IndiaAI Mission's 12 sovereign LLM partners: who they are, what they are building, and what it means for Indian and UK builders

The selection — the full list of 12

Per statements in the Rajya Sabha and official MeitY communications, the IndiaAI Mission's Foundation Models pillar has designated the following 12 organisations as sovereign LLM development partners:

The mix is deliberate: three tiers of model scale (large multilingual LLMs, mid-size specialised models, small domain-specific systems), three institutional types (pure startups, academic consortia, established IT/enterprise firms), and coverage across key verticals (healthcare, education, manufacturing, finance, creative media).

The two anchor institutions: Sarvam AI and BharatGen

Sarvam AI

Sarvam AI is the most visible of the 12 partners. Founded in 2023 by Vivek Raghavan and Pratyush Kumar (both formerly at AI4Bharat / IIT Madras), the company has raised approximately US$41 million across seed and Series A rounds and received a direct compute allocation of 4,096 H100 GPUs from the IndiaAI national portal — one of the largest single allocations on the portal.

Sarvam's current model stack includes the Bulbul ASR speech recognition system and a family of Indian-language text models covering Hindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati, and Odia. Their stated goal under the Foundation Models pillar is to develop a 70B+ parameter multilingual model trained predominantly on Indian-language data — a capability that does not currently exist at that scale outside of closed US/Chinese labs.

BharatGen

BharatGen is a government-academic consortium led by IIT Bombay and spanning eight institutions, with a confirmed budget allocation of ₹988.6 crore — by far the largest single allocation among the 12 partners. The consortium's mandate is to build open-weight multilingual models on Indian datasets that can be freely used by any Indian organisation, effectively creating a public-good foundation layer beneath the commercial offerings of the startup partners.

The open-weight commitment distinguishes BharatGen from Sarvam: Sarvam's models are commercial (API access, enterprise licences), while BharatGen's outputs are intended to be released under open licences for the Indian research and startup ecosystem.

The compute picture: 38,231 GPUs at ₹65/hr

The 12 selected partners do not receive cash grants in the conventional sense — they receive preferential access to the national compute portal. As of March 2026, the portal has provisioned 38,231 GPUs across government-procured hardware, at a subsidised rate of approximately ₹65 per GPU hour compared to ₹330+ per hour on AWS, Azure, or GCP in India.

For context: training a 70B-parameter model from scratch at reasonable quality requires roughly 2–5 million GPU hours. At ₹65/hr, that is ₹130–325 crore in compute — a cost that would be prohibitive for any Indian startup at commercial cloud rates. The subsidised portal makes frontier model training economically viable for Indian organisations for the first time.

What this means for the startup ecosystem

The 12 selected partners create a new layer of infrastructure that other Indian AI startups can build on top of. The most immediately relevant implications:

• API access to Indian-language models — Sarvam and several others are expected to offer API access to their models, similar to OpenAI's API. Startups building vernacular applications (regional e-commerce, voice agents, EdTech in Indian languages) will have a domestic model alternative to GPT-4 class APIs with better Indian-language coverage and lower latency from Indian infrastructure.
• Open-weight models via BharatGen — The BharatGen open-weight models, when released, will let any Indian team fine-tune on their own data without API cost or data-sovereignty concerns. This is particularly relevant for healthcare, legal, and financial applications where data cannot leave the organisation.
• Domain-specific verticals — Intellihealth (healthcare), Shodh AI (education), and Zenteq (manufacturing/logistics) represent pre-built domain layers. Startups in those verticals should evaluate partnership before building from scratch.

The UK parallel: what this means for British builders

For UK-based AI builders, the IndiaAI Mission's 12-partner list is primarily a partnership map. Direct access to government funding is restricted to Indian entities, but commercial partnerships are open.

Sarvam AI and Fractal Analytics both have established international commercial operations. Fractal Analytics, in particular, has significant UK enterprise presence (financial services, FMCG, retail) and the Foundation Models pillar work is expected to feed new analytical LLM capabilities into their enterprise products.

UK startups evaluating South Asian market entry should note that the 12 partners represent both potential competitors (if you are building Indian-language AI) and potential distribution channels (if you have capabilities — evaluation tooling, safety testing, fine-tuning infrastructure — that the 12 partners will need).

For comparison, the UK Sovereign AI Fund's £500m programme and the new UKRI £40m Fundamental AI Research Laboratory represent the British government's parallel sovereign AI investments. The two ecosystems are complementary rather than competitive — India's focus is multilingual capability and data sovereignty; the UK's focus is reliability research and enterprise AI infrastructure.

What to watch over the next 12 months

• First model releases from BharatGen — the open-weight multilingual models from IIT Bombay consortium are the most consequential output to watch. Expected in H2 2026 or H1 2027.
• Sarvam's 70B announcement — if Sarvam achieves its stated goal of a 70B Indian-language model, it will be the largest open or commercial Indian multilingual model by a large margin.
• API pricing from startup partners — Gnani AI, Soket AI, and GenLoop are expected to launch commercial APIs. Their pricing relative to GPT-4o mini and Gemini Flash will determine whether IndiaAI-trained models become the default for vernacular Indian applications.
• Dataset platform — the IndiaAI Datasets pillar is building a national data repository of Indian-language text, audio, and imagery. Access policies for this repository will determine how quickly the 12 partners can improve their models and whether external startups can also train on it.