No one thinks that the path to artificial general information (AGI) is smooth for investors, but the appearance of Deepseek clearly throws plot twisting the AI story.
nvidia (NVDA -3.67 %) In addition, on Monday, January 27, other AI shares appeared at a 1 -minute cost, as investors responded to the threat from Deepseek, a Chinese AI chatbot, comparable to the top models like Chatgpt. 。
NVIDIA lost 17 % in a single session and wiped out the $ 600 billion market value, the largest day in market history. Since then, NVIDIA has regained some of these losses. Investors may believe that selling is an overreaction.
Nevertheless, the AI’s stock remains remarkably low, and NVIDIA itself tilts a hat to a Chinese emerging company, and the spokesman calls it a “excellent AI progress.”
DeepSeek seems to have gained reliability because Microsoft, considered to be Openai’s largest investor, has already added a model to the Azure Cloud Infrastructure Service.
So how much threats are the deep threat to the AI ecosystem? To answer that question, let’s outline some facts about Deepseek.
Image source: NVIDIA.
What we know about DeepSeek
DeepSeek is a Chinese emerging company established in May 2023 by hedge fund chief Liang Wenfeng. Unlike Openai’s Chatgpt and Alphabet Gemini, DeepSeek uses a major open source language model. In other words, developers can update them and adapt to their needs.
DeepSeek is important because the R1 model is comparable to O1 in Openai in categories such as mathematics, code, and inference tasks.
According to some estimation, O1Nai’s O1 model is $ 60, generating 1 million tokens output, but Deepseek’s R1 can be distributed the same quantity for only $ 2.19.
DeepSeek was impressed by the industry insider in a research paper on page 22, which explains the modeling mechanism, but the company has been accused of building a model using Openai using a method called distillation. , More skilled. Doing so constitutes a violation of Openai’s terms of use.
Distillation is generally used in AI, but if the accusation is true, it seems to impair many Deepseek’s reliability, like a Chinese new company that steals at least a part of the model. You can see it.
There is also a debate on how much DeepSeek actually paid to the infrastructure because it costs only $ 5.6 million to train the V3 model. V3 has been built in the NVIDIA H800S, which was designed to be executed in the same way as H100, the GPU of NVIDIA, which is widely used for building US AI infrastructure and models, avoiding US export rules. 。
Analysts have questioned the $ 5.6 million numbers, and they do not seem to include important costs such as surveys, architecture, and data, and need billion dollars -based AI. It is difficult to compare directly with the model. 。
What it means for NVIDIA
It is too early to know what Deepseek’s meaning is for NVIDIA and a wider AI sector, and there is still a lot of uncertainty about DeepSeek’s achievement.
The company seems to have gained efficiency seriously, but if the model is partially built by borrowing from Openai, they are not very impressive. The true cost of the model is not completely clear.
The launch of the Deepseek-R1 is called “Sputnik Momon” by Silicon Valley Hincho Mark and Lisen, and the new chatbot geopolitical meaning is as meaningful as technical.
The United States has been able to respond by strengthening the technology Cold War with China, tightening export rules, and taking other measures. Furthermore, it seems incompatible to allow Deepseek on a US smartphone while banning Tiktok. US companies and governments may be skeptical to deliver data to emerging companies.
For NVIDIA investors, this is one episode of long -standing technology evolution, and it is worth remembering that it is probably meaningless enough to see a $ 600 billion day. Even if DeepSeek moves the industry as a more efficient open source architecture, this can be positive for NVIDIA for a long time. According to Jevons Paradox, lowering prices to execute AI models may increase demand, increase the total consumption and purchase AI chips from NVIDIA, but maybe low. It is cost.
It also makes sense that DeepSeek was built in the NVIDIA chip. No one is challenging its advantage there.
Why am I worried about NVIDIA?
Investors should not miss the wooden forest here. Here, you need to remember that a competent AI chatbot is not the goal. The ultimate goal is artificial general information, including applications such as autonomous driving cars and robot engineering, and it is unknown whether DeepSeek will dramatically change the amount of it around it.
These technologies are so powerful and valuable that the race to AGI continues, and the high -tech jirates, which are competing there, will continue to pour billions of infrastructure necessary to build it. Efficiency is important, but technical leadership is the real award here.
NVIDIA’s shares have already dealt with DeepSeek, which can apply to NVIDIA’s business, but it has been proven that the company has previously been agile.
Mainly providing video games has evolved to migrate to cryptocurrency, AI, automatic driving cars, 3D rendering, etc. CEO JENSEN HUANG is considered to be a foresight in the industry and continues to innovate rapidly on the new Rubin platform under development.
AI Frontier continues to evolve, and NVIDIA adapts to the market situation as needed. Regardless of the effect of DeepSeek, the race to AGI does not disappear and NVIDIA does not disappear.
