On January 27, the emerging Chinese AI large model DeepSeek surpassed ChatGPT in downloads for the first time, topping the US APPStore rankings. This has sparked widespread attention and reporting in the global tech, investment, and media circles.
This event not only suggests the possibility of a shift in the future tech development landscape between China and the US, but also transmitted a brief sense of panic in the US capital market. As a result, NVIDIA fell by 5.3%, ARM by 5.5%, Broadcom by 4.9%, TSMC by 4.5%, as well as Micron, AMD, and Intel. Even the NASDAQ 100 futures fell to -400 points, potentially marking the largest single-day decline since December 18. According to incomplete statistics, the US stock market is expected to see a market value evaporation of over $1 trillion on Monday, wiping out a third of the total cryptocurrency market capitalization.
The cryptocurrency market, which closely follows the US stock market trend, has also experienced a sharp plunge due to DeepSeek. Among them, Bitcoin fell below $100,500, a 24-hour decline of 4.48%. ETH fell below $3,200, a 24-hour decline of 3.83%. Many are still puzzled as to why the cryptocurrency market has experienced such a rapid crash, which may be related to reduced expectations of Fed rate cuts or other macroeconomic factors.
Where does the market panic come from? DeepSeek is not developed like OpenAI, Meta, or Google, with massive capital and GPU resources. OpenAI was founded 10 years ago, has 4,500 employees, and has raised $6.6 billion in funding. Meta spent $60 billion to develop an AI data center nearly the size of Manhattan. In contrast, DeepSeek was founded less than 2 years ago, has 200 employees, and was developed at a cost of less than $10 million, without the need for a large investment in NVIDIA GPUs.
People can't help but wonder: how can they compete with DeepSeek?
DeepSeek has not only broken through the cost advantages in terms of capital and technology, but also the traditional concepts and ideologies that people have previously held.
The VP of Product at DropBox marveled on social media X that DeepSeek is a classic disruptive story. Existing companies are optimizing existing processes, while disruptors rethink the fundamental approach. DeepSeek asks: what if we do this smarter, instead of just throwing more hardware at it?
It's worth noting that the current cost of training top-tier AI large models is extremely high. Companies like OpenAI and Anthropic spend over $100 million just on the computing power, requiring large data centers with thousands of $40,000 GPUs, akin to needing an entire power plant to run a factory.
DeepSeek suddenly appeared and said, "What if we do this for $5 million?" And they didn't just talk about it, they actually did it. Their models are on par with or even surpass GPT-4 and Claude in many tasks. How did they do it? They rethought everything from scratch. Traditional AI is like using 32-bit floating-point numbers to write each digit. DeepSeek is like, "What if we only use 8-bit floating-point numbers? It's still accurate enough!" This reduces memory usage by 75%.
The DropBox VP said the result is shockingly, the training cost has been reduced from $100 million to $5 million. The required GPUs have been reduced from 100,000 to 2,000. API costs have decreased by 95%. They can run on gaming GPUs without the need for data center hardware. More importantly, they are also open-source. This is not magic, just incredibly clever engineering.
Some have also stated that Deepseek has completely disrupted the traditional notions in the field of artificial intelligence: "China only does closed-source/proprietary technology. Silicon Valley is the global center of AI development with a huge lead. OpenAI has an unparalleled moat. You need to spend billions or even hundreds of billions of dollars to develop SOTA models. The value of models will continue to accumulate (the fat model hypothesis and the scalability hypothesis suggest that model performance is linearly related to training input cost (computing, data, GPUs)). All these traditional views, even if not completely overturned overnight, have been shaken.
Archerman Capital, a well-known US venture capital firm, evaluated DeepSeek in a briefing, stating that first, DeepSeek represents a victory for open-source over closed-source, and its contributions to the community will quickly translate into the prosperity of the entire open-source community. I believe that the open-source forces, including Meta, will further develop open-source models on this basis, as open-source is a matter of many hands making light work.
Secondly, the path of OpenAI's miraculous efforts may appear a bit crude for the time being, but it cannot be ruled out that a qualitative change will occur when it reaches a certain scale, and the gap between closed-source and open-source will widen again. Based on the 70-year history of AI development, computing power remains crucial and will likely continue to be so in the future.
Then, DeepSeek has made open-source models just as good as closed-source models, and even more efficient, reducing the necessity of paying for OpenAI's APIs. Private deployment and autonomous fine-tuning will provide more development space for downstream applications. In the next one or two years, we are likely to witness a more diverse range of inference chip products and a more prosperous LLM application ecosystem.
Finally, the demand for computing power will not decrease. There is a Jevons paradox that during the first industrial revolution, the increased efficiency of steam engines actually led to an increase in the total consumption of coal in the market. Similarly, from the era of large mobile phones to the era of Nokia's widespread popularity, it was precisely because they became cheaper that they could be popularized, and because they were popularized, the total market consumption increased.
Welcome to join the official BlockBeats community:
Telegram subscription group: https://t.me/theblockbeats
Telegram discussion group: https://t.me/BlockBeats_App
Twitter official account: https://twitter.com/BlockBeatsAsia
