The AI wave sweeps the globe, and DeepSeek rises to prominence, causing market turbulence.

Recently, a remarkable event occurred in the field of AI. A Chinese AI large model named DeepSeek has surpassed ChatGPT in download numbers on the US App Store for the first time, reaching the top spot. This phenomenon has not only attracted widespread attention from the global technology, investment, and media sectors, but has also caused a brief panic in the US capital markets.

As a result, the stock prices of several tech giants have fallen to varying degrees. Nvidia has dropped by 5.3%, ARM by 5.5%, Broadcom by 4.9%, and TSMC by 4.5%. Other companies such as Micron, AMD, and Intel have also experienced corresponding declines. Nasdaq 100 futures fell by -400 points, likely marking the largest single-day drop since December 18. According to incomplete statistics, the U.S. stock market may have lost over $1 trillion in market value during Monday's trading, equivalent to one-third of the total market value of the cryptocurrency market.

The cryptocurrency market has also not been spared from this wave of decline. Bitcoin's price fell below $100,500, with a 24-hour decline of 4.48%. Ethereum dropped below $3,200, with a 24-hour decline of 3.83%. Many investors are confused by this sudden market volatility, and some believe it may be related to a decrease in expectations for Federal Reserve interest rate cuts or other macro factors.

The rise of DeepSeek has prompted a rethinking of AI development models. Unlike companies such as OpenAI and Meta, DeepSeek has not developed relying on substantial capital and a large amount of hardware resources. In contrast, OpenAI was founded 10 years ago, has 4,500 employees, and has raised $6.6 billion in funding. A certain social media company even spent $60 billion to develop an AI data center comparable in size to Manhattan. DeepSeek, on the other hand, was established less than 2 years ago, has only 200 employees, development costs of less than $10 million, and has not purchased a large number of high-end GPUs.

This contrast has led industry professionals to wonder: how can traditional tech giants compete with DeepSeek? The success of DeepSeek not only reflects cost advantages in terms of capital and technology but also challenges people's inherent perceptions of AI development.

The Vice President of Products at a well-known technology company commented on social media that the story of DeepSeek embodies a typical disruptive innovation. Existing enterprises are optimizing current processes, while disruptors are rethinking fundamental approaches. DeepSeek proposes a new idea: what if we do this smarter, rather than just investing more hardware, what difference would it make?

Currently, the cost of training top AI models is extremely high. Some leading AI companies spend over $100 million just on computing, requiring large data centers equipped with thousands of $40,000 GPUs. However, DeepSeek has proposed a surprising solution: to accomplish this task with $5 million. Even more astonishingly, they not only came up with this idea but actually realized it. Their model is comparable to or even surpasses the industry-leading AI systems on many tasks.

The success of DeepSeek stems from their rethinking of everything from the ground up. Traditional AI models use 32-bit floating-point numbers to represent each digit, while DeepSeek attempts to use 8-bit floating-point numbers, finding that the accuracy is still sufficient. This change reduced the required memory by 75%. As a result, training costs decreased from $100 million to $5 million, the number of GPUs needed dropped from 100,000 to 2,000, and API costs were reduced by 95%. More importantly, their model can run on regular gaming GPUs without the need for specialized data center hardware.

DeepSeek's success challenges multiple traditional notions in the AI field, including the belief that China can only produce closed-source code, Silicon Valley's absolute dominance in AI, and the idea that developing top-tier AI models requires massive investment. Even if these views are not completely overturned, they have been seriously shaken.

A well-known American private equity investment firm commented on DeepSeek in its briefing, pointing out that this represents a victory for open source over closed source. The contributions of the open source community can quickly translate into the prosperity of the entire ecosystem. At the same time, they also believe that while the development path of traditional AI companies may seem straightforward and crude, it does not exclude the possibility of a new qualitative change occurring after reaching a certain scale. From the perspective of the 70-year development history of AI, computing power remains crucial, and this may still apply in the future.

The emergence of DeepSeek has brought open-source models to a level comparable to closed-source models, and even superior in terms of efficiency. This reduces the necessity for enterprises to purchase commercial AI APIs, providing greater development opportunities for downstream applications. In the next year or two, we are expected to witness a richer array of inference chip products and a more prosperous ecosystem of large language model applications.

Despite DeepSeek demonstrating high efficiency, experts believe that the demand for computing power will not decrease. This aligns with the Jevons Paradox in economics, which states that an increase in technological efficiency can lead to an increase in overall resource consumption. Just like during the transition from the bulky mobile phones to the widespread adoption of Nokia phones, it was the reduction in costs that enabled this proliferation, which in turn led to an increase in total market consumption.

This significant breakthrough in AI technology will undoubtedly have a profound impact on the global technological landscape and provide new ideas for the future development of AI. As technology continues to evolve, we look forward to seeing more innovations and breakthroughs that drive the AI field towards greater efficiency and accessibility.