What's the deal with DeepSeek AI?
Andrew Stein '27
On January 27th, Wall Street saw the most significant single-day drop in its history as nearly $600 billion of Nvidia stock went up in smoke. Tech stocks as a whole reeled from the introduction of a new, shockingly powerful AI assistant. The Chinese startup DeepSeek burst onto the international scene with the launch of their R1 chatbot, an AI model that offers functionality comparable to ChatGPT and Llama. DeepSeek-R1 even outperformed OpenAI-o1 on established math and computer coding benchmark assessments. The new model’s impressive capabilities are “wild and totally unexpected” according to Elvis Saravia, a UK-based AI researcher. DeepSeek developed R1 despite US limitations on exporting key AI technologies like the chips Nvidia manufacturers.
Perhaps more impressive than R1’s functionality is its efficiency. According to estimates, R1 costs around $6 million dollars to develop, significantly less than other models. The creation of Meta’s Llama 3.1 used eleven times as many computing resources and cost six times as much. Compared to Open AI’s o1 LLM, R1 is up to fifty times cheaper to use.
Market reactions to R1 were widespread. The NASDAQ fell by 3.1 points, the Philadelphia Semiconductor Index fell by 9.1 points, and tech companies like Alphabet and Microsoft also experienced significant losses. DeepSeek’s efficiency gains raise important questions about the future of tech, including U.S. dominance in AI, the semiconductor market, and future energy demand for LLMs.
Traditionally, American companies like Meta, Microsoft, and Alphabet have dominated AI research and pioneered the creation of LLMs. U.S. policy to protect this dominance prevents the exportation of AI technology to countries like China and Russia. DeepSeek’s progress startled policymakers, prompting the government to pass new legislation further restricting the number of chips that can be exported to China and expanding the restrictions to 120 other nations. Despite these fears, Chinese competition could be a boon to American tech companies. OpenAI and other U.S.-based models have been facing slow rates of improvement. Most of the data on the internet, the largest source of training material for LLMs, has already been used. DeepSeek used “chains of thought” reasoning to train their models around this problem. American companies have already begun replicating this innovative technique, paving the way for even more significant gains in AI performance than expected.
Experts hope that Chinese innovation will lead to more collaboration in the industry, although this seems unlikely in the short term. U.S. lawmakers have already banned DeepSeek from government devices. Alvin Wang Graylin of the Taiwanese tech firm HTC describes the government’s current approach as a “no-win arms-race.” Still, whether through collaboration or competition, foreign advances in AI technology can also benefit American companies.
The market for semiconductors is another area of uncertainty. As AI models become more efficient to develop and operate, they require fewer semiconductors. This potentially devastates companies like Samsung, Nvidia, Intel, and Qualcomm.
Nvidia significantly benefited from the rise in AI over the past five years. Investors have seen its super-advanced chips as central to the success of American LLMs. And since many AI companies like Anthropic and OpenAI are not publicly traded, investors have used Nvidia stock as a proxy for industry-wide AI optimism. DeepSeek’s new release puts a lot of this into doubt. Because of U.S. export restrictions, DeepSeek must not have access to Nvidia’s advanced technology. Because the R1 model is so impressive without using this technology, investors are starting to think that large quantities of advanced chips may not be that necessary.
Future energy demand is also questioned because AI uses immense amounts of electricity. The industry accounted for an estimated 2% of global electricity consumption in 2022, predicted to double by next year. This has stressed existing electricity infrastructure and necessitated the construction of new power plants. Three Mile Island, the Pennsylvania nuclear reactor infamous for its 1979 meltdown, is set to resume operations for the first time in decades to exclusively power new Microsoft data centers. Growth in AI has led to parallel growth in the power industry. Two of the nation’s largest electricity providers, Vistra and Constellation Energy, have plans to expand operations and construct new generators, and Vistra’s stock has grown over 700% since March 2023. While the short-term outlook for electricity companies remains stable, more efficient AI can lead to long-term reductions in demand and cause significant disruptions in the power industry.
These questions have renewed interest in Jevons Paradox, a centuries-old economic theory on efficiency gains and demand. In the 1860s, many people predicted that increasingly efficient engines would reduce the demand for coal. William Stanley Jevons hypothesized that the opposite would happen: as engines became more efficient, they would be used more frequently, and the demand for coal would rise to match. While it is easy to view DeepSeek’s innovation as a negative sign for chip manufacturers and competitive power companies, Jevons Paradox offers a more optimistic view. DeepSeek could propel further growth in AI by forcing innovation and allowing more people to use LLMs regularly. Chip companies, U.S. tech giants, and electrical utilities might be panicking right now, but they may all benefit from competition in the long run.
Disclaimer: The views and opinions expressed in the articles published in this newsletter are those of the respective authors and do not necessarily reflect the views, opinions, or positions of the newsletter, its staff, or its editors. This newsletter is a nonprofit recreational organization that serves as a platform for diverse perspectives and open discussion. It does not endorse any particular viewpoint; all content is intended only for informational and recreational purposes.
 | A guest post by
|