Stock markets were being rocked Monday by a flurry over an AI chatbot created by the Chinese tech startup DeepSeek, which was also igniting discussions about the geopolitical and economic rivalry between China and the United States in the development of AI technology.
Driven by interest in the ChatGPT rival, DeepSeek’s AI assistant rose to the top of the free software download charts on Apple’s iPhone store on Monday.
The notion that the Chinese startup has caught up to the leading American companies in generative AI at a fraction of the cost is one of the things that worries some observers of the U.S. tech industry.
If accurate, that raises doubts about the enormous sums of money that American tech companies claim they intend to spend on the computer chips and data centers required to support future developments in AI.
However, misunderstandings and exaggeration over DeepSeek’s technological innovations also caused uncertainty.
Several stock experts, like Stacy Rasgon of Bernstein, who tracks the semiconductor business, described Wall Street’s response as exaggerated, saying,
“The models they built are fantastic, but they aren’t miracles either.”
According to Rasgon, “They’re not using any innovations that are unknown or secret or anything like that.” “Everyone is experimenting with these things.
DeepSeek: What is it?
Later that year, DeepSeek, a Chinese business based in Hangzhou, unveiled its first AI large language model. Its CEO, Liang Wenfeng, was a co-founder of High-Flyer, a leading hedge fund in China that specializes in AI-driven quantitative trading.
According to a post that summer on the Chinese social networking site WeChat, the fund had accumulated a cluster of 10,000 of California-based Nvidia’s high-performance A100 graphics processor chips by 2022, which are needed to create and operate AI systems.
Soon later, the United States banned certain chips from being sold to China.
According to DeepSeek, its latest models were constructed using Nvidia’s less powerful H800 chips, which are legal in China.
This suggests that advanced AI research may not require the most expensive technology.
After releasing a new AI model last month that it claimed was comparable to models from American companies like ChatGPT maker OpenAI and was more economical in its use of pricey Nvidia chips to train the system on massive amounts of data, DeepSeek started to garner more attention in the AI industry.
When the chatbot first surfaced on the Google and Apple app stores earlier this year, it became more widely available.
However, the hysteria that ensued was triggered by a follow-up research article that was released last Friday, the same day that President Donald Trump took office.
That research discussed another DeepSeek AI model, R1, which was substantially less expensive than an OpenAI model of the same name, o1, and demonstrated sophisticated “reasoning” abilities, such as the capacity to reconsider how it approached a mathematical issue.
“I don’t know what their economics look like,” Rasgon remarked. “But I believe that people were alarmed by the price points.”
The background of “Sputnik”
The controversy surrounding DeepSeek’s technical prowess is a result of a U.S. discussion about how to effectively compete with China in the field of artificial intelligence.
In a Sunday post on social media site X, venture entrepreneur Marc Andreessen referred to the 1957 satellite launch that sparked a Cold War space race between the Soviet Union and the United States, saying, “Deepseek R1 is AI’s Sputnik moment.”
Andreessen, a tech policy advisor to Trump, has cautioned that excessive U.S. government regulation of the AI sector could hurt American businesses and help China advance.
However, limiting the sale of AI semiconductors made in the United States to China is another important U.S. foreign policy initiative that could be jeopardized by the focus on DeepSeek. That is not an accident, according to some U.S.-China relations experts.
“The technology innovation is real, but the timing of the release is political in nature,” said Gregory Allen, director of the Wadhwani AI Center at the Center for Strategic and International Studies.
Allen likened DeepSeek’s disclosure last week to the 2023 export limitations announced by the Biden administration by the U.S.-sanctioned Chinese business Huawei.
“A really important goal of Chinese foreign policy right now is trying to show that the export controls are counterproductive or futile,” Allen added.
Trump praised DeepSeek’s innovation on Monday, saying it was “good because you don’t have to spend this much money.”
If the DeepSeek report is correct, Trump described it as “positive” since “you won’t be spending as much and you’ll get the same result” while addressing House Republicans in Miami on Monday.
A “wakeup call for our industries that we need to be laser-focused on competing to win,” he said of the development.
Trump is likely to stick with and toughen Biden’s strategy, as seen by the order he signed on his first day in office last week, which stated that his administration would “identify and eliminate loopholes in existing export controls.”
Laptops 1000Without the same level of funding, DeepSeek’s advancements in AI could jeopardize the $500 billion AI investment that Trump announced at the White House from OpenAI, Oracle, and SoftBank.
Despite a 17% decline in its stock on Monday, Nvidia praised DeepSeek’s work in a statement, calling it “an excellent AI advancement” that made use of “widely-available models and compute that is fully export control compliant.”
What distinguishes DeepSeek?
Although the company hasn’t revealed the data it used for training, DeepSeek’s models are “open source,” which means that important components are freely accessible and modifiable by anybody. This sets it apart from rivals like OpenAI.
The feature of DeepSeek’s R1 model that has garnered the most praise, however, is what Nvidia refers to as a “perfect example of Test Time Scaling”—the process by which AI models successfully demonstrate their line of reasoning and then utilize that for more training without requiring them to be fed new data sources.
“It’s just thinking out loud, basically,” said Lennart Heim, a researcher at Rand Corp.
The same is true of OpenAI’s reasoning models, which begin with o1, and Heim stated that other American rivals like Anthropic and Google probably possess comparable but unreleased capabilities.
On the other hand, “This is the first time we’ve seen a Chinese company close that quickly. “That’s probably why so many people are interested in it,” Heim remarked.
I thought that OpenAI was the industry leader and no one could overtake it. It turns out that this isn’t entirely true.
