Fill India’s idle GPUs in 90 days
The GPUs exist. The demand does not reach them. Mission GPUs reportedly run at about 20% utilisation while founders wait months for compute. This field plan sets out the numbers, the five frictions, a three-phase fix and four moves a founder can make now.
What you need
- Your expected training and inference workload
- Your current cloud bill
- One hour
What you will have
- A clear view of why the capacity sits idle
- A rupee cost plan for your compute at the subsidised rate
- An email ready for Indian GPU providers
The problem in numbers
- 38,000+ GPUs onboarded under the IndiaAI Mission (IndiaAI and MeitY, 2025–26)
- ₹10,372 crore total IndiaAI Mission outlay approved (Union Cabinet, March 2024)
- ~20% utilisation reported for mission GPUs: roughly 8 in 10 sit idle (Rhodium Group, November 2026)
- ₹67 an hour subsidised GPU rate on the IndiaAI compute portal (IT Minister Ashwini Vaishnaw, March 2025)
- ≈₹4.9 crore a day of subsidised compute unused: 30,400 idle GPUs × 24 hours × ₹67. Our estimate, about ₹1,780 crore a year
- 3 to 6 months wait for GPUs reported by Indian AI founders and investors (Elets CIO, 26 October 2026)
Five frictions between a founder and a GPU
- Slow allocation Approval cycles built for grants, not for engineers who need a cluster this week.
- Delayed payments to providers When providers wait to be paid they ration capacity or send it to buyers who pay faster, including overseas.
- Demand stays on global clouds Enterprises and government teams default to the clouds they already use.
- Discovery and paperwork Most early-stage founders, colleges and MSMEs do not know the portal exists or stop at the forms.
- Skills to use the compute Raw GPU hours help only teams that can train, fine-tune and deploy.
The 90-day plan: one target, 50% utilisation
- Days 1–30 · Open the doors Auto-approve credits for every DPIIT-recognised startup, AICTE college and Udyam MSME. One-click UPI-simple sign-up and billing. A live public utilisation dashboard by day 30.
- Days 31–60 · Anchor the demand Route central and state government AI workloads to mission capacity first. Pay providers within 30 days every time. Open sector challenges in agriculture, health and Indic language with compute as the prize.
- Days 61–90 · Build the habit Compute grants to university labs tied to published models. Migration credits for enterprises moving AI work from global clouds. Hands-on training through IITs and startup hubs.
Step 1: Apply on the IndiaAI compute portal
15 minutes · IndiaAI compute portal
Check whether Sharma Dental Studio is eligible for subsidised GPU hours before you sign a global cloud contract.
Step 2: Price your compute in rupees
20 minutes · Claude
Model training and inference cost at the subsidised rate. It changes your burn and your runway.
I run Sharma Dental Studio. Here is my expected AI workload: [training hours, inference calls a month, model size]. Estimate my monthly GPU cost at ₹67 per GPU hour on the IndiaAI compute portal and compare it with my current cloud bill of [amount]. Show the working and what it does to my runway.
Step 3: Ask Indian providers for credits
15 minutes · Claude
Empanelled providers want steady customers. A clear workload plan gets better terms.
Write a short email from the founder of Sharma Dental Studio to an IndiaAI empanelled GPU provider. Describe our workload [details] and the steady monthly usage we expect, and ask for startup credits and a three-month price. Keep it under 150 words.
Step 4: Start with one agent, one use case
10 minutes
Prove value on a narrow problem before you scale compute. Utilisation follows usefulness.