IndiaAI Mission Crosses 38,000 GPUs, Targets a Sovereign LLM

Why in News?

By late 2025 the IndiaAI Mission‘s shared Common Compute facility had crossed 38,000 GPUs, making large-scale AI training power available to Indian startups, researchers and academia at heavily subsidised rates — around Rs 65 per hour. The Mission, run by the Ministry of Electronics and IT (MeitY) through its IndiaAI Independent Business Division, is the country’s flagship push for AI self-reliance.

The headline goal is a sovereign Indian LLM — a homegrown large language model trained on India-specific data and the country’s many languages, rather than depending on foreign foundation models. Several Indian firms have been selected under the Mission’s Foundation Model pillar to build these indigenous models.

• Common Compute pool crossed 38,000 GPUs by late 2025, up from about 34,000 GPUs in mid-2025, built with empanelled industry partners.
• GPU access is subsidised for startups, researchers and academia at roughly Rs 65 per hour after government support on the commercial rate.
• Approved in 2024 with an outlay of Rs 10,371.92 crore (over Rs 10,300 crore) spread across the Mission’s pillars.
• Sarvam AI was selected to build India’s sovereign LLM ecosystem with an open-source 120-billion-parameter model; Soket AI, Gnani AI and Gan AI were later picked to build further indigenous foundation models.
• The AIKosh platform hosts datasets and AI models — over 3,000 datasets and hundreds of models — as shared public infrastructure for builders.
• IndiaAI FutureSkills has supported more than 13,500 scholars, with the AI talent pool targeted to reach about 1.25 million professionals.

The development matters in the context of:

• Compute power, data and talent are the three scarce inputs for frontier AI — and India had little sovereign capacity in any of them before this Mission.
• A national LLM trained on Indian languages and contexts is framed as both a strategic-autonomy and a digital-public-goods question, not just a commercial one.

UPSC Relevance

Prelims Relevance

• IndiaAI Mission — approved 2024, outlay Rs 10,371.92 crore, implemented by IndiaAI (an Independent Business Division under MeitY)
• Seven pillars — Common Compute, Foundation Models, AIKosh datasets, application development, FutureSkills, startup financing and Safe and Trusted AI
• Common Compute — shared GPU pool that crossed 38,000 GPUs; subsidised access (~Rs 65/hour)
• AIKosh — national platform of datasets and AI models for developers
• Sarvam AI — selected to build the sovereign LLM ecosystem (open-source 120-billion-parameter model)
• BharatGen — government-funded multimodal LLM supporting Indian languages; Bhashini — language-translation platform
• A GPU (Graphics Processing Unit) is the hardware that does the parallel maths AI models need for training and inference
• Difference between training a model and running inference on it
• Foundation model and large language model (LLM) — meaning of ‘parameters’

Mains Relevance

GS Paper 3

• Building sovereign AI capability — compute, data and talent as strategic infrastructure for self-reliance.
• Public investment in shared GPU infrastructure as a way to democratise access and lower the entry barrier for startups and researchers.
• Indigenous foundation models trained on Indian languages and data, and what ‘sovereign LLM’ means for technology security.

GS Paper 2

• AI governance — the Safe and Trusted AI pillar, responsible-use frameworks and the state’s role in steering an emerging technology.
• Digital public infrastructure and equitable access to technology as a development and welfare question.

Essay

• Technology should not be left in the hands of a few — democratising the tools of the AI age.
• Self-reliance in the age of artificial intelligence: who controls the compute controls the future.

Background and Context

What the IndiaAI Mission is

The Mission is India’s umbrella programme to build a full AI ecosystem rather than a single product.

• Approved in 2024 with an outlay of Rs 10,371.92 crore, it is implemented by IndiaAI, an Independent Business Division under MeitY.
• Its stated aim is to democratise AI access, foster technological self-reliance and ensure the ethical, responsible use of AI.
• It is built around seven pillars — Common Compute, Foundation Models, AIKosh datasets, application development, FutureSkills, startup financing and Safe and Trusted AI.
• The design treats compute, data and talent as shared public infrastructure that any Indian builder can tap, not just large firms.
• Union Minister Ashwini Vaishnaw has framed it around the idea that technology should reach a wide section of society, not a privileged few.

The compute pillar — crossing 38,000 GPUs

The most concrete achievement is a large, shared pool of AI compute power.

• A GPU is the specialised processor that performs the massive parallel calculations AI training and inference demand; frontier models need thousands of them.
• The Common Compute pool rose from about 18,400 empanelled GPUs to roughly 34,000 in mid-2025, then crossed 38,000 GPUs by late 2025.
• Capacity is built with empanelled industry partners — Indian data-centre and cloud firms such as Yotta, Sify, Netmagic, Cyfuture, Locuz, Ishan Infotech and Vensysco.
• Access is heavily subsidised — startups, researchers and academia pay around Rs 65 per hour after government support on the commercial rate.
• This shared model lowers the single biggest barrier to AI — the prohibitive cost of owning GPU clusters.

The push for a sovereign Indian LLM

The Foundation Models pillar funds Indian firms to build large models trained on Indian data.

• A foundation model is a large model trained on broad data that can be adapted to many tasks; a large language model (LLM) is the text-and-language version.
• Sarvam AI was selected to build India’s sovereign LLM ecosystem, developing an open-source 120-billion-parameter model for governance and public-service use.
• Three more firms were added — Soket AI (a 120-billion-parameter open-source model for India’s linguistic diversity), Gnani AI (a 14-billion-parameter voice model) and Gan AI (a 70-billion-parameter multilingual model).
• The goal is models trained on India-specific data and many Indian languages, reducing reliance on imported foreign models.
• Earlier efforts like BharatGen (a government-funded multimodal LLM) and Bhashini (language translation) feed the same ambition of language sovereignty.

Data and skills — AIKosh and FutureSkills

Models are only as good as the data and the people behind them, so two pillars target both.

• AIKosh is a national platform pooling datasets and AI models — over 3,000 datasets and hundreds of models — as open building blocks.
• Curated, high-quality Indian-language and domain datasets are essential to train models that work for Indian contexts.
• IndiaAI FutureSkills has supported more than 13,500 scholars across AI courses and labs.
• The Mission aims to grow India’s AI talent pool toward roughly 1.25 million professionals.
• Together these tackle the data and talent bottlenecks that compute alone cannot solve.

Applications and Safe and Trusted AI

Alongside infrastructure, the Mission funds real-world applications and a governance pillar.

• More than 30 India-specific AI applications have been approved across healthcare, agriculture and cybersecurity.
• A joint hackathon with the Indian Cyber Crime Coordination Centre (I4C) produced AI tools to classify cybercrime complaints on the national reporting portal.
• The Safe and Trusted AI pillar funds work on responsible-use frameworks, bias and safety standards.
• Startup financing support is meant to take indigenous AI from research into deployable products.
• This breadth is meant to make AI useful for India — solving local problems — not just technically impressive.

Why sovereignty matters here

The framing of ‘sovereign’ compute and models is strategic as much as commercial.

• Depending on foreign GPUs and foreign LLMs creates supply, cost and policy dependence on a handful of overseas firms and governments.
• Models trained abroad may under-represent Indian languages, contexts and values.
• Owning the compute, data and models stack is treated as a question of strategic autonomy, akin to energy or defence.
• Subsidised shared compute also addresses equity — letting small teams and universities build, not only big corporations.
• The Mission positions India to be a builder, not just a consumer, in the global AI economy.

Way Forward

Scale compute without dependence

• Keep expanding the Common Compute pool while moving toward domestic chip and data-centre capability.
• Keep access genuinely affordable so startups and universities — not only large firms — can train serious models.

Build models that work for India

• Fund the selected firms to deliver usable Indian-language foundation models, not just demonstrations.
• Strengthen AIKosh with high-quality, well-governed Indian-language and domain datasets.

Govern as you build

• Pair the build-out with a working Safe and Trusted AI framework covering bias, safety and accountability.
• Convert FutureSkills numbers into deep talent so India can both build and audit advanced AI.

Conclusion

Crossing 38,000 GPUs is a real milestone, but the IndiaAI Mission’s larger bet is structural: treat compute, data, talent and foundation models as shared national infrastructure rather than the property of a few firms. Subsidised access and government-funded indigenous models are meant to let any Indian team build serious AI.

The test now is delivery. Empanelled GPUs and selected firms must turn into deployable Indian-language models, a working safety framework and a deep talent base. If that happens, India shifts from consumer to builder in the AI economy; if the sovereign LLM and the governance scaffolding lag the hardware, the country risks owning the compute while still importing the intelligence.

UPSC Practice Questions

Prelims MCQ 1

With reference to the IndiaAI Mission, consider the following statements:

1. It is implemented by IndiaAI, an Independent Business Division under the Ministry of Electronics and IT.
2. Its Common Compute pillar provides a shared pool of GPUs for AI training and inference.
3. AIKosh is the Mission’s pillar dedicated solely to subsidising GPU access.

How many of the above statements are correct?

(a) Only one (b) Only two (c) All three (d) None

Answer: (b) Only two

Explanation:

Statements 1 and 2 are correct. Statement 3 is wrong: AIKosh is the platform for datasets and AI models; subsidised GPU access falls under the Common Compute pillar.

Prelims MCQ 2

In the context of AI, what does the term ‘GPU’ most directly refer to?

(a) A government policy unit for AI regulation (b) A processor that performs the parallel computation needed to train and run AI models (c) A dataset repository for machine learning (d) A type of large language model

Answer: (b) A processor that performs the parallel computation needed to train and run AI models

Explanation:

A GPU (Graphics Processing Unit) does the massive parallel maths that AI training and inference require; clusters of thousands of GPUs underpin large foundation models.

UPSC Mains Questions

1. Compute, data and talent are the strategic inputs of the AI age. Examine how the IndiaAI Mission seeks to build sovereign capability in each, and discuss why public investment in shared GPU infrastructure matters for an emerging economy.
2. What does a ‘sovereign large language model’ mean for India, and why is it being pursued? Discuss the opportunities and the risks of building indigenous foundation models trained on Indian languages and data.
3. Discuss the governance challenges of a state-led artificial-intelligence push. How should a Safe and Trusted AI framework balance innovation, safety and accountability?

Sources: PIB, Ministry of Electronics and IT (MeitY) and DD News.

Frequently Asked Questions