AI Infrastructure in India: Compute, Data Centres, Chips and the IndiaAI Mission

The AI stack

• Training a large AI model needs thousands of GPUs running for weeks. A GPU is a chip that does many calculations at once.
• Those GPUs sit in data centres. These are large buildings with servers, power supply and cooling.
• Chips come from fabs, which are factories that make them. Models and apps are built on top.
• If any layer is foreign-owned, India depends on others for it. This is why the idea of sovereign AI is in the news.

For how models work, see Large Language Models (LLMs).

The basics

• The Cabinet approved the mission in March 2024 with an outlay of ₹10,371.92 crore over five years. (Source: PIB explainer, Dec 2025)
• It is run by IndiaAI, an independent business division under the Ministry of Electronics and Information Technology (MeitY).
• The aim is “Making AI in India and Making AI Work for India”.

The seven pillars

• Compute: affordable GPUs for all.
• Application development: AI tools for health, farming, climate and governance.
• AIKosh: a national dataset platform with over 5,500 datasets and 251 models by December 2025.
• Foundation models: India’s own large models.
• Future skills: support for PhD, postgraduate and undergraduate students, and data labs in smaller cities.
• Startup financing: funds for AI startups.
• Safe and trusted AI: work on bias, privacy and testing. (Source: PIB explainer, Dec 2025)

A public compute pool

• Buying GPUs is costly for a small startup. The mission therefore lines up GPUs from private cloud firms and lets users rent them.
• Over 38,000 GPUs have been onboarded. They are offered at a subsidised rate of about ₹65 per GPU hour. (Source: PIB explainer, Dec 2025)
• At the AI Impact Summit in February 2026, the government said 20,000 more GPUs would be added in the coming weeks. (Source: PIB, AI Impact Summit)
• Read about the summit in AI Impact Summit 2026.

Models built for India

• The mission got over 500 proposals for foundation models. Twelve startups and groups were picked in the first two phases.
• They include Sarvam AI, Soket AI, Gnani AI, Gan AI, Fractal Analytics, and BharatGen, a consortium led by IIT Bombay. (Source: PIB explainer, Dec 2025)
• The goal is models that understand Indian languages and Indian data. This is the “sovereign” part: India does not depend on a foreign model for public services.
• Example: Sarvam AI works with the UIDAI to make Aadhaar services easier to use with generative AI.

Supercomputers

• The National Supercomputing Mission (NSM) builds a network of high-performance computers in institutes across India. It is run by MeitY and the Department of Science and Technology, with C-DAC and IISc as implementers.
• The PARAM series is the Indian-made supercomputer line. PARAM Pravega at IISc, Bengaluru is described by PIB as the largest supercomputer in an Indian academic institution. (Source: PIB document, 2025)
• AIRAWAT is an AI-focused supercomputer at C-DAC, Pune. It ranked 75th on the global TOP500 list in 2023. (Source: PIB document, 2025)
• Weather systems named Arka and Arunika run at IITM Pune and NCMRWF Noida.
• For one system in detail, see PARAM Siddhi-AI.

Capacity and policy

• Industry estimates quoted by PIB put India’s cloud data centre capacity at about 1,280 MW. It may grow four to five times by 2030. (Source: PIB, Budget 2026-27)
• Union Budget 2026-27 proposed a tax holiday till 2047 for eligible foreign cloud firms that serve global customers from India-based data centres. (Source: PIB, Budget 2026-27)
• It also proposed a 15% safe harbour margin for related data centre entities. Domestic sales stay under normal tax rules.
• At the February 2026 summit, firms made big pledges. Reliance spoke of USD 110 billion over seven years. Adani spoke of USD 100 billion by 2035. Tata announced a partnership with OpenAI. (Source: PIB, AI Impact Summit)
• Pledges are not built capacity. Watch for actual megawatts switched on.

Making the chips

• The India Semiconductor Mission (ISM) started in 2021 with an outlay of ₹76,000 crore to attract chip fabs and packaging plants.
• A fab makes the chip on a wafer. An OSAT plant assembles, tests and packs finished chips.
• The Tata fab at Dholera, Gujarat is planned as India’s first commercial foundry. Reports differ on its start date, so wait for official word.
• Micron’s plant at Sanand, Gujarat is an assembly and test unit. Reports say it has begun shipping memory modules.
• Budget 2026-27 launched ISM 2.0 with ₹1,000 crore for FY 2026-27. It focuses on equipment, materials, chip design and skills. (Source: PIB, Budget 2026-27)
• Reports say about 12 manufacturing projects had been approved by mid-2026. Exact totals vary by source.

The hidden cost

• AI data centres use huge amounts of electricity. The servers also need cooling, which can use a lot of water.
• Power must be steady. A data centre cannot stop when the grid fails, so backup systems are needed.
• Using solar and wind power cuts the carbon cost. States compete to offer green power to data centre firms.
• Water-saving cooling and heat reuse are options to lower the load.
• The 2026-27 PIB note links data centres to energy systems and cooling as major costs. (Source: PIB, Budget 2026-27)

IndiaAI Safety Institute

• The IndiaAI Safety Institute works on the “Safe and Trusted AI” pillar. It follows a hub-and-spoke model with partner institutions.
• Thirteen projects were selected on topics like machine unlearning, bias, privacy and explainability. (Source: PIB explainer, Dec 2025)
• For the wider debate, see Ethics of Artificial Intelligence and The Paris AI Action Summit, 2025.

Side by side

• Stargate and InvestAI figures are announced targets, not money already spent. (Source: The Register) (Source: Euronews)
• India’s public outlay is far smaller. Its approach is to share compute widely and attract private money.

Four gaps

• Chip imports: advanced GPUs are still imported. Supply can be hit by export controls abroad.
• Power: steady, clean electricity at scale is hard to secure.
• Talent: India needs more researchers who can train frontier models.
• Data: good datasets in Indian languages are still thin. AIKosh is one fix.

Quick recall

• IndiaAI Mission: Cabinet, March 2024, ₹10,371.92 crore, seven pillars, under MeitY.
• Over 38,000 GPUs at about ₹65 per hour; 20,000 more announced in February 2026.
• AIKosh is the dataset platform; AIRAWAT is the AI supercomputer at C-DAC Pune.
• NSM builds a network of supercomputers; PARAM is the Indian-made series.
• ISM started with ₹76,000 crore; fab makes chips, OSAT packs them.
• Budget 2026-27: tax holiday till 2047 for foreign cloud firms using Indian data centres; ISM 2.0.

Newest first

• February 2026: Budget 2026-27 gives a tax holiday till 2047 for foreign cloud firms and launches ISM 2.0. (Source: PIB)
• February 2026: AI Impact Summit announces 20,000 more GPUs and large investment pledges. (Source: PIB)
• December 2025: PIB reports 38,000 GPUs under the IndiaAI Mission. (Source: PIB)