About the author
Philip Pilkington
Senior Research Fellow at the Hungarian Institute of International Affairs
In July 2023 Liang Wenfeng founded Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., in Hangzhou, Zhejiang. Liang was known in China for running a quantitative hedge fund called High-Flyer, which he co-founded in 2015. While the transition from hedge fund to artificial intelligence may seem unusual, it becomes less so when we understand that High-Flyer was inspired by the work of Jim Simons at Renaissance Technologies.1 Simons, an American mathematician and former National Security Agency (NSA) code-breaker, made a fortune applying cutting-edge mathematics and algorithms to financial trading. Many of these techniques were precursors to the pattern recognition technology that is used in artificial intelligence.
DeepSeek did not garner much attention from the Western press when it was first founded. Programmers, however, were paying attention to the company a few weeks after it was founded. In November 2023, co-founder of Answer.AI and fast.ai Jeremy Howard posted on Twitter/X about the progress that DeepSeek was making. "Early reports from people using this are that it's the real deal," Howard wrote, "Strong coding. Good multilingual. Consistent over long contexts. Looks to be yet another brilliant Chinese model." CEO of Abacus.AI Bindu Reddy also weighed in on X. "DeepSeek is from China and is proof that the Chinese don't need our LLM tech," she wrote, "they can develop their own and are enlightened enough to open-source it!!"
DeepSeek Roils American Markets
Because only the coding community was paying attention to DeepSeek, Western analysts were taken by complete surprise when the company launched its R1 model in January 2025. It appears that The Financial Times was the first major Western newspaper to cover the company. The newspaper, which often portrays China in a negative light, recognized how large an achievement the new model was. The article stated that "Liang Wenfeng [built a] model on [a] tight budget despite US attempt to halt China's high-tech ambitions," linking the development of DeepSeek to the attempts by the US to hobble the Chinese technology industry through sanctions on semiconductors.2
The week before the article was published, Twitter/X was buzzing about the implications of DeepSeek's new R1 model. Posts on Twitter/X as early as January 23rd were highlighting the potential efficiencies of DeepSeek's new model and how it might disrupt the American technology sector. There were multiple layers to this potential disruption, but the most obvious, and the easiest to digest, was the extremely low cost of development. DeepSeek claimed that the V3 model cost only 5.6 million USD to develop, which is a fraction of what American companies like OpenAI and Google were spending to train comparative models. These efficiencies also implied that AI models developed by companies like DeepSeek would require vastly fewer graphics processing units (GPUs) to train the models. DeepSeek claimed that it required 2,048 Nvidia H800 GPUs to train the R1 model. OpenAI's GPT-3 model which was comparable in size was trained on around 10,000 Nvidia V100 GPUs.3
In addition to the greater efficiency in training, DeepSeek claimed their model was much cheaper than alternatives previously on the market. As we will see, the cost of training the program would go on to be the subject of intense debate, but the cost of using it was not. When DeepSeek R1 was released, users found that it cost up to 95% less than its previous competitors, like OpenAI.4 In a market economy, it is not unusual that a new entrant will enter a market and drive down prices with new technology. But it is quite rare that a new entrant will drive down costs in such a radical way. Price declines of 95% are of a magnitude that can destroy entire business models. A few days after DeepSeek was released, Anjney Midha, a board member of multiple companies working on AI, wrote on Twitter/X: "From Stanford to MIT, DeepSeek R1 has become the model of choice for America's top university researchers basically overnight." Whatever would later be said about the development costs, there was no denying that DeepSeek's R1 model drastically drove down end user costs and became a favorite in research departments across the United States.
The Monday after these discussions had taken place online, the market reacted, with the stock prices of Nvidia falling from 142.62 USD per share to 118.42 USD per share - a decline of around 17%. Nvidia had seen its stock price soar in the previous months as investors had been led to believe that the AI boom would create an almost infinite demand for GPUs. Since Nvidia was a world leader in GPU production, market analysts assumed that this meant enormous future revenue growth for the company. If DeepSeek is to be taken at their word, their R1 model showed that a highly advanced AI program could be trained on around 20% of the number of chips that were used by American companies. Intuitively, this meant that future deliveries of Nvidia chips were being overestimated by markets by a factor of five.
A Bubble Exposed?
This market downturn enormously impacted the portfolios of a variety of wealthy and powerful people in the United States. The AI boom had been sold to the American public as an almost sure thing, and this led to a stampede of investors into stocks like Nvidia. During this period, Nvidia and related stocks - nicknamed the Magnificent 7 - became the key drivers for the NASDAQ index and, some would argue, the key drivers for the continued upward march of the American stock market in general.5 After DeepSeek triggered a decline in the price of Nvidia, the Magnificent 7 started to tumble, which hinted at just how important the AI boom was for the American stock market as a whole. By March 2025, investment researchers like Jim Bianco were highlighting the fact that the Magnificent 7 stocks, triggered by the DeepSeek emergence, were driving down the entire NASDAQ.
After DeepSeek triggered a decline in the market, it started to become clear how much of the wealth of many powerful people in the United States was in fact paper wealth reliant on buoyant stock market valuations. In the first two months of 2025, Elon Musk saw his net worth decline by around 81 billion USD due to the impact DeepSeek's new model had on markets, bringing his net worth down to around 350 billion USD.6 This raised obvious questions about the nature of wealth in the American economy in general: was so much of this notional wealth so fragile that a small Chinese company could develop an AI product that would expose it is phony?
Obviously, those with their wealth tied up in these vehicles had a strong personal vested interest in trying to prop up the stock valuations. Yet beyond this, the United States as a whole has a strong interest in convincing the rest of the world that these valuations are real. If they are not, this may mean that much of the paper wealth in the United States is the result of a large bubble in financial markets. There is certainly evidence to support this theory. The typical way to gauge how overvalued stock markets are uses a cyclically adjusted price-to-earnings ratio, typically known as the CAPE or the Shiller P/E ratio. In December 2024, a month before DeepSeek's new release roiled the market, the CAPE ratio stood at 37.9. We have only seen a higher CAPE ratio twice before: toward the end of 2021 after the massive post-pandemic rally and in mid-2000 at the height of the Dotcom Bubble.
The Empire Strikes Back
As the dust started to settle around DeepSeek, some American investors and entrepreneurs started to try to play down its importance. This effort was led by Elon Musk himself, who raised questions about whether DeepSeek had truly succeeded in training its model on a much smaller number of GPUs than previous models. Musk and others suggested that because the United States had imposed restrictions on the amount of chips that Nvidia and similar companies could send to China, it was strongly in the interest of the Chinese to lie about how many of such chips they possessed.7 The source of this claim appears to be a semiconductor consultancy firm called SemiAnalysis.8 A few days after the collapse in market valuations triggered by the launch of DeepSeek's R1 model, SemiAnalysis released an article claiming that they had internal evidence that suggested DeepSeek's claims of efficiency in training were likely false.
In the article, SemiAnalysis took aim at the claim that DeepSeek made that they had trained their new AI model for six million USD. SemiAnalysis claimed that the number was closer to 500 million USD. They further claimed that their analysis showed that the total server CapEx for DeepSeek was around 1.6 billion USD and stated that they believed that the company had access to around 50,000 hopper GPUs. This suggested that DeepSeek's claim to have trained the R1 on 2,048 GPUs may be misleading or even false. Everything hinges on this claim. If DeepSeek is telling the truth, then the demand for GPUs moving forward will shrink relative to current projections, and companies like Nvidia will see their future earnings forecasts contract as this new reality is factored in. If SemiAnalysis is correct, this suggests that current forecasts for GPU demand may be more robust.
SemiAnalysis has not stated publicly how they arrived at their conclusions. Logically, there seem to be only two ways of arriving at these conclusions. The first is that they have direct access to DeepSeek's capacities. That is, they have seen what DeepSeek is doing first-hand. The second is that they closely track GPU exports to China and have worked out a way to figure out which companies are buying these GPUs. If SemiAnalysis has direct knowledge of DeepSeek's operations, then it is hard to refute their claims. But they have not provided any evidence of this. In this circumstance, it is simply their word against DeepSeek's - and since we know DeepSeek definitely has access to their own internal operations, they seem like a more trustworthy candidate. If SemiAnalysis is inferring the number of GPUs that DeepSeek possesses from export data, this introduces multiple layers of uncertainty into their analysis. We cannot highlight all these layers, of course, because we do not have access to their methodology. SemiAnalysis may provide their private clients with this analysis, but for researchers in the public domain, SemiAnalysis is in effect asking us to trust their word over DeepSeek's.
China Becomes a Player in the Future of AI
The fact that DeepSeek released its product open source is also causing ripples in the AI market. In this space, there has been a long-running debate about whether AI should be an open source sector wedded to companies that produce profit or whether it should be treated like any other product and used to extract rent. Elon Musk has long rhetorically supported the open source approach, as suggested by the name of the company he founded: "OpenAI," despite OpenAI never being open source, and Elon Musk's own AI, Grok, not releasing its weights and measures. On the other side of this debate is Sam Altman, the current CEO of OpenAI, who has taken a more traditionally commercial approach. After the DeepSeek debacle, Musk called Altman a "swindler" on social media, and made an unsolicited 97.4 billion USD bid to buy OpenAI.9 It seems that no matter what happens next, DeepSeek's approach to AI is having an enormous impact on the American AI market.
Moving forward, it will become clear to what extent the initial hype around DeepSeek will prove itself to be accurate. At a minimum, DeepSeek's R1 platform has proven that Chinese companies can produce world-leading AI products and that they can blow their competitors out of the water based on end user cost. At the maximum, DeepSeek has shown that it can train a world-leading AI platform on much fewer GPUs than previously thought possible and at a much lower investment cost. If this maximalist interpretation proves correct, it could throw the American technology sector into chaos and expose a gigantic stock market bubble in the NASDAQ and related markets. We will have to wait and see.
1. Gregory Zuckerman, "The Guy Behind DeepSeek Blurbed My Book in China," The Wall Street Journal, January 27, 2025, https://www.wsj.com/tech/ai/the-guy-behind-DeepSeek-blurbed-my-book-in-china-ce08bd4f.
2. Eleanor Olcott and Zijing Wu, "How Small Chinese AI Start-Up DeepSeek Shocked Silicon Valley," Financial Times, January 24, 2025, https://www.ft.com/content/747a7b11-dcba-4aa5-8d25-403f56216d7e.
3. "DeepSeek Sharpens Its Reasoning," The Batch, January 27, 2025, https://www.deeplearning.ai/the-batch/DeepSeek-r1-an-affordable-rival-to-openais-o1/.
4. Brian Buntz, "This Week in AI Research," R&D World, January 23, 2025, https://www.rdworldonline.com/this-week-in-ai-research-a-0-55-m-token-model-rivals-openais-60-flagship/.
5. Taylor Sohns, "Dominance of the 'Magnificent Seven' Stocks," May 9, 2024, Nasdaq, https://www.nasdaq.com/articles/dominance-of-the-magnificent-seven-stocks.
6. "Elon Musk Loses 90 Billion USD: How China's DeepSeek Torpedo Hit Where It Hurts Most," The Economic Times, March 4, 2025, https://economictimes.indiatimes.com/news/international/global-trends/elon-musk-loses-90-billion-how-chinas-DeepSeek-torpedoes-hit-where-it-hurts-most/articleshow/118702268.cms?from=mdr.
7. "Elon Musk Not Convinced with China's DeepSeek That Rattled the US Tech Market. His One-Word Reply Is the Proof," The Economic Times, March 4, 2025, https://economictimes.indiatimes.com/news/international/global-trends/us-news-elon-musk-DeepSeek-ai-elon-musk-not-convinced-with-chinas-DeepSeek-that-rattled-the-us-tech-market-his-one-word-reply-is-the-proof/articleshow/117678549.cms?from=mdr.
8. Dylan Patel et al., "DeepSeek Debates: Chinese Leadership on Cost, True Training Cost, Closed Model Margin Impacts," SemiAnalysis, February 18, 2025, https://semianalysis.com/2025/01/31/DeepSeek-debates/.
9. Jessica Toonkel and Berber Jin, "Elon Musk-Led Group Makes 97.4 Billion USD Bid for Control of OpenAI," The Wall Street Journal, February 10, 2025, https://www.wsj.com/tech/elon-musk-openai-bid-4af12827.
This article is from the March issue of TI Observer (TIO), which explores the AI-powered digital economy, analyzing how nations navigate the balance between development and governance, while examining the impact of technological advancements on global competition and the broader international order. If you are interested in knowing more about the March issue, please click here:
http://en.taiheinstitute.org/UpLoadFile/files/2025/3/31/14372768c45e0ae0-6.pdf
——————————————
ON TIMES WE FOCUS.
Should you have any questions, please contact us at public@taiheglobal.org