Quick answer
Economic Survey 2025-26 treats AI compute as physical infrastructure: finance, GPUs, electricity and water jointly constrain data-centre expansion.
AI compute is the processing capacity required to train models and run inference, supported by hardware, memory, networks, power and cooling. Box XIV.2 shows why financing a data centre is not enough: execution also requires hardware access and a ready grid.
Concept diagram
Compute needs power, water and networks
Server racks sit between electricity, cooling and network systems, each a possible bottleneck.
- Reliable power, cooling and connectivity complement compute investment.
- Local resource constraints matter when deciding where and how to build.
Basics: the physical stack behind software
GPUs and memory
- GPUs perform parallel calculations used in AI training and inference. High-bandwidth memory supplies data quickly enough to keep specialised processors working.
- A shortage of memory, storage or advanced chips can make the entire compute installation more expensive or delay commissioning.
Power versus energy
- MW measures instantaneous power capacity or load; kWh measures energy used over time. A tariff in ₹/kWh prices electricity consumption, not GPU performance.
- Training can demand intensive bursts of processing; repeated inference also requires reliable energy and hardware availability.
Cooling and grid stability
- Cooling creates water and electricity requirements that vary with site, design and operations. A reported upper bound is not a universal centre-level average.
- Sudden changes in large loads can challenge grid operation, so planning must include dynamic stability as well as total supply.
Resource pressures cited by the Survey — paras 14.18–14.21 and footnotes
| Indicator | Reported number | Scope and caution |
|---|---|---|
| AI data-centre water use | Up to 20 lakh litres/day | Cited upper bound, not every centre’s consumption |
| Global annual water use | 56,000 crore litres | Figure cited in the Survey’s discussion |
| Northern Virginia load-loss incident | 1,500 MW; 10 July 2024 | Illustrates grid stability risk, not India’s capacity |
| Revised U.S. electricity growth projections | 3% in 2025; 5% in 2026 | Survey-cited EIA projections; not realised Indian growth |
Box XIV.2: an agent-based infrastructure stress test
Interacting constraints
- Agent-based modelling represents individual operators and their decisions, allowing financing, grid access and GPU supply to interact through feedback loops.
- The 40-quarter exercise identifies vulnerabilities and coordination failures; it is not an optimal-capacity estimate or a precise investment forecast.
Why subsidies are insufficient
- Cheaper finance can bring projects to the execution stage sooner without ensuring chip deliveries or power readiness.
- Investment competes with households and industry for scarce power, finance and water; opportunity costs must enter project appraisal.
Coordinate the investment pipeline
Secure hardware
Diversify access and track delivery uncertainty before treating proposals as executable capacity.
Synchronise infrastructure
Align land, power connections, cooling arrangements and financing milestones.
Improve efficiency
Evaluate task-specific models and decentralised compute alongside large installations.
Common exam traps
Statements examiners like to twist — learn the correct version.
A sanctioned loan guarantees compute capacity.
Hardware deliveries, memory and grid connections may still prevent execution.
MW and kWh are interchangeable.
MW is power; kWh is energy. A tariff cannot be read as capacity.
All data centres consume exactly 20 lakh litres daily.
The Survey cites an upper-bound figure; actual consumption depends on technology and location.
Prelims focus
- Distinguish GPU, compute, memory, MW and kWh.
- Box XIV.2 uses an agent-based model over 40 quarters.
Mains angle
- Explain AI infrastructure as a coordination problem rather than only a financing gap.
- Discuss energy–water opportunity costs and grid resilience in India’s AI strategy.
Flip & recall
Answer in your head first, then tap the card to check.
Practice MCQs
Original Prelims-style questions with explanations.
0 / 3 correct · 3 left
Q1.Which unit measures electricity consumption?
Q2.Box XIV.2 should be interpreted as:
Q3.Which is a complementary input to GPUs?
