HOW TO WIPE OUT WIPE OUT $1.5 TRILLION

On 29th January 2025, Nvidia stock ended trading down 10% in a turbulent session that wiped out about $130 billion in market value. And it more than rattled markets the following week. More than 70% of the 69 stocks in the Technology Select Sector SPDR ETF (XLK) dropped on the news. And not by a little. Nvidia (NVDA) alone lost $596 billion in value, which is more than more than 485 S&P 500 stocks are individually worth (Forbes and Guardian).

The reason was Deepseek. Akin to Chatgpt, DeepSeek is a Chinese AI large-language model believed to be on par with some of the best models released by OpenAI and others. The company Deepseek has developed this AI model despite not having access to leading AI accelerator chips, such as Nvidia's best Hopper and Blackwell products. In less than a week, the US and European stock markets lost $1.5 tn. President Donald Trump is considering placing more restrictions on Nvidia’s semiconductor chip sales to China beyond what the Biden administration placed in an effort to limit AI advancement in the Asian country, Bloomberg reported, citing anonymous sources.

DeepSeek is a Chinese artificial intelligence (AI) company based in Hangzhou. It emerged a couple of years ago from a university startup. Its stated goal is to make an artificial general intelligence – a term for a human-level intelligence that no technology firm has yet achieved. It’s not there yet, but this may be one reason why the computer scientists at DeepSeek have taken a different approach to building their AI model, with the result that it appears many times cheaper to operate than its US rivals.

Liang Wenfeng, who used to run a Chinese quantitative hedge fund that now funds DeepSeek. The arrival of DeepSeek R1, an AI language model built by the Chinese AI lab DeepSeek, has been nothing less than seismic. The system only launched in January 2025, but already the app has shot to the top of download charts, sparked a $1tn (￡800bn) sell-off of tech stocks, and elicited apocalyptic commentary in Silicon Valley. The simplest take on R1 is correct: it’s an AI system equal in capability to state-of-the-art US models that was built on a shoestring budget, thus demonstrating Chinese technological prowess. But the big lesson is perhaps not what DeepSeek R1 reveals about China, but about western neuroses surrounding AI. R1 is a "reasoning" model. These models produce responses incrementally, simulating a process similar to how humans reason through problems or ideas.

For AI obsessives, the arrival of R1 was not a total shock. DeepSeek was founded in 2023 as a subsidiary of the Chinese hedge fund High-Flyer, which focuses on data-heavy financial analysis – a field that demands similar skills to top-end AI research. Its subsidiary lab quickly started producing innovative papers. After graduating from Zhejiang University, Wenfeng co-founded the quantitative hedge fund High-Flyer in 2015.

The company has been quietly impressing the AI world for a while with its technical innovations, including a cost-to-performance ratio several times lower than that for models made by Meta (Llama) and OpenAI (Chat GPT). It hasn’t been making as much noise about the potential of its breakthroughs as the Silicon Valley companies.

This model uses a different kind of internal architecture that requires less memory use, thereby considerably reducing the computational costs of every search or interaction with the chatbot-style system. It has been praised by researchers for its ability to tackle complex reasoning tasks, particularly in mathematics and coding and it appears to be producing results comparable with rivals for a fraction of the computing power. DeepSeek has said it took two months and less than $6m (￡4.8m) to develop the model, although some observers caution this is likely to be an underestimate. Nevertheless it is vastly less than the billions that the Silicon Valley tech companies are spending to develop AIs and is less expensive to operate.

Why did US tech stocks fall?

Hundreds of billions of dollars were wiped off big technology stocks after the news of the DeepSeek chatbot’s performance spread. The timing was significant as in recent days US tech companies had pledged hundreds of billions of dollars more for investment in AI – much of which will go into building the computing infrastructure and energy sources needed, it was widely thought, to reach the goal of artificial general intelligence. DeepSeek’s performance has done all that at a fraction of the cost and who knows what it has in store for the future?

Nvidia is one of the companies that has gained most from the AI boom. It went from being a maker of graphics cards for video games to being the dominant maker of chips to the voraciously hungry AI industry. It has been compared to a modest trader in pickaxes and buckets in 19th-century California, which happened to be on the spot when the gold rush happened and so it became a massive supplier to the world’s richest industry. Nvidia a bellwether for the AI revolution, bore the brunt of the selloff, losing 17% in one day. As a result, funds and ETFs with an exposure to Nvidia had a torrid time, with one leveraged Nvidia ETF shedding 51%. Tech companies looking sideways at DeepSeek are likely wondering whether they now need to buy as many of Nvidia’s tools. Its market value fell by $600bn on one day.

What has been Deepseek’s strategy?

Deepseek researchers claim it cost $6m (￡4.8m) to train this AI, a fraction of the "over $100m" alluded to by OpenAI boss Sam Altman when discussing GPT-4. In 2021, Wenfeng began buying thousands of Nvidia chips as part of a side AI project—well before the Biden administration started limiting the supply of cutting-edge AI chips to China. Some experts believe this collection - which some estimates put at 50,000 - led him to build such a powerful AI model, by pairing these chips with cheaper, less sophisticated ones. By the time the US realised, Wengfeng had already built up his stockpile.

DeepSeek’s first AI model, DeepSeek Coder, was released in November 2023 as an open-source model designed for coding tasks. In May 2024, DeepSeek-V2 was released, which was well-received due to its strong performance and low cost. The company’s latest models, DeepSeek-V3 and DeepSeek-R1, further established DeepSeek as a leading AI research lab in China. However, it was DeepSeek-R1, released in January 2025, that focused on reasoning tasks and challenged OpenAI’s GPT-4 model with its advanced capabilities, making everyone take notice of DeepSeek. In no time, the DeepSeek app made it to the top of the US app store, dethroning ChatGPT.

The Future

Deepseek hasn’t reached artificial general intelligence, the threshold at which AI starts to reason and which OpenAI and others in Silicon Valley are pursuing. Sam Altman, OpenAI’s chief executive, has cautioned that breakthrough is unlikely to be imminent. But it does seem to be doing what others can at a fraction of the cost.

Buoyed by the rise of Deepseek many countries including India and Japan have announced joining of the global AI race by launching their own foundational artificial intelligence (AI) models on the lines of OpenAI’s ChatGPT, Google’s Gemini, and China’s DeepSeek R1.

Even as Deepseek avoids taboo topics like what happened at Tiananmen Square on 4 June 1989 (BBC), you should expect to see more of DeepSeek’s cheery blue whale logo as more and more people around the world download it to experiment.

(Disclaimer: Views expressed are the author’s own).