India must aim higher: cost optimization must not define the country’s tryst with artificial intelligence

The rest of us should be worried.

Costs have always been the ground on which vendors or firms compete, and that matters more in AI than their capability on benchmark tests. For the enterprise buyer trying to capitalize on latent demand for its services and boost productivity, raw frontier capability is not the binding consideration; it is the price per token.

Most firms do not need ‘Opus-grade’ reasoning to summarize a meeting, schedule executive meetings, handle customer service or run a basic agentic workflow. What they need is a ‘good enough’ tool that scales cheaper than hiring labour for those tasks.

This was China’s playbook in manufacturing; it has been brought to AI. Chinese manufacturers scaled up by supplying ‘good enough’ products at floor prices, eventually pushing competitors out of the market. Now, they are releasing capable models for free, setting the price floor and letting the market do the rest.

Even those who won’t deploy Chinese models for security reasons will use DeepSeek’s pricing to negotiate with Western providers. The pricing pressure will do its work.

These developments matter for India in ways the price tag on foreign AI services don’t capture.

India keeps its markets relatively open and does not ban services from any particular country, giving buyers a wide range of choice. With Chinese models getting cost-effective, it is likely that India’s private firms will reach for the most cost-effective option, making ‘indigenous’ options unattractive.

The incentive for Indian firms to invest in foundational AI research, already thin given the stiff competition they face from Western model providers with deep pockets, fades further with each release. If the world’s cheapest capable model is open-weight and Chinese, the business case for an Indian foundation model fails even before it can be made.

Few venture investors will fund a firm building a domestic frontier model when an almost-frontier alternative can be downloaded free. The talent that might have done that work will leave for firms abroad that are still investing in pre-training, or will get redirected to building applications on top of someone else’s model. Neither produces the foundational expertise that could turn a country from a customer into an AI leader.

The deeper cost is harder to see on a profit-and-loss statement. Free trade and the pursuit of cheap goods delivered enormous benefits, but they also concentrated ‘implicit knowledge’ in the places that did the work.

This is the kind of know-how that most don’t appreciate. It is the engineer who can tell by feel how many millimetres the drill needs to be adjusted, the team that has debugged software at scale enough times to predict where it will fail, the operator who has maintained a fab clean-room long enough to know which procedures cannot be cut. You acquire this by doing the work repeatedly at scale, not by reading about it.

Manufacturing know-how has not settled in China by accident and today there is barely a domain of modern manufacturing where China does not command both underlying knowledge and a cost advantage. Reshoring out of strategic necessity is not impossible, but it takes far longer than the timelines that policy incentives assume because a knowledge base has to be built by people who have never done such work. In that gap, the lead only widens. The same logic now applies to AI.

There is also no ecosystem effect for the importer. In complex systems, one capability seeds another. A country with foundation model labs develops a downstream pipeline of fine-tuning specialists, evaluation researchers, agent infrastructure firms, safety teams and inference optimisation startups, because each of these depends on engineers close to the actual work. If we strip out the foundation, the chain does not even begin to form. We end up with the integration layer alone, forever a customer.

The economics of that position has been kind to us in the past. It is unlikely to be as kind in an era of shifting geopolitics, though.

AI differs as an arena from manufacturing, which is why we should act now. Manufacturing shifted bases over decades. The AI stack is still being built, applications are being defined and implicit knowledge is being formed.

The window in which a country can take a strategic stance, before dependence becomes permanent, is narrow. Tooling, agent frameworks, evaluation suites, fine-tuning pipelines and research infrastructure get built around whichever models become the default.

Once Chinese open-weight models run inside India’s enterprises, the surrounding ecosystem will grow around them and switching costs will become structural rather than financial.

It is time to let go of the notion that everything must be optimized solely on the basis of costs and short-term profitability. It may be ‘cheap’ in the short run, but can prove costly in the long run. Intangibles such as implicit knowledge and strategic indispensability don’t show up in quarterly reports, but their benefit to firms and the country create long-term dividends.

The private sector needs to recognize this.

Firms foraying into AI need to raise their research and development budgets and focus on building foundational AI capabilities rather than relying on imported solutions. Such investments need to be measured not against profit statements but rather against everything India may not be able to do in 10 years if we optimize only for what we monetarily save today.

These are the authors’ personal views.

The authors are, respectively, chief economic advisor to the Government of India, and consultant, ministry of finance.