India's IndiaAI Kosh: 38,000 GPUs at ₹100/hour - The DPI Approach to AI Compute
India is applying its digital public infrastructure playbook to artificial intelligence. The IndiaAI Mission, backed by ₹10,300 crore in government funding, has deployed 38,000 GPUs nationwide through its compute✦ platform, with pricing starting at ₹67 per hour for NVIDIA H100 access. Phase 2, announced at the India AI Impact Summit in March 2026, will add another 20,000 GPUs within six months.
The strategy mirrors the approach that made Unified Payments Interface (UPI) and Aadhaar global benchmarks: government sets the rails, private enterprises compete on top. IndiaAI Kosh, the national AI compute and data repository, is designed to ensure startups and researchers can train models without burning through venture capital on cloud bills.
Who's Building the Infrastructure?
The compute backbone is a public-private partnership spanning multiple providers:
- Yotta Data Services leads with 9,216 GPUs, including 8,192 H100s, with plans for 20,736 NVIDIA Blackwell Ultra GPUs by August 2026
- E2E Networks deploys Blackwell clusters at L&T Vyoma Data Centre in Chennai
- Jio Platforms contributes 208 H200 and 104 AMD MI300X GPUs
- AWS-managed partners provide 1,200 lower-tier GPUs for lighter workloads
The IndiaAI Compute Portal, launching imminently, will let startups, ministries, state governments, and researchers request capacity through a unified platform. Eligible users receive up to 40% subsidies on compute costs.
India is no longer waiting for permission to build AI. We are building the infrastructure so the next generation of Indian AI companies can scale without dependence on foreign cloud providers.
Early adopters demonstrate the model's potential. Sarvam.ai trained its Sarvam-3 multilingual LLM✦ on Yotta's H100 cluster, then open-sourced models covering 22 Indic languages. The company's success shows that sovereign AI capacity does not mean isolation from the global ecosystem✦.
By The Numbers
- 38,000 GPUs deployed as of April 2026, with 20,000 more planned within six months
- ₹67-100 per GPU✦ hour for H100 access, roughly 33% below global cloud rates
- ₹10,300 crore (~USD 1.2 billion) total IndiaAI Mission investment through 2030
- 1,000+ machine-readable datasets available through AI Kosh repository
- 40% subsidy on compute costs for eligible Indian startups and researchers
| Provider | GPUs Deployed | Key Hardware |
|---|---|---|
| Yotta Data Services | 9,216 | NVIDIA H100, Blackwell Ultra (planned) |
| E2E Networks | TBC | NVIDIA Blackwell clusters |
| Jio Platforms | 312 | H200, AMD MI300X |
| AWS Partners | 1,200 | Lower-tier GPU instances |
| Others (Phase 2) | 20,000+ | Mix of H100/Blackwell |
Why This Matters for Asia
India's DPI approach to AI compute carries implications beyond its borders. If the model succeeds, it offers a template for countries across South and Southeast Asia that lack the capital for independent AI infrastructure. The combination of subsidised compute, open data repositories, and government coordination addresses three barriers simultaneously: cost, data access, and coordination failure.
The risk is that subsidies drive consumption without proportional innovation. Phase 2 will test whether affordable GPUs translate into competitive AI products or simply cheaper experimentation. DeepSeek's open-source models have already shown that raw compute is necessary but insufficient; quality training data and engineering talent matter as much as hardware access.
The DPI model works when you combine infrastructure with ecosystem development. GPUs alone do not produce AI companies.
Frequently Asked Questions
How does IndiaAI Kosh pricing compare to global cloud rates?
AWS charges approximately USD 1.20 per GPU hour for on-demand H100 instances. India's ₹67 rate (under USD 0.80) represents roughly a 33% discount, significant for startups training large language models where compute costs can exceed millions of dollars.
Can overseas companies access IndiaAI Kosh?
Policy remains focused on Indian startups and researchers. Multinational companies with R&D centres in India may gain access under bilateral arrangements, but the primary beneficiaries are domestic AI developers.
What datasets are available through AI Kosh?
The repository hosts over 1,000 machine-readable datasets spanning healthcare, agriculture, governance, and financial technology. The education sector is a particularly active contributor, with datasets supporting Indic language model development.
Drop your take in the comments below.
