America’s $500 Billion Stargate “Fails,” China’s DeepSeek Strikes Back and Shatters the Computing Power Myth? [Master Wang Zijing’s Lecture 1.1]

```

Content of This Issue

Hello to all Wallstreetcn users, I am Wang Zijin, Chief Analyst for the Computer Industry at Soochow Securities. Welcome to my masterclass "The China-US AI Investment Paradigm." I am honored to be sharing my views in this class and hope you find it rewarding. In this course, we’ll first discuss the currently asymmetric strategic deployments between China and the US, then explore future investment opportunities based on China's four major comparative advantages, and finally talk about the research framework for the computer industry. China and the United States have become the only two major players in the global artificial intelligence industry. Globally, the increasing R&D threshold is accelerating the concentration of innovation power in China and the US. In 2024, over 90% of new models worldwide come from these two countries. Other nations can only pick a side, becoming enablers and part of the US or China AI ecosystems. In Chapter 1, let’s first look at the cards that China and the US have played in the AI game. At the beginning of this year, the US launched a major initiative — the “Stargate” project. It aims to establish an absolute leadership position in the AI era at the infrastructure level through unprecedented investment. Launched directly by top levels of the US government, it is considered as important as the Manhattan Project or the Apollo Moon Landing, positioning itself as a national strategic project. The initial plan is to invest $500 billion over the next four years, far more than most historical tech projects, aiming to build 20 super data centers in Texas. The purpose of “Stargate” is very clear: to reinforce the US's advantage in computing power and cement its hegemony in the AI era. Trump also stated directly that this massive investment is to ensure the US leads the world in AI. We can see that such an enormous plan hopes to limit the rapid development of Chinese AI by creating a computing power gap. The conclusion is: the US is a staunch believer in computing power, convinced that it is the core resource. Currently, global capital markets are also stuck at this stage, with the whole world investing around computing power. Is the US’s “computing power absolutism” right? No one knows — I don't either. But precisely because it’s an unknown, one must participate. Tech giants cannot afford the risk of being left behind in this round of the AI revolution, and are forced to join this super-scale arms race. According to the initial plan, “Stargate” Phase I should establish 10 data centers, deploying 16,000 GB200 chips. But in fact, there has been nearly no progress. The current target has been reduced: completing just one small data center by the end of 2025 will count as a success. It is noteworthy that although “Stargate” is a national-level project for the US, there is no government funding at all; it is entirely privately funded, inevitably leading to irreconcilable conflicts of interest. This is also why progress is far behind expectations. There are three specific problems: First, disagreements among partners, especially between OpenAI and SoftBank, leading to slow progress; Second, a funding shortage — the current total raised is only 10% of the target, and the rest remains highly difficult; Third, energy constraints; the planned data centers require enormous power, while the US's aging grid is hard to support. The Wall Street Journal also pointed out that the project essentially made no substantial progress after six months, its target was greatly reduced, and even basic issues such as site selection and energy supplies haven’t been resolved. Moreover, the US launched the “AI Action Plan” on July 23, which in some ways continues Cold War thinking and is filled with moves to suppress competitors. It attempts to use the power of the state and large-scale investment to maintain a US-centered global tech order and hegemony. The US’s push behind “Stargate” is actually quite complex, including both arrogance and anxiety. On the surface, it tries to embolden itself with massive investment and grand narratives, but in reality the technology route for AI is far from settled or clear. This bet is more like a political gamble. It reflects US strategic anxiety in tech competition and a lack of confidence in maintaining its dominance. Facing the US’s move, China made a swift countermove at the start of the year with “DeepSeek”. Its emergence directly broke the US's attempt at a computing power blockade. In many of the world's authoritative benchmark tests, DeepSeek matches OpenAI’s o1 in math programming and reasoning, but its training and API call costs are 95% lower than o1’s. The core innovation in lowering costs lies in DeepSeek’s use of new mechanisms, such as MLA (Multi-head Latent Attention Mechanism) and MoE architecture. For example, the traditional multi-head attention mechanism is like a person reading and having to think through the relationship between each word. For example, for “I am going to eat an apple,” the model needs to compute “I and am”, “I and going”, “I and eat”, “I and apple”, “eat and apple,” etc. For long texts, the computations rise exponentially. MLA, on the other hand, reads the prompt as a whole, divides it into segments, extracts local information and then combines it, more like the human reading process, significantly reducing computing consumption. MoE architecture works on a similar principle. A traditional large model is like a generalist expert; no matter the question, it must use its entire knowledge base—even for something as simple as “what is 1+1?”, it requires a lot of computing. MoE is like a library of experts — calling the relevant expert for the question, thus requiring fewer resources and naturally saving computing power. The latest update from DeepSeek: last month, it just released V3.1. It was originally going to launch R2, but due to adaptation issues with domestic computing power and China-US competition, the launch was delayed. They may wait for a more favorable time. That concludes this section of the course. Feel free to leave comments and discussions. See you in the next session. New course online! Click the image or here to join the learning. Risk Warning: Masterclass selects experienced professionals from third-party compliant institutions to teach theory courses on this platform. The content does not constitute buy/sell or investment advice for any specific product. The views expressed in platform classes are for learning and reference only, do not represent Wallstreetcn's opinions, and do not address users’ special investment objectives, financial status, or needs. The market has volatility and uncertainties. The platform is not responsible for any losses you may incur based on the views or information contained in the course. Investing carries risks — make decisions prudently. ```