Zhang Hongjiang's speech: There is a large gap in AI computing power reserves between China and the United States, and Agents will change the composition of human organizations.

According to a report by Zhidx on September 11th, today, at the opening ceremony of the 2025 Inclusion·Bund Summit, Zhang Hongjiang, an investment partner at Source Code Capital and a foreign member of the US National Academy of Engineering, delivered a keynote speech. He believes that this year is a turning point for the popularization of large model applications.

Since the release of ChatGPT two years ago and the emergence of DeepSeek at the beginning of this year, there have been several obvious trends in the global large model industry: large models will impact many important applications in the future, with search being the first to be affected; large model-related applications are evolving from problem-solving tools to AI assistants and AI companions; the emergence of the DeepSeek model is driving up the demand for computing power, and the computing demand of large models will continue to scale up; the ability to build large-scale computing power systems in a single IDC is becoming increasingly important.

Zhang Hongjiang refers to the ecosystem in which models drive the development of underlying chips, the entire cloud computing industry, power, and energy industries as the computing power system ecosystem chain (AI’s Industrial Scaling Up). Currently, the construction of IDCs for overall computing power in the United States is progressing rapidly, which reminds us that we need to rapidly expand our chip and IDC construction.

Finally, he shared his thoughts on the development of Agents. When a person and a group of Agents form a company, it will lead to an unprecedented leap in efficiency. However, this also requires us to think about the impact on youth employment, social structure, and the education system.

The following is a summary of the highlights of Zhang Hongjiang's speech:

01.

DeepSeek Drives Up the Demand for Computing Power

Large Models Have Evolved from Tools to AI Companions

Zhang Hongjiang mentioned that looking back at the progress in the past six months, the most exciting development was the emergence of DeepSeek. DeepSeek broke the record for the shortest time to reach 100 million users for an application in human history, achieving this milestone within just seven days.

Another aspect is that DeepSeek combines high performance with low cost. When it was first released, the performance of the DeepSeek model was comparable to that of the best models in the world at the time, while its cost was only a fraction of others.

Does this mean that chips or computing are no longer important? Zhang Hongjiang believes that those who think so have clearly been taken advantage of. Resources follow a basic economic law: when the cost decreases, the demand will increase on a larger scale. That is to say, the emergence of the DeepSeek model will drive up the demand for computing power.

Since the release of ChatGPT over two years ago, we can measure user stickiness through the ratio of daily active users to monthly active users. When ChatGPT was first released, this ratio was 14%, and it exceeded 30% in March this year. This indicates that ChatGPT has become a tool that users often rely on to solve problems.

In July and August this year, another change occurred. The weekly active users of ChatGPT reached 700 million, and a large proportion of them started chatting and interacting with ChatGPT, treating it as a companion. This is an important sign of the change in user stickiness.

From a corporate perspective, the penetration rate of large models in companies is also increasing rapidly. After the release of inference models, the number of users of OpenAI and Anthropic has been growing rapidly. This year is a turning point for the popularization of large model applications.

The popularization of large models has first affected the traffic of search engines. A study by The Economist shows that due to the impact of AI, web search traffic has decreased by 15%, and the search traffic in the health field has dropped by over 40%.

Therefore, Zhang Hongjiang believes that this means large models will impact many important applications in the future, with search being the first to be affected.

Previously, the Scaling Law was the principle for the growth of large models. However, after the emergence of inference models, the industry has discovered another inference curve. In the future, new dimensions will be introduced, including "Learning from Experience" mentioned by Richard Sutton, the 2024 Turing Award winner and a Canadian computer scientist, which is essentially about memory and context.

From this perspective, the computing demand of large models will continue to scale up.

In the past few years, the cost of large language model inference has been continuously decreasing, and this decrease is quite significant. Zhang Hongjiang mentioned the law of LLMflation, which states that when the performance of a model improves rapidly, the usage cost will decrease rapidly.

In the industry, on the one hand, the emergence of DeepSeek can rapidly reduce costs. On the other hand, the improvement in the performance of chips and large models themselves also proves that the cost reduction will continue as large models develop.

02.

Large Models Drive the Scaling Up of the Entire Industry

Accelerate the Construction of AI Infrastructure

Currently, the large model ecosystem has driven the scaling up of the entire industry.

Initially, cloud providers only offered models as a service. Subsequently, models began to define the entire cloud platform and establish new platforms, similar to the PC, iOS, and cloud platforms in the past. This ecosystem will surely give rise to a complete industry.

Currently, the chip and cloud computing industries have been driven by large models, and they are gradually driving the development of the entire power industry and even the economy. This has also been the biggest driving force for the US economy in the past two years.

Elon Musk built the world's largest AI data center for xAI, deploying 200,000 graphics cards in a single cluster. The significance of this for data centers lies not only in achieving a certain scale of overall computing power but also in the ability to build large-scale computing power systems in a single IDC. OpenAI has also taken similar measures. Since Microsoft could not provide sufficient computing power for OpenAI in a timely manner, OpenAI launched the "Stargate" project.

Comparing the changes in computing power between China and the United States in the past five years, we can see that the construction of IDCs for overall computing power in the United States is progressing rapidly, which reminds us that we need to rapidly expand our chip and IDC construction.

Zhang Hongjiang refers to the ecosystem in which models drive the development of underlying chips, the entire cloud computing industry, power, and energy industries as the computing power system ecosystem chain (AI’s Industrial Scaling Up).

03.

Enterprises Need to Redesign Workflows for Agents

A Person + Agents Achieves a Leap in Efficiency

At the beginning of this year, Sutton won the Turing Award, which proves that reinforcement learning has become a core technology in AI.

Agent technology can only emerge when inference models gradually mature, and its growth rate is currently very fast. In the past 12 months, the thinking time of inference models has increased rapidly, doubling every six months, which also means a doubling of thinking performance.

As the performance of models and Agents improves, the relationship between humans and machines, and humans and AI is changing. AI has evolved from a pure tool to a human assistant and then to a partner. Moreover, the time for AI to become a human assistant will be short, and it will quickly become a partner, capable of independent thinking, initiative, planning, and action.

Once Agents have the capabilities of computing, thinking, planning, and acting, they will replace corporate processes. Currently, Agents are replacing the processes that companies previously designed for humans. In the future, enterprises must redesign the entire process for Agents or let Agents redefine the workflow. This is the driving force behind the change in the relationship between humans and machines.

Once there are more and more Agents in society, they will form an Agent group. Agents will operate the entire workflow, exchange information, share decisions, and conduct transactions, while humans will become providers of resources and data. This is the state of the Agent economy. In the Agent economy, each Agent is a node in the vast social neural network, similar to a neuron today.

The impact of the Agent economy on society is that currently, the most important asset of institutions is talent, but in the future, it will be computing power, models, and data. Currently, companies hire more talent to expand their business, but in the future, they will need to expand their computing power scale, have more powerful models, and richer data. When a person and a group of Agents form a company, it will lead to an unprecedented leap in efficiency.

This will also bring about social changes. For example, AI has already had an impact on youth employment.

In July this year, Mark Zuckerberg spent a large amount of money to hire AI engineers. This may not only be because Meta is lagging behind in the AI field. Zhang Hongjiang believes that it is more likely because Zuckerberg sees the future, and the combination of super individuals and Agents may be the main body of future scientific and technological research and development.

Due to these structural changes, we should start thinking about the future social structure, tax system, and education.

This article is from the WeChat official account "Zhidx" (ID: zhidxcom), written by Cheng Qian and edited by Xin Yuan. It is published by 36Kr with authorization.