The competition among large models is accelerating, but the challenges for AI startups are becoming increasingly specific: how to transform technical capabilities into orders? And once the first order is secured, how can one continue to win customers?

As the first large-model innovation ecosystem community in the country, since its establishment three years ago, ModSpeed Space has gathered more than 300 model enterprises. From initially exploring on their own to delivering products, and from tackling challenges individually to collaborating with upstream and downstream partners, an increasing number of enterprises have successfully completed the "last mile" from technology to market at ModSpeed Space – achieving revenue growth, customer repurchases, and deepening of use cases. The market is now verifying the true value of AI technology, and the ecological value of ModSpeed is also being tested by the market.

Recently, Yicai's "Unknown Territory" in collaboration with Mosa Speed Space's "Mosa Speed Open Mic" launched a special program titled "The Theory of Commercial Evolution," which reviewed the commercial achievements of outstanding AI enterprises and restored their most authentic growth paths.

The program has invited “Moso Alumni” enterprises that have achieved impressive revenue generation, as well as ecological partners with rich industry insights. Through a format of “diverse perspectives + in-depth discussions + case analyses,” it delves into these real choices, turning points, and strategies to help newcomers avoid detours and to promote the Moso community’s transition from “physical aggregation” to “chemical connections.” From model development to the first order, from repeat purchases to ecological networking, this special episode of “The Commercialization Evolution Theory” focuses solely on orders, without discussing any other topics.

The following is an excerpt from the conversation between "The Unknown Realm" and three guests. (For the full content, please click on the video to watch.)

Conversants

• Zhou Qi: Vice President of Qian Shi Technology
• Wen Deliang: Founder of Red Bear AI
• Fanyang: Capital O Partner
• Model Speed: Model Speed Employee No. 001
• The Unknown Realm: Xin Zi, the person in charge of Yicai's "The Unknown Realm"

Three guests: from infrastructure, AI applications, to investment

The Unknown Realm: Today, our main focus is on the commercialization of this model. We’re not discussing how powerful the technology is, but whether it can generate profits, and how the first order was achieved, as well as how to maintain the continuity of commercialization for subsequent orders. This year, for ModSpeed Space, which has been in operation for three years now, there is a very important issue that needs to be addressed. That is, within our ecosystem, which specific links can help our entrepreneurs and scientists form effective connections with capital providers to facilitate successful transactions? Let me briefly introduce our three guests: Zhou Qi, the Vice President of QianShi Technology; Mr. Wen Deliang, the founder of Red Bear AI; and Mr. Fan Yang, a partner of Capital O. At the beginning of the program, I will briefly explain what each of them does.

Zhou Qi: Qianshi Technology mainly focuses on developing the next generation of artificial intelligence infrastructure. They have a self-developed super-node system based on the domestic GPU. They offer two implementation options to the market: cloud-based Cloud AI and privatized Local AI. Their main task is to make good use of the domestic GPU, to run it efficiently on numerous models, and to provide excellent Token services.

Wen Deliang: Red Bear AI focuses on a brain-like memory engine driven by AI memory, and through its agent native architecture, it provides enterprises with personalized services and delivery of subscription products. The standardized products include artificial intelligence customer service, marketing, education, chat, BI, AIGC.

Fanyang: I am a partner of Frontier Technology Venture Capital Fund Capital O, focusing mainly on the intersection of artificial intelligence, biotechnology, and life sciences, which includes areas such as scientific factories, AI, for, science, as well as basic biological models.

The models are becoming increasingly powerful, bringing improvements in efficiency, but also presenting practical difficulties (Pros/Cons).

The Unknown Realm: We often say that model performance is getting stronger, but models can have illusions, just as people can. Don't assume that better performance will make commercialization easier. Over the past six months, as models have become more powerful, what improvements have they brought to everyone's business? And in what areas have they created greater difficulties?

Zhou Qi: Over the past six months, I have personally felt a significant improvement. The integration of chip adaptation and model architecture has been well accomplished through AI, which has led to higher R&D efficiency. Our Token factory provides cloud services, allowing us to produce more graphics and text per unit of time, resulting in greater returns for the company's commercialization efforts. Throughout the entire R&D process, our employees have been equipped with AI and agent to provide Coding assistance.

There is a tangible sense of improvement. Previously, adapting a model to different chips would take months; now, we can launch it as early as Monday after releasing it on Friday. The pace is incredibly fast, and this is all thanks to our self-developed agent system drive. This model is capable of reading its own structure, completing the development work on its own, conducting verifications, tests, performing evaluation-level scoring, and assessing performance. In essence, it represents a process of enhancing efficiency.

On the B side, we focus on building the infrastructure. As the performance of the models improves, everyone's iteration speed will get faster and faster. However, with the increasing iteration speed, it also requires those who build the infrastructure to constantly adapt to it. For us, this means that we will have even more work to do.

Wen Deliang: I might have a point that goes against common sense: although it’s better for a model to have stronger capabilities, in a corporate context, that doesn’t necessarily mean the model will be more effectively implemented. For example, recently when a certain model was upgraded from version 3.5 to 3.8, the delivery results and logic required some adjustments (represented by update), but customers didn’t appreciate it. Instead, they preferred the previous version of the model. In a corporate setting, what’s really needed is a model that can be stably implemented.

Additionally, from the practice of a AI company, the first immediate observation is that the cost of Token is becoming more expensive: with each model upgrade, the consumption of Token increases, directly affecting our costs. Therefore, it's not necessarily the case that newer model upgrades are always better; we need to be more cautious. Moreover, the speed of model iteration can have a reverse effect on team and organizational behavior—if the models become faster, better, and more accurate, people may struggle to keep up.

However, on the C-side, the stronger the model capabilities, the better the user experience. Therefore, the enhancement of model capabilities is tailored to different scenarios, paths, industries, and customers. In the manufacturing industry, for example, it follows standard processes, and model upgrades have little to do with the underlying infrastructure.

From the perspective of early investors, no matter what stage a company has reached, one must still view it as if it is in its early stages. For example, in the large industry of biotechnology, today, less than 5% has undergone digitalization and intelligent transformation using AI; the remaining 95% is the much larger portion that truly needs to be transformed with AI. This same phenomenon is also occurring in fields such as pharmaceuticals, materials, and chemical energy.

The challenging part, I think, is that today many people often find themselves in this state of both pain and joy, because so much is unknown. Take the profession of an investor, especially a venture capitalist, for example. First and foremost, one must be a person with insight and foresight; moreover, one needs to have taste. One also has to be experienced and act more like a researcher or scholar. To achieve this, one cannot simply focus on investing alone. In fact, one needs to understand a wide range of things, learn about these new tools, seek advice from younger people, and constantly update one's own understanding. Sometimes, it even means having to overturn one's own worldview.

From the first order to the Nth order: Product strength is the foundation of a business model.

The Unknown Territory: Two entrepreneurs, could you share how you secured your first client? What were the key reasons for that first deal to close? And what are the key factors for maintaining the continuity of securing subsequent deals (the Nth deal and beyond)?

Zhou Qi: Why did we decide to build the infrastructure at that time? Around the end of 2023 and the beginning of 2024, we looked at the architecture of DeepSeek MoE and noticed that the scale of the models was indeed growing larger and larger. We wondered if the MoE architecture could solve the problem of inference costs as the model scale continued to increase. Since there was no good MoE infrastructure solution in the industry, we decided to create a prototype. We also showed it to some upstream and downstream partners, who found our approach to be very advanced and innovative. After completing the first-generation delivery and achieving good results, we were able to achieve a large-scale expansion from 1 to 10, and then to 100 in 2025.

What is truly tested is organizational capability—whether one can serve such a large number of users with a relatively small scale of systems. It is feasible to create prototypes in the laboratory, but when it comes to mass production, what really needs to be verified are organizational and engineering capabilities.

Wen Deliang: When Red Bear first entered the AI arena, our goal was to find scenarios for long-term interaction with people. Therefore, we chose artificial intelligence customer service. This way, we could determine what information needed to be remembered and what didn't, what to do if something was remembered incorrectly, and how to handle it if it was remembered correctly. Our first project involved collaborating with the four major cloud service providers for PK; we provided POC for free, and I conducted result-based verifications. My approach proved more effective than theirs, and I was also cheaper. We managed to secure that project. At that time, our company only had 10 people, and those 10 people were working on just this one project.

With our first order, our reputation within the operator became useful, and we received multiple orders from the government and enterprise services department one after another. At that time, it wasn't us who went out to seek business (BD), but rather various departments within the operator helped us with the sales (BD). However, the channel customers were scattered and not well-targeted, so later we shifted to focusing on brand customers. By targeting specific industries and scenarios, we were able to quickly reach them through brand advertising and convert them into customers.

The Unknown Territory: Your first order was with many well-known brands on PK. What is the reason you were able to win so convincingly?

Wen Deliang: None of these relationships are as important as the effectiveness of your product. Your product OK is the root of everything; it is the foundation of the business model, which is essentially the first principle. In terms of cash flow, the best option is to's subscription service, where payment is made in advance followed by the provision of services. to and to's models definitely involve payment terms, and going through the financial process takes 1 to 2 months. Although the cash flow is not ideal, they allow for scaling. However, to and to generate profits and gross margins, but the gross profit margin on subscriptions is extremely low, with the highest being only around 5%. Therefore, we can only combine these two approaches: localize the subscription products and continue to cultivate small and medium-sized customers. We need to walk on two legs and not rely solely on one.

The Unknown Realm: Mr. Fan, from an investment perspective, when evaluating the quality of commercialization for a technology-driven company, which indicators would you focus on?

Fanyang: I completely agree with what President Wen said; in the end, it's essential to speak with products as the ultimate proof. Here are some examples: There was a team that developed automated laboratory systems. At first, they didn't place as much emphasis on being customer-oriented, but they started by replacing imported products with domestic alternatives and improving the integration of software and hardware. Then they moved on to enhancing their capabilities at the model layer. This small team of about ten people was able to rapidly grow its revenue within two to three years because they understood their customers' needs more deeply and focused on providing customized solutions. Once they built a good reputation, word of mouth helped them spread quickly. Their next step is to improve their modeling capabilities and create a closed loop for data production by integrating data flows. Today, the vast majority of companies are in this transitional phase.

The Unknown Realm: Therefore, the true commercialization of AI may not merely depend on what the model can do, but rather on whether users continuously perceive it to be worthwhile.

"Moving up and down stairs is like moving between upstream and downstream": For an ecosystem to be effective, we must reject noisy and ineffective social interactions.

The Unknown Realm: We often talk about the "model speed space," where the concept of "going up and down stairs" represents "upstream and downstream." This not only brings people closer physically but also facilitates collaboration in actual orders and business operations. So, let's focus on this concept of "upstream and downstream." What is the real value of this ecosystem?

Zhou Qi: From the perspective of our company, we very much need an aggregated system: for infrastructure, we rely on computing power vendors and supply chain systems; for Token factories, we depend on the needs of the community, developers, and users. I'm more interested in understanding the cutting-edge model companies within this system and their ideas for solving the next generation of models. I also hope to communicate more with agent and other companies that create end-to-end closed-loop applications. Therefore, we greatly need a large ecosystem like Mosa Speed Space to facilitate coordination, aggregation, and sharing.

The Unknown Territory: Since the launch of the model speed, which links have been effective, and which ones seem lively but actually haven't been put into practice?

Zhou Qi: For effective collaboration, for example, when working with a model company on a release, if both parties agree to keep it confidential, they can share the content with us in advance so that we can prepare accordingly. Then we can jointly carry out promotional activities and launch the product together. We have also collaborated with a well-known company from ModSpeed Space in two areas, helping them to build a basic large-scale model operating system. This kind of coordination is very effective.

Wen Deliang: One of the shareholders in our Series A round was recommended by Modus Space, which is quite direct; several of our more important KA clients were also recommended by Modus Space. So you see, there aren’t many places that offer both funding and resources like this.

However, creating an effective link between the products and business models of each company is quite challenging in itself. Quantifying this difficult task effectively poses a significant challenge for Modu Speed Space. Nevertheless, this can actually be implemented quite straightforwardly. Companies are willing to pay for the services needed, and we are also capable of providing those services. Therefore, my suggestion to Modu Speed Space is: charge for some of these services. Essentially, it's a matter of supply and demand: once you identify a need and provide a service, generating revenue will then enable you to discover that more people have the same needs.

Unknown Territory: Colleagues in the mod speed space, it's time to collect your payments!

Fanyang: Modus Space is located on the West Bank of Shanghai. When people arrive by the water, their states seem to change. This is my first visit here, and it was brought to us by a company that our fund invested in. Before coming, I thought this place would be more focused on AI and hard technology, but instead, it turned out to be a very young team with a stronger focus on creativity. The biggest gain for me was discovering some excellent talents and meeting some very interesting people. This community is indeed very attractive, especially for many of our founders today. In the end, as a space, what truly makes it sustainable is the network of people within it; Modus Space itself is what truly ensures its longevity.

The Unknown Realm: So, which ones are invalid links?

Wen Deliang: I am currently in the Model Speed Space and have not participated in any invalid links.

Unknown Territory: What are some ineffective links on the market that entrepreneurs should avoid?

Wen Deliang: For many events, it's important to understand the background in advance. You need to filter out those events whose background doesn't match your current stage of development; attending such events that seem impressive but are irrelevant to your current situation is of no use at all. For example, a roadshow organized by Modu Space might not be very useful for us at this stage, but it can be very beneficial for many startups, as it allows them to connect with 20 investors at once, thus accelerating their financing process.

The Unknown Realm: Those suggestions for replicable content in the mock speed space, what other links do you hope to see in the mock speed space in the future?

Fanyang: For the model speed space, we could organize more activities like researcher and day. This would encourage cutting-edge researchers to share their ideas and research, and listening to everyone's feedback would be very appealing.

Looking ahead to the next three years: Who will become the winner in the commercialization of AI?

The Unknown Territory: Looking at the next three years from today's perspective, what will the successful commercializers of AI look like?

Zhou Qi: This year, I define it as a year of blooming diversity: everyone is working on AI applications and agent, but at present, there is no complete and unified agent platform or standard – there is no aggregation, no classification, no norms. I hope that in the next three years, we can develop such a platform-based, categorized, and standardized agent ecosystem, defining it as a AI OS. The infrastructure needs to consider how to adapt to different forms that operate according to these standards.

Wen Deliang: If I were to start a business again from today, I might still pursue quantitative methods of investment, because AI is much better than humans in making investment decisions. AI security could become a new field of opportunity, as any application must be built on a secure foundation. Moreover, there are still significant issues unresolved in China's foundational research. By leveraging AI, we can quickly make up for the shortcomings compared to Europe and America, and in the future, a new wave of entrepreneurs will emerge in this area. At the same time, with the sharp decline in Token costs, consumer applications will see a new surge. Those who can produce higher-quality content and innovative social connection methods will shape the future of products.

Fanyang: Many tracks will turn into tracks and wildernesses; it's hard to predict what will happen. All you can do is participate in and create that future. So, what I imagine is a future where science becomes increasingly mature due to AI, becoming like infrastructure, just like Taobao. With creativity, ideas, and aesthetic sense, you can engage in scientific exploration in the same way that companies today are able to develop social networks based on Taobao and REDnote. I really hope that the next generation of younger student entrepreneurs will grow up and be able to pursue what they want to do, whether or not there is financial reward. Just as people in creative industries today, they can do what they love regardless of the rewards.

The Unknown Territory: In the next 3 years when investing in some AI projects, what abilities of a person will you value more?

Fanyang: In fact, people are the best asset. On one hand, as entrepreneurs, one must possess a fast and strong learning ability; on the other hand, it becomes increasingly important to have an idealistic outlook. You need to attract talented individuals and investors, and you need to tell compelling stories. Without that idealistic side, it's difficult to tell a story that touches people, feels genuine, and seems to involve real, living individuals. However, you also need to be pragmatic; you need to bring in orders, just as you need to hunt every day to sustain the morale of your team. Therefore, it's even more like being a heroic figure.

A most genuine commercial experience

What if only the most genuine commercial experience were left to the alumni of Moshu Space?

Zhou Qi: Find a traditional industry with a large scale, a stable customer base, and stable upstream and downstream connections. If we can use AI to increase its efficiency by 10 to even 100 times, then its future value potential will be even greater.

Wen Deliang: In the future, commercialization is definitely the key. Therefore, when you choose a track to pursue, you must take into account the growth rate.

Fanyang: For the alumni of Modu Space, it is essential to fully leverage Shanghai as an international metropolis. From technology to AI, we should extend our connections to the entire business ecosystem, engage with more upstream and downstream partners, as well as companies of various sizes both domestically and internationally, discover more like-minded individuals, and achieve greater business results.

Unknown Territory: If you can't find a new continent on an old map, you can't keep circling around in the same old world.

Expectations for the Mod Speed Space Ecosystem

Miao Xiaosu: In the next three years, who do you most hope that the Miao Speed ecosystem will help you connect with first, and what issues do you hope it will solve?

Zhou Qi: More cutting-edge researchers, research teams, as well as more innovative developers of AI and agent, to understand what they are thinking and what they hope the next generation of infrastructure will look like.

Wen Deliang: In fact, we really prefer more creative young geniuses to come and communicate with us. You're also always welcome to approach me.

Fanyang: Since I am an early-stage investor, I hope to connect with young entrepreneurs at an earlier stage, or with those who have just come to Shanghai from overseas or other cities, so that they can get in touch with such a great community.

Written at the end

Through our discussions, we have seen that this industry is moving away from the era of focusing on conceptual storytelling and towards a stage that values delivery results, commercialization, and real value. It's not that the newer the model, the better; however, there are always customers willing to pay for certainty—stability is in itself a form of product strength.

"Moving up and down stairs is like moving between upstream and downstream" – the true ecological value lies in facilitating the matching of supply and demand, turning ideas into actual orders;

"Unable to find new lands on old maps" – AI The tide rolls forward, and technology is merely a tool; it is the people who are in the midst of that tide who truly determine the future direction.

As more and more AI entrepreneurs successfully complete the “last mile” from their first order to the Nth order, and as the vision of “walking up and down stairs representing upstream and downstream partners” transforms into actual orders and growth, the “commercialization” moment for AI companies is approaching.

Together with our predecessors, venture into the unknown.

(For the full conversation, please click to view the video)

When more AI assistants are integrated into local services, e-commerce, and other areas, 12% of DouBao will become an important reference point.

The AI model has entered the fast lane of diminishing marginal costs.

The explosive growth in computing power represented by AI is forcing a transformation of the energy system, with cutting-edge technologies such as nuclear fusion and solid-state batteries accelerating their development. However, transitioning these technologies from the laboratory to commercialization still requires overcoming multiple challenges, including manufacturing, talent, and economic viability.

Yi Rating: Small models establish a high cost-performance paradigm, AI commercialization accelerates

Physical AI is at a critical turning point from being a technical concept to large-scale mass production. The degree of coordination among the computing power foundation, simulation platform, communication network, terminal entity, and algorithm model will directly determine whether AI can truly overcome the "last mile" and establish a commercial closed loop in the real economy.