India must forge its own AI path amid a foundational tug of war
Summary
The country would do better by focusing on developing AI applications for specific purposes than by playing catch-up with large foundational models. As with UPI, we could lead the world on making artificial intelligence work at scale for large numbers.
It’s no secret that there’s a worldwide tug of war for artificial intelligence (AI) supremacy among leading economies. America has the upper hand by default, China is a disruptor par excellence, France has jumped in and India wants a seat at the top table, too.
Amid all of this, there’s one aspect that keeps coming up every now and then—foundational models. In February, Ashwini Vaishnaw, the Union information technology minister, said that India will build and showcase its own foundational AI models by the end of this year. Many took this as a knee-jerk reaction to the hype around Chinese upstart DeepSeek’s release of open-for-all AI models. But, all of this begs a question: Does India really need to focus its investments and efforts on foundational models?
Also Read: Nilesh Jasani: Snap out of the DeepSeek delusion and invest big in basic research
Tracing the AI story back to OpenAI’s 2022 launch of ChatGPT, the world’s first mainstream and popular Generative AI application, one can see why foundational models are pivotal. They are incredibly complex in nature and are trained on vast sets of data to make them generally good at nearly everything. But foundational models have a pitfall—they’re not really ‘excellent’ at anything.
Over the years, advancements in AI research have shown us that while foundational models form the backbone of a GenAI applications, their biggest drawback is their lack of excellence in particular subjects. As a result, algorithms would get their data sources jumbled up and create responses that are often hallucinatory in nature.
Now, more than two years later, we know a little more about how the modern contours of AI work. It also means there’s a case to be made that as far as India’s sovereign AI push is concerned, investing the biggest chunk of available capital in building foundational models may not be necessary.
What India’s race to the top of the AI industry really needs are high-quality data-sets and an innovative set of startups building second-layer or smaller AI models based on large-scale foundational models. This approach offers us multiple advantages.
Foundational models require huge amounts of compute power. Even when subsidized, the use of such compute capacity requires long stretches of time and subsequent testing to prove itself against the likes of OpenAI’s GPT, Google’s Gemini and Meta’s Llama.
Instead, using open-source foundational AI models as the basis for building smaller custom models can help startups across India collaborate with large institutes, procure targeted data and deliver applications that are purpose-built to serve specific tasks. Project Vaani is one such innovative name that comes to mind, for instance; it leverages a significant trove of speech data to create applications such as automatic speech recognition in Indic languages.
Also Read: India must wake up on basic R&D for technology before it gets too late
The advantage here is that by doing so, Indian startups can start ramping up the country’s presence in the overall AI ecosystem. With foundational AI, we would be looking at long turnaround times of development before Indian models could gain global attention.
There are two key points on the other side of the argument: One, that foundational models will give India its own AI architecture that is not reliant even on an open-source framework created by any entity in the West; and two: just building applications to sell to the world could lead us to the same domestic inadequacy risks as seen before in fields like semiconductors, electronics manufacturing and even technology services.
But foundational models may not be the answer to those concerns. Instead, purpose-built AI applications that address focused pain points hold high potential.
Take, for instance, Silicon Valley’s Anysphere—whose AI code-editing software, Cursor, is a great success story. Within 12 months, this little-known service leapt to popularity, reporting an annualized revenue run-rate of $100 million. In February, it bagged a $105-million funding round from investors that included Andreessen Horowitz, among others, to scale its service. A fellow startup, Codeium, has also reached an impressive valuation of $2.85 billion. Both these startups are less than four years old.
Taking cues from these success stories, it is plausible to argue that instead of investing $100 million in building a foundational AI model from scratch, using the fund to incentivize 10 projects implementing AI applications across various segments could lead to greater results for India. As these applications scale up, India could showcase its success with on-ground AI implementation. Much like India’s success with the Unified Payments Interface (UPI), the country could lead the way in making AI work at scale.
Also Read: Jaspreet Bindra: The K-shaped trajectory of AI offers India a big opportunity
Do note, though, that developing AI applications for India is not a one-dimensional journey. To achieve success, we must fire on all cylinders in every aspect of the challenge. But it is essential to work out which part of the AI development journey we must put more capital behind and do so as a matter of priority.
Even as a global tug of war intensifies for foundational AI supremacy, India can show the way ahead by developing applications that leverage this nascent technology—and build for India to scale for the world.
The author is partner, Accel.
