The first batch of AI hardware that became extremely popular is quietly exiting the market.
In November 2025, Rabbit, once regarded as the "most dazzling new species" in the AI era, finally collapsed due to a broken capital chain.
The Rabbit R1 was once regarded as the "iPhone" of the AI era, setting a miracle of selling 100,000 units in four days. However, this popular product, prematurely packaged as an "entry - level product", raised users' expectations too high. Later, due to its inability to meet the product's capabilities, the return rate was extremely high, and its reputation declined severely. Eventually, this team of only 26 people fell into a dilemma of unpaid salaries and depleted cash flow.
"There were too many negative feedbacks on this product," said Silvia, who has years of experience in hardware investment. "The more it was sold, the more negative feedbacks it received from users, and the high return rate ultimately dragged down the company's cash flow."
Among all the people we talked to, Silvia was the one most curious about the overall development of hardware. In the past few years, she has been actively involved in the investment and startup circles, and finally got involved in the research and development process of specific hardware. This is because of a distressing reality: current AI hardware products simply cannot meet users' expectations.
Rabbit's experience is not an isolated case. We once visited a store selling AI hardware on the Digital Culture Street in Hangzhou and tried out several products priced between 399 and 1,499 yuan, which mainly focused on interaction and English learning. They felt more like toys and were basically unable to achieve real - time voice interaction. Even those with visual viewing functions were just reading from the script and could not compare with software assistants like Doubao, Qianwen, Yuanbao, and KiMi on mobile phones.
In the mainstream narrative of single - product popularity and capital betting on AI hardware, Rabbit is undoubtedly the most extreme example. Its enlightenment is obvious: it's not that the path of AI hardware lacks imagination, but that the current AI hardware is far from reaching the "iPhone moment". Most AI hardware is just a transitional product and is destined to be eliminated by the era.
So, in this field with the most certain opportunities, where are the opportunities for entrepreneurs to enter the game?
01
AI for the Sake of AI
As AI companion toys gradually move from concept to the public's life, the real - life experience always comes with a bit of a gap.
Last week, Jianzi spent 1,000 yuan to buy an AI picture - book robot online. After the experience, he was disappointed but also relieved. He is a veteran in the hardware industry. After leaving Hema, he started his own business and is now developing an English enlightenment product for children called Baobao Long. To understand the market, he has bought no less than seven AI educational robots in the past few months, but none of them satisfied him.
Either the AI's voice was mechanical when reading picture books and it had no ability for random interaction, or the voice recognition was inaccurate, with a reaction time of up to 3 - 4 seconds. "People, especially children, can tolerate a duration of one second. Currently, the time for large models to process pictures or for cloud - based large models to respond to voices can basically be within 2.5 seconds," said Jianzi. "A reaction time of 3 - 4 seconds is a bit excessive."
After market research, he felt that most products only adopted the concept of AI but did not really solve the problems. Of course, the real reason is that the computing power cost is too high, and edge - side chips cannot be popularized in consumer hardware products. Only cloud - based large models can be used, and the audio and text conversion in cloud model solutions has a high rate of hallucinations.
In the specific scenario of reading picture books, the AI picture - book robot he bought was relatively timely in voice interaction, but the AI's voice lacked warmth, making it difficult to create a happy learning atmosphere for children, and it could not be interrupted midway. Otherwise, if these functions were to be realized, the token computing power consumption would be huge, and the business model would not work.
"This kind of AI experience is very strange, more like AI for the sake of AI," said Jianzi. He wants to deeply polish the product to truly solve users' problems.
Silvia also thinks the same way. Due to the limitations of software and hardware technologies, most AI hardware does not meet users' expectations. Even most of the popular AI products in the current market rely on other elements.
The most representative case is fuzozo, which was launched in June 2025. It created a paradigm for hanging AI hardware and was called the "AI version of Labubu". It sold thousands of units in just a few days after its launch and a total of 200,000 units in half a year.
Silvia believes that people are more attracted to the trendy toy elements. Some users buy AI hardware on a whim, and their ultimate fate is to gather dust. Even Meta's AI glasses, which sold 2 million units, are more like high - end sunglasses with AI photography and voice interaction capabilities, and users' core experience is closer to the concept of "sunglasses +".
However, the current problems will not prevent entrepreneurs and investors from looking towards a longer - term direction. The story of betting on the next iPhone in the midst of bubbles and chaos is still very appealing.
02
Betting on the Next New Species
In October 2024, Silvia participated in a communication meeting of less than a hundred people in Silicon Valley. The meeting was full of engineers from large companies such as Apple, Meta, Google, and Microsoft, as well as active investors and entrepreneurs like Kai - Fu Lee.
The theme of the meeting was quite forward - looking. At that time, the concept of intelligent agents was not yet popular, but a group of engineers had already predicted what the new hardware entrance would be, and that there would be a "hundred glasses war" in China in 2025.
This closed - door meeting strengthened her belief in the AI hardware field: AI technology has created a new interaction paradigm that urgently needs a new "body" to carry it, and traditional mobile phones and computers are not suitable carriers. This interaction paradigm will definitely bring about a major hardware revolution.
In this once - in - a - century revolution, the opportunities for hardware are destined to be in China. There are practical reasons for this. China has a mature supply chain and a large number of experienced engineers.
One number that impressed Silvia deeply is that the engineers in the United States are seriously aging. The average age in 2024 was 43, and the proportion of young engineers under 30 was only one - tenth. Most of these engineers are concentrated in large companies, and there are few engineers in startups.
"This is the best opportunity for Chinese entrepreneurs," said Silvia. Since new and unique AI hardware is still in the exploratory stage, large companies usually do not invest enough energy and resources to set up special departments for it. They focus more on major fields.
The investment boom in AI hardware soon broke out in China, and the valuations of some companies changed dramatically. At the end of 2024, an AI glasses company held a product launch, and the products sold well at the event. Some investors would probably regret it if they didn't invest in this company before the launch because two hours later, the company's valuation increased by 240% from 500 million yuan.
At the beginning of 2024, an AR glasses company that only had samples and whose specific functions were not yet launched was valued at only 200 million yuan. Its valuation increased tenfold after one year.
2025 also became a big year for AI glasses. In addition to startups in the AR glasses field such as Xreal, Rokid, and Leiniao, which have successively launched AI glasses, companies like Alibaba, Baidu, Huawei, Xiaomi, and ByteDance have also followed suit. Similar stories have extended from wearable devices to more niche fields, such as companion robots, AI toys, AI recording cards, AI earphones, AI rings, and pet AI monitors.
According to data released by 36Kr, in May 2025 alone, the funds flowing into AI hardware accounted for more than half of all investment and financing.
From the second half of 2025 to the present, the enthusiasm of capital for AI hardware has remained undiminished, but it is more obviously inclined towards star entrepreneurs at the top, such as high - level executives from large companies. In June 2025, after the cumulative sales of Robopoet's AI companion hardware exceeded 120,000 units, it quickly completed an angel round of financing worth tens of millions of yuan.
In January, Looki, an AI multi - modal wearable device startup with a luxurious team, received more than 100 million yuan in Series A financing in less than a month after selling more than 10,000 units. After Wang Teng, a former high - level executive at Xiaomi, announced his startup, he quickly received a seed round of financing in the tens of millions from Hillhouse Capital, Zilliz, and Yunjiu Capital without even having a prototype.
Not only entrepreneurs with backgrounds from large companies but also those with high educational backgrounds are very popular. Entrepreneur Feichai currently has three AI hardware projects. In December last year, the project he newly joined only had four people. The founder is a computer doctor from Tsinghua University, and another is the supply chain management director of a large company. Even when the team only had a vague direction, there were more than a hundred investors contacting them.
Entrepreneurs without a glamorous background also have tricks for quick financing. A post - 00s entrepreneur in Shenzhen told us that after they got good data feedback at the CES, dozens of investment institutions extended olive branches. The initiative has completely shifted to the entrepreneurs.
03
AI Entrepreneurship is Losing the Standard Answer
Although the valuations of companies in the AI hardware field are much higher than those in other fields, there is currently no standard answer as to what the future form of AI hardware should be.
"Most traditional hardware can solve specific problems in specific scenarios, such as lawn - mowing robots, coffee makers, and bean grinders," said Silvia. "They can be compared in terms of parameters and the ability to solve problems."
However, the problem - solving ability of current AI hardware may be diminished with the evolution of large model capabilities. Investors simply cannot conduct deductions through pure logic.
In the past decade, a quantifiable standard has been abstracted for the commercialization path of traditional hardware. After having a product prototype, it is verified for PMF (Product - Market Fit, meaning the market will quickly absorb the product without active promotion) on global crowdfunding platforms like Kickstarter and Indiegogo, then it raises funds and goes into large - scale mass production.
This logic soon became invalid. The AI recording pen Plaud, which was popular globally last year, actually did not raise a large amount of money on the crowdfunding website for its first - generation product. This did not prevent it from achieving a global sales volume of more than one million units in the next two years.
A partner of a VC firm with years of experience in the hardware field said that Plaud's story of becoming popular without financing almost "shocked the investment circle".
Plaud is not the standard answer, and the story of becoming popular overnight does not mean the end. In addition to Rabbit mentioned above, AI Pin also performed well on Kickstarter at first but has now stopped selling due to reasons such as a high false - touch rate in gesture recognition, high learning costs, and a voice interaction delay of 5 - 10 seconds.
Many investors are in a dilemma. On the one hand, they believe that the AI era has broken the traditional investment logic and there is no basis to follow. But when the company valuations soar, they unconsciously use traditional experience for evaluation, such as sales data.
Facts have proven that short - term sales do not prove that a product has the potential to survive in the long term.
Looking back at Plaud's success, it is mainly about finding a differentiated scenario in the recording niche field. Different from the recording pens of large companies, it mainly serves users who focus on conversations rather than document - driven users. This positioning determines the product's definition logic, which is to help users who mainly engage in conversations, such as doctors, lawyers, salespeople, and investors, explore the boundaries of intelligence from unstructured daily conversations.
In addition to finding the right scenario, faster iteration is also needed. "The difference in the final results delivered at the application layer may only be 20%," said Silvia. "For users, such a difference is difficult to truly translate into continuous growth. What really makes a difference is whether the founding team can respond quickly and continuously iterate in a highly uncertain technological cycle."
Li Yunzhou, the CEO of Xuanyuan Technology, is a veteran in the robot industry. He first worked on fighting robots and then switched to companion robots. Different from most entrepreneurs, Li Yunzhou decided from the very beginning that Xuanyuan Technology would not rely on external suppliers. It maintains self - control in software development, hardware development, and even front - end development, and pays more attention to engineering capabilities.
This also ensures that Xuanyuan Technology's product iteration speed is faster than that of other companies. Li Yunzhou once said in an interview, "Because our technical architecture is complete and our content library is becoming more and more abundant. The content can be completed in as fast as one day, and hardware R & D can basically be completed within 1 - 3 months."
For most AI hardware entrepreneurs, the risks are still direct and cruel. Due to the uncertainty of technology, product R & D is ahead of its time. Entrepreneurs are anxious about how to develop a product suitable for the market in the next year based on the expectation of future technology. Feichai said that you also have to pray that no other entrepreneurs will launch the same product before yours.
An entrepreneur described the process from the product demo to the crowdfunding prototype and then to the mass - production prototype as "a nerve - wracking journey".
Pan Zhidong, the head of the Global AI Hardware Business Group at Dreame, told us that even if entrepreneurs find the right scenario, they still need to constantly balance the adaptation between hardware and software. "Take Dreame's AI smart ring as an example. Integrating the vibration motor into a tiny volume and making the vibration frequency match the user's perception requires countless rounds of debugging and optimization."
There are still unicorns emerging in niche fields in the market. In addition to Plaud, there is also Oura, a Finnish company that represents AI rings. It sold three million smart rings last year, and its valuation has soared to 11 billion US dollars.
The significance of these unicorns is that although no one can be completely sure that the direction they are betting on is the correct answer for the next - generation product, entrepreneurs still have great and certain opportunities. Before becoming a world - class company, entrepreneurs need to take advantage of China's supply chain advantages to get their products up and running and then try their best not to fall behind.
This article is from the WeChat public account "Baijing Lab", author: Liu Jia, editor: Ba Chi. Republished by 36Kr with permission.