Zhou Hongyi: I see an opportunity for Chinese chips to overtake others.

The Zhou Hongyi who shuttled between live - streaming rooms and auto shows has disappeared.

In the past year, Zhou Hongyi seems to have deliberately avoided the so - called "traffic code", which forms a sharp contrast with his previous actions such as "forming a CP" with Lei Jun at the auto show, holding an auto show downstairs of the company, and constantly hosting live - streams and posting videos around artificial intelligence.

Recently, Zhou Hongyi appeared in front of the public again at an exchange meeting. He seemed a bit "awkward" and self - mocked as being like a robot. "I had an eye surgery and replaced the lens with an artificial one. So maybe I look more like a robot now in terms of my eyesight."

After replacing the lens with an artificial one and no longer wearing glasses, Zhou Hongyi said it was for the convenience of wearing "AI glasses".

"This thing is quite difficult to make," Zhou Hongyi said. "Currently, we haven't found a particularly good scenario for it."

AI glasses are one of the connections between Zhou Hongyi, 360 Group, and artificial intelligence. When asked about related businesses, Zhou Hongyi euphemistically responded to the progress. He believes that for smart glasses to be attractive to wear, they need to have powerful functions. However, the more powerful the functions, the more power they consume, and the heavier the glasses become. This is a contradiction that cannot be reconciled in the short term.

Compared with AI hardware, Zhou Hongyi pays more attention to the evolution of the underlying models, agents, and computing power. During the "Two Sessions" in 2026, Zhou Hongyi brought three proposals focusing on the field of artificial intelligence, which revolve around three major directions: "dual - line empowerment" of agent technology and talent, optimization of the layout of inference computing power, and support for the wide application of security agents.

Zhou Hongyi said that Anthropic has solved many problems in security that couldn't be solved before through AI programming and AI vulnerability detection. So he put forward a suggestion to pay attention to AI (security) agents.

The security - related suggestions are also closely related to his summary of the industry in 2025. He said, "When agents start to work, it also means they may make mistakes. At the same time, the output of AI - generated code has increased sharply, but AI will inherit human errors, and the vulnerability problem has been magnified exponentially."

Despite the existence of problems and risks, Zhou Hongyi emphasizes that AGI is being steadily realized.

Regarding the more fundamental computing power base, Zhou Hongyi believes that computing power should be divided into training computing power and inference computing power. "I think there is still some room for the development scale of training computing power, but the development scale of inference computing power has unlimited potential."

In his prediction of the core trends in 2026, Zhou Hongyi also emphasized the value of inference computing power. "Inference computing power has become a strategic high - ground."

Zhou Hongyi believes that in 2026, large - scale models and agents will evolve simultaneously on two lines. As the core carrier for the implementation of AI capabilities, agents will become digital employees of enterprises and personal assistants for individuals, shifting from "chatting" to "working" and helping you complete complex tasks 24/7. This evolution will also reshape interpersonal relationships. Humans will change from doing things themselves in the past to setting goals, making plans, and checking results for agents.

This change in interpersonal relationships easily triggers discussions about "AI replacing humans", but Zhou Hongyi gives a negative answer. He said, "It places higher requirements on people, and a large number of compound - type talents who understand both business and how to manage agents are needed."

The following is the transcript of Zhou Hongyi's exchange, with deletions and adjustments without changing the original meaning:

01

China's Chips to Overtake in a Curve

Zhou Hongyi: Hello, everyone! I'd like to suggest that the questions be more specific. I'm not good at answering macro - level questions.

I had an eye surgery. It wasn't a double - eyelid surgery. I replaced the lens with an artificial one. So, maybe I look more like a robot now in terms of my eyesight.

Question: What proposals or suggestions related to AI technology do you have this year? What changes can they bring to people's lives?

Zhou Hongyi: I'm focusing on three directions.

First, it's the AI - empowered security I mentioned earlier. Take Anthropic as an example. Through AI programming and AI vulnerability detection, many problems in security that couldn't be solved before have been resolved. So I put forward a suggestion to pay attention to AI agents. 360 has developed dozens of types and tens of thousands of AI security agents. These agents can use AI to discover software vulnerabilities, run automatically, and resist hacker agents from other countries. They can also be used to solve AI - related security problems. Currently, two million small and medium - sized enterprises in China are using our AI agents for free to provide real - time protection and operation for their enterprise security.

Second, I'm more concerned about how to implement AI in China. I've developed a "six - force" model. China has done very well in energy and power. With the support of electricity and energy, computing power can be generated through chips. I suggest that computing power should be divided into training computing power and inference computing power. I think there is still some room for the development scale of training computing power, but the development scale of inference computing power has unlimited potential. Therefore, by distinguishing the two types of computing power, I hope that local areas can focus more on inference computing power in the development of computing power.

Third, I'm concerned about how enterprises and individuals can quickly use AI. OpenClaw has given us an inspiration that we should simplify things. So, I proposed the concept of building an open platform for agents, hiding the infrastructure of agents behind, so that ordinary enterprises and individuals can easily build their own agents and learn their own skills.

Fourth, this year, we're going to carry out agent training across the country.

Question: You mentioned that inference computing power in China may be a future necessity, but currently, more emphasis is placed on training computing power. How should the deployment be adjusted?

Zhou Hongyi: It was reasonable to focus on training computing power before because two years ago, large - scale models hadn't been well - trained and hadn't reached the passing score, so applications were out of the question. At that time, everyone was talking about training, and training computing power became very important during the "hundred - model war".

Since last year, large - scale models have passed the passing score and can perform inference. The capabilities of the basic base models are sufficient, and there's no need to train models repeatedly. Now, for applications in various industries, even industry - specific large - scale models don't need to be trained because the amount of data for training is too small and doesn't make much sense. Instead, people should focus on building excellent professional agents on the base large - scale models.

In this case, agents consume a lot of computing power. How many tokens can you use in a chat? But for creating a short drama, it can easily consume millions of tokens. That is to say, when agents really help enterprises work, the energy and computing power consumption is extremely astonishing.

So, Jensen Huang said that once we enter the application era, the growth of computing power will increase by a factor of 100 million.

Why did OpenAI later invest in AMD? It hopes to use AMD's chips for inference rather than training. Because finally, people found that the requirements for inference chips and training chips are different.

Jensen Huang recently spent $20 billion to buy a company (Groq) that makes inference chips, a company specializing in dedicated chips, which proves that NVIDIA is also taking this path. Even Broadcom has received many orders. Amazon, Microsoft, Google, and Facebook are all making inference chips because everyone has to reduce the cost of inference.

For China's chip industry, it may still be a long way to reach the capabilities of B200 in the short term, and it will take some time to disrupt the CUDA ecosystem. However, it's relatively easier to develop dedicated chips with a solidified Transformer inference algorithm that has lower requirements than NVIDIA's GPUs. Moreover, mass - producing these chips can bring about several revolutionary changes.

Many enterprises need to make private deployments. Their large - scale models and agents should be deployed within their enterprises, and their computing power should be localized. With inexpensive inference chips, enterprise deployments will be very convenient.

In the future, many smart hardware devices need to be deployed at the terminal or on the edge. For example, there are so many smart - city cameras. If you transmit all the data from these cameras to a central location, the bandwidth will be unbearable, and the storage will be very costly. If the cost of each computing power inference chip drops, say a 500T or 100T chip only costs a few dozen yuan, each camera can be upgraded to have simple AI processing capabilities.

Chinese people can make anything very cheap. So, developing inference chips is of great strategic significance. With local chips and local computing power support for robots, complex tasks can be handed over to cloud computing power.

Question: Can you briefly summarize the overall development of the artificial intelligence industry in 2025 with a few keywords and predict several core trends in 2026?

To summarize the development of AI in 2025, I have three keywords:

First, AGI is being steadily realized. AI is developing rapidly. When combined with specific businesses and developed into industry experts, you'll find that it's much smarter than you.

Second, it's the year of agents. Various vertical agents have begun to emerge. Especially the popularity of OpenClaw before the Spring Festival has made ordinary people more concretely aware that agents are their own digital assistants that can work on the computer.

Third, there are new challenges in AI security risks. When agents start to work, it also means they may make mistakes. At the same time, the output of AI - generated code has increased sharply, but AI will inherit human errors, and the vulnerability problem has been magnified exponentially.

These new challenges began to emerge in 2025. Looking forward to 2026, I think the following are the core trends:

First, inference computing power has become a strategic high - ground. As the training computing power stabilizes and agents are combined with industrial scenarios, the demand for inference computing power has begun to explode. This is an opportunity for China's chip industry to overtake in a curve and is also the key to making agents affordable for small and medium - sized enterprises and individuals.

Second, there is a dual - line evolution of large - scale models and agents. While the models are evolving, agents are also evolving. As the core carrier for the implementation of AI capabilities, agents will become digital employees of enterprises and personal assistants for individuals, shifting from "chatting" to "working" and helping you complete complex tasks 24/7.

Third, the security confrontation has been upgraded, and AI needs to be used to fight against AI. On the one hand, hackers have started to use agents to launch attacks, which are continuous 24/7 and can even write attack tools on the fly. It's impossible to defend against them manually. On the other hand, AI itself can also cause trouble - hallucinations, being manipulated, deleting files by mistake, and in the future, hacked robots may cause physical harm. So in 2026, security agents must be used to fight against hacker agents, and AI should be used to solve the new problems brought about by AI.

Fourth, the human - machine relationship has been reshaped. AI enters the workplace as a digital employee. Humans change from doing things themselves in the past to setting goals, making plans, and checking results for agents. This actually places higher requirements on people, and a large number of compound - type talents who understand both business and how to manage agents are needed.

02

The Model War and the "Little Lobster" (OpenClaw)

Question: How do you think about and judge the model war during the Spring Festival? Why is the iteration of large - scale models so intense?

Zhou Hongyi: The core reason is the success of the open - source strategy.

Domestic open - source models such as DeepSeek and Qianwen have entered the international first - tier. Since the capabilities of the base models have reached the standard, there's no need to train models repeatedly. Naturally, the competition has shifted from "comparing parameters" to "comparing implementation" and from "training models" to "using models".

Intense competition is a good thing as it brings out the unique advantages of Chinese AI.

We're using "scene density" to offset the "chip gap" - you have high - end chips, and I have rich application scenarios.

Ultimately, the intense iteration of large - scale models is because everyone has realized that in the second half of the AI era, what matters is not who has a more knowledgeable model, but who can truly use and implement AI.

Question: The "Little Lobster" (OpenClaw) was very popular before the Spring Festival. Please also talk about the positive impact it has brought to the industry, as well as the problems and deficiencies it shows. Everyone says that tokens are expensive, and just saying "Hello" may cost a few cents. Is this important? Is it a key issue?

Zhou Hongyi: I think there may be some misunderstandings about the high cost of tokens because the backend of large - scale models can be flexibly configured.

Currently, the computing power cost in China has dropped significantly, and the cost of daily chat conversations is actually very low. What really consumes tokens are complex tasks, such as generating videos, creating short dramas, or writing novels.

The key reason why OpenClaw has attracted attention is that it has introduced a new concept: "raising" an agent on the computer to make it your personal assistant.

In the past, when developing agents, due to security concerns, we were hesitant to open up computer permissions easily - worried that if the agent made a mistake, it might delete files randomly. However, OpenClaw dared to take this step, like the first person to eat a crab. By giving the agent the ability to operate the computer, it can call tools and build personal data memory, thus continuously evolving.

What's more worthy of attention is that it has connected with Anthropic's skill packs.

In the past, when using an agent, we had to search in the store first, which was a bit troublesome. But if it becomes a skill pack that can be called by simply checking a box, the experience will be much smoother. Currently, there are thousands of skills online, and the "Little Lobster" you "raise" can unlock new abilities every day - this shows that agents can continuously grow.

Of course, we also need to view it rationally.

Many people may mistakenly think that its data is very secure. Although OpenClaw stores data locally, as long as it calls the cloud - based large - scale model, your prompts and some data will still be uploaded for processing, so it's not a truly local closed - loop system.

Question: Whether it's the agents from OpenClaw being able to post on forums or the videos generated by Seedance that are hard to distinguish from real human - generated content, it seems that it's becoming increasingly difficult to tell the difference between AIGC and human - generated content. What's your attitude towards this increasingly difficult - to - distinguish situation?

Zhou Hongyi: I'm more optimistic about this issue than you. First, it shows that AGI is being steadily realized.

AGI doesn't necessarily mean the emergence of an Einstein - like figure. It means that AI is now smarter than an average person. With the support of computing power and the generation of tens of billions of agents, just imagine how powerful AI will be when working with us in society?

OpenClaw is still an agent system, but it gives you the illusion that the agent is running on your computer, which is not the case.

The agents and instructions still run on the large - scale model. They just assemble the instructions on the computer. However, they don't just go to a forum to chat. Just as the Internet has connected humans and generated a lot of co - created content, when agents are connected, they may also inspire more agents, which will lead to emergence.

Regarding Seedance, you can clearly see that the gunfight scenes are a complete imitation and re - creation of "John Wick", which shows that it has studied all four seasons of "John Wick" many times. This verifies what I said earlier: the capabilities of agents have reached the level of AGI, which will bring great help to the film and entertainment industries.

As for people using this to create fake videos, humans have many technical methods to solve this problem, such as adding secret fingerprints. When users receive such a video, it will be marked, and people will know that it's generated by AI.

Question: You mentioned that as the capabilities of agents improve, they may replace APPs as the new core entrance for services. Can you talk about the impact on the existing Internet business models and how existing enterprises and entrepreneurs can seize this opportunity?

Zhou Hongyi: Many things will be redone.

Current websites and APPs are designed for humans because humans are the operators. In fact, an agent also has