IndiaAI Mission's Compute Pool Crosses 38,000 GPUs at the World's Cheapest Subsidised Rates

The single most consequential number in India's AI strategy is not a model benchmark or a funding round. It is a price: the per-GPU-hour rate at which the IndiaAI Mission's common compute pool now rents accelerators to domestic developers. That pool has scaled past 38,000 GPUs, and the rates attached to them are, by design, among the cheapest subsidised AI compute available anywhere.

This is the infrastructure layer beneath every sovereign-model bet India is making. Without it, the funding announcements and open-weight releases would not pencil out.

The pool, the providers and the price

The IndiaAI common compute facility now aggregates more than 38,000 GPUs across ten empanelled providers, including Yotta, Jio, Tata Communications, E2E Networks and CtrlS. The structure is an aggregation model: rather than the state building and operating one national supercomputer, MeitY empanels qualified commercial providers, who contribute capacity to a shared pool that approved users can draw on at controlled rates.

Those rates are the lever. Standard access is priced at ₹115.85 per GPU-hour, with H100-class accelerators at around ₹150 per GPU-hour. For projects designated as being of national importance, a subsidy of up to 40% drives the effective rate below ₹100 per GPU-hour. Set that against commercial hyperscaler pricing of roughly $2 to $4 per GPU-hour on AWS or Azure, and the gap is an order of magnitude. A startup or research group training on the IndiaAI pool pays a small fraction of what it would on global cloud.

The roadmap is to push the pool to 100,000 GPUs by the end of 2026, which would make it one of the largest pools of subsidised AI compute under a single national programme anywhere in the world.

Why cheap compute is deliberate industrial policy

The instinct is to read the subsidy as a giveaway. It is more precisely an industrial-policy instrument, and the mechanism is worth understanding.

Frontier and near-frontier model training is gated by compute cost. Sarvam's two open-weight models were trained on a six-month allocation of 4,096 H100s; at commercial rates that bill alone would have run into the tens of millions of dollars, capital that an early-stage Indian lab would have to raise purely to rent hardware before writing a line of differentiated code. By collapsing that cost, the state lowers the capital barrier to frontier work and lets domestic talent compete on architecture and data rather than on balance-sheet size.

The empanelment design does double duty. It seeds demand for India's commercial data-centre operators, the same Yotta, CtrlS and Tata Communications building out capacity, while giving the government a pricing lever it would not have if it simply handed out cloud credits. The MeitY empanelment process sets the eligibility bar, and the national-importance subsidy tier lets policymakers direct the deepest discounts toward the projects they most want built: sovereign foundation models, Indian-language systems, and public-interest AI.

The honest caveat is dependence. The economics of India's sovereign-AI push currently rest on compute the state is subsidising. That underwrites the bets, but it also means the unit economics of the labs drawing on the pool are not yet proven at full market cost. The wager is that subsidised compute now buys durable capability, talent, models, an ecosystem, that outlasts the subsidy. On current evidence, with a unicorn-status sovereign lab already trained on this pool, the bet is being validated faster than sceptics expected. The cheap GPU-hour is not charity; it is the cheapest line item in an industrial strategy, and arguably the highest-leverage one.