Thanks to China’s unknown AI startup DeepSeekThe United States and the world have been thrown into a huge shock by the generative AI model “DeepSeek-R1” developed by . Some test results show that this is similar to or exceeds the most advanced American AI models such as Open AI’s latest chat GPT-o1 or Antropic Cloud3.5, Google Gemini 2.0, etc., so Silicon Valley is almost in a state of panic. This is considered a fatal blow to the United States’ hegemony over the world’s AI.
On the 27th (local time), the share prices of Nvidia and other major technology giants plummeted on the New York Stock Exchange, and global media, including the United States, reported the news extensively. It is becoming big global news, even more shocking than OpenAI’s development of ChatGPT in 2022.
Eventually, Trump grudgingly conceded this at another press conference that day. He could not hide his anxiety, saying, “This is a great achievement for China, and it also sounded the alarm for the resurgence of the U.S. technology community (to maintain its hegemony).”
DeepSeek-R1 ” ranked first in iPhone 16 Pro App Store
DeepSeek’s first product, “DeepSeek-R1”, caused a sensation due to its outstanding performance and ranked first on the App Store for iPhone 16 Pro. What’s even more shocking is how cheap it is compared to existing LLM models like ChatGPT or Gemini. This world-leading AI was developed at a ridiculously low cost of just $6 million (about 8.5 billion Korean won). It is estimated that its development cost is only 1/50 of ChatGPT.
Even more surprising is that China is allegedly using Nvidia ‘s outdated H800 chips. The company hit hardest on the New York Stock Exchange was Nvidia. “If such cutting-edge AI can be developed using an old H800 and $6 million, then hasn’t Nvidia been ‘cheated’ all along?” This statement has sparked doubts and opposition. As a result, the company’s stock price plummeted 20% in an instant. The company’s ranking also dropped to fourth place, with its market value evaporating by more than $500 billion.
It’s not just Nvidia. Google’s parent company Alphabet and Microsoft also fell sharply, as did the tech-heavy Nasdaq and S&P 500 overall. Even European and Asian markets showed signs of weakness and consolidation. This is consistent with the Western world’s concerns that global AI hegemony may shift from the United States to China.
In addition, DeepSeek’s “R1” model uses an older chip, which also makes people question whether it is really necessary to invest so much money in Nvidia’s latest AI chip platform “Blackwell”. The Blackwell-based RTX5080 and RTX5090 are estimated to cost at least $70,000 (about 100 million won) each. It will also be worth watching how American AI companies, which have to use expensive Nvidia products, react in the future.
NVIDIA, has it always been “blackmailing” others for profit?
Although Chinese technology continues to advance, at least in the short term, this is tantamount to a “bolt from the blue” for Silicon Valley, which believes that its advantage in the field of AI will not be shaken. What’s more, this is not the result of using the most advanced Nvidia chips based on H100 and Blackwell, but the outdated H800, so the impact is even greater.
“R1”, which ranks first in the App Store, is evaluated as second only to the most advanced GPT-o1, which prides itself on being an open AI close to AGI, or Google Gemini FlashSync, at least in terms of inference capabilities. Silicon Valley analysis said that although an old media model called “Deepseek V3” was used in the learning and inference process, “the expulsion of chat GPT from the App Store benefited from the power of “R1″”
It is said to outperform not only closed models such as GPT-o1, but also high-performance open source models such as Meta “Rama”. Most importantly, the biggest shock to the United States is that China did not use the most advanced Nvidia chips that the United States was very confident in and sanctioned, but instead used inferior chips to create this almost god-level product.
For years, the United States has strictly restricted exports of powerful artificial intelligence accelerators and chip manufacturing equipment to China. This is to prevent China from using the technology for military purposes.
Nvidia only supplies second- and third-rate products such as the old H800 and H20. But there was something wrong with the US strategy. The R1 they built using the H800 cost only $6 million, which is actually a “cheap” machine. It also reportedly consumes less power than OpenAI, Google, and Meta.
DeepSeek’s “R1” model is available on the web, in apps and APIs, and has a variety of features, including AI assistant (secretary) functions such as ChatGPT and coding content creation. However, it is open source and provided under the MIT standard, so it can be used free of charge without restriction.
Speculations about “secretly smuggling Nvidia’s cutting-edge chips” are also rampant on the Internet.
DeepSeek is a startup based in Hangzhou, China, founded by Liang Wenfeng last year. Naturally, it is at a disadvantage compared to competing models such as OpenAI and Google. Therefore, they must use AI models more efficiently and work on technologies that can build and train models at ultra-low cost.
Of course, other analysts, including Citigroup, have questioned these achievements, saying that “China has a more restrictive environment for the development of artificial intelligence than the United States, and is more stringent (politically, socially and technologically).”
Nevertheless, a sense of crisis is growing among OpenAI and large technology companies such as Silicon Valley, which have invested billions of dollars in artificial intelligence infrastructure.
However, on the day when DeepSeek became the focus of the world, a cyber attack was also carried out on it, although it was not known what was behind it.
According to CNBC, DeepSeek also issued an announcement on the same day saying: “Due to a large-scale malicious attack, the service is temporarily restricted,” “but existing users can log in as usual.”
Under such circumstances, the Internet is full of speculations such as “How can China train such complex LLM at a low cost?”
Reliable source MIT Technology Review said that “DeepSeek initially used high-performance Nvidia chips that were secretly acquired” and that “before the US export ban was further strengthened, it stockpiled a large number of Nvidia A100s and then speculated that they tried to merge technology with H800.
Alexander Wang, founder and CEO of Silicon Valley startup Skeleton AI, said the same thing. He asserted in an interview: “DeepSeek already has about 50,000 H100s.” As if he had been waiting for this, Silicon Valley’s “evil man” Musk also joined the rumor party. He argued that “China must have secretly stored a large amount of H100, as Wang said,” “otherwise, we cannot believe China’s achievements.”
However, the controversy continues as the rebuttal that “Nvidia’s H100 GPU is simply not sold to China” emerges. Some netizens further speculated that DeepSeek may have purchased the H100 chip via Singapore.
Among them, it is said that “Singapore accounted for 22% of Nvidia’s revenue last quarter.” However, according to Nvidia’s actual filing with the U.S. Securities and Exchange Commission, the amount was “extremely small.”
The result of China’s accumulated artificial intelligence technology capabilities
Experts believe that “all these speculations are due to the unwillingness to believe in China’s achievements.” Instead, he commented that DeepSeek R1 “is the crystallization of China’s long-term accumulation of excellent AI technology and the result of patient development and testing.”
Analysts in Silicon Valley have also been candid in their assessments of China’s technological prowess. This is a “new era” that shows how to create new models by fully utilizing and expanding existing models in a limited environment and regardless of sanctions.
Some say, “Of course, that requires lots of Nvidia GPUs and high-performance networks,” however, it is now possible to break this stereotype and achieve this using efficient pre-learning and post-learning (continuous validation and upgrading) as well as persistent dialectical testing expansion.
On the contrary, the conclusion is that “it is a miracle in itself that such a groundbreaking achievement could be achieved while fully complying with U.S. export controls.” On the one hand, doubts and questions about the products, performance, and prices of Nvidia, which has so far monopolized the “tyranny,” are becoming increasingly intense.
Therefore, the general view is that the company that suffered the most fatal blow in this “DeepSeek” incident is Nvidia.