How do you stop Nvidia’s golden run on Wall Street? A team of young Chinese developers with little to no experience, a hedge-fund owner with disdain for conventionalism, and some clever programming shortcuts – that was DeepSeek’s recipe anyway.
Nvidia’s prolonged rally on the stock market came to a tumbling halt on January 27th, with the release of a new AI language model ‘R1’ by Chinese firm ‘DeepSeek’. The American chip giant suffered a $588.8bn market value decline, the largest single-day market value loss in history, while the tech-heavy Nasdaq slid 3.3% in a single day.
The scale of this loss may sound surprising to some. At face value, DeepSeek is simply another AI language model among many, including ChatGPT, Gemini and LLamMA. How, then, did such an ostensibly familiar product lead to a $1 trillion dollar sell off on Wall Street and what does its introduction mean for the future of AI?
The profound market reaction stems not from the product’s features but its revolutionary development process. Upon the model’s release, the DeepSeek development team claimed to have trained the DeepSeek R1 foundational model for a mere $5.6m, using only 2000 Nvidia H800 GPU’s. While there has been some skepticism cast about the cost, if true, it is a tiny fraction of what it cost to develop OpenAI’s ChatGPT, which cost over $100m to train and used as many as 27,000 H100 GPU’s, a more advanced and sophisticated GPU than the chip used by DeepSeek.
AI researcher Lennart Heim recently explained how DeepSeek managed to achieve the technology for such a low price in the Wall Street Journal, using a librarian analogy. Imagine the earlier versions of ChatGPT as a librarian who has read all the books in the library. When asked a question, it gives an answer based on the many books it has read. This is a costly process; the analogy’s librarian requires a lot of chips, electricity and cash to read so many books.
DeepSeek took another approach. Its librarian hasn’t read all the books, but it is trained to hunt out the right book for the answer after it is asked a question. The developers combined this with another approach called ‘mixture of experts’. Rather than trying to find a librarian who can master questions on any topic, DeepSeek’s R1 model delegates questions to a roster of experts in specific fields. Each of these ‘experts’ needs less training, lowering the demand on chips to do everything at once.
This approach requires less time and power before a question is asked but uses more time and power while answering. These innovative shortcuts helped DeepSeek to train its AI model at a fraction of the cost of competing models and it is these remarkable developments that have turned the AI world on its head.
The DeepSeek revolution has shown that the cost of training AI models may be much lower than initially anticipated, opening up the AI playing field to smaller developers. The increased competition has signalled that Big-Tech may not be able to deliver the oligopolistic profits investors had hoped for, igniting the big-tech sell off.
Furthermore, DeepSeek’s ability to train such a model with less chips has also created concerns among the investment community that the anticipated demand for semiconductors could be lower than expected, explaining Nvidia’s sudden downturn.
The improvements in energy efficiency of the AI model has also ruptured investor anticipations surrounding future energy demands, leading to a fall in the energy market also.
DeepSeek has provided a revelation that has dramatically shaken-up the AI market. The market believed high-end chips, huge energy demands and a generous budget were necessary to develop complex AI algorithms. DeepSeek told them they might be wrong.
