Express View on DeepSeek: Small and smart

If artificial intelligence is the new battlefield, when was the first shot fired? Was it last week, when a Chinese AI lab released a Large Language Model (LLM), developed at a fraction of the cost of the ones built by Silicon Valley, and yet capable of beating these same models on several benchmarks? Or was it this week, when Wall Street opened to a bloodbath, with US stocks losing a reported $1 trillion in value, including chip-maker Nvidia’s $600-billion drop? Perhaps it can be traced way back to October 2022, when the US, in an effort to retain its edge, banned the export to China of advanced chips used for AI development? As questions pile up on the implications of China’s DeepSeek R1 and the future of AI, one thing is clear: America’s lead in the race has shrunk considerably. It seems to have happened, moreover, in the blink of an eye, catching the US and Silicon Valley — as indeed, the world — completely by surprise.

Yet, it is not as if DeepSeek, with a logo featuring a blue whale, surfaced out of nowhere. A minnow compared to the Big Four in China’s tech sector (Baidu, Alibaba, Tencent and Xiaomi), DeepSeek made ripples last year when it released its V3 model, which was described by Andrej Karpathy, one of the founders of OpenAI, as a “frontier-grade LLM, trained on a joke of a budget” (a “mere” $6 million, compared to the over $500 million that OpenAI has reportedly spent on ChatGPT). The January 20 debut of R1, a reasoning model developed on a shoestring budget and with second-rate chips which can, nonetheless, perform as well as OpenAI’s most advanced model, shook one of the central tenets of AI development in the US — that it takes billions of dollars to produce cutting-edge technology.

Advertisement

That lack of access to first-rate chips couldn’t hinder DeepSeek also shows the limits of America’s “small yard, high fence” approach, and that smaller, nimbler competitors could easily get within heel-snapping distance. The open-source strategy adopted by Chinese labs, including DeepSeek, as opposed to the proprietary model favoured by those in the US (except, notably, Meta) also hints at another way in which the rules of the game could be rewritten by China, as more global talent — and, indeed, other countries nursing their own AI ambitions — gravitate towards free-to-use models.

India, which talks of building “sovereign AI”, may find hope in the DeepSeek story, that it is possible to do more with less. Yet, even as it puts its weight being the ambition and talent of the homegrown tech sector, and puts in the Rs 10,300 crore committed under the IndiaAI mission in March 2024, the government must also find a way to ensure access to the best AI hardware. If the future will be defined by AI, India must be a part of it.