Liang Wenfeng has indeed not disappointed, as his recent moves have made waves in the AI industry. Nvidia CEO Jensen Huang can no longer sit still! Just a few days ago, during an interview, Huang showed visible anxiety when asked about Deepseek, stating, “If Deepseek launches on Huawei’s platform first, it would be catastrophic for the U.S.” Just days after his comments, Liang Wenfeng dropped a bombshell—Deepseek version 4 was officially released, and it fully embraces Huawei!

Why is this update of Deepseek attracting global attention? What does the partnership with Huawei signify, and why does it cause unease among global AI giants?

Understanding the Current State of US-China AI Competition

To clarify this, we must first understand the current state of AI competition between the US and China. Both sides have their strengths; in China, we have Wenxin Yiyan, Tongyi Qianwen, Doubao, and Deepseek, while the US boasts OpenAI’s GPT-4, Google, and Meta.

On the surface, it seems like a level playing field, but the consensus among industry players is clear: whether Chinese or American, if you want to develop AI and large models, you cannot bypass Nvidia. In fact, Nvidia is not just a company; it is the pathway to AI intelligence.

For example, training a large model is akin to constructing a skyscraper. The algorithms are the blueprints, while Nvidia’s GPUs are the steel and concrete. Without these materials, no matter how good the design, it remains theoretical. Nvidia is the only company that can mass-produce these essential components. Others either lack the technology or the production capacity.

The reason the US has maintained its lead in AI over the years is not just due to superior algorithms, but primarily because of Nvidia’s chips. They not only utilize their own chips but also restrict exports to China, banning the sale of advanced GPUs like the A100 and H100, effectively putting Chinese enterprises in a “no rice to cook” situation. This is akin to a race where one runner is barefoot while the other wears shoes—can this be considered fair competition?

In this context, many Chinese AI companies have resorted to secretly purchasing Nvidia cards, some even paying exorbitant prices through third-party channels. The real dynamics of US-China AI competition are not just about large model contests but also about monopolization and anti-monopolization of computing power, with the US’s “chokehold” against China’s quest for breakthroughs.

Deepseek’s Impact on Nvidia’s Monopoly

The emergence of Deepseek, especially with the launch of Deepseek 4, directly challenges Nvidia’s monopoly and provides a groundbreaking solution for China’s AI large models.

On April 24, Deepseek’s official website announced the preview of the updated V4 version. The most shocking aspect for the market was not the model’s system parameters or response speed, but rather a small line on the website: “Due to limited high-end computing power, the current service throughput is very limited. It is expected that the Ascend 950 super nodes will be mass-produced in the second half of the year, significantly reducing prices.” In layman’s terms, this update signifies that Deepseek has completely moved away from Nvidia and is now compatible with Huawei chips, sending a strong signal that domestic large models are collectively breaking free from Nvidia’s grasp.

Liang Wenfeng’s brilliance lies in his ability to achieve four significant objectives with this new version: first, rewriting the global computing power landscape; second, reshaping the future of AI agents; third, initiating a reverse efficiency revolution; and finally, establishing China’s own AI rules. Let’s break down these points.

1. Rewriting the Global Computing Power Landscape

For over a decade, the global AI industry has operated under an unspoken rule: no matter how strong your model algorithm or technology, you ultimately have to rely on Nvidia’s GPUs. In the realm of large models, computing power is life, and that lifeline is firmly in Nvidia’s hands. Simply put, whoever has more Nvidia cards holds the power and advantage in AI competition.

Nvidia acts as a “toll booth”; all companies wanting to engage in AI must pass through it and pay the toll. If Nvidia decides to raise prices or restrict access, companies are left in the lurch. However, Liang Wenfeng chose to go against this trend, actively bypassing Nvidia and focusing on adapting to Huawei’s Ascend chips.

2. Reshaping the Future of AI Agents

Currently, most people use AI at a basic level, such as chatting, searching for information, or writing drafts. In essence, AI is viewed as an advanced chat tool or intelligent search engine. However, the release of Deepseek 4 transforms AI from merely a conversational agent into a versatile employee capable of independently completing entire projects.

How is this achieved? The key lies in expanding the context window of the large model to 1 million tokens, with Liang Wenfeng clearly stating that this will become the standard for all official Deepseek services. This means you can hand over an entire project plan, a complex workflow, or even a complete system requirement to Deepseek, which can fully understand and execute your needs step by step, continuously validating and optimizing its outputs until they meet your expectations.

3. Initiating a Reverse Efficiency Revolution

Globally, there is a shortage of computing power. Previously, training a high-end large model required substantial resources, often costing millions or even billions, making it accessible only to large corporations. In contrast, Deepseek 4 has sparked a “reverse efficiency revolution” by making 1 million tokens the standard for all official services without requiring more CPU power or higher costs. Instead, it achieves a significant enhancement in capability while drastically reducing computing power consumption.

According to Deepseek’s official data, the computing power usage rate for 1 million tokens is only 27%, with a cache usage rate of just 10%. This marks the first occurrence in AI development history where increased capability coincides with reduced costs, breaking the old rule that enhancing capability necessitates higher expenses.

4. Establishing China’s Own AI Rules

Why is Liang Wenfeng remarkable? Why does the release of Deepseek 4 excite us? It’s not just because it breaks Nvidia’s computing power monopoly or enhances AI capabilities while lowering costs. More importantly, Liang is tackling the most challenging issues while adhering to two core principles essential for breaking the US’s dominance in AI discourse.

The first principle is to insist on open-source large models, sharing core technologies with everyone. In the global AI industry, there is a common belief that the top-tier large models must be closed-source. This is because these models require substantial investment and resources, and opening them up would expose a company’s core competitive advantage. However, Liang Wenfeng defies this norm by making Deepseek’s core technology open-source, allowing anyone to use the model code for free and develop their own AI products.

The second principle is to adhere to domestic production, no longer relying on foreign technologies and equipment. Many domestic AI companies claim to pursue domestic solutions but secretly purchase Nvidia cards, believing that domestic technology is not mature enough. Liang, however, has steadfastly pursued a domestic path, willing to spend extra time to adapt to Huawei’s Ascend and Cambricon chips.

The release of Deepseek V4 is not just a technical breakthrough but a strategic one. It demonstrates that China’s AI industry can produce world-class large models without relying on American technology and signals that China is building a fully autonomous and controllable AI industry chain. The pressure is now on Jensen Huang to consider how to respond to the rise of China’s AI industry—whether to maintain a monopoly or seek collaboration. This choice will undoubtedly determine Nvidia’s fate over the next decade.