://www.kollmorgen.com/en-us/solutions/robotics/humanoid-robots

31. maxon at the SPS 2025: New drive solutions for robotics, automation and mobility, accessed March 31, 2026, https://www.maxongroup.com/en-us/news-and-events/news/maxon-at-the-sps-2025-new-drive-solutions-for-robotics-automation-and-mobility-288150

32. Humanoid Robots 2025-2035: Technologies, Markets and Opportunities - IDTechEx, accessed March 31, 2026, https://www.idtechex.com/en/research-report/humanoid-robots/1093

33. US lawmakers to introduce bill to ban government use of Chinese ..., accessed March 31, 2026, https://virginiabusiness.com/us-lawmakers-to-introduce-bill-to-ban-government-use-of-chinese-robots/

34. USTR Issues Affirmative Determination in China Semiconductor Section 301 Investigation, accessed March 31, 2026, https://www.thompsonhinesmartrade.com/2026/01/ustr-issues-affirmative-determination-in-china-semiconductor-section-301-investigation/

35. The U.S. Imposes New Section 301 Tariffs on Chinese Semiconductors, accessed March 31, 2026, https://www.internationaltradeinsights.com/2025/12/the-u-s-imposes-new-section-301-tariffs-on-chinese-semiconductors/

36. Top Quadruped Robot Safety Standards and Risk Assessment for Industry 2026, accessed March 31, 2026, https://www.oxmaint.com/blog/post/quadruped-robot-safety-standards-2026

37. China to Invest 1 Trillion Yuan in Robotics and High-Tech Industries, accessed March 31, 2026, https://ifr.org/news/china-to-invest-1-trillion-yuan-in-robotics-and-high-tech-industries/

38. Horizon Europe: Digital — 2025 call results: 20 proposals funded, accessed March 31, 2026, https://hadea.ec.europa.eu/news/horizon-europe-digital-2025-call-results-20-proposals-funded-2026-02-20_en

39. Horizon Europe - funding digital technology and research, accessed March 31, 2026, https://digital-strategy.ec.europa.eu/en/activities/horizon-europe-funding-digital

40. Inside DARPA’s High-Stakes Challenge to Build Life-Saving Robots - eWeek, accessed March 31, 2026, https://www.eweek.com/news/darpa-challenge-life-saving-robots/

41. DARPA Triage Challenge, accessed March 31, 2026, https://www.darpa.mil/research/challenges/darpa-triage-challenge

42. Challenge Events - DARPA, accessed March 31, 2026, https://www.darpa.mil/research/challenges/darpa-triage-challenge/events

43. Sunny Cheung Testimony - U.S.-China Economic and Security Review Commission, accessed March 31, 2026, https://www.uscc.gov/sites/default/files/2025-02/Sunny_Cheung_Testimony.pdf

44. 2026 AI Laws Update: Key Regulations and Practical Guidance - Gunderson Dettmer, accessed March 31, 2026, https://www.gunder.com/en/news-insights/insights/2026-ai-laws-update-key-regulations-and-practical-guidance