India's AI strategy: We should not invest in the whole stack but focus instead on areas of expertise

The question for India is: where to place its bets? The complexity of the AI stack makes this harder. Policymakers must think simultaneously about chips, data centres, AI models, applications and the industries that will use them.

The US and China account for 75% of AI firms valued at more than $50 million and also have the most diversified AI industries. A correlation between diverse AI ecosystems and market value may seem a reason to increase AI stack presence using policy levers.

But India is far from the AI frontier. Perhaps the best way to think about an enduring comparative advantage, oxymoronic as the concept may be today, is to consider existing rather than future capability as a starting point.

From this perspective, national spending on subsidizing AI infrastructure, through graphic processing units (GPUs) or otherwise, will not create meaningful capability; but focusing on the knowledge and IT layer might.

Data centres, for example, are highly vulnerable and reliant on the types of chips on offer, and don’t substitute for manufacturing capability at the frontier. G42, one of the biggest data centre companies of the UAE, decided to drop Chinese hardware in 2024 rather than face a ban on the use of American chips.

This showed the dependence of data centres on elusive semiconductor technology. Also, promising large language model (LLM) companies like Sarvam recognize the need to scale up continually to keep up with an advancing frontier.

Google Deepmind research shows that pushing the frontier requires ever-larger training datasets and compute power for post-training inference. Not to mention the costs of developing this technology safely.

In other words, data centre capacity and resource-constrained LLM development do not determine AI capability. However, the use of AI to reinforce comparative advantages in knowledge services like consulting and finance, with IT at the core, may just prove to be.

This will be especially so if the world remains reliant on humans to support humans and transform ideas into execution. Handholding is what India excels at. For instance, armies of accountants are on call for personalized services to micro, small and medium enterprises (MSMEs) like mine, and are likely to remain in demand despite multinational competitors.

But if knowledge services are where India’s comparative advantage lies, policymakers need to recognize the real resource trade-offs involved in reinforcing it. Instead of spending limited resources on subsidizing access to chips, India should support services.

Besides, anecdotal evidence shows many of India’s 38,000-odd subsidized GPUs are busy servicing global demand, since there simply isn’t enough domestic usage and local users are asking for hefty discounts vis-à-vis hyper-scalers.

The complexity of AI shouldn’t overwhelm policymakers. What they can be certain of is that the challenges of competitiveness and job creation can be addressed by building on India’s existing services capability stack. Radical direct tax exemptions for small firms that absorb and skill new knowledge workers could be a useful place to begin.

Today’s knowledge-service firms like mine are increasingly in a bind about whether to hire new talent or unleash AI’s efficiencies in their business. But this needn’t be a binary choice if we have the requisite abundance and incentives to do both. And there is no better way to achieve this than higher cash flow. India needs a fresh wave of services growth, fuelled by a combination of AI and new talent with high-levels of exposure to global markets.

For instance, India would do well to anticipate the need for enterprise AI to be trusted, which would require knowledge workers embedded in the relevant AI supply chains. The world will still need humans to certify all manner of enterprise AI, starting with closely regulated industries.

RBI already wants its regulated entities like banks to ensure they know the ins and outs of the AI systems they deploy. This is easier said than done for an industry steeped in legacy systems.

And while global audit standards such as the ISO’s 42001, subjective as they may be, are now available as starting points for trusted deployments, there’s no audit workforce to speak of that can actually carry out the task of reviewing over 10,000 regulated institutions just in the domestic financial ecosystem.

Policymakers may find it tempting to double down on India’s current strategy of acquiring capabilities throughout the AI stack. But the pace of change requires that they let the market find its way and allocate resources based on proven capabilities.

By the time India produces usable AI chips or world-beating models, frontier capabilities would look exponentially more daunting than they seem today. A temporally grounded perspective is the need of the hour.

The author is a policy expert at Koan Advisory Group.