The global AI contest has become a scale spectacle. India's advantage lies somewhere else entirely.

The current AI race is being scored like a construction project. Who stacked the most GPUs. Who trained the largest model. Whose data center campus covers the most acreage and draws the most megawatts. The United States and China are committing tens of billions of dollars to mega-clusters, and the scoreboard rewards the spending itself as much as anything it produces.

For India, the tempting question is whether to enter that contest on the same terms.

The answer is no. Not out of modesty, and not because the ambition is misplaced. It's because the terms of that contest are set by an abundance of capital and energy that India does not have, and the contest measures something India does not particularly need to win. India's advantage lies in frugal AI: lean, efficient, purpose-built systems that solve real problems without mega-clusters or the theatre that surrounds frontier-model releases.

The compute trap

Chasing raw scale is a losing game for a capital- and energy-constrained economy. Four constraints make this concrete.

• Capital intensity: Training a frontier model runs into the hundreds of millions of dollars in compute alone, before talent, tooling, and physical infrastructure. That number is a floor, not a ceiling, and it resets every eighteen months as the frontier moves. A country can spend that money once and find itself a generation behind by the time the run finishes.
• Import dependency: India has no domestic high-end GPU industry. Every accelerator in every planned cluster arrives through a supply chain that runs through a small number of foreign firms and is subject to export controls that can change with an administration. Building a national strategy on top of that is building on someone else's permission.
• Energy cost: Mega-clusters draw enormous, continuous power. India's grid cannot casually absorb that load, and every megawatt routed to a training run is a megawatt negotiated against industrial demand, household reliability, and the country's renewable transition targets. In the US, the energy question is a cost line. In India, it is a genuine trade-off.
• Diminishing returns on the actual work: This is the constraint that matters most and gets discussed least. Vernacular voice interfaces, crop advisories, claims triage, fraud detection, document understanding for a state welfare department — almost none of these need frontier-scale parameters. Past a certain point, additional scale buys marginal accuracy on the specific task while multiplying inference cost on every single query. For a use case serving a hundred million people at near-zero revenue per user, that arithmetic decides everything.

What frugal AI actually means

Frugal is not a euphemism for underpowered. It describes a specific engineering discipline, and it has a real toolkit:

• Small, task-specific models trained or tuned for one job, rather than general-purpose models asked to do everything adequately.
• Edge and on-device inference, so a system works on a mid-range phone in a district with unreliable connectivity, and costs nothing per query once deployed.
• Data-efficient training — transfer learning, distillation, quantization, LoRA-style adaptation — which turns a compute problem into a technique problem.
• Building on open-weight foundations instead of training from scratch, which converts a $200 million capability into a $200,000 one.
• Ruthless scoping, which is the least glamorous item on the list and probably the most valuable. Most deployed AI failures are not model failures. They are problem-definition failures.

None of this is a compromise position. It is the same instinct that produced Mangalyaan, which reached Mars orbit in 2014 on a budget smaller than the marketing spend of a Hollywood film about space. It is the instinct behind Indian cardiac stents priced at a fraction of their imported equivalents, and behind vaccine manufacturing that supplies a large share of the world's doses because someone solved for cost per dose rather than margin per dose.

Jugaad has always been a slightly patronising word for this. What it actually describes is engineering under constraint, and constraint is a design input, not a handicap.

The objection, and the honest answer

The obvious criticism: isn't this a rationalisation for falling behind? If frontier capability keeps improving, doesn't a frugal strategy just lock India into permanent dependence on models built elsewhere?

It's a fair challenge, and the honest answer has three parts.

First, frugal AI depends on open-weight models continuing to exist and continuing to close the gap with closed frontier systems. That is a genuine strategic risk, and it argues for India investing in and supporting the open ecosystem rather than treating it as free infrastructure.

Second, this is not an argument for zero frontier investment. It is an argument about proportion. A modest sovereign compute capability for research, security-sensitive workloads, and Indic-language pretraining is defensible. A national strategy built around out-spending Nvidia's largest customers is not.

Third, and most importantly, the frugal path is not the consolation prize. India's constraint set — cost-sensitive users, thin margins, patchy connectivity, dozens of languages, low willingness to pay — is the constraint set of most of the world's population. Systems that work under those conditions are exportable to Southeast Asia, Africa, and Latin America in a way that a bigger cluster in Hyderabad never would be. Solving for the hardest deployment environment produces something other people want to buy.

The metric that matters

The real AI race is not about who has the most GPUs. It's about who delivers the most value per rupee, per watt, and per person reached.

Those are three separate denominators, and they don't reward the same behaviour that a leaderboard does. Optimising for them looks unimpressive in a press release and considerably better in a district hospital, a mandi, or a mid-tier bank's fraud queue.

India is well positioned to compete on those numbers. It is poorly positioned to compete on cluster size. Choosing the first contest is not settling. It is picking the fight worth winning.