Xiao Wang: From Tool to Agent, AI Ushers in the Era of Intelligent Everything | Unity Ventures

九合创投九合创投··0·3

The "Android Moment" for AI Has Arrived

Large language models have opened the true door to the intelligent era. DeepSeek has simultaneously changed the feasibility and cost of making everything intelligent, and we have arrived at an inflection point.

Xiao Wang, founder of Unity Ventures, recently spoke at the CEIBS EMBA New Knowledge Forum, starting from DeepSeek to share the underlying shifts and commercial opportunities of the new intelligent age.

Key Takeaways:

  • Only when cars, humanoid robots, home appliances, and more all possess considerable intelligence will the era of intelligent everything truly arrive — this is a massive opportunity.

  • DeepSeek's breakthrough is analogous to the "Android moment" — it provides a strong technical foundation upon which applications can be built directly.

  • The application layer will see significant explosion this year, with the potential to build moats and market share through data flywheels, yielding high gross margins.

  • Within three to five years, robots may scale into households at accessible prices. Embodied models are very likely to be achieved first in China, given the country's strong data collection capabilities, lower costs than abroad, and globally leading manufacturing capacity.

We are facing a structural transformation: the dawn of the intelligent age. Large language models have endowed machines with greater intelligence. Tasks that previously required humans can now largely be substituted by LLM-powered intelligence. Many industries are currently digitizing; the next step is the intelligent transformation of every industry. This trend is irreversible — whoever intelligenizes first gains the advantage.

So this is a deep structural shift, just as the internet opened the information age, large language models will open the new intelligent age. DeepSeek's emergence has simultaneously changed the feasibility and cost of this transition. LLM capabilities have risen while costs have fallen dramatically low — now the curtain has truly risen. Today's sharing will focus on the rise of large models, the impact of DeepSeek, and commercialization opportunities.

DeepSeek Is Only the Beginning: How to Embrace the Era of Intelligent Everything

The title of my talk is From Tool to Agent: AI Opens the Era of Intelligent Everything. Right now we treat large models as tools — asking them to write code, make PowerPoints. Eventually they will gradually become agents. A company's team structure might become one manager plus ten agent intelligent workers, with the manager overseeing these agent employees. One company we've invested in is already doing this; organizational structures will change enormously.

The concept of IoT has been around for over a decade, but those "things" still lack true intelligence. Only when cars, humanoid robots, home appliances, and more all possess considerable intelligence will the era of intelligent everything truly arrive — this is a massive opportunity.

With every wave that comes, the total market cap of companies born in that wave rises exponentially. Mobile internet companies were worth dozens of times more than the early BAT era of the internet. AI companies in the era of intelligent everything may increase total market cap by at least 10x compared to the previous era, and the number of companies will continue to grow.

Fortunately, as this great wave begins, we are doing early-stage tech investing, supporting innovators from 0 to 1, being their earliest investors.

This opportunity began with the popularization of large models brought by OpenAI, and recently DeepSeek achieved new breakthroughs. Essentially, it used incentivization to replace learning — rather than humans teaching the large model, it directly incentivizes the AI to find optimal strategies through continuous trial and error, through self-reflection. This conceptual breakthrough not only enabled rapid improvement in large model capabilities and dramatic cost reduction, but also brought new development opportunities for domestic GPUs.

The "Android Moment" That Opens the Large Model Era

DeepSeek shattered the carefully planned U.S. "compute quota" new order. The industry is now moving toward a higher cost-performance inference route, achieving AI democratization and reshaping the global AI ecosystem. Moreover, DeepSeek took the open-source route, opening its capabilities to the world.

This is a truly admirable open-source spirit, hence somewhat similar to the Android system. It was because of Android that smartphones developed so rapidly over the past decade and more. DeepSeek's breakthrough may be analogous to the "Android moment" — it provides a strong technical foundation. People can build applications directly on top of it, and the application layer will see significant explosion this year.

We have already invested in twenty to thirty AI-related companies. Some are SaaS companies transforming with large model capabilities, such as intelligent sales and customer service. Tungee, which we invested in, is a domestic leader in intelligent sales with tens of thousands of enterprise clients, and has already integrated DeepSeek to help sales teams improve efficiency.

We also have positions in AI companionship, short drama generation, recruitment, and other areas. CreativeFitting, which we invested in, is a global short drama platform that uses AI to generate short dramas without actors or filming — its short dramas launched in the United States have been well received.

AI will ultimately manifest on end devices, including cars, phones, glasses, humanoid robots, and more — these are the final carriers that will possess intelligence. So we have also invested in some robotics and intelligent hardware companies.

Disrupting Internet Logic: The Delivery Model of the Large Model Era

We did an analysis: why are internet companies so powerful? Because they have high gross margins, high concentration, and can establish monopoly positions. The internet mainly does connection — search connects people with information, e-commerce connects people with goods, ride-hailing connects people with cars, food delivery platforms connect people with restaurants. Once they reach considerable scale, connection costs become very low, and they profit through advertising with extremely high gross margins.

Commercialization of large models is easier than imagined. What do AI companies do next? Service. They do end-to-end delivery services, and their gross margin is 1 minus (token cost / service price).

Initially token costs were very high; now the cost per million tokens is less than $1, and in the future may even fall below 1 RMB. Then token cost approaches zero, and gross margin approaches 100% — such companies will certainly become large.

Moreover, they form data loops in service scenarios — users generate data, which is used to train models, making them better understand user needs and preferences, enhancing service capability, which then acquires more users. So data flywheels can form competitive moats and market share, alongside high profits.

This is the value generated by models, and it doesn't require burning money — profitability comes faster. For example, LynkSoul, which we invested in, does emotional companionship applications based on large models and AI digital humans. Founded one year ago, it already has millions of users. It can not only accompany you in gaming, but in the future can also provide emotional value and life planning management.

From "Brain in a Vat" to Freedom from Household Chores

Embodied Artificial Intelligence is a very important milestone for the next step of AI.

Current large models are somewhat like a "brain in a vat" — a brain in a box without the ability to act or observe the world. If we want robots to better understand human society, they must possess the ability to act, entering various life scenarios to collect data. Within three to five years, robots may scale into households, and not be too expensive.

Independent Variable Robotics, which we invested in, is dedicated to achieving general-purpose robots through developing embodied intelligent general large models. Previously, large models and embodied models were separate — the brain planned first, then manipulation executed. Independent Variable Robotics chose the "unified big-brain-small-brain end-to-end large model" route. Last year their robots could already complete complex tasks such as preparing beverages, pulling zippers, as well as drying, organizing, and folding clothes.

I believe embodied models are very likely to be achieved first in China, because our data collection capabilities are strong, costs are lower than abroad, and manufacturing capacity is globally leading.

To summarize, large models have entered an inflection point moment, especially with DeepSeek as the core inflection point. China will have more innovations in algorithms and engineering.

Over a year ago, when people asked me about the prospects for China's large model development, I said foreign countries couldn't restrict our technical exploration — our innovation and engineering capabilities are globally leading. A year later, DeepSeek has fully validated this judgment.

We invest at the early stage, and early investors have one particularly important trait: optimism. We remain optimistic no matter what. We bet on many companies when they were just three to five people. I have always been fully confident in the development of China's technology.

Roundtable Discussion

What is the current overall state of AI investment? What kind of exits are investors looking for in the future?

Xiao Wang: Currently, Chinese investment remains relatively concentrated in technology, especially in the primary market. And the biggest wave in technology right now is the AI breakthrough brought by large models, including hardware, software, and various services — this is the bright spot in the market.

We have maintained our own investment pace. Although there is uncertainty in capital markets and IPOs, we can judge the degree of technology development and the commercial opportunities it brings — this is deterministic. If you have relatively good understanding of how technology has developed in the past, when you see this company and this trend, you will know what kind of company will appear next and to what extent it can develop. I have been doing this kind of judgment work for over a decade, maintaining relatively high accuracy.

For example, with QingCloud, which I invested in, we were their first investor. At the time, I felt there needed to be a company specifically doing cloud computing — no one would need to buy servers, they could just buy services directly. And this centralized approach could save massive resources and electricity — this was the core logic of the investment. Later, QingCloud grew into a listed company in the fiercely competitive infrastructure market.

Regarding exits, not every company is suitable for IPO, but these companies still have value. M&A will gradually move forward. We have also exited dozens of companies through M&A, and help portfolio companies connect with acquirers. Some quality assets are relatively low-priced, and many companies' main businesses need new technology for synergy — there may be multiplier effects after M&A.

Which companies have such opportunities? I think it's new technologies applied in relatively universal industries that can cooperate with leading companies in existing industries. We are also investing in such companies.

Looking ahead, what investment areas do you think are most likely to explode?

Xiao Wang: Investment is essentially making judgments about the future, selecting targets in the present — it needs to see through time, looking at what will likely exist in five or ten years, then working backward to determine what layouts should be made now. Looking at what scenarios may appear in 5 or even 10 years, I have several judgments.

First, there will very likely be one to two humanoid robots in homes, handling housework more difficult than vacuuming — such as stir-frying, cooking, drying clothes, even walking dogs.

Second, in the future we may interact not with individuals but with AI Agents — AI customer service, AI family doctors, AI teachers. Many services may potentially be replaced.

Third, if AI model capabilities keep strengthening while compute requirements keep decreasing, everything within sight may possess a certain degree of intelligence. Manufacturing and other industries will all change. Previously, reliance was on human efficiency; through robots and AI, not only is the aging population problem addressed, but per capita GDP may rise another level. If nuclear fusion also achieves breakthrough progress, "nuclear power + chips + AI" will reconstruct social structure and development paths, and social resources will become greatly abundant.

Additionally, AI's push on biotechnology is also very significant. AI can simulate experimental environments, and biopharmaceutical R&D speed will increase greatly — possibly rare diseases and cancer cure rates will improve substantially.

From this perspective, compute, algorithms, and data are all very important, and in the future may be more critical than land, capital, and other means of production.