DeepSeek saga holds lessons for India

IN January this year, a little-known Chinese startup made a dramatic announcement which shook up the world of artificial intelligence (AI). The company was DeepSeek, which released its Large Language Model (LLM) to compete with models developed by American technology giants, such as ChatGPT.

What surprised everyone was the minuscule cost of its development in comparison to hundreds of billions of dollars spent by Silicon Valley behemoths. The development showed to the world that China was much ahead in AI and that America was not the only dominant force in technology development. Now, scientific journal Nature has named DeepSeek founder Liang Wenfeng among the top 10 people behind key scientific developments globally during 2025.

DeepSeek’s general-purpose V3 model and advanced model, R1, did not just make sensational headlines but have made a substantial difference to the AI landscape in just about a year. R1 is a reasoning LLM that can handle complex tasks in mathematics and coding. The Chinese firm went a step ahead of American companies by making DeepSeek an open model, allowing researchers to adapt the model’s algorithms for different fields. It has also revealed how it kept the ‘training’ costs so low.

All LLMs have to be fed with massive amounts of data or must be ‘data-trained’ to make them efficient. An industry survey has shown that Chinese models raced ahead in the AI markets this year, accounting for one-third of the global use of this technology. Chinese-language prompts ranked second in token volume behind English. It shows that the Chinese models are doing well globally and also being used widely within China. Tokens are units of data processed by AI models during training and inference and help in enabling prediction, generation and reasoning.

Soon after the DeepSeek splash, India announced a set of projects to develop foundational models, including LLMs, Small Language Models and problem-specific AI solutions suitable for Indian needs. The plan includes the development of Digital India Bhashini (AI language translation platform for Indian languages), BharatGen (multimodal LLM for public service delivery), Sarvam-1 AI Model (LLM for 10 Indian languages), Chitralekha (video transcreation platform) and Hanooman’s Everest 1.0 (a multilingual AI system for 35 Indian languages). A consortium of technology institutes, BharatGen, was formed with headquarters at IIT-Bombay; it was tasked with the development of ‘sovereign multilingual LLMs’ with funding from different ministries, including Rs 980 crore from the National AI Mission. The support makes BharatGen the first such venture mandated to develop LLMs as a national goal.

Since the Indian announcement followed soon after that of DeepSeek, BharatGen has been viewed as a direct response to the Chinese initiative. However, there are significant differences. DeepSeek is a commercial venture, built in the startup mode and meant to serve global markets. BharatGen is a strategic national mission, fully funded by the government and focused on achieving technological sovereignty. A key challenge for global models like ChatGPT and DeepSeek is understanding India’s complex linguistic and cultural diversity. BharatGen seeks to fill this gap as it will be trained on vast amounts of Indian-specific data.

DeepSeek models are general purpose and meant for global consumption, while BharatGen will be India-specific. The biggest difference is the fact that DeepSeek is already in the market, while BharatGen is yet to announce firm timelines.

Unlike companies that developed ChatGPT, DeepSeek and other such LLMs, BharatGen is not a commercial venture. It has been registered this month as a non-profit company by IIT-Bombay. Though it is fully funded by the government, it is a corporate entity. Ganesh Ramakrishnan, the IIT professor who heads the BharatGen Technology Foundation, has explained that autonomy and operational flexibility are necessary to take AI models from the lab to the market.

BharatGen is an interesting example of what can be called ‘government entrepreneurship’ — a model of the 1970s and 1980s. Two of the most successful technology ventures of that period – Computer Maintenance Corporation (CMC) and Centre for Development of Telematics (C-DOT) — followed this model. Both were formed to assert technological sovereignty and promote self-reliance in strategic areas. In both cases, research institutions like IITs and Tata Institute of Fundamental Research contributed significantly.

The absence of Indian technology companies in the development of foundational models is intriguing, though the private sector has been enthusiastic about collaborating with global firms in creating AI infrastructure such as data centres. The Indian technology services industry, with revenue close to $200 billion, should not remain just a bystander or mere adopter in the fast-changing AI landscape. It has the necessary talent, experience and financial power that must be used for national ventures like BharatGen.

DeepSeek is owned and funded by a Chinese quantitative hedge fund, High Flayer. The founder and CEO of both companies is Liang Wenfeng. It is not supported by the government, but the models it has developed follow government rules of censorship. The current Chinese regulations require all AI services to “reflect socialist core values” and refrain from delivering content that might “subvert state power” and “undermine national unity”. The Chinese models go silent or provide the official view on topics like the Tiananmen Square protests, status of Taiwan, human rights violations and criticism of the Chinese Communist Party leadership.

The India AI Governance Guidelines, released last month, talk of “responsible AI development” and underscore the need for accountability, fairness, safety and human oversight. The state-funded BharatGen will have to walk a tightrope by balancing government support and public good. Its models must avoid algorithmic bias seen in Western models and also steer clear of misinformation as well as possible political influence as seen in the Chinese models. Given that it is being funded from the taxpayer’s money, BharatGen must follow principles of transparency and accountability while developing its LLMs for different language models. Wider engagement with linguists, social scientists and experts from different fields would go a long way in developing robust AI models and avoiding mistakes of others.