Jaspreet Bindra: The K-shaped trajectory of AI offers India a big opportunity

While doubts around DeepSeek’s unbelievably low cost of $5.6 million have only grown, it is still pivotal for two reasons: an open-source model is almost as good as a proprietary one, and Chinese large language models (LLMs) are as good as US ones. 

Also Read: Parmy Olson: The DeepSeek AI revolution has a security problem

While US tech stocks, especially Nvidia’s, tanked the week after, investors have kept their faith in US Big Tech firms that have announced around $350 billion of AI capex just this year, even as OpenAI and SoftBank double down on America’s $500 billion Stargate project. So, what will happen to AI foundational models, and where will India be in this game?

These models seem to be following a K-curve. Economists spoke of a K-shaped economic recovery after the covid pandemic: an uneven revival of the economy with different sectors, industries or groups of people on different trajectories, some moving up while others fared badly. That seems to be happening to LLMs too. 

Big models are getting bigger and more expensive, as Big Tech companies spend billions on the latest Nvidia graphics processing units (GPUs), training and building trillion-scale parameter models, massive data centres and the gigawatts of power that drive them. 

Also Read: Silicon Valley’s blind spots have been exposed by China’s DeepSeek

Meanwhile, there is DeepSeek and a rash of Chinese models like the Kimi K1.5, Kwen 2.5, Doubao 21.5 Pro, with their open-source-led innovation driving the downward cost slope of the K’s lower arm. Using older and fewer chips, together with innovative algorithmic techniques, models are being developed with only millions of dollars rather than billions. 

Much like in economics, this K-shaped curve will prevail for a while—with rich countries and even richer Big Tech companies pouring resources into the K’s upper arm to build models housed in giant clouds, while scrappy startups and emerging countries drive the lower arm further down with models hosted by mobile handsets and other devices.

Upper-K players will endeavour to finesse reasoning models and achieve artificial general intelligence and super intelligence as key differentiators, with control, scaling and monetization as priorities; they will offer high-end, closed-source AI products. 

Meanwhile, lower-K players will strive for open source models, the democratization of AI and delivery of good-enough AI to every citizen, helping solve population-scale problems. They will prioritize cost efficiency and distribution. 

Also Read: DeepSeek’s breakthrough is a pivotal moment for the democratization of AI

Lower-K success is what countries like India should aim for. IT minister Ashwini Vaishnaw seemed to be going along that path when he announced plans to develop indigenous foundational AI models for India. The question of whether we need our own LLMs has gained salience lately, and was hard to answer when the upper-K narrative prevailed. However, with the open-source democratization path that DeepSeek has shown, India can confidently envision its own LLM. 

We need a locally developed model to cater to India’s diverse linguistic and cultural landscape. Non-English speaking populations need to be served well. Such a model must be trained on India-specific data-sets, with our cultural and social nuances taken into account. Our healthcare and education related problems deserve special attention.

Also Read: Nilesh Jasani: Snap out of the DeepSeek delusion and invest big in basic research

It will also serve the purpose of data sovereignty and assure us minimal AI self-sufficiency. Given our expertise with frugal engineering, we can build a model that drastically reduces the cost of building applications. We need it to be citizen-centric, designed to serve over 1.4 billion Indians, much like our digital public infrastructure (DPI).

There are challenges. Building a slew of AI models will still be resource intensive. India will face AI talent and research challenges. We will still be dependent on the US for GPUs and chips. Also, competing with global models will be difficult. But we have taken similar strides before. The government, academia and industry jointly created our DPI, for example, that has led to a digital transformation.

AI is now a national mission, just as DPI was. Or the Indian Space Research Organisation’s space endeavours, or even the green revolution in farming. India set world-beating precedents in each of these and can do the same with AI. An indigenous AI model is a good start, but it will just be the beginning.

The author is a founder of AI&Beyond and the author of ‘The Tech Whisperer’.