With DeepSeek, are India’s foundational AI model dreams closer to reality? Premium

DeepSeek-R1's emergence from China disrupts AI landscape, sparking debate on cost-effective foundational models in India.

The emergence of DeepSeek-R1 from China — a highly advanced Artificial Intelligence large language model (LLM) — has jolted technology stocks that were valued on the assumption that highly expensive Graphics Processing Units (GPUs) and the investments that accompanied them were key to winning in the global AI race. R1’s capacity to compete with cutting-edge models from OpenAI’s ChatGPT on a relatively shoestring budget of $6 million has transformed the debate over building a foundational LLM in India, mere months after such an undertaking was deemed too expensive to bother with. 

Foundational models are the most resource-intensive LLMs to make, as they require a massive database of content on which a generative pre-trained transformer (GPT) model bases its functioning. GPTs like OpenAI’s are by far the most common type of LLMs, and their costs have raised prohibitive barriers to starting from scratch, especially for Indian firms with little cash to expend in exploratory efforts.

“Foundation models are not the best use of your money,” Infosys co-founder Nandan Nilekani reportedly said in December. “If India has $50 billion to spend, it should use that to build compute infrastructure, and AI cloud. These are the raw materials and engines of this game.” As DeepSeek R1 emerged with capabilities that seemed the reserve of highly capitalised firms like OpenAI just a few weeks ago, the template may have well changed. 

Shortly after Mr. Nilekani made those observations, he had drawn opposition. Aravind Srinivas, the CEO of Perplexity, the AI-powered answering engine that has emerged as a competitor in the search engine space, said in mid-January, posted on X (formerly Twitter) that Mr. Nilekani was “wrong on pushing Indians to ignore model training skills and just focus on building on top of existing models,” and that it was “[e]ssential to do both.” 

India’s cost arbitrage has been its key advantage in technology. From business outsourcing firms leveraging lower labour costs for specialised IT workflows to telecom firms like Reliance Jio leaping into the industry after network infrastructure costs sufficiently decreased, innovating on cost has been both a strength and a practical approach for Indian technical innovation. R1 being built and deployed on an open source basis gives India both a signal and a key tool in proceeding down the foundational AI model path. 

“We have a ray of hope that LLM training and usage can be democratised,” Paramdeep Singh, CEO of the Gurugram-based Shorthills AI said. “It’s not people sitting in ivory towers, but talent with frugal hardware that can train the best model.” HCL founding chairman Ajai Chowdhry said in a statement that the capacity to build a homegrown LLM “exists within the country and we must create own GPUs and develop state-of-the-art models, such as LLMs, using Indian languages  considering we possess an abundance of data that can be employed as a tactical advantage.” 

Dr. Chowdhry, who now heads the EPIC Foundation, an organisation seeking to boost domestic electronics manufacturing, said that China’s advances were “alarming,” and cited that country’s mushrooming number of AI labs and initiatives, and said that India should keep pace. “Imported chips and enormous data centres aren’t necessary for innovation,” Mr. Chowdhry said, concurring with the growing assessment that costs are no longer the barrier that they appeared just days ago.