India’s sovereign AI push is changing its job description. According to ET Government, the effort that began with building home-grown foundation models is now being judged on whether those models do useful work for Bharat, meaning citizens who use Indian languages, public services and low-cost devices. The story’s headline names a Ganesh Ram…, and the most prominent figure by that name in this field is Ganesh Ramakrishnan, the IIT Bombay professor who leads BharatGen. We could not read the full ET Government text, so the facts below come from other named outlets covering the same programme.
The shift matters because the first phase has largely shipped. Forbes India counted five new sovereign models at the India AI Impact Summit in February 2026, and the government has since lined up 20 model proposals for support under the IndiaAI Mission. What buyers, ministries and hospitals want now is evidence that a 17-billion-parameter model can summarise a patient visit or route a grievance in Marathi without sending data abroad.
This analysis walks through the numbers, the models, the compute behind them and the risks. It also sets India’s approach next to open-weight efforts elsewhere, such as Mistral’s open-weight Large 4 release, because the sovereign question is no longer unique to one country.
From Foundation Models to Applications: What the Shift Means
Foundation-model work is slow and expensive, and its output is hard for a non-specialist to judge. An application is different. A clinic either saves ten minutes per patient or it does not. A court clerk either finds the right clause or does not. By moving the conversation to applications, India’s AI planners are choosing a scoreboard that officials, voters and startups can all read.
The Federal’s interview with Ganesh Ramakrishnan frames the same turn. He argues that India has been a willing adopter of technology, with UPI as his example, and that the harder step is becoming a producer of the technology itself. In his telling, the country should treat AI models as inspectable systems, with documented data and observable behaviour, rather than opaque services rented from abroad.
That framing lines up with how the mission is described in official and press material. The IndiaAI Mission, launched in 2024 with a budget of Rs 10,371.92 crore, spans compute, application development, datasets, skills, financing and safe AI. Models are one pillar among several, and the application pillar is now the one drawing attention.
Who Is Ganesh Ramakrishnan and Why BharatGen Matters
BharatGen is described as India’s first government-funded initiative to build indigenous, multimodal, multilingual large language models. It launched on September 30, 2024, under the National Mission on Interdisciplinary Cyber-Physical Systems run by the Department of Science and Technology, and IIT Bombay leads the consortium. Ramakrishnan serves as its founding director and principal investigator.
An Academic Consortium With a Company Wrapper
The Federal reports that BharatGen is structured as a Section 8 company based at IIT Bombay, with roughly 60 full-time engineers and a consortium of nine academic institutions. That mix is unusual. Most national AI efforts either sit inside a ministry or hand the work to a single private lab. BharatGen tries to keep academic transparency while hiring like a product team.
Funding That Dwarfs Its Peers
Funding figures vary by date and by what is counted. Tribune India reported Rs 988.6 crore for BharatGen under the IndiaAI Mission. MediaNama later reported, citing an April 2026 Lok Sabha reply, that BharatGen received a total of Rs 1,058.52 crore, the highest among 12 funded organisations and about four times the next-highest allocation. The Federal puts combined Department of Science and Technology and IndiaAI backing above Rs 1,235 crore. We use the most conservative figure when a single number is needed.
The IndiaAI Mission Numbers Behind the Push
Public figures on the mission fall into four groups: money, compute, models and deployment. The table below collects the ones that named outlets have reported. Where sources disagree, the table shows the more conservative number and flags it.
| Metric | Reported figure | Source outlet |
|---|---|---|
| IndiaAI Mission budget (2024) | Rs 10,371.92 crore | Coverage of the mission |
| BharatGen total support | Rs 1,058.52 crore (earlier reported: Rs 988.6 crore) | MediaNama, Tribune India |
| Funded organisations (April 2026) | 12 | MediaNama |
| Model proposals identified for support | 20 (12 LLMs, 8 SLMs) | Business Standard, Hans India |
| GPUs onboarded | 38,000+ at Rs 65 per hour | Government press material, DQ India |
| Additional GPUs announced | 20,000 (announced Feb. 17, 2026) | BW Businessworld, EE Times |
| BharatGen trillion-parameter target | March 2027 | Analytics India Magazine |
Two details stand out. First, most of the money flows to compute rather than salaries, a point MediaNama makes about the 12 funded projects. Second, the 20-proposal list is a mix of large and small language models, which signals that India is not betting everything on a single giant system.
Compute: 38,000 GPUs at Rs 65 an Hour
Compute is the quiet engine of the sovereign strategy. Union IT Minister Ashwini Vaishnaw said at the India AI Impact Summit on February 17, 2026 that India would add 20,000 GPUs on top of the 38,000 already onboarded. Orders were due within a week and deployment was expected over about six months. Vaishnaw also said he expected more than $200 billion of investment to flow into the Indian AI ecosystem over two years, which is his projection rather than a committed figure.
The Rs 65 hourly rate for subsidised access is what makes the compute story relevant to applications. A startup that cannot afford a cluster can still fine-tune a 17B or 30B model for a district hospital workflow. Compare that with the capital race among US hyperscalers, where Big Tech’s $720B data center bet is draining free cash flow. India’s state-pooled model spreads cost across many small users instead of concentrating it in four balance sheets.
The trade-off is utilisation. Subsidised GPUs only create value if teams can schedule them, find data and ship products. Compute without datasets, evaluation suites and deployment partners ends up as idle racks. That is exactly why the mission’s emphasis is migrating from chips to use cases.
The 20 Sovereign Model Proposals: 12 LLMs and 8 SLMs
Business Standard reported in July 2026 that the government would support 20 indigenous model proposals, split between 12 large language models and eight small language models. Named projects include Sarvam AI’s 30-billion and 105-billion parameter models, Gnani.ai’s speech-to-speech model, BharatGen’s multilingual foundation models and Avataar AI’s video generation model.
MediaNama’s earlier list of the 12 funded organisations adds the rest of the field: Sarvam AI, Soket AI, Gnani AI, Gan AI, Avataar AI, the IIT Bombay consortium behind BharatGen, GenLoop, Zenteiq, Intellihealth, Shodh AI, Fractal Analytics and Tech Mahindra’s Maker’s Lab. The mix covers speech, video, health, science and enterprise. No single vendor owns the stack.
Small models deserve more attention than they get. An 8B model that answers in Hindi on a modest server is often a better fit for a state agency than a frontier system priced per token in dollars. The inclusion of eight SLM proposals suggests planners understand that Bharat’s constraint is cost and latency, not benchmark glory.
Param2, Sarvam and Indus: The Models Carrying Applications
Forbes India’s summit coverage gives the cleanest side-by-side of what actually shipped. The table summarises the five models it profiled. Performance claims come from the builders and have not been independently verified here.
| Model | Builder | Size | Focus | Reported detail |
|---|---|---|---|---|
| Sarvam LLMs | Sarvam AI | 30B and 105B (MoE, 10.3B active on the larger) | 22 languages, reasoning, real-time speech | Open-sourced Feb. 2026, linked to UIDAI Aadhaar services |
| Param2 | BharatGen | 17B Mixture of Experts | 22 Indian languages, public services | Sixfold step up from earlier 3B model |
| Vachana TTS, Inya VoiceOS | Gnani.ai | Not specified | Speech-native, 12+ languages | Voice cloning in under 10 seconds, MOS 4.23 |
| Project Indus (upgraded) | Tech Mahindra | 8B | Hindi-first education | About 500M synthetic training tokens, launched Feb. 20, 2026 |
| Vaidya 2.0 | Fractal Analytics | Not specified | Healthcare triage and diagnostics | Builder claims it beats GPT-5 and Gemini Pro 3 on HealthBench |
The Vaidya claim is the one to treat carefully. A domain model beating a general model on a single health benchmark is plausible, but it says little about safety in a real clinic. Procurement teams should ask for third-party evaluation on Indian patient data before they read much into it.
Sarvam’s deployment list is more concrete. Forbes India ties the company to UIDAI Aadhaar services, an Odisha hub with a 50MW footprint and a Tamil Nadu research park called Digital Sangam. Those are infrastructure and government commitments, which is a stronger signal than a demo.
Where Applications Are Landing: Health, Education, Governance, Voice
BharatGen’s launch summit introduced four model families: Shrutam for speech, Param for text, Patram for document vision and Sooktam for text to speech. Each maps to a public-service task. Speech recognition feeds helplines and clinics. Document vision reads forms, land records and court filings. Text to speech gives low-literacy users an audio route into the same systems.
The Federal reports one live example. BharatGen is working with Amrita Hospitals on MedSum, a pipeline that turns speech into text and then into a clinical summary, with patient privacy and consent rules built in. Coverage of the sovereign effort also points to legal and administrative uses, such as easing the interpretation of complex documents and routing grievances to the right office.
Speech is the common thread. India’s users often prefer talking to typing, and many speak a language that general-purpose models handle poorly. Global open releases are moving in the same direction, as seen in Nvidia’s Nemotron 3.5 Transcribe with open weights for 40 languages. The difference is that Indian builders are collecting speech data at pin-code level, so dialect coverage can be measured rather than assumed.
Data Provenance as a Product Feature
The Federal says BharatGen prioritises provenance at a granular level and has started speech collection with five initial vendors. For enterprises, that is not an academic nicety. A bank or hospital that must prove where training data came from will prefer a model with a paper trail over a larger one without.
This is also where sovereignty becomes a sales argument. Keeping data inside national borders satisfies a regulator. Documenting its origin satisfies a lawyer. Publishing weights satisfies a researcher. India’s open-source approach, with Sarvam releasing weights and BharatGen describing itself as open and multilingual, lets it make all three arguments at once.
The weak point is quality control. Pin-code-level speech data is valuable, yet it is costly to label and easy to skew toward urban speakers. If the next phase leans on thin or uneven data, application accuracy will lag the headline parameter counts.
How India’s Approach Compares With Frontier Labs
Indian models are not trying to win general reasoning contests against the newest US systems. Frontier labs compete on scale and cost per token, as in the head-to-head between Gemini 4 Carbon, Claude Opus 5.5 and GPT-6.1 Terra. India competes on language coverage, local deployment and trust. The two strategies overlap less than headlines suggest.
Sovereign Versus Frontier: Different Games
| Dimension | India’s sovereign models | US frontier APIs |
|---|---|---|
| Primary goal | Indian-language public services | General reasoning and coding |
| Typical size | 8B to 105B, some MoE | Not disclosed by most vendors |
| Weights | Often open (Sarvam) | Mostly closed |
| Data residency | In-country by design | Depends on contract |
| Funding model | State-backed, subsidised compute | Private capital |
| Weak spot | Evaluation, data quality at scale | Local-language depth, sovereignty |
The table simplifies, since some US vendors now offer regional hosting and several Indian firms use foreign models in products. Still, the contrast explains why a government can fund a 17B model even when a larger foreign one exists. The goal is control and fit, not a leaderboard position.
Historical Context: From UPI Thinking to Sovereign Stacks
India has run this playbook before. UPI turned payments into public infrastructure that private apps build on, and Ramakrishnan cites it as proof that Indians adopt new technology readily. The sovereign AI programme tries to repeat the pattern at the model layer: public rails, private products on top.
The timeline is short. BharatGen launched in September 2024. The IndiaAI Mission followed with its Rs 10,371.92 crore budget. By February 2026 the summit showcased five models. By April, 12 organisations had funding. By July, 20 proposals were identified. In roughly 22 months the effort moved from a research launch to a funded portfolio.
Indian talent also shapes frontier labs abroad, a point underlined by the report on four Indian-origin leaders at Anthropic as it eyes an IPO. Sovereign programmes partly exist to keep more of that expertise working on problems at home.
Market Impact: Who Gains and Who Gets Squeezed
Application-stage funding favours integrators, health and education startups, and firms that already hold government relationships. Fractal Analytics and Tech Mahindra sit in that group, with Sarvam close behind thanks to its Aadhaar and state-hub work. Pure model labs without a distribution channel face a harder road once the question becomes who pays.
Foreign vendors will adjust. Expect more local hosting, Indian-language fine-tunes and partnerships with domestic firms. State-linked funding rounds elsewhere show the pattern, such as DeepSeek’s $16B round with 27% from state funds. India differs in its transparency and its focus on public services. For wider context, see our AI and machine learning coverage.
Hardware vendors benefit as well. Every added GPU and every 50MW hub is an order for chips, cooling and power gear. That demand lands while memory supply is already tight, so cost pressure will reach Indian buyers too.
Risks and Open Questions
Several questions remain unanswered in public reporting. Independent benchmarks for Indian-language tasks are thin, so builder claims dominate. Utilisation of the subsidised GPU pool has not been published in the sources we reviewed. The trillion-parameter BharatGen target for March 2027 is a stated goal, and nothing we found shows it is on track.
There is also a coordination risk. Twelve funded organisations and 20 proposals can produce overlap, with several teams building similar Hindi chat models while rural speech and legal text stay under-served. A strong application phase needs a clear map of which model handles which job.
Finally, privacy and safety rules must keep pace. The MedSum example shows consent being designed in early. Other deployments, especially those touching Aadhaar-linked services, will face closer scrutiny as volumes rise.
Four Predictions for the Next 12 Months
These are our analysis, not reported facts.
- Application announcements will outnumber model launches. Expect health, agriculture and legal pilots to dominate mission updates as the application pillar matures.
- Small models will win more government contracts than large ones. Cost and latency favour 8B to 30B systems for district-level work.
- Independent Indian-language evaluations will become a procurement requirement. Builder-reported benchmarks like the HealthBench claim will face pressure for third-party checks.
- BharatGen’s trillion-parameter goal will slip or shrink in scope. The March 2027 date is ambitious, and compute and data constraints make a staged release more likely.
Frequently Asked Questions
What is the news about India’s sovereign AI push?
ET Government reports that India’s sovereign AI effort is moving from building foundation models to delivering real-world applications for Bharat. The story names a Ganesh Ram…, and BharatGen’s Ganesh Ramakrishnan has made the same adoption-to-production argument in The Federal.
What is BharatGen?
BharatGen is an IIT Bombay-led initiative launched on September 30, 2024, to build indigenous multimodal, multilingual language models. It is funded by the Department of Science and Technology and the IndiaAI Mission.
What is Param2?
Param2 is BharatGen’s 17-billion-parameter Mixture of Experts model covering 22 Indian languages, according to Forbes India. It is positioned for education, governance and agriculture.
How many GPUs does the IndiaAI Mission have?
Government material reports more than 38,000 GPUs available at Rs 65 per hour, with 20,000 more announced on February 17, 2026 by Minister Ashwini Vaishnaw.
How many sovereign models is the government backing?
Business Standard reported 20 proposals, made up of 12 large language models and eight small language models. MediaNama reported 12 organisations funded as of April 2026.
Which sovereign models are already in use?
Forbes India links Sarvam’s models to UIDAI Aadhaar services and state hubs, and The Federal describes BharatGen’s MedSum pilot with Amrita Hospitals. Other models, such as Project Indus and Vaidya 2.0, were launched at the February summit.
Will India build a trillion-parameter model?
BharatGen has stated a goal of a one-trillion-parameter model by March 2027, according to Analytics India Magazine. No independent confirmation of progress was found.