India’s bid to become a serious builder—not just a consumer—of artificial intelligence is picking up pace. The government is poised to name a new cohort of roughly eight homegrown companies that will receive IndiaAI Mission incentives to develop foundational AI models (large language models and other base models). IT & Electronics Minister Ashwini Vaishnaw is expected to announce the list soon.
Below is a crisp explainer of the scheme, progress so far, and why the next phase is likely to emphasise inference-ready models and GPU access.
The Incentive Program: A Quick Recap
- Budget window: Earlier this year, MeitY announced an ₹1,500 crore incentive pool for entities and individuals proposing to build AI models from the ground up in India.
- Application momentum: By mid-February, the program had received dozens of applications from Indian and global startups as well as researchers, with total applications reportedly reaching around 500 soon after—indicating strong domestic and diaspora interest.
- Round-1 beneficiaries: In the first tranche, four startups—Sarvam, Soket Labs, Gnaani.ai, and Gan.ai—were approved to receive subsidised GPU compute to train indigenous AI models.
- Open-source signal: In July, BharatGen released Param-1, a bilingual (English/Hindi) open-source LLM reportedly trained on trillions of tokens—an early proof that Indian researchers can ship credible base models tuned to local languages.
What’s Coming in Round 2
Government sources indicate that the next shortlist will include close to eight firms focused on building foundational models (LLMs, vision-language models, speech models). The intent is twofold:
- Strategic capability: Ensure India can design, train, and control base models aligned with Indian languages, culture, and regulatory priorities.
- Ecosystem flywheel: Create downstream opportunities for application developers, SMEs, and public-sector use cases (governance, education, health) by making indigenous base models readily available.
Expect emphasis on:
- Indian-language coverage and multimodality (text, speech, vision)
- Practical benchmarks (accuracy, safety, cost)
- Deployment readiness (latency, memory footprint, guardrails)
The GPU Bottleneck—And a Shift Toward Inference
Even as funding flows, compute remains the choke point. Access to high-end GPUs through the national compute facility and private providers is improving, but demand far outstrips supply. That is shaping policy thinking in two ways:
1. Prioritise inference efficiency: A “school of thought” within the ministry suggests that upcoming GPU procurement should also focus on inference-optimised hardware/software stacks, not only on training clusters.Why it matters: Many public- and private-sector deployments don’t need massive retraining; they need fast, cheap, and predictable inference for at-scale usage (citizen services, call centers, translation, document workflows).
2. Hybrid access model: Alongside government clusters, India has empanelled multiple GPU-as-a-Service providers, enabling startups to rent compute at competitive domestic rates. This reduces capex and allows teams to scale workloads elastically.
Bottom line: Training matters, but serving models to millions of users at sustainable cost is where impact meets economics. Expect grants and reviews to increasingly ask, “What is your cost per 1,000 inferences? What’s the latency on a mid-range GPU? How will you localise guardrails?”
Why Foundational Model Incentives Matter
- Language equity: India’s linguistic diversity is vast. Foundation models tuned for Hindi and regional languages unlock inclusion across citizen services, education, and commerce.
- Data sovereignty & safety: Building in-country allows tighter control over data pipelines, safety policies, and compliance.
- Economic leverage: A robust local model layer catalyses startup applications (fintech, agritech, healthcare, retail), public-sector efficiency, and enterprise productivity, keeping more value capture within India.
- Talent retention: Access to compute and grants helps Indian researchers and founders build at home rather than relocating purely for infrastructure.
What Startups Should Do Now
- Show measurable utility: Tie your model to clear use cases—translation quality gains, call-center AHT reduction, document processing accuracy, or TCO improvements for enterprises.
- Design for inference first: Optimise for quantisation, speculative decoding, distillation, and LoRA-style adaption so the model runs well on cost-effective GPUs.
- Guardrails & evaluation: Demonstrate safety layers, bias checks, hallucination control (retrieval-augmented generation, constrained decoding) and transparent evaluation on Indian data.
- Open vs closed strategy: If you open-source, plan for community contributions and license clarity; if you go closed, articulate a sustainable unit-economics path and customer value.
- Partnerships: Explore co-development with universities, public-sector units, and industry consortia for datasets, domain expertise, and distribution.
What to Watch Next
- Official shortlist announcement by the IT minister and specifics of grant/compute allocations.
- GPU capacity additions and the mix between training and inference infrastructure.
- Regional-language benchmarks that become de-facto standards for Indian deployments.
- Open-source momentum (e.g., Param-series and others) and how enterprises adopt them with fine-tuning and RAG.
- Public-sector pilots that move beyond POCs into stable, measured rollouts.
The Takeaway
India’s AI journey is moving from policy intent to execution. Backing a new cohort of foundational model builders, while simultaneously fixing compute access and pushing for inference efficiency, is a pragmatic path: build a few strong engines, make them affordable to run, and let thousands of applications bloom on top.