Pageviews:
Introduction
In late January 2025, the artificial intelligence industry experienced a shock that nobody had fully anticipated. A Chinese AI research company called DeepSeek, which had been founded in 2023 as a subsidiary of a Chinese hedge fund, released a new large language model called DeepSeek R1. Independent evaluations showed that R1 matched or outperformed the leading models from American companies including OpenAI, Google, and Anthropic on many standard benchmarks. This would have been notable enough on its own. What made it extraordinary was the cost at which DeepSeek had achieved this: the company claimed it had trained its model for approximately six million dollars, compared to the hundreds of millions or even billions of dollars that American companies have reportedly spent training comparable systems. The reaction from financial markets was swift and severe. Nvidia, the company that makes the specialised chips used to train AI models, lost approximately six hundred billion dollars in market value in a single trading day, the largest single-day market cap loss in US stock market history at that point. Other American technology companies also fell sharply. The assumption that enormous spending on specialised chips and enormous electricity bills were the only path to frontier AI performance had been challenged in a direct and public way.
Why Chips Matter So Much
To understand why DeepSeek’s announcement mattered so dramatically for Nvidia specifically, it helps to understand the current economics of AI development. Training a large language model requires performing a staggering number of mathematical calculations in parallel. The most efficient chips for this type of computation are graphics processing units, or GPUs, originally designed for rendering video game graphics but repurposed for AI because of their ability to handle many calculations simultaneously. Nvidia has a near-monopoly on the high-performance GPU market for AI training, and its most advanced chips, the H100 and A100 series, have become so valuable that they are rationed and sold for tens of thousands of dollars each. American AI companies have been building enormous data centres containing tens of thousands of these chips. This capital-intensive model of AI development has benefited Nvidia enormously and has also been used to justify US export controls on advanced chips to China, which were intended to prevent Chinese companies from being able to train competitive AI models. If DeepSeek had achieved competitive performance while using far fewer chips or less efficient chips than its American competitors, then both the investment case for Nvidia and the effectiveness of the export controls were called into question simultaneously.
How DeepSeek Did It
DeepSeek’s technical reports, which the company made publicly available, described several approaches that allowed it to train effectively with fewer computational resources. One key technique was mixture-of-experts architecture, which means that only a fraction of the model’s total parameters are used for any given calculation. Rather than running the entire model for every query, the system routes different types of problems to different specialised sub-networks, which is more computationally efficient. Another technique, called reinforcement learning from chain-of-thought reasoning, allowed the model to improve its performance on reasoning tasks without requiring labelled training data that is expensive to produce. Researchers around the world reviewed DeepSeek’s technical papers and largely confirmed that the company had genuinely innovated in making efficient use of limited computational resources. There was debate about whether all of DeepSeek’s claims about training costs were accurate, but the basic achievement of matching frontier performance with greater efficiency was not seriously disputed.
What This Means for India
India is positioning itself as a significant player in the global AI landscape, with government programmes supporting AI infrastructure development and a large pool of technical talent in the field. The DeepSeek story has two implications for India. First, it suggests that the massive capital requirements previously assumed to be necessary for frontier AI development may be lower than they appeared, which makes it more plausible that countries and companies outside the United States can participate meaningfully at the frontier. Second, it reinforces the importance of software innovation as a source of competitive advantage. India’s technology industry has historically focused more on software services than on hardware or chips. A world where software efficiency can compensate for hardware limitations is, in some ways, a more favourable environment for India’s particular strengths.
Final Thoughts
The DeepSeek story does not mean that computing resources no longer matter for AI development, or that American AI companies have lost their lead. But it does demonstrate that technological leads in any field are rarely permanent, and that the assumption of large, established players in any industry that their advantage is defensible indefinitely is often wrong. For students watching the AI industry develop, the most important lesson from DeepSeek may not be about China or the United States but about the nature of innovation itself: breakthroughs often come from people working under constraint, finding clever solutions to problems that those with abundant resources never had to solve.