India runs 1.8 GW of data centres; AI alone may add 26.3 GW of grid load by FY32

01What India runs today: about 1.8 GW, half of it in Mumbai

The trackers broadly agree on the stock and disagree on the pace. Savills counts 1.8 GW of operational IT capacity after 258 MW was added in the first half of 2026, 59% more than in the first half of 2025; it expects more than 7 GW by 2030.¹ Cushman & Wakefield counts 1,789 MW across the seven largest markets, with 178 MW added in the same half, 147 data centres run by 33 providers, a colocation vacancy of 14.9% and a pipeline of 3,860 MW to 2030, of which 509 MW is under construction.² The government’s own figure, given to the Rajya Sabha in March, is “around 1500 MW by 2025”, up from about 375 MW in 2020.³

The half-year additions differ by 80 MW between the two firms, mostly because they draw the market differently (Savills counts the country, Cushman & Wakefield its top seven cities), so we chart one firm’s city data rather than mix the two.

02What AI will ask for: 26.3 GW, and what that number is

On 27 July, in a written reply in the Rajya Sabha, Minister of State for Power Shripad Naik said: “An additional load of 26.3 GW from Artificial Intelligence (AI) Data Centres is projected by 2031-32, which is expected to be integrated into the grid and primarily served by renewable energy capacity.”⁴ Energy Watch’s report of the same reply attributes the estimate to the mid-term review of the 20th Electric Power Survey, the Central Electricity Authority’s five-yearly demand assessment.⁵

Four months earlier, answering a question on the energy and water footprint of the IndiaAI Mission, the IT ministry said that “as per information available with the Ministry of Power, electricity demand from data centres is estimated to reach 13.56 GW by 2031−32.”³ Coverage called the July figure almost twice the March one.⁶ That reading needs care. March gives total data-centre demand; July gives additional load from AI data centres. Neither answer states its assumptions about utilisation, cooling overhead or how much of the announced pipeline is counted.

Industry forecasts sit on either side. Wood Mackenzie, as reported by 5paisa, projects India’s data-centre capacity rising from 2.2 GW in 2025 to 12 GW in 2030, and AI capacity from 275 MW to 6,546 MW.⁶

03How the rest of the world is handling it

The International Energy Agency put global data-centre electricity at about 415 TWh in 2024, 1.5% of the world’s consumption, and projected about 945 TWh by 2030 in its base case, just under 3%. It expected the United States and China to account for nearly 80% of the growth.⁷ A year later it reported that data-centre demand grew 17% in 2025, against 3% for electricity overall, and that supply of gas turbines, transformers and chips had tightened while project pipelines were “holding up grid connections and other necessary approvals”.⁸

Utilities are making data centres carry the risk. In Ohio, the state commission adopted AEP Ohio’s data-centre tariff on 9 July 2025. A new site of 25 MW or more pays a load-study fee, signs a contract of its ramp-up period plus eight years, and is billed for at least 85% of its contracted capacity whether it uses it or not. A customer without strong credit ratings posts collateral of half its minimum charges for the whole term, and a project cancelled or delayed by more than a year before power-on repays the utility’s build-out cost in full.⁹

Developers are building the generation themselves. At the PORTS-Pike campus in the same state, SB Energy and SoftBank say they will build at least 10 GW of new generation for 8 GW of IT capacity leased to OpenAI, and put at least US$4.2bn into regional grid infrastructure with AEP Ohio, with capacity coming online in phases from 2028; NVIDIA guarantees the land, power and shell.¹⁰ The IEA reports that the technology sector signed about 40% of corporate renewable power-purchase agreements in 2025.⁸

“The IEA was early in recognising that there is no AI without energy – and that countries that provide secure, affordable and rapid access to electricity will be one step ahead.”

04India's constraints: the grid, the water, and what four states offer

The grid. The same mid-term review projects India’s peak demand rising from 289 GW in 2026–27 to 388 GW in 2031–32.⁶ The AI load would be about 7% of that later peak, on our arithmetic. The power ministry says the transmission plan to 2032 envisages about ₹9,16,142 crore of investment and that 154 inter-state transmission projects are under construction.⁴ The IT ministry told Parliament in March that the national grid “is adequately prepared to ensure reliable power supply across regions.”³ Savills, for its part, lists access to reliable power and suitable land as “a key challenge” in the major markets.¹

Water. The government’s answer on water points to general rules, not data-centre ones. Asked for city-wise water needs, the IT ministry answered that “the water requirement of data centres depends on the type of cooling technologies deployed” and pointed to the Jal Shakti ministry’s groundwater notification of 24 September 2020.³ The operators describe their own choices. Microsoft says its Hyderabad region uses effectively zero water for cooling.¹⁶ Amazon says none of its operational Indian data centres uses water for cooling, because they run air-cooled chillers, and that its planned Thane site will use 125 kilolitres a day, 115 of them treated sewage for landscaping.²¹²²

The states. Tamil Nadu’s Data Centre Policy, G.O.(Ms.) No. 33 of 23 November 2021, offers projects above ₹500 crore that meet 30% of their energy from renewables a five-year electricity-tax waiver, concessional open-access and cross-subsidy charges, and dual grid feeds for loads of 50 MW or more; it promises water “on payment of cost” and encourages recycling. Its window covered data centres incorporated from 1 April 2021 to 31 March 2026, and we found no successor notification.¹¹ Telangana’s Data Centres Policy 2016 promises dual grid networks, renewable open access for up to a third of a data centre’s power, power “at the cost of generation” to eligible sites, and “continuous water supply for heavy duty cooling systems” on a state campus.¹² For Andhra Pradesh and Maharashtra we could not retrieve the governments’ own policy texts, so we report only what the investors have said: Google will work with partners on new transmission lines, clean generation and storage in Andhra Pradesh, and Amazon’s Thane site will connect to the high-voltage grid through its own substation.¹³²²

05Who is paying: three hyperscalers, counted three ways

Google announced “approximately $15 billion (USD) over the next five years (2026-2030)” for a gigawatt-scale AI hub in Visakhapatnam with AdaniConneX and Airtel, and broke ground on 28 April.¹³¹⁴ Microsoft committed US$17.5bn for calendar 2026 to 2029, on top of US$3bn announced in January 2025, and described the total as US$20.5bn when it opened its fourth Indian region, in Hyderabad, on 21 September.¹⁵¹⁶ Amazon said in December 2025 it would invest more than US$35bn across its Indian businesses through 2030, then in June raised the plan to US$48bn, including more than US$21bn for AI and cloud infrastructure in 2026–30; AWS broke ground on a new Hyderabad data centre on 15 July.¹⁸¹⁹²⁰

Public compute. The IndiaAI Mission had onboarded about 38,231 GPUs through 14 providers by March, offered at a subsidised average of ₹65 an hour, which the ministry called about a third of the global average.³ By August the government reported 93.18 lakh GPU-hours sanctioned to 237 projects and 15 providers, against a five-year outlay of ₹10,371.92 crore; Communications Today gives the portal rate as ₹100 an hour, and reports, unconfirmed by the ministry or IndiaAI, that about 30,000 GPUs are deployed against 45,000 committed.²³ On our arithmetic, the hours sanctioned equal about 3% of one year of the onboarded fleet.