AI gains in US, China put India’s goal under scanner: Should the country build an LLM?

These two developments in the span of a few days have put the spotlight squarely on India, and the path it chooses in the new AI arms race. Some questions are crucial: Does India have the economic bandwidth to fuel procurement of AI-specific hardware at scale? Should India even focus on building a foundational model, or just work on use-case specific wrappers that are built on top of other companies’ models?

To LLM, or not to LLM

One of the key questions facing India and the country’s businesses and researchers is whether to build a foundational model from scratch or rely on some already available open source LLMs to build wrappers. Even within this question, there are further doubts. If an LLM is to be made, should it be India-specific, filled with information about specific Indian languages and related knowledge, or if should be a complete foundational model?

Infosys co-founder Nandan Nilekani has said that India should not focus on building large language models. “Our goal should not be to build one more LLM. Let the big boys in the (Silicon) Valley do it, spending billions of dollars. We will use it to create synthetic data, build small language models quickly, and train them using appropriate data,” Nilekani said at the Meta AI Summit last October.

However, some in the AI industry believe this could be a flawed approach. Aravind Srinivas, the founder of Perplexity AI, a conversational search engine which is being seen as a potential Google rival, said that Nilekani’s comment “is wrong.”

“…he’s (Nilekani’s) wrong on pushing Indians to ignore model training skills and just focus on building on top of existing models. Essential to do both,” Srinivas said in a post on X. Speaking about the breakthrough made by DeepSeek, Srinivas added in another post: “I hope India changes its stance from wanting to reuse models from open-source and instead trying to build muscle to train their models that are not just good for Indic languages but are globally competitive on all benchmarks.”

To be sure, the model released by China’s DeepSeek has shown that training an AI model may not be as expensive an endeavour as previously thought, with cutting edge models possible at a fraction of the cost of what it took companies like OpenAI and Meta.

“It is a good start for Indian companies to build wrappers as a first step. But there is a fundamental challenge there. In building wrappers, you are dependent on the foundational model of someone else. That model might be open source, and easy to access today. What if it isn’t tomorrow,” a senior government official said.

Computing costs

That brings one to another fundamental requirement for training LLMs: the hardware. As of now, Nvidia has a virtual monopoly on the graphics processing units (GPUs) needed to train sophisticated foundational models.

Companies like OpenAI and Meta have procured thousands and thousands of these GPUs to build their models. These GPUs are a major cost centre of AI operations today, and an area where Indian companies are challenged.
The government has announced the Rs 13,370 crore IndiaAI Mission, under which it plans to subsidise private companies to help them procure 10,000 GPUs and set up AI data centres, which could be accessed by other start-ups at a nominal cost to build their wrapper models.

There could, however, be a challenge to these plans, coming from the United States. Days before demitting office, the Joe Biden administration released an expansive regulatory framework on the export of AI hardware, which could have far-reaching consequences for India’s AI ambitions.

In an “interim final rule”, titled ‘Framework for Artificial Intelligence Diffusion,’ the US government has proposed to create three tiers of countries with specific restrictions on the export of AI chips and GPUs for each. India is in the middle tier of this classification and will face some restrictions on the number of GPUs it can import from the United States.

These countries will face a limit on how much computing power they can import from American companies, unless that computing power is hosted in trusted and secure environments.

There are caps on the levels of computing power that can go to each of these countries: roughly around 50,000 advanced AI chips through 2027, although that can double if the state reaches an agreement with the US.

Apart from this tiered classification, the law also envisions a special review called the General Validated End User. This list includes only two countries: India and China.

Indian companies that get this authorisation can use the exported items for civilian and military purposes, but not for nuclear use. Chinese companies with this authorisation can only use the technology for civilian use.