Innerhalb von zwei Jahren erreichte die Schätzung des Unternehmens Qianxun Intelligence über 10 Milliarden Yuan, aber das Unternehmen wagt es nicht, sich zu entspannen.

The "Billionaires' Club" for humanoid robots has welcomed its sixth player.

On February 24, Spirit AI announced that it had raised nearly 2 billion yuan in two consecutive financing rounds and achieved a market value of over 10 billion yuan.

Getting money at this time is in itself a signal.

The list of investors is also interesting: New shareholders such as Yunfeng Capital, Sequoia China, TCL Venture Capital, and 360 Fund have participated, while old shareholders such as Shunwei Capital, Prosperity7, and Fortune Capital have continued to invest. Those familiar with the primary market know that this structure of "new and old investors together" means that the institutions are not just watching but clearly taking positions.

Before 2026, there were only three members in the "Billionaires' Club" for humanoid robots: the robots of Zhiyuan, a well - known Bilibili UP member, Unitree Technology, which appeared at the Spring Festival Gala, and Galaxy Universal stood side by side like "three pillars". In 2026, Xinghaitu, Zhipingfang, and Spirit AI successively stepped onto the "playing field" of billion - dollar companies.

In other words, a market value of 10 billion yuan is no longer an isolated case in the field of humanoid robots, but is beginning to form an industry effect.

Founder Han Fengtao said quite directly in an interview with LatePost AI: "The year 2026 for Embodied AI will be very similar to the year 2023 for Large Language Models. If you don't get enough money and your model performance doesn't make it into the top group, you have no chance to get to the table."

To put it another way: In the field of humanoid robots, financing has changed from an "accelerator" to an "admission ticket". Without sufficient capital, it directly means elimination.

The fact that Spirit AI got 2 billion yuan at this time reflects to some extent a subtle market sentiment: Investors are still willing to bet on the future, but only on the premise that you are already in the first group.

A billion is not just a market - value label.

It is an admission ticket and also a pressure gauge. The moment you get it, the countdown begins.

The crazy record time of 2 years for a market value of 10 billion

Among the members of the "Billionaires' Club", Spirit AI, founded in 2024, is not only one of the youngest companies, but it only took 26 months to reach a market value of over 10 billion yuan.

The financing curve has hardly any pauses.

In January 2024, the company was founded, and in 7 months, nearly 200 million yuan was raised in the Seed Round and Angel Round. In 2025, 528 million yuan was raised successively in the Pre - A Round and nearly 600 million yuan in the Pre - A+ Round. At the beginning of 2026, nearly 2 billion yuan was raised again. In total, 6 financing rounds in two years, about 3.328 billion yuan in total.

This pace is more of a "race against time" than "high - speed".

In comparison, Unitree Technology, founded in 2016, took 9 years to reach a market value of over 10 billion yuan. Spirit AI has shortened this period to a little over two years. Essentially, it is not just an increase in efficiency but a different environment: It stands in the time - window after the outbreak of Large Language Models.

But Spirit AI didn't come out of nowhere.

Founder Han Fengtao is a twice - entrepreneur in the robotics industry. In 2015, he founded the company Luoshi, which manufactures collaborative robots, with others and served as CTO there. He led the team in manufacturing several dozen models and a total of over 20,000 industrial robots.

According to him, in the past ten years, he has seen how "the local production rate of industrial robots has risen from 3% to over 50%, but most companies don't make money."

The core problem of not making money lies not in the fact that there are too few robots, but in their limited working ability.

Industrial robots are highly dependent on the adaptation and fixation of scenarios. For a new task, the logic has to be rewritten. The generalization ability is weak, the scale effects are limited, and the profit margin is constantly reduced. Han Fengtao also tried to take the path of "hardware upgrade + early deep learning" at that time, but due to the limitations of AI capabilities at that time, he could never develop a real "general robot brain".

The real turning point came in 2023. When ChatGPT appeared and the generalization ability of Large Language Models was shown, he realized that the variable had changed.

"Gears, integrated joints, motors, complete machines... but that's not the biggest opportunity for humanoid robots. The biggest opportunity is the brain. So I'm looking for someone who can build a brain."

In the summer of 2023, Han Fengtao met more than 100 people one after another and finally found Gao Yang, the director of the Laboratory for Visual and Humanoid Robotics at Tsinghua University and one of the "four researchers who returned from Berkeley". "Gao Yang has been researching end - to - end self - driving vehicles since 2017 and knows how to use Internet videos for training. Technically, he is absolutely reliable."

This was a decision - making moment.

Investors once evaluated that the team of Spirit AI is comparable to those of Pi and Google DeepMind. The evaluation may stand, but the key point is: It bets on the path of integrating "Large Language Models + Embodied AI".

In January 2024, Spirit AI was founded. In July of the same year, the first robot Moz0 of Spirit AI was born. In August, nearly 200 million yuan was invested successively in the Seed Round and Angel Round.

In March 2025, the preview version of the Spirit V1 VLA model was released. The humanoid robot equipped with it can perform long and complex tasks such as "folding clothes". As soon as the technical milestone was reached, financing followed directly. 528 million yuan in the Pre - A Round was quickly realized.

In January 2026, the company further opened the Spirit v1.5 model. In the Table30 list of the RoboChallenge, the Spirit v1.5 scored 66.09 points with a success rate of 50.33%, achieving a better result than Pi0.5 and becoming the first Chinese open - source Embodied model to outperform Pi0.5 in a public benchmark test. Subsequently, nearly 2 billion yuan in financing was obtained.

If you look at the rapid financing of Spirit AI, you can clearly see a pattern: Every leap in model performance leads to an increase in market value.

This is a typical "technology - driven financing curve". But it also means a higher risk: As soon as the performance evolution slows down, the patience of investors will quickly wear thin.

The bet on data collection

Unitree's show "WuBOT" at the Spring Festival Gala has once again ignited the emotional value of humanoid robots; the market value of Spirit AI, which has risen to over 10 billion yuan in two years, has quietly raised the upper limit of the entire industry. In 2026, the popularity of humanoid robots is undisputed.

But Han Fengtao, who has received the financing, doesn't feel relaxed. In his opinion, this popularity is more of a watershed than a celebration: "The year 2026 for Embodied AI is very similar to the year 2023 for Large Language Models. If you don't get enough money and your model performance doesn't make it into the top group, you have no chance to get to the table."

The subtext is clear: The time - window is closing quickly. The double pressure of capital and technology means that the probability of making a comeback once you fall behind is almost negligible.

After getting the money, Spirit AI has set an even more aggressive goal: To reach the global top 3 of Embodied AI brains and increase the size of effective training data to 1 million hours. The best open - source model so far has only used 10,000 hours of data. Now it's about a 100 - fold expansion.

In contrast to language models based on Internet texts, robot models need real interaction data from the physical world: the repetition of action execution, environmental feedback, force control errors, and error correction. Every failed grasp, every path deviation becomes a training example for the model. The size and structure of the data directly determine the upper limit of generalization ability.

The problem is that this is a very costly way. Real - world data can't be "collected like with a web crawler". It requires hardware, space, personnel, and time. As the volume increases, the costs rise exponentially.

So Spirit AI is putting 80% of its resources into the data system. Essentially, it's a strategic bet. Han Fengtao is very direct in his assessment: The reason why humanoid robots haven't exploded yet is not that the algorithms aren't smart enough, but that there aren't enough data. If he had enough capital early on, he would unhesitatingly bet "All in on data".

But the crucial difference lies not in "whether you need data" but in "how you get data".

Spirit AI didn't choose the path of "small and refined" and didn't rely on a few laboratory samples. They chose an industrial method that prioritizes scale and keeps costs controllable.

The team uses Gao Yang's experience in self - driving technology and Internet video training. Instead of building a complex world model step by step, it first uses a large amount of Internet videos for pre - training to reach a higher starting point with fewer parameters and at the same time significantly reduce computing power costs. This is actually the typical thinking of Large Language Models: First, broad coverage, then physical adaptation.

For offline data, multiple ways are also used in parallel: remote control, data collection with portable devices, and reproduction of real scenarios. The goal is not to achieve "perfect samples" but to reduce data costs. According to reports, compared with traditional methods, costs have been reduced by about 90%. Only when costs are reduced can mass expansion be supported.

What's even more interesting is that the team doesn't deliberately clean up "dirty data". Failed grasps, chaotic movements, and non - standardized operations are kept. Their logic is: The real world is chaotic anyway. If the data is too clean, the model won't learn to deal with complex scenarios. Compared with highly - selected laboratory samples, this noisy, diverse data is more conducive to the generalization ability of the model and can even approach zero - shot generalization.

This training logic has already been shown at the model level. The Spirit v1 released in March 2025 has achieved long - term manipulation of flexible objects - folding clothes is no longer just a single grasp but a complete process of continuous motion planning and execution. The focus is not on "having grasped something" but on "constantly acting correctly".

In December of the same year, Moz performed continuous tasks such as tidying up the table, throwing away the trash, putting the chair back, and heating in the microwave oven in the field of rehabilitation and elderly care at the Intelligent Robot Competition 2025. It shows not only a single ability but a decision - making chain over multiple steps.

Han Fengtao, who comes from the era of industrial robots, knows better than most that it's never a one - time spectacular demonstration that determines the commercial value, but the ability to repeat stably 1,000 times, 10,000 times. Only the repeatable, transferable, and scalable ability can bring real commercial value.

"All in on data" is currently the most rational choice but also a very risky bet.

Because once the scale is maximized but the performance doesn't increase linearly, the patience of investors will quickly run out. The market can pay for technical visions, but not indefinitely for efficiency problems.

Hanging on to Ningde Times isn't enough

Nevertheless, Spirit AI, which can fold clothes, has first decided to enter the factory.

Although Gao Yang proposed the "Double - Ten Plan" in 2024, stating that "10% of the world's population should have their own robots in 10 years", in reality, Spirit AI didn't first bet on the private household sector but decided to "enter the factory and work".

Behind this decision lies a pragmatic calculation of the risk - return ratio.

The private household sector has huge imagination potential, but there are many variables, the willingness to pay is unclear, and the cost structure is not clear. In the industrial sector, the tasks are clearly defined, the standards are known, and the testing is quantifiable. For a company founded less than two years ago, cash flow and testing speed are more important than the story.

Han Fengtao doesn't have a romantic view of industrial robots. After ten years in the industry, he sees the reality of a "hard - to - make - money" sector: Traditional industrial robots have only a single function and can only perform fixed work processes. The penetration rate is rising very slowly. There are about 100 million industrial workers in China, but the number of industrial robots is only about 3 million, and only about 300,000 new ones are added each year. The structural substitution has not yet...