The Forged Leaders and Blunted Track: Five Capital Dilemmas Facing Embodied AI

In the first half of the year, there was 46 billion yuan in financing, 226 companies received funds, and the top 20 received 70% - the "Matthew effect" and "capital barrier" are objective descriptions of the current situation. These terms are all correct, but I want to summarize the deeper information.

The Underlying Color of the Matthew Effect: Capital Involution

In the first half of 2026, there were 288 financings in the embodied intelligence track, with a disclosed total of over 46 billion yuan, involving 226 companies.

The distribution of funds is extremely unbalanced -

The top 5 companies absorbed about 17.1 billion yuan (accounting for 37%), the top 20 companies together took about 33 billion yuan (over 70%), and the remaining more than 200 companies shared about 12.4 billion yuan, with an average of only tens of millions per company.

For example, Qianxun Intelligence raised a cumulative 4.5 billion yuan in 3 months, which is equivalent to about one-third of the total of more than 200 companies.

The most intuitive interpretation is the "Matthew effect of the strong getting stronger", but the more I think about it, the more I wonder if this interpretation is problematic?

The Matthew effect can be driven by two forces:

One is the positive feedback after market verification - good products, many customers, increasing revenue, and capital addition;

The other is the herd mentality under capital inertia - the leading companies have received funds, so they should receive more funds.

In the current Matthew effect of embodied intelligence, the latter factor is significantly more prominent than the former.

An inescapable fact is that in all public reports, we can hardly find quantifiable revenue data, shipment volume, or number of customers of these leading companies.

The label of "leading" is more defined by the amount of financing, and there are very few verified by product - market fit (PMF).

The rankings of financing, technology, and business are originally three different things, but in the current market narrative, they are quietly equated.

The fact that "49 companies completed two or more rounds of financing in half a year" is also worth pondering.

Frequent financing itself does not mean that a company is strong - it may mean that the company burns money quickly, capital is competing for market share, or it may mean that the valuation anchor is unstable.

Then, when the mentality of capital changes from "invest after understanding" to "rush in for fear of missing out", the Matthew effect becomes a bubble accelerator.

Valuation Acceleration: A Jump Detached from Fundamentals

I was deeply impressed by the valuation change of a star company:

In January 2026, its valuation in the A+ round was 3 billion yuan, and in February, its valuation in the B round was 10 billion yuan - a 3.3 - fold jump in one month.

Even for SaaS companies with high growth in the past, the valuation doubling from the A round to the B round usually required 6 to 12 months of product iteration and revenue growth support.

And the valuation change of the current embodied intelligence in one month is actually just "getting more money".

Does it have any actual technological breakthroughs or large - scale orders?

Recently, many media have reported that a dexterous hand company had a valuation of 20 billion yuan last year, and its target valuation in the new round is over 40 billion yuan. This figure is almost equal to the IPO issuance valuation of Unitree Technology, which is 42 billion yuan - and Unitree is recognized as the number one in terms of shipment volume in the track.

Phoenix Finance has a statistic that reveals a more macroscopic deviation:

As of the end of the first quarter of 2026, the total market value of the A - share humanoid robot concept sector reached 11.89 trillion yuan, while the actual global shipment volume of humanoid robots was only about 2,000 units.

The price - to - sales ratio of the industry leader, Unitree Technology, is 24.7 times, but 73.6% of its revenue comes from scientific research and education customers, and the industrial rigid - demand scenario only accounts for about 9%. In Q1 2026, its revenue growth rate dropped sharply from 332.64% in the same period last year to 68.49%, and its non - recurring net profit was halved year - on - year;

The PS of the chasing company, Deep Robotics, reached 41 times. It only sold 1 humanoid robot in the whole year of 2025, corresponding to a PE of about 485 times.

Where exactly is the valuation anchored?

When "gambling on growth expectations" replaces "performance verification" as the pricing logic, valuation has changed from a condensation of information to a projection of imagination.

The Hidden Cost of State - owned Capital Entry: Capital for Localization, Localization for Shackles

In the article "In the first half of 2026, 46 billion yuan was invested in embodied intelligence, only 'enriching' 20 companies?", the data shows that in the heavy - weight transactions of hundreds of millions of yuan, the participation rate of investment institutions with state - owned background is as high as 42%.

The 700 million yuan in the A round of the Beijing Humanoid Robot Innovation Center was almost entirely taken by local state - owned platforms in Beijing; Zhipingfang received state - owned capital from Changzhou and Chengdu, and it is certain that it will set up a second headquarters or a regional R & D center in these places.

"Capital for localization, localization for orders" is the standard model of local state - owned capital for investment promotion now.

To put it more bluntly: State - owned capital investment comes with triple shackles of geographical lock - in, gambling constraints, and exit restrictions.

The invested enterprises must establish production capacity and pay taxes in the city where the investor is located.

However, for the enterprises, the supply chain is scattered in multiple places, and the management cost increases sharply; capacity adjustment changes from a business decision to a political decision; when it comes to IPO or mergers and acquisitions, state - owned capital pays more attention to "whether the enterprise stays locally".

The exit return has become a secondary consideration - for the entrepreneurial team, this is a hidden dilution of control.

The photovoltaic industry is undoubtedly a lesson from the past.

Under the "dual - carbon" narrative, local governments flocked in. In 2025, the nominal production capacity of China's photovoltaic industry exceeded 1,100 GW, while the global demand was only about 600 GW, with a production capacity redundancy of nearly twice; the leading enterprises suffered a total loss of over 28 billion yuan, nearly 60% of the 77 listed photovoltaic enterprises suffered losses, and more than 150 went bankrupt and liquidated; the total liabilities of the photovoltaic and energy storage industries reached 6.5 trillion yuan. Even more shocking is that after the projects supported by local governments in the form of "investing in real estate" and "building factories on behalf of others" went bankrupt, the state - owned capital of the governments was sued for debt collection.

When "nanny - style blood transfusion" turns into a "government - enterprise debt quagmire", free exit is no longer possible.

The current participation rate of state - owned capital in the embodied intelligence track is highly similar to that in the early stage of the photovoltaic industry. A 42% participation rate in large - scale financing means that nearly half of the leading enterprises have bound geographical commitments. Once the commercialization fails to meet expectations, these commitments will become sunk costs.

The Fate of the "Water Sellers": The Safety Margin of Components Is Not Safe

The data in the original article shows that the average financing per company in the component track is 320 million yuan, which is the highest in the entire embodied intelligence industry chain.

The investment logic for component companies seems impeccable: no matter which complete machine enterprise wins, it cannot do without dexterous hands, sensors, and joint modules - "selling shovels during the gold rush".

But this logic has a fatal blind spot: Complete machine enterprises have the ability and motivation to conduct self - research.

ZhiYuan Robotics has invested in or incubated Fullive.AI, ZhiShen Technology, and ZhiDing Robotics, and carried out vertical supply chain integration in the track; the core joint modules of Unitree Technology are also self - developed. When the shipment volume of these enterprises reaches a certain scale, the marginal cost of self - research will be lower than that of external procurement - this is an iron law in the manufacturing industry.

The more cruel evolution path for component enterprises is: in the early stage, complete machine enterprises rely on external procurement. After the scale increases, they start self - research, the demand for external procurement shrinks, and component enterprises are forced to reduce prices or transform.

The smartphone industry chain has completely gone through this path - touch chips, screen modules, and camera modules were all the territory of independent suppliers in the early stage, and most of them were finally vertically integrated by complete machine manufacturers.

The only moat for component enterprises is "low - priority components that complete machine enterprises cannot or do not want to make". But "not wanting to do" and "not being able to do" are two different things. When the market is large enough, complete machine enterprises may change their strategies at any time.

The premise of the safety margin of the "water sellers" is that gold diggers always need to purchase shovels externally. But in the high - capital - density track of embodied intelligence, complete machine enterprises have enough funds and motivation to make their own shovels.

The Lesson from Autonomous Driving

According to IT Juzi data, in 2021, the annual financing in the Chinese autonomous driving track was nearly 100 billion yuan, reaching a historical high; in 2022, it dropped sharply by 74%.

Haomo.AI is the most complete specimen of this bubble:

In 2021, it raised nearly 1 billion yuan in the A round, and its valuation exceeded 1 billion US dollars, making it a unicorn;

In 2025, its cash flow completely broke down, its accounts were frozen, and it was unable to pay the 31,500 - yuan execution target, and all employees stopped working.

It only took four years from being a unicorn to shutting down.

The bubble path of autonomous driving is extremely clear: Capital frenzy → Leading companies attract capital → Valuation detaches from products → Commercialization fails to meet expectations → Capital retreats → Market clearance.

The current embodied intelligence track is precisely replicating the first three stages.

However, the challenges faced by embodied intelligence are more severe than those of autonomous driving.

Autonomous driving at least has a clear PMF - Robotaxi, with clear user needs, a clear business model, and a converging technical route. Embodied intelligence has none of these: PMF is undefined, the business model is unvalidated, and the technical route is not converged (VLA or VLM? Bipedal or wheeled? General humanoid or scenario - specific?)

A decrease in cost does not mean that the business closed - loop is established. Currently, the core tasks of robots in the industrial scenario are still repetitive tasks such as material box handling and loading and unloading, and there is still a huge gap from "replacing skilled workers".

Assuming that a humanoid robot worth 170,000 yuan replaces a worker with an annual cost of 80,000 - 120,000 yuan, the rough pay - back period is 8 - 14 months - but this does not include hidden costs such as maintenance, production change debugging, and scenario adaptation. After including these costs, the pay - back period may be extended to 18 - 24 months.

For most small and medium - sized enterprises, this ROI is still not enough to support large - scale procurement decisions.

Bubbles Are Not Terrible, Blunting Is

After witnessing the rise and fall of countless trends, I still think that the bubble itself is not terrible.

On the contrary, every major technological revolution is accompanied by a bubble - railways, the Internet, new energy vehicles, and large AI models are no exception.

The bubble is a necessary cost of technological revolution. It provides a capital buffer for the early high - cost technological exploration.

What really deserves our vigilance is whether the bubble is "blunted" - whether the capital is still flowing in an effective direction.

The symptoms of blunting in the current track are already very obvious.

First, look at the funds. 70% of the money flows into 20 companies that already have sufficient ammunition, and the marginal utility is decreasing. Which contributes more to industrial progress: a company with 4.5 billion yuan in cash getting another 1.5 billion yuan, or a small and medium - sized enterprise with technology but lacking 50 million yuan getting this money? The answer is self - evident, but the market is choosing the former.

Then, look at the signals. The valuation detaches from the product progress, and the financing ranking replaces the product ranking as the main variable of industry discourse power. When "who gets more money" is more important than "whose product is more useful", the pricing function of capital fails.

What worries me most is the innovation pipeline.

The total of the seed round and the angel round is less than 1.3 billion yuan, only accounting for 3% of the entire track. The entry channel for grass - roots entrepreneurs is almost closed - without a background in large companies or academic titles, they can't even get an angel round.

However, there is still a possibility of grass - roots counter - attack in the track.

Especially recently, the narrative of grass - roots counter - attack in 2026 is in vogue - Zhangxue Motorcycle was established just over two years ago, and it has already won 6 stage championships in this year's World Superbike Championship, ending the monopoly of giants such as Ducati, Yamaha, and Kawasaki.

Its path is very simple - not following the hottest concepts, conducting full - stack self - research on core technologies, first achieving large - scale operation in the consumer market, and iterating technologies with real user feedback.

This kind of "scale flywheel" - obtaining users and data through real product sales, driving technological iteration, reducing costs, and expanding sales - forms a sharp contrast with the "financing flywheel" of leading companies - burning money to make demos, attracting the next round of financing, and making bigger demos.

However, the current 3% share of early - stage funds means that these grass - roots break - through players have to face a thick wall of capital at the starting stage.

Finally, there is the issue of exit. A 42% participation rate of state - owned capital means that nearly half of the leading enterprises have bound geographical commitments. Once the commercialization fails to meet expectations, they can neither shrink their front lines, nor be easily merged and acquired, nor freely relocate.

Judgment and Coordinates

I think the current embodied intelligence track is in the bubble expansion period.

This does not mean that the track has no future - Embodied intelligence is the inevitable path for AI to enter the physical world, and its long - term value is beyond doubt. But "the right direction" and "the right rhythm" are two different things.

The typical characteristics of the bubble expansion period are: the financing growth rate is much faster than the commercialization growth rate, the valuation growth rate is much faster than the revenue growth rate, and the production capacity planning is much faster than the demand verification.

All three are true in the current track.

The time window given by history is not generous.

It took about 3 years for autonomous driving to go from frenzy to retreat, and about 4 years for the photovoltaic industry to go from explosion to overcapacity. The "year of mass production" of embodied intelligence is 2026. According to a similar rhythm, 2028 - 2029 will be the critical time point when the verification period window closes. By then, if the leading enterprises still cannot present large - scale revenue and a replicable business model, the retreat of capital will be inevitable.

Several inflection - point signals worthy of continuous tracking:

The valuation/revenue ratio of leading enterprises drops below 20 times (currently, Unitree's PS is 24.7 times, far from meeting the standard);

The participation rate of state - owned capital drops from 42% to below 25% (market - oriented financial capital regains the pricing power);

The proportion of self - developed components by complete machine enterprises exceeds 40% (the window period for independent component suppliers begins to narrow);

The first event of significant lay - offs or valuation reduction by a leading enterprise (a leading indicator of the bubble bursting