Skip to content
Gallery
p+ai: Wiki
P+ai: Wiki ToC
People+ai Theses
Volunteer Wiki
Knowledge Hub
Share
Explore
People+ai Theses

icon picker
Enable Compute to enable DPI^AI

Addressing India's Compute Needs for a Digital Future
Written by
@Tanvi Lall
⁠
megaphone

Create Q&A for this.

How much compute does it take to run an AI search? Read on to find your answer!
As much as a normal Google Search
⁠
⁠
10 times a normal Google search
⁠
⁠
50 times a normal Google search
⁠
⁠
megaphone

Create infographic for compute capacity and engineering talent in India vs. China and USA

Focus area rationale

India is pursuing a multi-pronged approach to AI with a focus on applied AI solutions for key sectors and foundational AI research to build indigenous capabilities. With a growing number of students enrolling in deep learning courses, research scholars needing dedicated GPU access, and large-scale mission-mode projects at startups and research institutions requiring substantial compute resources, the need for robust AI infrastructure is evident.
While compute provides the necessary power, speed and scalability necessary to build and consume AI, the domestic compute supply market in India is fragmented and a diverse set of players play in it - there are infrastructure providers building data centres and providing basic hardware resources; PaaS and SaaS providers providing services/applications and then support businesses helping with cybersecurity, privacy, compliance in this stack. Being a provider in this space requires subject matter expertise and long-term partnerships in your value chain. Many of these smaller players exist but don’t have a significant market share.
India is gearing up to further enhance the impact of DPI and, in a similar “tech-for-nation” format, enable new and innovative AI applications that could help alleviate population-scale problems. India's focus on enhancing the impact of its Digital Public Infrastructure (DPI) and enabling innovative AI applications necessitates substantial compute power. The DPI stack, comprising Aadhaar, DigiLocker, and UPI, has already addressed large-scale challenges unique to India, such as financial inclusion, efficient government service delivery, and data-driven decision-making. These systems are heavily utilised, with UPI processing 10-14 billion transactions per month as of 2024. The Indian Government is further leveraging AI in population-scale programs like Bhashini for digital payments, PM-KISAN for direct benefit transfers, e-Courts for judicial digitization, and DIKSHA for personalised education. These initiatives demonstrate India's commitment to using AI for significant social impact. To continue this trajectory, and to deploy AI capabilities in Natural Language Processing, Computer Vision, and Speech Recognition across diverse sectors such as education, healthcare, agriculture, and climate, a robust compute infrastructure is essential. This will support millions and billions of AI-driven daily inferences, driving innovation and addressing population-scale problems effectively.
While opportunities exist, India currently has far less compute capacity than is necessary to meet its ambitious goals. The country faces significant challenges, including a lack of practical engineering skills, brain drain, a low number of published research papers, heavy reliance on external cloud providers, and a relatively small compute market. As diverse AI applications and users emerge, the need for robust and scalable compute infrastructure becomes even more critical.
⁠
megaphone

How much power is required to train an LLM?

As much as that of 5 cars during their lifetime
As much as it takes for a flight from Bengaluru to San Francisco
As much as Thor’s Mjolnir
Our belief is that there are multiple ways to enable the impact and reach of India’s domestic compute supply. Out of the many ideas, people+ai is focused on initiatives that enable the compute system as a whole and plays in areas where standalone entities/players don’t venture. At the ecosystem level, we have efforts to do the following:
Acknowledge the diversity of compute form factors available today
Formally recognise smaller players who serve in this market as a category of providers enabling them to achieve prominence
Policy actions to increase the volume and spread of domestic compute providers
Increase discoverability for domestic providers by creating necessary channels and incentive schemes for customers to consume compute from them
Solve for low-cost inferencing-at-scale
edison-bulb
Yes, access to high-end compute needs to be prioritised by India.
⁠
1
⁠
⁠
⁠
⁠
⁠

Current projects and ideas

A program called Open Cloud Compute (OCC) that includes:
Watch this video to understand OCC’s scope better.
⁠
⁠
⁠
“connecting the grid” design an open network that democratises access to affordable, and reliable domestic compute resources.
“making the grid” scaling the micro data centre model with public and private players
“enabling the grid” policy actions to improve ecosystem with MeitY and STPI
“build partnerships” with actors of the Indian compute ecosystem
Exploring synergies: Study niche models that could increase access to compute for specific consumer personas

⁠

Prioritisation for next 6-12 months

India needs to be set up for inferencing-at-scale. Underlying projects include:
Baselining inferencing costs for population-scale domain applications
Strategies to do high-volume, low-cost AI inferencing in India
Design a financial instrument to invite investments in data centre and computing resources in India
info
Add this button to gauge who has weighed in on your doc. Great for live review or when sharing with a large audience. Delete this callout when ready!

Done reading?
⁠
⁠



Create Q&A for this.
Create infographic for compute capacity and engineering talent in India vs. China and USA
Focus area rationale
How much power is required to train an LLM?
Current projects and ideas
Prioritisation for next 6-12 months
Done reading?
Want to print your doc?
This is not the way.
Try clicking the ⋯ next to your doc name or using a keyboard shortcut (
CtrlP
) instead.