India's Cheap-Compute Gambit: 38,000 GPUs at a Third of Global Prices

The numbers behind the IndiaAI Mission read like a procurement footnote: more than 38,000 high-end GPUs onboarded, available to startups and researchers at ₹65 (about $0.78) per hour — roughly a third of the global average — plus 1,050 TPUs added to the national pool. But buried in that footnote is the most distinctive AI industrial strategy of any major economy, and this week's inauguration of the CG Semi chip assembly plant in Sanand makes it worth taking seriously as a system.

A Different Theory of the Race

The US strategy is frontier-first: concentrate capital in a handful of labs and let capability trickle down. China's is champion-first: state capital behind national models and domestic silicon. India is running a third experiment — access-first. Subsidise compute until it's nearly free, spread it across 12 selected model-builders (Sarvam AI, Gnani, BharatGen, Fractal and others), fund 500 PhD scholars and thousands of graduate researchers, push AI labs into Tier-2 and Tier-3 cities, and let the ecosystem decide what to build.

The theory has an honest premise: India will not out-spend OpenAI or out-subsidise Beijing. Total IndiaAI compute is a rounding error against a single US hyperscaler's quarterly capex. What India has instead is the world's largest pool of English-speaking engineering talent, a billion-user digital public infrastructure stack (Aadhaar, UPI, ONDC) that AI can plug into, and domestic demand across 22 official languages that no foreign model serves well.

The Dependence Problem

June's Fable 5 export-control episode gave the strategy its sharpest argument yet. When Washington suspended Anthropic's frontier model for 19 days, Indian startups and enterprises discovered overnight that their AI roadmaps had a single point of failure in another country's Commerce Department. TechCrunch's reporting captured the mood: a national debate about whether India's AI future can be rented.

The answer emerging from Delhi is layered self-reliance rather than autarky. Rent the frontier where it's cheapest, build sovereign capability in the layers that matter most — Indic-language models, government-workflow AI, and now, with Sanand's OSAT line running, the unglamorous back end of the chip supply chain. Nobody in Delhi pretends assembly-and-test is fab-grade sovereignty. But it is a foothold, and India's semiconductor strategy has been consistent about sequencing: packaging first, mature nodes next, advanced fabs last.

The Honest Scorecard

Can it work? The risks are real. Cheap compute without frontier-scale clusters produces fine-tuned derivatives, not breakthroughs — none of the IndiaAI model projects is training anything near the frontier. Talent leakage continues: the researchers India educates still ship for San Francisco salaries. And subsidised GPU hours have a way of becoming a permanent entitlement rather than a bridge.

But the strategy's ceiling may matter less than its floor. Even in the pessimistic case, India ends up with tens of thousands of AI-fluent engineers, working Indic-language models embedded in government services, and a startup ecosystem whose unit economics were bootstrapped by public compute. The optimistic case is more interesting: that the next hundred million AI users — voice-first, vernacular, price-sensitive — look nothing like the customers US labs are building for, and the ecosystem closest to them wins that market by default.

The US is betting on capability. China is betting on control. India is betting that in a decade, distribution beats both. It's the cheapest bet on the table — which is precisely the point.