IndiaAI Mission: 34,000 GPUs at Rs 150/Hour for Startups

India now has 34,000 GPUs available through the IndiaAI Mission at Rs 115 to Rs 150 per GPU-hour — roughly 42% below market rates. The government has committed to scaling this to 100,000 GPUs by end of 2026.

What Is the IndiaAI Mission?

The IndiaAI Mission is the Indian government's Rs 10,371.92 crore ($1.25 billion) programme to build sovereign AI infrastructure. It funds shared GPU compute, an open Indian dataset platform, application development grants, and a programme to support Indian AI startups building on domestic infrastructure.

The compute component — Rs 4,563.36 crore over five years — operates as a national shared resource. Researchers, startups, MSMEs, and academic institutions can apply to access the GPU cluster at subsidised rates. Eligible projects in areas of national importance receive up to 40% cost reduction administered by a government committee under MeitY (Ministry of Electronics and Information Technology).

How Much Do These GPUs Cost?

The lowest successful bids came in at Rs 115.85 per GPU-hour for standard GPUs and Rs 150 per GPU-hour for high-end GPUs (likely H100-class). After government subsidies, effective rates for approved projects fall below Rs 100 per GPU-hour.

For context: comparable H100 GPU cloud access from AWS, Google Cloud, or Azure costs roughly $2.50 to $4.00 per GPU-hour, or Rs 210 to Rs 335 at current exchange rates. The IndiaAI Mission rate is significantly cheaper, though access is gated through an application and approval process rather than self-serve provisioning.

The compute cluster runs across two facilities. Chennai is live at 30 MW. A new 40 MW facility is planned for Mumbai.

Who Is Building on This Infrastructure?

Sarvam AI is the first startup selected under the IndiaAI Mission to build India's sovereign LLM. On February 18, 2026, Sarvam launched two open-source models: Sarvam-30B (a mixture-of-experts architecture) and Sarvam-105B (which activates approximately 9 billion parameters per token with a 128,000-token context window).

Sarvam also launched a startup programme on March 5, 2026, offering AI credits and developer tools to Indian startups building on its models. The company has raised $53 million in its Series A from Lightspeed, Peak XV (formerly Sequoia India), and Khosla Ventures.

Krutrim, Bhavish Aggarwal's AI company, is running a parallel track. Its Krutrim-1 (7B) and Krutrim-2 (12B) models focus on Indian language support. Aggarwal committed Rs 10,000 crore to Krutrim by 2026 and announced a partnership with Nvidia to build India's largest supercomputer.

Why the Government Is Subsidising AI Compute

The reasoning is straightforward. AI training is compute-limited. A startup in Bengaluru pays the same AWS rates as a startup in San Francisco. But an Indian AI company competing against OpenAI or Google also faces higher funding constraints and a smaller domestic VC market.

Subsidising compute is a direct way to lower the barrier. It also keeps sensitive AI workloads on domestic infrastructure — important for sovereignty arguments when the AI handles government data, defence applications, or critical national services.

The 100,000 GPU target by end of 2026 would give India one of the larger national AI compute clusters outside the US and China. It is not comparable to what OpenAI or Google train on, but it is sufficient for fine-tuning, inference at scale, and research-grade training runs.

What Developers Should Know

If you are building in India — or building multilingual AI for Indian languages — the IndiaAI Mission compute programme is worth applying for. The approval process adds friction compared to spinning up an EC2 instance, but the cost savings on longer training runs are significant.

Sarvam-105B is now publicly available. It is the most capable open-source model trained specifically on Indian language data, with support for 10+ Indian languages. For any application targeting Indian users — voice assistants, customer service, document processing in Hindi or regional languages — this is now the baseline to evaluate.

The DPDP (Digital Personal Data Protection) Rules, notified November 13, 2025, set a compliance timeline that developers should track: Consent Manager registration launches November 2026, full compliance requirements including breach notifications kick in May 2027. AI systems processing Indian user data will fall under Significant Data Fiduciary obligations if they cross certain thresholds.

Key Takeaways

• 34,000 GPUs — current IndiaAI Mission national compute capacity, targeting 100,000 by end of 2026
• Rs 115-150/GPU-hour — subsidised rate, 42% below market; eligible projects get an additional 40% off
• February 18, 2026 — Sarvam AI launched Sarvam-105B open-source, India's most capable sovereign LLM
• Rs 10,371.92 crore ($1.25 billion) — total IndiaAI Mission budget across compute, data, and applications
• For developers: Apply to IndiaAI Mission compute access if building AI products in India — the cost advantage over commercial cloud is real. Evaluate Sarvam-105B for Indian-language use cases before defaulting to GPT-4 or Gemini.
• What to watch: Whether the 100,000 GPU target is met by December 2026 and who wins the next ATMP contract under India Semiconductor Mission 2.0 (Rs 1,000 crore allocated in Budget 2026-27 for equipment and materials)