When others can't do it, and you can, that's when you become the most valuable.
Source: Zhang Peng Technology Business Observation
I received a WeChat message from the co-founder of Manus early this morning: "Brother Peng, we have a new development today. Previously, it was indeed difficult to disclose too much. Now it's officially released, and I wanted to let you know right away."
The new round of financing for Manus, with a valuation of $2 billion, has been confirmed in the industry, but not long ago, I heard about the possibility of a deal involving Meta worth nearly $4-5 billion (unconfirmed rumors). I initially thought it was a bit far-fetched, and I never expected it to happen so quickly; it was truly lightning-fast.
More than a year ago, founder Xiao Hong and his team rejected a multi-million dollar acquisition deal from a certain giant. He told me, "We hesitated, but ultimately realized that there are not many opportunities in life worth going all in for, and we didn't want to miss this chance." Now, going all in has paid off; with the Manus product, they have achieved over a hundredfold value growth in less than a year, hitting a "home run" this year!
We must admire their courageous decision to continue exploring, and congratulate them on their remarkable returns.
We should also thank them for validating the value and opportunities of AI application innovation before the end of 2025, which will greatly boost the confidence of all entrepreneurs and investors.
I previously wrote an in-depth analysis of Manus, and the subsequent developments and today's deal basically confirm that analysis, which I would like to share again. (The original article was published in May 2025)
With this, I wish AI entrepreneurs that your upcoming 2026 will be just as exciting!
In recent days, news has circulated in the community: Manus has nearly $100 million in ARR and has reached a valuation of $2 billion.
After Manus was launched, the contrast in evaluations between domestic and international perspectives was significant. In March of this year, it gained significant attention domestically, but soon faced a wave of skepticism and criticism. Since Geek Park reported on Manus early and gave it a high rating, more than one person told me, "You at Geek Park are betting your 15-year reputation on promoting a company!"
As the saying goes, "Three men can create a tiger," and this baseless talk made me quite anxious, even reflecting on whether our judgment was indeed "amateurish."
Later, I realized there was really nothing to reflect on, because when I turned my attention overseas, especially to Silicon Valley, I found that although Manus was not being discussed as fervently there, the overall sentiment among people in the AI community was positive.
In particular, during my communications with internal personnel from OpenAI, Microsoft, Google, and others in the U.S., I discovered that these giants are taking Manus very seriously. For example, Google places great importance on Manus, with engineers almost permanently stationed with the Manus team to assist in better integrating with the Gemini model. Meanwhile, Microsoft’s CEO Nadella has already had face-to-face discussions with the Manus team, expressing considerable praise and actively promoting collaboration. It is not an exaggeration to say that they are currently one of the most favored entrepreneurial teams by overseas giants.
Why can such a relatively early-stage startup from China, with a product that was once viewed as "just okay" in the domestic entrepreneurial circle, be questioned domestically but achieve rapid global recognition in the AI industry ecosystem? This is worth some deep thinking.
01 "No Model" Yet Bringing Incremental Games
The point that Manus is most often criticized for domestically is "not having its own model." But from another perspective, from the standpoint of giants like Google and Microsoft, they might actually be pleasantly surprised: "Hey, this guy has no model, yet can create such things! It gives my model another outlet for token consumption."
If the path of AI Agents ultimately can only be played by giants with their own models, then this game is indeed too narrow, merely a "stock game + zero-sum game" among a few giants.
However, the truly big companies in this world do not rely solely on a certain business or product for "self-sufficient smallholder economics" to become giants; rather, they have built an ecosystem that includes "trade and division of labor." Some of their core capabilities drive external value that is tenfold or even a hundredfold greater than their own business, meaning that when they distribute value to others, they can also gain greater returns.
Thus, the giants who have invested enormous resources in developing models also hope to see a thriving application layer ecosystem. If someone can use their models to solve more practical problems, create richer application scenarios, and consume more tokens, it is naturally beneficial for everyone.
Products like Manus connect with all the top models, and with each task execution, they consume tokens from the underlying large models and the computing power of cloud vendors. If Manus's ARR is indeed close to $100 million, then consider how much of that value is actually going to the giants behind the models? It makes sense that they are being taken seriously.
This realization made me suddenly think that sometimes entrepreneurs shouldn't hesitate too much about the ultimate question of "Will the giants also do this?"
Thinking is a product of the environment. Over the past 15 years, Geek Park has accompanied entrepreneurs through too many instances of "giant PTSD," and I feel that my thinking may have also become accustomed to the "winner takes all" mentality of the internet era, with a significant inertia of "stock thinking."
In fact, it seems that overseas giants, including OpenAI, which internally clearly considers products like Manus to be worthy of attention as "peers," generally have an open mindset, actively helping in the early stages and observing closely in the long term. After all, Manus is still in its early stages, and utilizing APIs and cloud services is not a bad thing.
Given the giants' control over the model field, which is indeed a core focus for them, and their overall approach to productization being "cautiously bold," they would be pleased to see a vibrant new seedling grow in the ecosystem. If this seedling can grow into a large tree that expands the ecosystem, it would benefit the entire ecosystem.
If this tree is truly recognized in the future for its value of standing out, then whether these giants observing closely will offer sufficiently favorable conditions to turn this "increment" into their "new stock" will depend on the growth speed of this new seedling during its development phase, the barriers it explores, and the giants' judgment of its future value ceiling.
This incremental thinking is not only worth learning for entrepreneurs but also merits reflection for the large domestic companies.
In fact, the Manus team was already seen by domestic giants when they were working on Monica.im. They likely know about their exploration of the general Agent concept, and there have even been clear acquisition offers from giants before. However, based on the internal feedback from domestic giants to the Geek Park team, it seems they either want to recruit early and do it themselves or seek maximum control, wanting to take a large share of the value both now and in the future.
This may need to change. In the AI era, large companies need to focus on what large companies should do, rather than immediately engaging in stock games with entrepreneurs. It is necessary to rethink the relationship with entrepreneurs with a more open "incremental thinking."
02 "Quantum Tunneling" and "Barrier Changes"
If Manus receiving praise from industry giants can be understood as a "business logic," then why can it, as a still immature cutting-edge product, quickly achieve such a high ARR, and why do overseas investors give it such high valuation recognition?
Regardless of whether Manus's product is sufficiently refined today or whether others can replicate it, it must be acknowledged that it has indeed gained significant "first-mover advantages." Even when other relatively better similar products emerge later, as long as they are not "generational" improvements, it will be difficult to replicate such excess returns.
I believe the "quantum tunneling effect" in quantum physics and the resulting "barrier changes" can serve as a good analogy for the answer to this question.
First, let's talk about "quantum tunneling." Imagine a small ball trying to climb over a high mountain. According to classical physics, if the ball's kinetic energy is insufficient, it cannot get over. However, in the quantum world, particles exhibit "wave-particle duality"—they are both a physical entity and a probability wave. Therefore, even with insufficient energy, there is still a certain probability that it can "tunnel" through, appearing on the other side of the mountain as if by magic. This seemingly counterintuitive phenomenon can explain the breakthrough paths of many startups: they have limited resources and seem unlikely to shake up the industry landscape, but certain innovations can allow them to "penetrate" barriers and achieve market breakthroughs.
Moreover, once a particle successfully tunnels through, the entire competitive landscape undergoes structural changes—this is what quantum physics refers to as "barrier changes." First, the "height" of the barrier decreases—the first mover validates market demand and technical feasibility, making it easier for later entrants to replicate similar products. For example, after OpenAI launched ChatGPT, the barriers to entry for large model startups significantly decreased, and various companies quickly followed suit. However, at the same time, the "width" of the barrier increases—the advantages in users, capital, and ecosystem accumulated by the first mover make it difficult for later entrants to truly disrupt their position unless they achieve "generational innovation." Tesla is also a typical case: after it first broke into the electric vehicle market, although new players have emerged more quickly, it still remains difficult to shake its industry dominance.
Manus's path is similar. When the general AI Agent was not yet mature, it did not wait for the giants to act but instead "tunneled" through technical barriers with its engineering capabilities, gaining first-mover advantages.
So how can entrepreneurs, who inherently have limited energy, achieve such a "quantum tunneling effect"? In fact, quantum physics also provides an explanation, which is similar to the concept of "probability clouds"—because particles exhibit "wave-particle duality," it may seem that a single particle lacks the energy to penetrate (a small team lacks the energy to break through the giants), but sometimes it can miraculously bypass barriers in the form of "waves" (by creating technologies or products that the giants did not anticipate or produce). The smaller the particle's mass, combined with higher initial energy and narrower energy barriers faced, the greater the probability of penetration.
Isn't this exactly the "efficient + sharp + focused" innovative breakthrough approach that Geek Park has witnessed countless times among entrepreneurial teams over the years?
Returning to Manus, I believe the achievements of Manus stem from their courage to tackle a goal that others are still hesitating over. Their resolute choice of objectives, full commitment to engineering investment, combined with the effective practical experience accumulated from their work on Monica, has brought a high level of "initial energy" to the entrepreneurial team.
I specifically checked the articles and discussions in the Geek Park community, and it turns out that discussions about Agents in the industry had already begun as early as last spring. Throughout 2024, the progress in Coding and Computer Use was quite evident, and vertical field Agents even started to have ARR. However, the vast majority of people were waiting for the giants to create a general Agent, as they believed that without their own models and world-class engineering capabilities, it was impossible to achieve this.
But in reality, the "energy barrier" at that time was not as high as imagined. The rapid development of model capabilities, while not yet able to directly realize the capabilities of a general Agent, was already close to the "concept machine" of a general Agent by early 2025, with just a significant amount of engineering issues left to solve. No entrepreneur could penetrate with "particles" (model capabilities), but whoever first used "waves" (engineering reinforcement) to get through would be on the verge of "quantum tunneling."
It can be said that teams like Manus and Genspark were among the first to "overreach" by choosing this goal that most people were waiting for the "giants" to achieve, and then they began to work tirelessly, "handcrafting" and "replacing magic with engineering," providing clear results at various stages. This naturally led to strong positive feedback from the market.
As I write this, I suddenly recall a line from the movie "Batman v Superman," where Batman tells Superman: "You are not brave; men are brave."
What he means is that Superman's "bravery" is a byproduct of his near-divine superpowers, while the courage of mortals to face difficulties is the greater bravery.
In front of global AI leaders, and even the Chinese internet giants with their super abilities, Deepseek is undoubtedly the "Batman"—a superhero of ordinary people (also fitting for Liang Wenfeng, who, like Bruce Wayne, has certain resources supporting his beliefs). Meanwhile, teams like Manus and Genspark can be considered truly impressive "real civilian heroes." They certainly deserve applause.
In the past, Chinese entrepreneurial teams rarely achieved such high-level treatment in the core areas of global technology and business ecosystems at such an early stage. This should show Chinese entrepreneurs another possibility. This is even a meaningful contribution to the community of Chinese entrepreneurs. For example, Silicon Valley's increasing interest and confidence in the product and engineering capabilities of Chinese entrepreneurs in the AI field is, in effect, paving a new path for future entrants.
Therefore, what Chinese entrepreneurs should see here is not just a "tactical-level" approach of carving a boat to seek a sword, but also recognize that this is indeed an opportunity to leverage the transformative changes of the era, aiming for a higher "energy leap" goal.
This requires some "mortal courage" and the ability to think from a broader worldview within the global technology ecosystem.
03 What Should Manus's Next Goals Be?
Next, we need to talk about the challenges, as Manus still faces significant hurdles. I believe the key moving forward is to continuously shape blockbuster scenarios that allow users to see tangible results and actively participate, all built on their general AI Agent foundation.
This reminds me of how Douyin (TikTok) became popular. It sparked continuous imitation of certain trending dances or challenges, drawing in new users wave after wave. Then, new gameplay continuously emerged on the platform, attracting even more participation. From the initial active engagement to the later systematic emergence.
Today's technology is still advancing and needs to continue improving, which means that user conversion cannot be completed in a single "perfect moment"; it must be a gradual process. Therefore, what is needed next is the ability to engage users, bringing in more users in batches.
In fact, before Manus, when Xiao Hong and I discussed their AI browser plugin Monica at the AGI Playground conference in 2023, it felt more like a "feature phone"—adding features simply meant increasing pipelines. Each new trend might signify the development logic of a new product and a completely new project.
But today, Manus has a general foundation, resembling a "smartphone"—it can create a plethora of great applications more efficiently on this general base. It doesn't require hiring a large number of engineers or building countless project lines; instead, it should observe in which scenarios users find it useful, to do "subtraction," optimizing these validated paths to deliver better and more certain results, enhancing operational efficiency.
In this way, the first-mover advantage combined with user feedback can create a positive cycle—when a scenario becomes popular and breaks out, it drives the growth of the entire platform. Moreover, it can continuously break out and grow.
If we observe high-user-volume "general AI products," such as ChatGPT or the user Q&A demands on DeepSeek in China, we can see that the vast majority of demands are not that deep or complex. The same goes for the Agent field; in fact, not many users have such high-frequency complex tasks in mind. It is likely that 80% of users' 80% most commonly used tasks can be converged to some extent. If you can deliver well on both of these overlapping 80% tasks, you become the "general Agent" in their minds.
The practical shocking result of this demand convergence model is that covering just half of the core scenarios can trigger a sense of "generality."
So, although Manus's ARR revenue has reached $100 million, I believe we should not view today's revenue significance through the lens of traditional ARR. Greater revenue certainly represents more user engagement, and the revenue generated from repeated token consumption of similar tasks is more meaningful, as it effectively locks in users' "workflows" and "lifestyles," and this retention is key.
Today, we should not engage in a "self-sufficient smallholder economy." For instance, at this stage, you need to have the ability to enhance meaningful token consumption by users, rather than constantly thinking about how to reduce token consumption to increase your own profits; only then do you play a positive role in the AI industry relative to other forces.
The level of AI technology and the cost of technology will inevitably rise and fall over time. The result is that optimizing costs today has little significance for the future. Meanwhile, users' mindsets, as well as the user prompt habits, personalized data, workflows, and lifestyles brought about by the era of large models, are resources that are easily obtained with first-mover advantages but will become increasingly expensive to acquire in the long run.
Therefore, for general AI products, as long as there are resources, the only correct strategy is to form user engagement through continuous innovative results and better delivery based on the aforementioned "demand convergence model." Only users are the barriers and continuously appreciating assets.
Thus, while the $75 million that Manus has secured may seem substantial, it is certainly not enough. The less you have, the more effectively you need to spend it. The least advisable way to spend might be to directly buy large amounts of traffic or invest in advertising, paying "entrepreneurial taxes" to the giants. Effective investment should be to "cost no object" in delivering experiences that exceed user expectations, continuously achieving those "amazing goals."
Ultimately, the simple logic of business is: when others can't do it, and you can, that's when you become the most valuable, and you can more easily acquire users at a lower cost. After all, entrepreneurs always need to find opportunities at the intersection of the technology diffusion curve and the market demand curve.
04 The Discussion on "Shelling" Can Be Turned the Page
Finally, let's talk about the issue of "shelling."
A couple of days ago, I was chatting with Li Zhifei, the founder of DuerOS, and he mentioned a very good point. As a system, a computer, besides the CPU, also requires a series of management systems such as process management, memory management, and peripheral management to ensure its effective operation. However, if we view large models as a new CPU today, there are still many unresolved issues with these surrounding systems, which is actually a significant obstacle at present.
This sparked our thinking: if we treat AI Agents as a revolution in personal computing, meaning that the purpose of personal computing is no longer to provide a workstation for various tools in a digital world, but rather to input demands and directly output final results for delivery. Then relying solely on large models (analogous to CPUs) is not enough; there are still many related management systems that need to be built. And there are numerous engineering problems that need to be addressed seriously. For example, better virtual machines, longer contexts, a large number of MCPs, and even smart contracts… a series of engineering issues represent huge demands.
After more than two years of industry frenzy, we should have a clear understanding that the progress of large models themselves remains the biggest driving force. However, as always, after every technological breakthrough, humanity discovers that "improving engineering precision" still holds immense value for technological development.
Manus can completely disregard the term "shelling." You could say that every Apple phone is a shell for a CPU, but that shell can also be a complex and exquisite product engineering. This is certainly meaningful and will undoubtedly experience a flourishing of diverse approaches, with some companies emerging that hold sufficient value.
In this worldview, opportunities also belong to more entrepreneurs.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
