India's AI data centres: IndiaAI Mission, power and capacity in 2026

India has roughly 1.7 to 1.8 gigawatts (GW) of operating data centre capacity, a government GPU pool of more than 38,000 chips offered at subsidised rates, and a tax holiday aimed at foreign cloud firms. The hard part is electricity: official demand projections for AI data centres have already been revised sharply upward.

This article sets out what India is doing, what the numbers say and where sources disagree. Every figure carries its source and date. Nothing here is investment advice, and many company figures are pledges over several years, not money already spent.

Key numbers at a glance

• IndiaAI Mission outlay: Rs 10,372 crore (Government of India, Digital India page, last updated 14 October 2026). A crore is 10 million, so this is about Rs 103.7 billion.
• GPUs onboarded on the national AI compute portal: more than 38,000 (MeitY statement to the Lok Sabha, 25 March 2026, as reported by TelecomTalk on 30 March 2026; same figure on the Digital India page, 14 October 2026).
• Additional GPUs announced: 20,000 "in the coming weeks" (summit announcement, February 2026, as reported by IANS on 2 March 2026).
• Projects approved under the mission: 190 (MeitY, 25 March 2026), rising to 237 with 93.18 lakh (about 9.3 million) GPU hours sanctioned (Lok Sabha disclosure in August 2026, as reported by Communications Today, 1 October 2026).
• Compute pool in May 2025: 34,333 GPUs, made up of 18,417 from the first empanelment plus 15,916 added (PIB, 30 May 2025).
• Operating data centre capacity: about 1,700 MW at end 2025 (CBRE, reported by IANS, 1 April 2026).
• Operating capacity in H1 2026: 1,789 MW (Cushman and Wakefield, reported by ANI, 16 September 2026).
• Installed data centre power capacity: 1.57 GW, up from about 375 MW in 2020 (official factsheet, reported by IANS, 14 September 2026).
• New capacity expected by 2030: 3,860 MW, of which 509 MW under construction (Cushman and Wakefield, September 2026).
• Capacity added in 2025: a record 440 MW (CBRE, April 2026).
• Data centre investment commitments in 2025: 56.4 billion US dollars, cumulative 126 billion (CBRE, April 2026).
• Mumbai share of operating inventory: over 50 percent (CBRE, April 2026).
• AI data centre power demand by FY32: 26.3 GW (Minister of State for Power, written reply in the Rajya Sabha, reported 29 to 31 July 2026), against 13.56 GW in a March 2026 estimate and 17 GW from the Central Electricity Authority (factsheet, 14 September 2026).
• Planned transmission investment to FY32: about Rs 9.16 lakh crore (Rs 9.16 trillion) under the National Electricity Plan (Construction World, 31 July 2026).
• Google Visakhapatnam hub: about 15 billion US dollars, 1 GW (Google announcement 14 October 2025; foundation stone 28 April 2026).

What India is trying to do

India's approach has three parts. The state pays for some AI compute itself. It offers tax incentives to pull in foreign capacity. And it relies on large private groups to build the buildings and the power supply.

The state part is the IndiaAI Mission. The Digital India page, last updated 14 October 2026, says it started with an outlay of Rs 10,372 crore and has onboarded more than 38,000 GPUs through a compute portal, offered to startups and academia at affordable rates. A GPU is the graphics processor chip that trains and runs AI models. "Onboarded" means the government has contracted capacity from private providers and listed it on the portal. It does not mean the government owns the chips.

The 25 March 2026 reply to the Lok Sabha, as reported by TelecomTalk, said 190 projects had been approved. Of those, 78 were government entities, 46 were startups and small firms, 30 were early stage startups, 27 were researchers or academia, 5 were students and 4 were early stage researchers. That mix is useful. It shows that a large share of the subsidised compute goes to public bodies, not only to private startups.

By August 2026 the count had grown. Communications Today, reporting on 1 October 2026 a government disclosure in the Lok Sabha, gives 237 approved projects, 93.18 lakh GPU hours sanctioned and 15 compute service providers empanelled. A lakh is 100,000, so that is about 9.3 million GPU hours. A GPU hour is one chip running for one hour, which makes it a better measure of what is actually handed out than a count of chips.

How the subsidised compute works

The mission does not build its own data centres. It runs tenders. Private cloud and data centre companies bid to supply named classes of chips, and the lowest bid in each class is published as the L1 price, a standard term in Indian public procurement. Empanelled providers supply the capacity, and approved users apply through the portal and pay a subsidised rate.

The clearest primary record of that process is a Press Information Bureau release of 30 May 2025. It said 15,916 GPUs had been added to the 18,417 already empanelled, for a total of 34,333. Seven bidders quoted in that round, including Yotta Data Services, Sify Digital Services, Netmagic IT Services and Cyfuture India. The release published the L1 prices per hour in rupees. Two examples show the range: an NVIDIA B200 SXM at Rs 323 an hour on demand (Yotta), and an NVIDIA H100 NVL at Rs 168.74 an hour on demand (Netmagic). The same table covers AMD MI300X and MI325X, Intel Gaudi 2 and Gaudi 3, and AWS Trainium and Inferentia2 units, so the pool is not only one chip maker's hardware.

Those L1 figures are what the government pays providers before any subsidy, as quoted in 2025. They are not the price a startup sees. Communications Today reports a subsidised rate of Rs 100 an hour on the portal, set against a global market rate of 2.5 to 3 US dollars an hour, citing a PIB factsheet. That is a reported average, not a published rate card for each chip, so check the portal for the rate that applies to a given unit.

The pool also serves a second goal: home grown foundation models, meaning large general purpose AI models trained from scratch. The same May 2025 release said 506 proposals had been received by 30 April 2025. Sarvam AI was selected on 26 April 2025 to build an open source model of 120 billion parameters. Three more were named in May 2025: Soket AI (an open source 120 billion parameter model), Gnani AI (a 14 billion parameter voice model) and Gan AI (a 70 billion parameter multilingual model aimed at text to speech). Parameters are the adjustable numbers inside a model; more parameters usually means more compute to train. These teams draw on the same subsidised pool, which helps explain why the government keeps adding chips.

The GPU count: from 18,417 to 38,000 plus

The official series runs like this. The first empanelment brought 18,417 GPUs. The May 2025 round took the total to 34,333 (PIB, 30 May 2025). Official pages in March and October 2026 say "more than 38,000". An extra 20,000 was announced at the February 2026 summit as coming "in the coming weeks" (IANS, 2 March 2026).

Higher totals circulate in media coverage. Communications Today noted on 1 October 2026 that reports of about 30,000 GPUs deployed against 45,000 committed had not been confirmed by MeitY, IndiaAI or C-DAC. As of 19 October 2026 no official page confirms that the additional 20,000 are live, so treat them as announced.

If you see a higher figure, ask three questions. Is it onboarded or only tendered? Is it GPUs or GPU hours? Is it dated? A tender result, an empanelled count and chips in active use are three different things, and each can be quoted as "India's GPUs".

Data centre capacity: three trackers, three numbers

Market trackers disagree slightly because they count differently. Some count IT load (the electricity reaching servers). Some count gross power. Some count only colocation (space rented to many customers).

• CBRE (reported by IANS, 1 April 2026): about 1,700 MW at end 2025, with roughly 500 MW of new supply expected in 2026, around 30 percent growth.
• Cushman and Wakefield (reported by ANI, 16 September 2026): 1,789 MW operating in H1 2026, up from about 400 MW in 2019. The top seven markets added 178 MW in H1 2026, and colocation vacancy was 14.9 percent.
• Official factsheet (reported by IANS, 14 September 2026): 1.57 GW installed, projected near 8 GW by 2030.

The projections to 2030 spread even wider. Cushman and Wakefield adds 3,860 MW by 2030, which would put the total near 5.6 GW if added to its 1,789 MW base. The factsheet says nearly 8 GW. These are not contradictions so much as different methods and different dates. The fair summary is that India sits at under 2 GW now and the forecasts for 2030 run from about 5.6 GW to about 8 GW in these two sources, while news reports cited by Angel One News on 29 July 2026 give 12 GW.

Capacity is concentrated. CBRE puts Mumbai at over half of operating inventory, and says Mumbai, Chennai, Delhi NCR and Bengaluru together hold nearly 90 percent. The city by city detail, including Hyderabad's pipeline and the new Visakhapatnam hub, is in our India AI data centre map. For the wider region, see the Asia Pacific regional overview.

Money: pledges and the tax holiday

Headline numbers are large, and they are mostly commitments over many years. An official statement at the India AI Impact Summit in February 2026, reported by IANS on 2 March 2026, put expected AI related investment at over 200 billion US dollars, built from company pledges that run to 2035 in some cases. The largest single site is Google's planned 1 GW hub in Visakhapatnam, announced in October 2025 at about 15 billion US dollars, with its foundation stone laid on 28 April 2026. What each pledge covers, and over what years, is broken down in India's AI investment pledges decoded. An announced pledge is not a contract, a building permit or a grid connection.

The policy lever is tax. The Union Budget on 1 February 2026 proposed a tax holiday until 2047 for foreign companies that provide cloud services to customers worldwide using data centres in India. As reported by the Free Press Journal on the day, the relief applies only if services to Indian customers are routed through an Indian reseller entity. That report does not say which taxes are covered or how the reseller rule is enforced, so check the enacted Finance Act before relying on it.

Power: the real constraint

Data centres run all day, so they need steady supply, not just peak supply. This is called baseload.

India's official numbers on how much extra power AI data centres will need moved sharply this year:

• March 2026 government estimate: 13.56 GW by FY32 (FY32 is the financial year ending March 2032).
• Written reply in the Rajya Sabha by the Minister of State for Power, Shripad Naik, reported 29 to 31 July 2026: 26.3 GW by FY32, expected to be integrated into the grid and served primarily by renewable energy.
• Central Electricity Authority figure in the 14 September 2026 factsheet: demand could reach 17 GW by FY32.

The reports do not make clear whether the 17 GW and 26.3 GW measure the same thing. Inc42 quotes the reply as "an additional load of 26.3 GW from artificial intelligence (AI) data centres", while Construction World says "data centres are projected to consume 26.3 GW", and the factsheet's 17 GW may use a different scope or date. The safe reading is that official estimates range from about 14 to 26 GW for FY32, and they have been moving up.

For context, the same reply cited CEA figures that overall peak demand would rise from 289 GW to 388 GW between FY27 and FY32. Construction World adds that peak demand had already reached 270.8 GW. On that basis, data centres would be a meaningful but not dominant share of growth.

What the government says it is doing about it, according to Construction World's 31 July 2026 account of the reply:

• The CEA is reassessing how demand is mapped, because data centres clustering in a few states could skew consumption and affect grid stability.
• States and distribution companies will be asked to include data centre projections in their resource adequacy plans. Resource adequacy means having enough firm generation and contracts to meet expected demand with a safety margin.
• Emerging loads are being assessed in the 21st Electric Power Survey, and data centre demand will be factored into the National Electricity Plan for FY28 to FY37, which the ministry expects by March 2027.
• Transmission is being built in phases. The National Electricity Plan estimates about Rs 9.16 lakh crore of transmission investment by FY32.
• Grid support equipment under consideration includes STATCOMs (fast voltage control devices), battery energy storage systems and synchronous condensers, which steady the grid as more solar and wind connect.

Inc42, citing the Indian Express, adds that the ministry has discussed with industry how to absorb the spiky load of AI data centres and whether to site them away from urban centres to avoid transmission congestion. Local opposition is already visible: the same report says residents in Thane, Maharashtra, oppose a proposed 422 MW Amazon data centre over diesel generator noise, power demand and tree felling.

CBRE notes that India added a record 44.5 GW of renewable capacity in 2025 and that operators are focusing on renewable sourcing. Solar and wind do not run around the clock, so data centres still need storage, other generation or grid backup. No official plan published so far sets out how that gap will be filled for data centre load specifically.

Data rules and the sovereignty question

Where the data sits affects where the machines go. The Digital Personal Data Protection Rules were notified on 14 November 2025 by MeitY, as reported by SCC Online. Most operative rules come into force 18 months after publication, which points to around May 2027. According to SCC Online's summary, significant data fiduciaries (the larger or riskier data handlers) must ensure personal data and related logs are not transferred outside India unless permitted, and wider cross border transfers are allowed subject to conditions set by the Central Government. The summary does not list which data or which conditions, so check the Gazette notification. Our guide to what must stay in India covers these rules in more depth.

If you sell software into India, this is also a software supply chain matter. India's CERT In has published SBOM guidance, covered in our explainer on CERT In SBOM guidelines version 2.

What is firm and what is uncertain

Firm, because a government source states it:

• The IndiaAI outlay of Rs 10,372 crore and the 38,000 plus GPU figure, on the Digital India page dated 14 October 2026.
• The May 2025 tender round, its 34,333 GPU total, the L1 prices and the four foundation model selections, from the PIB release of 30 May 2025.
• The 190 approved projects and their breakdown, from the 25 March 2026 reply.
• The 1 February 2026 Budget proposal for a tax holiday to 2047.
• The DPDP Rules notification in November 2025 and the 18 month phase in.

Reported by market trackers or media, and method dependent:

• Total capacity (1.57, 1.7 and 1.789 GW) and the 2030 forecasts.
• City splits, and the August 2026 project and GPU hour totals, which come via a trade publication.

Uncertain or disputed:

• Whether the extra 20,000 GPUs are live.
• The investment amount for the Google hub, where two figures were given in October 2025 (see the pledges post linked above).
• The subsidised hourly rate users pay for each chip type.
• The 14 to 26 GW range for FY32 power demand, and how each number is defined.
• Whether summit pledges turn into built capacity. Company pledges are unaudited statements, and some span seven to ten years.

What to watch next

• A fresh MeitY or Lok Sabha statement on the GPU count and the status of the additional 20,000.
• The enacted text and notified rules for the 2047 cloud tax holiday.
• CEA planning guidance asking states to include data centre demand in their resource adequacy plans, and the National Electricity Plan for FY28 to FY37, expected by March 2027.
• Commissioning dates at the Visakhapatnam site and any energy supply agreements.
• The DPDP commencement date around May 2027 and the cross border conditions.
• The next CBRE and Cushman and Wakefield half year reports.

For related context on how large AI buyers contract for compute, see how Anthropic's compute commitments work and OpenAI's Stargate and 26 gigawatt commitments.

If your team builds or sells software for this market, Safeguard offers SBOM and dependency scanning, and a free scan is available at https://gold.safeguard.sh/scan.

Quick check

Sources

• IndiaAI Mission expands AI ecosystem with affordable compute and startup support, Digital India, last updated 14 October 2026
• Over 38,000 GPUs onboarded on AI compute portal under IndiaAI Mission, TelecomTalk, 30 March 2026
• India's data centre capacity to grow 30 pc in 2026, IANS reporting CBRE, 1 April 2026 (ianslive.in)
• India to add 3,860 MW data centre capacity by 2030, ANI reporting Cushman and Wakefield, Deccan Chronicle, 16 September 2026
• India's data centre capacity likely to touch GW by 2030, IANS reporting an official factsheet, 14 September 2026
• India's AI data centre power demand to reach 26.3 GW by FY32, Angel One News, 29 July 2026
• Union Budget 2026 tax holiday till 2047, Free Press Journal, 1 February 2026
• Over 200 billion dollars to be infused in AI related infrastructure in India, IANS via Morung Express, 2 March 2026
• Google to invest 15 billion dollars in Visakhapatnam AI hub, UPI, 14 October 2025
• Google's 1.35 lakh crore AI data centre in Visakhapatnam, Drishti IAS, 29 April 2026
• India's common compute capacity crosses 34,000 GPUs, PIB release of 30 May 2025 (PDF copy hosted by NeGD)
• IndiaAI Mission faces gap between capacity and deployed GPUs, Communications Today, 1 October 2026
• AI data centres to consume 26.3 GW power by FY32: MoS Power, Inc42, 29 July 2026
• Government projects 26 GW data centre demand by FY32, Construction World, 31 July 2026
• Digital Personal Data Protection Rules 2025, key highlights, SCC Online, 26 December 2025

Last reviewed: 19 October 2026.