India's AI Data Centres in 2026: Where Your AI Actually Runs, and How Much Power and Water It Uses
Short answer: India's data centres reached about 1.75 GW of operational capacity by mid-2026 and are on track to cross 2 GW by the end of the year, according to CBRE. Forecasts for 2029 to 2030 range from about 6 GW to 10 GW. Investment commitments since 2021 have reached about $173 billion, and the AI and cloud commitments announced by just four companies, Reliance, Amazon, Microsoft and Google, add up to more than $160 billion over the next few years.
That build-out has a physical cost. India's data centres used an estimated 10 to 15 TWh of electricity in 2025, and that could triple to 40 to 45 TWh by 2030. Water is the less visible cost: researchers estimate a water-cooled data centre uses about 26 million litres a year for every megawatt.
This report explains where India's AI actually runs, who is building it, how much power and water it takes, what the data protection law means for where your data lives, and what it changes for Indian developers and users.
What happens when you send a prompt?
When you type a question into ChatGPT, Gemini or an app built on an AI API, your text travels from your phone over the mobile network and the internet to a data centre that holds the model. There, a server full of GPUs reads your text, generates the answer one token at a time, and sends it back.
Where that data centre is matters for three reasons:
- Speed. Every 1,000 km of fibre adds roughly 10 milliseconds to a round trip, before any processing. A server in Mumbai can answer an Indian user much faster than one in Virginia.
- Law. India's data protection rules decide what personal data can leave the country, and some companies will be required to keep certain data in India.
- Cost and resilience. Local capacity means less dependence on undersea cables and foreign regions, and lets providers price in rupees.
How big are India's data centres in 2026?
Capacity is measured in megawatts (MW) or gigawatts (GW) of IT power, the electricity the servers themselves can draw. It is the standard way to size data centres, because AI hardware is limited by power more than by floor space.
| Measure | Figure | Source |
|---|---|---|
| Capacity in 2025 | About 1.2 GW | PTI, citing government goals |
| Operational capacity, mid-2026 | About 1.75 GW (CBRE) or 1.6 GW (JLL) | CBRE H1 2026, JLL |
| New capacity added, H1 2026 | About 130 MW (CBRE) or 85 MW delivered (JLL) | CBRE, JLL |
| Expected by end of 2026 | About 2 GW | CBRE |
| Forecast for 2029 | About 6 GW, needing about $110 billion | JLL |
| Goal for end of decade | Almost 10 GW | PTI, citing government goals |
| Vacancy, H1 2026 | 2.8%, a record low | JLL |
The two big property consultants count slightly differently, which is why CBRE says 1.75 GW and JLL says 1.6 GW for the same period. Both agree on the direction: capacity is set to grow three- to five-fold by the end of the decade, and almost everything built today is already rented.
Where India's data centres are
India's capacity is heavily concentrated. Mumbai, Chennai and Delhi-NCR together hold more than 80% of operational capacity, and Mumbai alone accounted for 94% of new supply added in the first half of 2026. Mumbai wins because it is where most international undersea cables land, and because it has the financial and enterprise customers that rent capacity first.
The next wave is spreading out. Google's AI hub is in Visakhapatnam, Microsoft's newest cloud region is in Hyderabad, Reliance's first AI data centre is in Jamnagar, and Amazon is expanding in Mumbai and Hyderabad.
Who is building India's AI infrastructure?
About 15% of all data centre commitments since 2021 are tied directly to AI, according to CBRE, but the newest announcements are AI-first. Here are the largest.
| Company | Commitment | Period | Where | Status (November 2026) |
|---|---|---|---|---|
| Reliance Industries | ₹10 lakh crore (about $110 billion) | Next seven years, announced February 2026 | Jamnagar first, gigawatt-scale plans | First facility expected to exceed 120 MW by end of 2026 |
| Amazon (AWS) | $48 billion total, including more than $21 billion for AI and cloud infrastructure | To 2030 | Mumbai and Hyderabad regions | $13 billion added in June 2026 to an earlier $35 billion plan |
| Microsoft | $17.5 billion, on top of $3 billion pledged in January 2025 | 2026 to 2029 | Hyderabad, Pune, Chennai, Mumbai | Hyderabad region live since 6 August 2026 |
| $15 billion | 2026 to 2030 | Visakhapatnam, with AdaniConneX and Airtel's Nxtra | Construction started in 2026. Nearly 1 GW planned at one location |
Two things are behind the rush. The government extended a tax holiday for foreign cloud service providers to March 2047, and demand for processing data inside India is growing. JLL counts more than $50 billion of foreign capital commitments attracted so far.
The public sector is building too. The IndiaAI Mission had onboarded more than 38,000 GPUs on its common compute portal by March 2026, rented to startups, researchers and students at subsidised prices. See our guide to getting an IndiaAI GPU.
How much electricity do India's data centres use?
AI hardware is power-hungry. AI servers pack many GPUs into each rack, and a data centre also needs power for cooling, networking and backup.
India's data centres used an estimated 10 to 15 TWh of electricity in 2025. Government goals cited by PTI put that at 40 to 45 TWh by 2030. For global context, the International Energy Agency estimates data centres used about 1.5% of the world's electricity in 2024 and projects that their use will roughly double to about 945 TWh by 2030.
The challenge for India is less the total than the location. A single gigawatt-scale campus needs a continuous gigawatt of power, delivered reliably to one spot, around the clock. That is why the Central Electricity Authority has asked states to factor data centre demand into their grid planning, why operators sign long renewable power contracts, and why the industry is already discussing small modular nuclear reactors, which JLL expects around 2033.
How much water do AI data centres use in India?
Data centres turn electricity into heat, and that heat has to be removed. Many facilities use evaporative cooling, which consumes water. The figures vary enormously by design:
| Scale | Estimated water use | Source |
|---|---|---|
| 1 MW of water-cooled capacity | About 26 million litres a year | Researchers cited by Down To Earth |
| A 30 MW facility | About 780 million litres a year, roughly 2.1 million litres a day | Down To Earth |
| A planned 160 MW data centre park (Greater Noida) | About 4.2 billion litres a year, comparable to a town of about 85,000 people | Down To Earth |
| One 100-word email written with an AI chatbot | 200 to 1,500 ml, depending on location | University of California researchers |
These are estimates for water-cooled designs. Operators are moving away from them in water-stressed areas. Microsoft says its new Hyderabad region uses air-cooled chillers that consume no water for cooling, and Yotta says its Greater Noida campus uses a closed-loop, air-cooled system. Where water is used, the debate is about where it comes from: groundwater, municipal supply or treated sewage. Local residents near some projects have raised concerns about groundwater, so the source matters as much as the volume.
For a single user, the water cost of a chat is small. For a country adding gigawatts of capacity in cities that already face water stress, it is a planning question that needs public data. Few Indian operators publish facility-level water figures today.
Does your data stay in India? What the DPDP rules say
The Digital Personal Data Protection (DPDP) Rules were notified on 13 November 2025. They phase in over 18 months:
- 13 November 2025: the Data Protection Board was set up.
- November 2026: consent manager registration opens.
- 13 May 2027: most obligations apply, and the Board's complaint and penalty powers begin.
On where data can go, the approach is a blacklist. Personal data can be transferred abroad unless the government notifies restrictions for specific countries. Separately, the rules allow the government to require significant data fiduciaries, large or sensitive data handlers, to keep specified personal data inside India on a committee's recommendation.
For most developers, that means three practical things:
- Know where your AI provider processes prompts. Many model APIs run in the United States or other regions by default. Check whether your provider offers an India region.
- Do not send personal data you do not need. Masking names, phone numbers, Aadhaar and PAN numbers before a prompt leaves your server reduces your risk wherever the model runs. Our Python tutorial on anonymising data before an LLM shows how.
- Keep a data-flow map. If transfer restrictions are introduced, you will need to know quickly which data goes where.
Why location matters for speed
Light in optical fibre travels at about 200,000 km per second, so every 1,000 km of cable adds about 10 milliseconds to a round trip. Real routes are longer than straight lines and add routing delays, so actual times are higher. The theoretical minimums from Mumbai are:
| Destination | Straight-line distance | Minimum round trip |
|---|---|---|
| Hyderabad | About 620 km | About 6 ms |
| Chennai | About 1,030 km | About 10 ms |
| Singapore | About 3,900 km | About 39 ms |
| Frankfurt | About 6,300 km | About 63 ms |
| US East Coast | About 13,000 km | About 130 ms |
For a chatbot, 100 milliseconds of network delay is noticeable but not critical, because generating the answer takes longer. For voice agents, real-time translation and coding agents that make dozens of calls per task, the delay adds up on every step. This is one of the strongest arguments for serving models from inside India. Our guide to local AI vs cloud AI for coding covers the trade-offs for developers.
What it means for Indian developers, students and users
- Developers: expect more models to be offered from Indian regions over 2026 to 2028, with lower latency and simpler compliance. Until then, choose providers and gateways that tell you where requests are processed, and design around the DPDP deadline of May 2027.
- Students: data centres need electrical, mechanical, cooling and network engineers, not just software developers. Mechanical and electrical graduates are directly relevant to the fastest-growing part of the AI industry.
- Users: more local capacity should mean faster AI responses and more products priced in rupees. It should also mean more public debate about power and water in the cities that host these campuses.
Frequently asked questions
What is India's data centre capacity in 2026?
About 1.75 GW of operational capacity by mid-2026 according to CBRE, or 1.6 GW according to JLL, with CBRE expecting about 2 GW by the end of 2026.
Which city has the most data centres in India?
Mumbai. Mumbai, Chennai and Delhi-NCR together hold more than 80% of India's operational capacity, and Mumbai accounted for 94% of new supply added in the first half of 2026.
Where is Google's AI data centre in India?
In Visakhapatnam, Andhra Pradesh. Google is investing $15 billion from 2026 to 2030 in an AI hub built with AdaniConneX and Airtel's Nxtra, with nearly 1 GW of capacity planned at one location.
How much electricity do data centres use in India?
An estimated 10 to 15 TWh in 2025, projected to rise to 40 to 45 TWh by 2030 as capacity grows towards 10 GW.
Do AI data centres use a lot of water?
Water-cooled designs can use about 26 million litres a year per megawatt of capacity, according to researchers cited by Down To Earth. Air-cooled and closed-loop designs, like Microsoft's Hyderabad region, use little or no water for cooling.
Does ChatGPT run on servers in India?
Major AI providers decide their own serving regions, and many run models outside India by default. Check each provider's data residency options. Under the DPDP rules, transfers abroad are allowed unless the government restricts specific countries.
When do the DPDP rules fully apply?
Most obligations under the DPDP Rules apply from 13 May 2027, 18 months after the rules were notified on 13 November 2025.
Limits of this report
- Capacity figures differ by source. CBRE and JLL count operational capacity differently, so we show both.
- Investment pledges are not spending. Commitments run over five to seven years and may change.
- Water figures are estimates. Few Indian operators publish facility-level water data, and use varies widely by cooling design.
- Latency figures are theoretical minimums. They assume straight-line fibre. Real network paths are slower.
Sources
- CBRE H1 2026: India data centre capacity hits 1.75 GW (CXOToday)
- India's data centre capacity expected to jump to 6 GW by 2029 (JLL) (IANS, 2 September 2026)
- India's data centre boom and grid planning (Outlook Business, 8 April 2026)
- India's digital thirst: data centres rising in water-scarce regions (Down To Earth, November 2025)
- Reliance to invest $110B in data centers across India (Built In, 19 February 2026)
- Amazon to invest additional $13 billion in India cloud, AI (Reuters via The Star, 25 June 2026)
- Microsoft brings Hyderabad cloud region online (The Next Web, August 2026)
- Google breaks ground on AI hub in Visakhapatnam (YourStory, April 2026)
- DPDP Rules 2025 implementation roadmap (TCSA)
- IndiaAI Mission onboards 38,000+ GPUs (Devdiscourse, March 2026)