With foundational models underway, public-private partnerships, GPU access, and talent retention efforts, India’s sovereign AI ambitions are taking root—one language model at a time
2025: The Year India Got Serious About Sovereign AI
Until late 2024, India’s stance on building its own large language models (LLMs) was hesitant at best. But by the end of 2025, that caution gave way to strategic urgency.
- Triggered by global geopolitical shifts and China’s DeepSeek-R1 breakthrough, the IndiaAI Mission intensified its efforts to back homegrown foundational AI models.
- The mission’s INR 10,300 Cr allocation became the fuel for identifying and supporting a new class of LLM builders, ranging from startups to established enterprises.
From skepticism to momentum, India’s sovereign AI journey has entered a decisive phase—not just to catch up, but to lead.
India’s Growing LLM Ecosystem: Who’s Building What?
The IndiaAI Mission’s first two rounds of selection have backed 12 foundational model builders. These include:
- Sarvam AI (open-weight hybrid LLM, Sarvam-M, built on Mistral Small)
- Soket AI (working on a 120B parameter open-source LLM)
- BharatGen (launched a 3B parameter model; building 14–20B from scratch using ~25% Indian data)
- Gnani.ai (building a 16B speech-to-speech model)
- Fractal Analytics (14B parameter model; now developing domain-specific reasoning models)
- Zeinteiq, Shodh AI, Avataar AI, NeuroDX, and others—each working on either general-purpose LLMs or specialised models (SLMs) for sectors like healthcare, science, and language.
Some models are built from scratch, others adapt existing open-source models—both approaches are moving India closer to LLM sovereignty.
From Foundation to Execution: What’s Actually Working?
1. GPU Access & Infrastructure:
- The cost of compute remains the biggest hurdle in LLM development. IndiaAI Mission has helped by allocating high-end NVIDIA H100 GPUs.
- Soket AI, for example, was allotted over 1,500 GPUs but is scaling usage gradually to avoid inefficiency.
- Gnani and BharatGen also confirmed receiving GPU and grant support.
2. Public-Private Model:
India’s model resembles neither the US’s corporate-led structure nor China’s closed ecosystem. Instead, it draws from successful frameworks like UPI—combining public support and private innovation.
3. Progress Despite Delays:
Most models are still in development—but milestones have been reached.
- Sarvam AI and BharatGen have released smaller models.
- R&D pipelines are now clearer, and early partnerships are bringing data, talent, and funding together.
Closing the Gaps: What’s Holding India Back?
1. Compute Scarcity & Resource Imbalance:
- With only 12 companies selected out of 500+, access remains limited.
- Newer startups may face delays or unequal access, depending on relationships and approvals.
2. Capital Limitations:
- VC appetite remains higher for AI applications over deeptech like LLMs.
- IndiaAI grants have helped, but India still lags behind US or China in scale of investment.
3. Talent Drain & Retention:
- Many Indian researchers are still moving abroad due to better compensation and global research access.
- Startups like BharatGen are working with IIT consortia to retain top students locally.
- Proposals like a unified “India AI Talent Mission” aim to build and retain an AI-native workforce, embedding AI in schools and higher education.
Building LLMs Isn’t Enough: The Case for SLMs
India’s LLM ecosystem isn’t just chasing size. Increasingly, there’s a push for mid-sized, domain-specific models that are:
- Faster to train
- Lower in cost
- Better aligned with India-specific use cases
Examples include:
- Fractal’s STEM and healthcare-focused reasoning models
- NeuroDX’s clinical LLM
- Zeinteiq’s scientific simulation models
As Pulkit Upperwal (Soket AI) noted, “Large models are necessary for generalisation. But SLMs are critical for real-world usability in India.”
The Talent Engine: India’s Real LLM Advantage?
India’s long-term edge may lie not in compute, but in talent.
- BharatGen’s unique academic collaboration has convinced many researchers to stay back in India instead of joining global PhD programs.
- Gnani.ai’s Gopalan sees potential in India becoming an AI education hub, especially as global universities plan campuses in India.
- Soket AI’s Upperwal calls this “an opportunity to learn from scratch,” like the West once did.
If India can build a self-sustaining ecosystem of AI talent, IP, and infrastructure, its sovereign AI vision can be more than just catch-up—it can be globally competitive.
What’s Next: From Experiments to Deployment
The mission doesn’t end at building LLMs. For India to lead:
- Public-private collaboration must deepen, with consistent GPU and grant access.
- Smaller companies must not be sidelined as compute allocation increases.
- Continued R&D investment, even post-model launch, is essential.
- Talent development needs to begin early, across school and university levels.
- Open ecosystems and interoperable tools must remain central to India’s LLM vision.
As BharatGen’s Rishi Bal puts it, “India needs its own model of building models.”
