Since the beginning of last year, the six major investment themes I identified—semiconductor chips, computing power and algorithms, humanoid robots, commercial aerospace, solid-state batteries, and biomedicine—have been validated sequentially based on their order of benefiting from AI. Today, I will provide a detailed breakdown of the second major theme: computing power and algorithms.
What is computing power? It is the water, electricity, and coal of the AI era. Regardless of which large model or application is running on top, the underlying layer requires computing power for support. AI is underpinned by computing power, and computing power is underpinned by electrical power. In the traditional era, opening a factory required access to water, electricity, and coal; in the AI technology era, computing power is considered the most essential of necessities—without it, nothing can be accomplished. The operation of AI software, and the training and inference of all large models, cannot proceed for a moment without computing power. This determines its unique position within the AI industry chain.
I have previously discussed an industrial pattern: in any major technological explosion, the infrastructure layer is often the first to develop. The infrastructure of the AI era is computing power. Breaking down computing power infrastructure reveals several layers: the foundational layer consists of chips, including various computing chips like CPUs, GPUs, ASICs, and FPGAs; above that are servers; further up are data centers; there is also network interconnection, such as optical module switches; and finally, scheduling platforms and algorithm frameworks. This industry chain has two characteristics: first, high technological barriers, and second, strong and stable demand. Once a data center is built, its operational cycle is very long, unlike some industries where demand fluctuates significantly.
From a geographical distribution perspective, China is vigorously promoting the "East Data, West Computing" strategy. The western regions have energy advantages, with photovoltaic power generation costs potentially as low as 0.13 yuan per kilowatt-hour; the eastern regions have massive demand advantages. Connecting them through a computing power network is the essence of the East Data, West Computing project, which holds very significant meaning. Through computing power exports, it is even possible to indirectly achieve electricity exports. China generates approximately nine trillion kilowatt-hours of electricity annually, and often there is a surplus. Traditional electricity exports require building ultra-high-voltage transmission lines, which are costly and face challenges in overseas construction. With computing power, electricity can be indirectly exported through computing power exports. For example, a researcher in Silicon Valley needing to compute complex data could send the data via trans-Pacific submarine fiber optic cables to a large model in China for processing. This computation would utilize large data centers in western China, such as in Guizhou, Inner Mongolia, or Qinghai, using local electricity costing only 0.13 yuan per kilowatt-hour. After processing, the results are sent back to the Silicon Valley researcher, who pays based on the amount of computing power and tokens used. This leverages China's advantage of abundant electricity supply—a major success of the new energy strategy over the past decade—and becomes a highly competitive aspect for China in the AI era.
From a policy background perspective, national policies regarding computing power infrastructure are very intensive. The "15th Five-Year Plan" explicitly proposes building a multi-level computing power infrastructure and a nationally integrated computing power network. When I was invited to participate in a program on CCTV's financial channel to discuss this, it was mentioned that this year, seven trillion yuan will be invested in constructing six major networks, including water networks, new-type power grids, computing power networks, communication networks, urban underground pipeline networks, and logistics networks, with investment in computing power networks included. The construction of computing power networks is receiving continuous policy and financial support, with even news broadcasts reporting on it. The nation places such high importance on computing power infrastructure because, similar to highways and broadband networks in the past, it represents the information superhighway, the foundational infrastructure of the AI technology era. Once the infrastructure is laid, applications can flourish on top of it.
An interesting public data point: as of March this year, China's daily average image processing volume has grown by over a thousand times compared to two years ago. This indicates that demand for computing power is experiencing exponential growth. Technological routes are not about one replacing another; rather, different solutions are used for different scenarios: GPUs for cloud-based training, and ASICs being more cost-effective for edge-side inference. Demand is stratified, and so is technology.
Reviewing the communication infrastructure cycle: from 3G to 4G to 5G, each round of infrastructure deployment has spurred an explosion of applications. 3G gave birth to the mobile internet, 4G spurred short videos and live streaming, and 5G is now fostering IoT and AI applications. This review helps understand a pattern: infrastructure paves the way first, followed by an application explosion. At what stage is computing power infrastructure currently? At what stage are downstream applications like AI large models and humanoid robots? Downstream applications are currently on the eve of an explosion.
The first to explode are AI large models. ChatGPT was the earliest large model launched. The United States, in an effort to outpace China, initially prevented our participation in testing. A few years ago, when I attended the Berkshire Hathaway annual meeting in the US, I first visited Silicon Valley to tour tech giants like NVIDIA, Google, and Apple. At that time, like many others, I felt anxious: ChatGPT was upgrading every six months, while our large models had not yet been launched. Would we truly fall behind the US in the AI era?
Subsequently, the "DeepSeek moment" arrived, directly breaking through the relevant US technological barriers. Following that, more large models emerged, such as Doubao, Qianwan, Yuanbao, Kimi, and others. A significant advantage of Chinese large models is their low cost. The usage price of DeepSeek is over 90% lower than that of ChatGPT. Currently, Chinese large models account for 61% of global usage, surpassing American large models. Even many laboratories in Silicon Valley are using Chinese large models. Ultimately, it comes down to the performance-to-cost ratio, or value for money. High value for money is our tremendous advantage. This is why some major technological breakthroughs are now referred to as "DeepSeek moments." Liang Wenfeng's investment of tens of billions in R&D for DeepSeek has proceeded to a new round of financing, which is a very meaningful case of industrial breakthrough. I proposed two years ago that China and the US would present a pattern of catching up and advancing side by side in the AI technology era. Sino-US competition and cooperation is the main theme. Other countries largely lack the capability and conditions to participate in this AI technology competition; they can only follow the lead of China and the US in certain areas. This is the advantageous position we occupy in the AI technology era. We must cherish this position, leverage our advantage of abundant electricity supply, and utilize our strengths of a complete manufacturing industry chain and low manufacturing costs in areas like computing power, large models, and future applications such as humanoid robots to achieve leadership, find a second growth curve for China's economy, and truly realize a new round of rapid development.
Comments