Guide

The data centre wave: what is actually planned at Dholera, tier by tier

Published 3 August 2026. Facts verified to 27 July 2026 unless dated otherwise.

Add the announced data centre numbers and Dholera hosts one of Asia's largest digital-infrastructure pipelines. The word doing all the work in that sentence is announced, and this page refuses to let it hide.

The three plans, ranked by evidence

L&T Vyoma leads on sourcing: a Rs 25,000 crore, 250 MW AI-ready green campus with operations around 2028, carried by the Government of India's newsonair channel REPORTED at MoU stage but government-channel confirmed, the strongest evidence class below construction. The company file holds the detail. Adani's reported hub is the giant: up to 7.5 GW, around Rs 3 lakh crore via AdaniConneX REPORTED, press-sourced and appropriately flagged. Tillman Global's Rs 83,000 crore plan is the caution case: enormous and single-sourced, which our ledger marks in its own row rather than averaging away REPORTED. Three plans, three evidence grades, one honest ranking.

Why data centre capital circles Dholera

• Power with a green label: about 300 MW of solar operational and 700 MW under development toward a 1,000 MW Phase 1 DURABLE. Hyperscale operators buy green power contractually and reputationally; a site that generates it adjacently is structurally attractive.
• Land at campus scale: gigawatt-class plans need contiguous serviced ground measured in hundreds of acres, exactly what the TP scheme machinery exists to produce and almost nothing near Indian metros can offer.
• The silicon adjacency: a fab cluster normalises the heavy utilities, water planning and administrative competence data centres also need; the anchor de-risks the geography for everyone after it DURABLE.
• The fibre question, honestly: long-haul connectivity and latency economics for a greenfield site are the part this record cannot yet document from its sources, so it says so plainly instead of gesturing at digital corridors. When documentable facts exist, this page gets them.

The gigawatt reality check

India's entire operational data centre base took two decades to reach single-digit gigawatts. A single site announcing up to 7.5 GW is therefore announcing a multi-decade ambition contingent on power markets, demand growth and capital cycles, not describing a construction schedule, and the difference is exactly what tier tags exist to hold. The honest frame: Dholera has one government-channel MoU of credible scale, and beyond it a pipeline of optionality whose conversion rate is the true story to watch. Our quarterly State of Dholera tracks conversions and expiries both, because a record that only counts announcements is a brochure with a spreadsheet.

What would upgrade this page

Ground-breaking with a dated announcement, a power purchase agreement, an equipment order, or a construction contract award: any of these moves a plan toward DURABLE and gets same-week recognition here. Until then the summary stands: the data centre story at Dholera is real, early, and running about two evidence grades behind its own press coverage, which is normal, survivable and worth saying out loud.

The size of the market Dholera is trying to enter

Any assessment of a proposed data centre cluster has to start with the addressable market, and India's is both large and awkwardly measured. Savills put operational stock at roughly 1.8 GW of IT capacity, with 258 MW added in the first half of 2026, a 59 per cent increase year on year. CBRE counted more than 1,700 MW at the end of 2025, 440 MW of it added during that year, and expected around 500 MW in 2026. Invest India works with a lower figure of about 1.5 GW operational and a conservative 5 GW by 2030. Knight Frank tracks a pipeline of 8.33 GW, of which only 322 MW is actually under construction and 2,920 MW is committed. Wood Mackenzie is the most bullish, modelling 2.2 GW in 2025 rising to 12 GW by 2030, a compound rate near 40 per cent.

These numbers do not contradict each other so much as measure different things. IT load, the power delivered to servers, is not the same as gross facility capacity, which includes cooling and electrical overhead and is typically some tens of per cent higher. Commissioned capacity is not the same as contracted capacity, and neither is the same as a signed memorandum. The gap between Knight Frank's 8.33 GW pipeline and its 322 MW under construction is the single most useful statistic in Indian data centre analysis, because it quantifies how much of the announced sector exists as intention rather than concrete. Roughly four per cent of the pipeline is being built. Anyone reading a Dholera announcement should hold that ratio in mind.

The demand case is not in doubt. The Observer Research Foundation notes that India generates around twenty per cent of the world's data while holding three to four per cent of global data centre capacity, a mismatch that data localisation rules turn into a construction requirement rather than a preference. Wood Mackenzie projects AI-dedicated capacity rising from 275 MW in 2025 to 6,546 MW in 2030, and data centre electricity demand climbing from 10 TWh in 2025 to 191 TWh by 2040, about seven per cent of national consumption. The IndiaAI Mission has allocated more than 38,000 GPUs with a further 20,000 planned. Announced investment across the sector runs to sixty to seventy billion US dollars over five years on Invest India's count.

Why power, not land, now decides where a campus goes

Wood Mackenzie's most consequential finding for Dholera is not a capacity number at all. It is the observation that reliable, cost-competitive power has overtaken land and capital as the defining site-selection factor for Indian data centres. This is an inversion of the logic that built Mumbai, Chennai, Delhi-NCR and Bengaluru into roughly ninety per cent of tier-one capacity. Mumbai alone holds more than half the operational inventory on CBRE's count, and Knight Frank puts 766.6 MW live in the city, the two figures resting on different definitions of what counts as inventory.

The mechanism is straightforward once the cost stack is laid out. Construction runs at single-digit dollars per watt: Turner and Townsend put Mumbai at US$6.64 per watt, the second lowest of fifty-two global markets surveyed. Land is typically a minor line in that total on a large campus. Electricity, by contrast, is a twenty-year operating cost on a load that never switches off. At Indian industrial tariffs, a large campus can plausibly spend more on power across its life than it did to build itself. When the marginal facility is an AI training cluster drawing tens of megawatts continuously, the tariff and the firmness of supply dominate every other variable.

This is the argument Dholera is built on. Gujarat has around 69 GW of installed generation capacity, of which 47 GW is renewable DURABLE. The Dholera Solar Park is planned at 1,000 MW, of which the 300 MW awarded to Tata Power Solar is commissioned DURABLE, with the remaining 700 MW under development by ReNew, SJVN, Vena and TEQ Green through Gujarat Power Corporation Ltd, targeted for March 2027 TARGET. The honest engineering caveat is that solar generation and data centre load have almost opposite profiles. A photovoltaic park delivers a bell curve peaking at midday with a capacity factor typically in the high teens to low twenties in western India. A data centre draws a flat line at close to full load for 8,760 hours a year. Co-located solar reduces the annual energy bill and the emissions accounting, but it cannot by itself keep servers running at three in the morning. Firm supply still comes from the grid, from storage, or from contracted round-the-clock renewable products, and none of those arrangements has been published for any Dholera data centre proposal.

What the Gujarat data centre policy actually pays

The Viksit Gujarat Data Centre Policy 2026-29, announced on 9 July 2026 by Chief Minister Bhupendra Patel, sets a target of roughly Rs 6 lakh crore of investment and 7.5 GW of capacity TARGET. Its qualifying threshold is a minimum 150 MW of IT load DURABLE, which is worth pausing on. That figure excludes edge facilities, enterprise captive rooms and most conventional colocation. It is a policy written for hyperscale and AI campuses only, and it implicitly concedes that Gujarat is not competing for the retail colocation market that Mumbai owns.

The incentive stack has five meaningful components. A 2.5 per cent capital subsidy applies specifically to projects inside Dholera SIR REPORTED, which is the state discriminating in favour of this site rather than Ahmedabad or Jamnagar. There is an interest subsidy of up to 4 per cent, a power tariff subsidy of Rs 1 per unit for twenty years, full reimbursement of electricity duty for twenty years, and a hundred per cent exemption from stamp duty and registration charges on land DURABLE.

The tariff subsidy is the item with real weight, and the arithmetic is worth doing openly. A facility with 150 MW of IT load, once mechanical and electrical overheads are included at a realistic efficiency ratio, draws somewhere near 190 to 200 MW of grid load. Running close to continuously, that is on the order of 1,600 to 1,700 GWh a year. A rebate of one rupee per unit on that consumption is a saving in the region of Rs 160 to 170 crore annually, sustained for two decades. Those are illustrative figures derived from the policy's own threshold and standard load assumptions, not published project numbers, and the rebate's exact base and caps would be set by the scheme's operating guidelines, but they show why a state-level tariff line can move a siting decision that no land subsidy could. Electricity duty reimbursement compounds the same effect. Stamp duty relief, against Gujarat's 4.9 per cent duty plus 1 per cent registration DURABLE, is a one-off saving that matters far less over a twenty-year hold.

The three Dholera proposals, sorted by what is actually evidenced

Three data centre plans are associated with Dholera, and they sit at very different evidentiary levels. The clearest is L&T Vyoma, which has an agreement with the Gujarat government for a 250 MW green AI-ready data centre at Dholera at around Rs 25,000 crore, reported by News on AIR on 20 February 2026 REPORTED. State broadcaster sourcing and a named counterparty put this above the rest, though it remains an agreement rather than a build.

Adani's plan is larger and thinner. The group is reported to be pursuing an AI-ready hub of up to 7.5 GW, in the region of Rs 3 lakh crore, through AdaniConneX REPORTED. The reporting is announcement-stage and the extent to which that capacity is committed to Dholera specifically is not established in any primary document we can point to. It is also worth noting that the figure is numerically identical to the entire state policy target of 7.5 GW, which should prompt caution about whether the two numbers are independent or whether one is an aspiration restated. Separately, the group has announced US$100 billion of investment to 2035 with a 5 GW data centre ambition REPORTED, which is a national rather than a Dholera commitment.

The third is Tillman Global, associated with a hyperscale plan of roughly Rs 83,000 crore REPORTED. This should be flagged hard: it is single-sourced and unverified, and Tillman's documented Indian commitment is a Rs 15,000 crore memorandum with Andhra Pradesh for 300 MW at Visakhapatnam REPORTED, a different state and a different project. Conflating the two is a common error.

What none of the three has, as far as public record goes, is a construction start, a grid connection agreement, a published commissioning date, a named EPC contractor, or an equipment purchase order. On the ratio established earlier, roughly four per cent of India's announced data centre pipeline is under construction, and there is no published evidence placing any Dholera proposal in that four per cent.

The competing geographies, including one inside Gujarat

Dholera is not bidding against Mumbai for the same tenants. It is bidding against a small set of greenfield AI campuses, and the competition is well funded. Google broke ground on a gigawatt-scale AI hub at Visakhapatnam on 28 April 2026 DURABLE, which is a construction event rather than an announcement and therefore ranks above every Dholera proposal on evidence. Microsoft has committed US$17.5 billion to India across 2026 to 2029 and AWS US$8.3 billion to its Mumbai region REPORTED, both anchored on established metros with existing subsea connectivity.

The sharpest competitor is inside Gujarat. Meta and Reliance announced a 168 MW facility at Jamnagar on 9 June 2026, renewable powered and cooled with desalinated seawater REPORTED. Jamnagar carries advantages Dholera does not yet have: an operating industrial complex, captive generation, existing seawater intake and outfall infrastructure, and an anchor tenant that is also the landlord. A hyperscaler choosing between two Gujarat sites under the same state policy will compare execution risk, and one of the two has an operating refinery complex around it. Reliance has separately outlined a Rs 10 trillion artificial intelligence programme over seven years including gigawatt-scale facilities at Jamnagar, and a Gujarat-wide renewables and green hydrogen memorandum of around Rs 5.95 lakh crore REPORTED. Dholera's counter-argument is the fab, the airport, the expressway and the semi high-speed rail line, which together promise a serviced industrial city rather than a fenced campus. That argument depends entirely on those assets being finished and used.

Water, cooling and the arithmetic of a desalination plant

Cooling is where a coastal, arid site becomes either an advantage or a liability. Published planning heuristics for data centre water use vary widely and do not reconcile with each other. One commonly quoted figure puts consumption in the order of 25 to 26 million litres per megawatt of capacity per year, which works out near seven million litres a day for a 100 MW facility. Another puts a 100 MW facility at roughly two million litres a day. The spread is a reminder that these are rules of thumb, not design numbers, and that the answer depends entirely on the cooling method chosen. Both figures assume evaporative or adiabatic cooling, which trades water for electricity: evaporating water removes heat cheaply in power terms, which is why it is used in hot climates. The alternative, closed-loop dry cooling with chillers, uses very little water but raises electrical load and therefore the power bill, and its performance degrades exactly when ambient temperatures peak. Direct-to-chip and immersion liquid cooling, which AI racks increasingly require because of rack densities that air cannot handle, change the internal heat path but do not by themselves eliminate the need to reject heat to the outside world.

Dholera's current supply is a desalination plant of roughly 20 MLD DURABLE. A tender was issued on 9 July 2026 to appoint consultants for a 200 MLD seawater desalination plant TARGET, which is a consultancy appointment and not a construction contract, and the gap between the two is usually measured in years. The Gujarat Budget 2026-27 proposed bulk water pipelines including 100 MLD from Bhaskarpura lake in Surendranagar REPORTED. Applying the heuristics above, a single 250 MW campus would want somewhere between roughly 5 and 17 MLD if evaporatively cooled, so the arithmetic is not obviously prohibitive against a 200 MLD plant, though it would be a material claim on the roughly 20 MLD available today. The competing draw is the one that is already funded: a 300mm wafer fab consumes very large volumes of ultrapure water, and ultrapure water production itself rejects a substantial fraction of feed as concentrate. The fab and any data centre cluster would be drawing on the same desalinated supply, and no published allocation splits that supply between them. Under the Gujarat Semiconductor Policy 2022-27 the state offers a 50 per cent capital subsidy on a desalination plant built within the first five years, alongside water at Rs 12 per cubic metre for five years DURABLE, terms written for the semiconductor investor rather than for data centres.

The connectivity gap nobody has closed on paper

Mumbai and Chennai dominate Indian colocation for a reason that has little to do with land or tariffs: subsea cables land there. International capacity, peering density and carrier neutrality accumulate around cable landing stations, and enterprise and inference workloads are sensitive to round-trip latency in a way that makes proximity to those landing points commercially decisive. Gujarat plans two cable landing stations REPORTED, which would begin to address this, though no landing date, cable system or operator has been published.

The mitigating technical point is that not all workloads care. Large-scale model training is throughput-bound rather than latency-bound: a training cluster ingests data, runs for weeks and writes checkpoints, and the additional round trip from an inland site to a coastal landing point is immaterial to it. That is precisely why remote, power-rich sites are viable for AI training globally in a way they never were for enterprise colocation. It is also why the "AI-ready" framing in the Dholera announcements matters technically, whatever one makes of the announcements themselves.

What remains genuinely unpublished is the terrestrial fibre position. There is no public information on how many carrier-neutral long-haul routes reach Dholera SIR, whether they are physically diverse, who owns the ducts along the NH-751 expressway corridor that opened on 31 March 2026 DURABLE, or what dark fibre is available to a prospective tenant. Trunk infrastructure works in the 22.5 sq km Phase I Activation Area are recorded as complete by the NICDC and DPIIT Delivery Monitoring Unit as of 31 May 2026 DURABLE, and utility ducting is normally part of such works, but a duct is not a lit route with two independent carriers on it. Until a named operator publishes a Dholera point of presence, the connectivity case rests on plans rather than on capacity.