(Yicai) May 20 -- The number of users using artificial intelligence every day will likely reach five billion by 2030, after standing at just one million in 2020 and exceeding one billion last year, according to the chief executive officer of US semiconductor giant Advanced Micro Devices.

The AI agent economy requires extremely substantial computing power support, Lisa Su said at AMD's AI Developer Day in Shanghai yesterday. What is needed is not just a single type of computing power, but full-stack computing systems that can seamlessly work together, she noted.

During a talk with Su themed "New Paradigm of Agentic AI" at the event, Kai-Fu Lee, founder and CEO of Beijing-based 01.AI, said that the programming ability of AI has crossed a critical point, with the tech only capable of assisting in writing code, functions, and other such tasks a year ago, but now delivering a complete set of functions end-to-end. Once its coding capabilities cross the threshold, autonomous agents will truly have the potential to become a reality, he added.

The more significant change lies in the fact that the industry is gradually realizing that the capabilities of a single AI agent have an upper limit, Lee stressed. No matter how large the model parameter scale is, relying solely on the reasoning ability of a single AI agent will eventually hit a bottleneck when dealing with real-world complex problems, he added.

However, the multi-agent architecture has broken through this limit for the first time, he said, adding that AI agents responsible for planning, evaluation, research, and execution have begun to collaborate with each other and engage in debates, continuing to evolve based on each other's results.

For multi-agent collaboration to truly be feasible in reality, the system must meet several conditions: local priority, edge processing, and a response delay of less than 100 milliseconds, Lee stressed. This is precisely where a winner in a hardware competition will be determined, he pointed out.

With the widespread use of AI agents, AI has entered the "CPU + GPU" era, according to Su. The ratio of central processing units to graphics processing units in major nodes might have been 1:4 from 2022 to 2025, but it is evolving towards a 1:1 ratio, she said.

AMD will focus on accelerating momentum this year, including extending its leadership in computing technologies and data centers, Su noted.

AMD has over 4,000 engineers at its research and development centers in Beijing, Shanghai, Shenzhen, and Taipei, with its Epyc processors supporting more than 700 cloud instances of China's leading cloud service providers, Su pointed out. The company also has ecosystem alliances with over 100 software providers, startups, and universities, she added.

Su expressed confidence in the expanded use of CPUs during an earnings conference call earlier this month, saying that as AI applications grow, more CPUs will be needed to support AI inference.

Editor: Martin Kadiev