IndiaAI Mission GPU compute: meaning, how it works in India

The IndiaAI Mission is a national programme under the Ministry of Electronics and Information Technology that aims to build computing power, models and skills for artificial intelligence in India. The Cabinet approved it in March 2024 with an outlay of ₹10,371.92 crore over five years.

Its compute pillar is the part most startups meet first. The idea is simple: AI training needs many GPUs, which are expensive, so the government arranges a shared pool and lets approved users pay less than the market rate.

The government invites private data centre and cloud companies to bid. Those that qualify are empanelled, which means listed as approved suppliers, and they offer GPUs at bid rates through an IndiaAI compute portal. Reports from 2025 counted 34,000 or more GPUs from 14 empanelled providers, including E2E Networks, Yotta, Tata Communications, CtrlS, NxtGen and Sify. Counts differ by source and change as bids are added.

A user applies on the portal, describes the project and, if approved, is allotted compute. Payment goes to the provider at the empanelled rate, and the government subsidy reduces the user's share.

The pool has three uses besides startups: researchers, public sector bodies and the teams building foundation models. Startups share it with these groups, which is part of why allotment is reviewed case by case.

Early reports said the average price was around ₹65 per GPU-hour, with an H100 at about ₹92, and the minister cited a figure near ₹67 an hour. Later reports based on bids and the portal calculator show higher numbers, such as roughly ₹115 for standard GPUs and ₹150 or more for high-end ones. One late-2026 report put an H100 at ₹153 an hour on demand and ₹117 on a 12-month booking, before GST.

These figures come mostly from secondary sources and differ by date and GPU type. Parliament records describe a subsidy of up to 40% of compute cost for eligible users. As of November 2026, the live portal is the only reliable price list.

• Cabinet approval: March 2024.
• Outlay: ₹10,371.92 crore over five years.
• Empanelled providers: about 14 to 15, depending on the report.
• Subsidy: up to 40% of compute cost for eligible users, as reported.
• Price: varies by GPU, from under ₹100 to above ₹150 per GPU-hour in different reports.

Eligible users include startups, researchers, academic institutions and government bodies working on AI. A startup usually has to show a project and a plan for using the compute. Allocation is by request and approval, not a fixed sticker price, so the cost can differ from the headline rate.

Startups should check whether they are recognised by DPIIT and what the portal asks for. The guide on the Startup India scheme explains recognition.

A foundation model is a large AI model trained on broad data, which can be adapted to many tasks. The Mission selected teams to build Indian ones. Sarvam was the first, chosen in April 2025. Soket AI Labs, Gnani.ai and Gan.AI followed, and in September 2025 eight more were added, including BharatGen, Tech Mahindra and Fractal Analytics, according to Inc42 and Business Standard. Twelve teams in all are reported.

Each team gets access to compute and support linked to its plan. Reports differ on the sizes of the planned models, so a figure in a headline should be checked against an official announcement.

Training costs are one of the largest bills for an AI startup. Cheaper GPUs can cut cash burn and stretch a funding round. The Mission also gives Indian-language and India-specific models a public base that startups can build on.

It does not remove the need for private capital. Compute is only one cost, and a startup still needs customers, data and talent. Investors also check whether a team is using the subsidy to train a model that has a market.

Subsidised compute is only one piece of an AI business. Reports have raised questions about how many GPUs are actually in use, how fast approvals move and whether smaller startups can compete with larger teams for allotments. Some of these points come from commentary and not from official data, so they are better read as open questions.

There is also a gap between the number of GPUs listed and the number allotted. A listed GPU is capacity that a provider has offered, while an allotted GPU is one a user has been cleared to use. Hardware moves fast as well, and a rate for one chip generation may not apply to the next.

For founders, the practical point is to compare the subsidised rate with private cloud offers, since a provider outside the Mission may match the price for a long booking. The right choice depends on the size, the length and the paperwork a team can handle.

Check whether a story talks about announced GPU counts or GPUs that are actually allotted. Check whether a price is a bid, a portal rate or a rate after subsidy. And check whether a company named is in the foundation model list or only a user of the portal.

Government programmes change terms, so this guide is general information. A professional should be asked about any specific application.