VC investor: "We are looking for companies of this kind."
In May this year, the new move of Zhu Xiaohu, the managing partner of GSR Ventures, caught people's attention. At that time, after publicly stating that he was not optimistic about humanoid robots and exiting a number of humanoid robot projects in batches, he turned around and invested in an underwater robot project - Shihang Intelligence (hereinafter referred to as Shihang).
At that time, the outside world had already interpreted Zhu Xiaohu's actions. Zhu Xiaohu exited the humanoid robot projects because of the unclear business prospects, and publicly declared that humanoid robots "are only used for research or display by customers and have not created continuous value." The company he invested in later belongs to "pragmatic investment", with in - depth scenario capabilities and the ability to achieve rapid commercialization.
Founded in 2023, Shihang is the earliest domestic enterprise to research and the first to commercialize underwater cleaning robots. Focusing on underwater general - purpose robots, its "Orca" series of products has built the marine productivity from 0 to 10,000 meters underwater, and is currently mainly used in the ship - cleaning field.
"Is there a person pushing your product underwater?" At first, the Shihang team also received a lot of doubts, and the founding team could only explain patiently, "It's completely not the case." Recently, Cao Ying, the co - founder and COO of Shihang, told me, "The company started commercialization this year, and the revenue has increased tenfold. It is expected to turn a profit next year."
Cao Ying revealed that the R & D costs for the underwater scenario are actually higher than those for humanoid robots in other scenarios. In the entire robot industry, it is rare for a company to achieve profitability in just over a year after a large amount of R & D investment.
Many people may not know that Shihang's rapid commercialization is quite related to its cutting - edge business model - "pay - for - results with AI".
Specifically, every time it cleans a ship, the shipowner pays a cleaning fee. The reason why shipowners are willing to pay is that after the large ships are cleaned, it can not only replace the untestable manual cleaning but also save fuel costs. "The ships cleaned by us can save 100,000 yuan in fuel costs per day."
This has two major benefits for Shihang: First, the products can more accurately meet customer needs. Second, it can break through the ceiling of revenue scale. This subverts the industry's previous revenue models of selling hardware, SaaS, or integrated solutions.
Actually, not only Shihang, but many well - known enterprises at home and abroad have started the era of the "pay - for - results with AI" business model, and this has become the consensus of global AI giants and capital magnates.
At the closed - door meeting of the Third Sequoia Capital AI Summit that ended in May this year, after 6 hours of in - depth discussions among 150 top AI founders from around the world, including Pat Grady, a partner at Sequoia Capital, Sam Altman, the CEO of OpenAI, and Jeff Dean, the chief scientist at Google, many cutting - edge views and consensuses emerged. Among them, Outcome - based Pricing and Outcome - as - a - Service became one of the most core views, which means: "In the next round of AI, what is sold is not tools, but benefits." Pat Grady called this a "trillion - dollar opportunity".
In China, there are also institutions that resonate with this concept. "We are looking for companies with this kind of business model and will also promote some potential enterprises to transform in this direction," said Wang Xiangyun, the managing partner of Shengjing Jiacheng Venture Capital. This charging model has enabled some invested enterprises to achieve tenfold growth in revenue and profit.
At a time when the assertion of the "AI bubble" is constantly warning the venture capital community, a clearer commercialization path has emerged. It's like a ray of light in the chaos.
Why was it proposed?
In China, Shengjing is one of the early institutions to propose this business model.
In March 2025, Shengjing proposed AI RaaS (Result - as - a - Service), an extreme result - oriented model that dares to use results as the basis for pricing, charging, or making a profit. It vividly compares it to being an "AI owner or AI client", advocating that only by providing end - to - end services and deeply embedding in the physical world can real value be created.
Actually, Shihang's business model was facilitated by Shengjing. At a business event, Chen Xiaobo, the founder of Shihang, met Peng Zhiqiang, the chairman of Shengjing Network and the founding partner of Shengjing Jiacheng Venture Capital. They shared the same ideas and values. At the beginning, Shihang followed Peng Zhiqiang's advice and adopted the pay - for - results business model, and Shengjing invested in its angel round.
The team conducted financial calculations on whether to provide equipment or results. "Finally, we found that providing results is the optimal solution because the number of times can be infinitely increased, and the final result will also be infinite. For customers, which has a higher decision - making cost, spending 100,000 yuan to clean a ship once or spending millions to buy a machine? Obviously, it's the latter," recalled Cao Ying.
This happened in 2023, a year when infinite opportunities coexisted with a cold capital environment. At that time, Chen Xiaobo, a young scientist in his early 30s, had explored the underwater robot world for 18 years. After the business model was verified, Zhu Xiaohu made an investment at the beginning of this year. It is reported that Zhu Xiaohu's resources and capabilities helped Shihang open up the situation, and Zhu Xiaohu got on board the most promising underwater robot project in China.
Why did Shengjing respond so quickly?
Wang Xiangyun told me that it was the bottleneck of the SaaS model that promoted this change. "During the previous SaaS investment cycle, we found that the SaaS industry in the United States was booming, and enterprises had high valuations. However, the domestic SaaS industry faced huge pressure and challenges in terms of revenue, valuation, and exit. In short, it 'didn't make sense financially'."
So the team concluded that simply imitating the US model may not work in China, and the software logic should be considered in the context of the industrial Internet and longer business chains. Following this logic, Shengjing also successfully invested in some enterprises.
In 2022, after the launch of ChatGPT, a new round of AI - driven industrial cycle began. Shengjing observed that with the iteration of AI basic models, the path of some "simple - shell" applications would become narrower and narrower. Therefore, the long - term and sustainable capital value of simple - shell projects was challenged. Then, what kind of capabilities need to be enhanced to avoid being eliminated?
So, Shengjing's conclusion was: having strong scenario capabilities and charging based on results. Through investment cases, it was found that the result - oriented charging model can lead to tenfold growth in revenue and profit. Besides Shihang, Shengjing's investment cases in this regard also include Lingyun Zhikuang, an AI mineral exploration company.
For this reason, the Shengjing Research Institute has continuously published articles and launched the "30 Global Cases of AI RaaS" series, systematically analyzing benchmark enterprises to provide reference paradigms for local entrepreneurs.
This idea has been recognized by many investors. Yun Ke, a senior investor, also told me that the SaaS model is very likely to end in the AI era, and there are two reasons behind this:
First, the payment logic has changed. The essence of SaaS is to let users pay for tools, but tools are only means and cannot solve the final result. However, AI can directly replace labor, so this is a much larger market than SaaS.
Second, the best AI models are all closed - source and controlled by giants. The new - generation SaaS with shells can hardly build a moat.
Currently, Yun Ke is also looking for enterprises with such commercial potential.
What are the standards?
Perhaps many people are confused about the measurement standards for paying for AI agents based on results.
Taking Shihang as an example, Cao Ying admitted that the company needs to gradually prove its value through actual operation effects, such as saving fuel for customers, and educate the market. This process is challenging.
Cao Ying summarized that the successful implementation of the "pay - for - results" business model ultimately depends on the following three major capabilities:
First, having a far - leading hardware and system integration ability; second, having the "fuel" for continuous iteration and building barriers. Through cleaning "thousands of ships", it has accumulated operation data under different ship types (bulk carriers, container ships, etc.), different sea areas (the North Sea, the East China Sea, the South China Sea), different water qualities, and different seasons. This is something that any company that only sells equipment or does not directly face end - customers cannot obtain; third, the service results can be quantified and verified.
Currently, Shihang has become the company with the most underwater scenario data in China. "It's difficult for latecomers to catch up." However, Cao Ying also introduced that during the process of implementing the business model, the company basically "competes in the domestic market and makes money overseas."
On the one hand, domestic shipowners are very sensitive to prices and tend to choose the lowest - cost solution, even if the service quality varies. In overseas markets such as Japan and Singapore, the labor cost is higher, and they recognize the value of technology more and are willing to pay a higher price. The overseas customer unit price can be more than "three times" that of the domestic market. Therefore, the company is currently actively exploring overseas markets.
Internationally, the RaaS model is being practiced in multiple fields. Companies such as Clay, Sierra, and 11X have developed from the traditional software subscription model to the task - based charging or a mixed pricing model based on tasks and results.
Sierra, an AI customer - service unicorn founded by Bret Taylor, the chairman of the OpenAI board, is particularly radical. It is not a simple customer - service system but a closed - loop sales agent platform that helps brands complete the entire sales process from the first inquiry to placing an order.
It not only contacts customers but also takes responsibility for the conversion results, truly following the path of "you give me a budget, and I'll bring you a certain amount of GMV."
There is a more intuitive detail: when the AI agent independently solves the needs of incoming calls or online consultations, Sierra will charge a fee; if it finally has to transfer to a human operator, it will be free this time.
"We really like this model, and I also think it will become the standard business model for agents," said Bret Taylor. Sierra was founded in 2023 and has now become a unicorn valued at tens of billions of dollars.
Ramp takes this idea to the extreme. It is a fintech company founded in New York, USA, in 2019. Starting with a corporate credit card, it aims to use technological means to subvert the traditional corporate expenditure management method and help enterprises save time and money. It doesn't sell a financial system to enterprises but directly promises to save a certain amount of fees. Its AI can automatically identify redundant subscriptions, negotiate price cuts, and predict risks, turning "the benefits of using this tool" into KPIs.
To summarize a more general measurement indicator for "result - oriented products", at Sequoia's closed - door summit, Sequoia gave three major judgment criteria: whether it can complete a full task process; whether it has persistence during task execution; whether it can deliver measurable business value.
In the view of the Shengjing team, the popularization of this model is a gradual process. According to the degree of intelligence, it can be divided into four levels from L1 to L4:
L1 represents short - process businesses that are mainly online digital applications, highly repetitive, with clear processes and a high degree of standardization. For example, it will be first applied in industries such as law and customer service. L2 often involves long - operation processes that require complex reasoning, tool invocation, and integration, and often requires the use of hardware tools for implementation. L3 focuses more on helping customers achieve a closed - loop sales of products and services and finally achieve revenue sharing based on results, which means a qualitative improvement in the external connection ability of AI services; L4 is upgraded to an "AI owner", which not only has AI service capabilities but also becomes the main or partial "owner" of core assets or company value with the help of AI advantages.
Wang Xiangyun believes that in the long run, businesses with a higher intelligence level need high - quality collaboration between AI and highly professional people, which is a healthier AI industrialization model. In the future, as technology matures, the proportion of AI will gradually increase; in terms of the overall promotion speed, the more market - oriented the supply chain and value chain are, the faster the promotion of AI RaaS will be.
The AI blue ocean and the AI bubble can coexist
AI agents emerged when large - scale AI models were being applied, and they are also in the midst of the extensive discussion about the "AI bubble".
In the second half of 2025, the AI capital market had its worst performance since April this year, with the Nasdaq index dropping by more than 3% in a single week, which further triggered a wider discussion about the AI bubble.
The main reason is the sharp contrast between the huge R & D investment and commercialization revenue of leading enterprises such as OpenAI: in 2024, its R & D investment exceeded 15 billion US dollars, but its commercialization revenue was less than 3 billion US dollars.
Even if, as Altman expected, the annual revenue will exceed 20 billion US dollars by the end of 2025 and reach hundreds of billions of US dollars by 2030, it will still be difficult to generate positive cash flow.
Meanwhile, the cost problem of AI has become increasingly prominent. The expectation that the cost of large - scale models will decrease by ten times a year has not saved the paid - subscription models of many AI enterprises.
A study by the Massachusetts Institute of Technology has also triggered extensive discussions. The study pointed out that although enterprises have invested 30 - 40 billion US dollars in generative AI, 95% of organizations have not achieved any business returns.
However, Yun Ke emphasized that there is a logic that should not be confused: "The general large - scale models themselves and making money by using the capabilities of large - scale models are two different things. The former requires huge investment and will surely see no returns in the short term. Their goal is not rapid profit - making but to seize the strategic high - ground of the next - generation technological revolution. The former is a game that only a few national - level teams and technology giants can afford to play. The latter is where ordinary entrepreneurs should focus."
He believes that the over - valued area this year is humanoid robots. "This field is still developing rapidly, and breakthroughs may be seen in the next few years. However, if we expect rapid large - scale implementation, I think it is a big over - estimation. The development of technology must go through several stages and cannot be skipped quickly. As we can see, Tesla also lowered its production expectation for Optimus this year."
So, how should we view the current bubble?
I think we can perhaps quote a clear but complex answer given by Bret Taylor, the co - founder and CEO of Sierra, in an interview:
AI will reshape the economy and create huge value; at the same time, the bubble does exist, and some people will lose a lot of money. Both can be true at the same time.
He believes that the current AI bubble is very similar to the Internet bubble: indeed, there were many failed cases during the Internet bubble period. However, if we look at it over a 30 - year period, we can see the emergence of giants such as Amazon and Google, and the cloud business of Microsoft has become an important pillar of its market value. We can also directly see the profound impact of the Internet on the global GDP - many of the "optimistic" views in 1999 were actually in the right direction. Even a company like Webvan (online fresh food delivery) reappeared in the form of healthy businesses such as Instacart and DoorDash after the popularization of smartphones and the maturity of the Internet scale. Many ideas were not bad; they just came too early.
In addition, the exploration of paying for AI agents based on results will gradually become a new trend in AI revenue - generation.
Generally speaking, in 2025, AI agents have bid farewell to the concept - verification stage in multiple industries and entered the value - realization stage. Amid the discussion about the AI bubble, the emergence of the RaaS model has pointed out a practical development path for the industry - AI technology must return to the essence of business and create measurable value for customers. This path may not be as eye - catching as unlimited financing and piling up computing power to pursue SOTA, but it may be more sustainable and solid.