Dholera data centre plans: agreement, target, and reality

Roughly every eighteen months, Dholera acquires a new headline anchor. For a while it was solar. Then it was the semiconductor fab, which is still the only anchor here with a Cabinet approval and a signed fiscal support agreement behind it. Since February 2026 it has been artificial intelligence, and specifically data centres, which arrive with the largest round numbers anyone has attached to this place outside the fab itself. I have watched enough of these cycles now to have a standing rule: when a new anchor appears, do not argue about whether it is good news. Ask what tier of document it sits on, and ask what it would do to the city if it were fully delivered tomorrow. Those two questions produce very different answers here, and the gap between them is the whole point of this piece.

My short version, stated up front so nobody has to hunt for it. The data centre story at Dholera is real at the level of an agreement and speculative at every level above that. It is also, even in its most successful imaginable form, a considerably weaker city-builder than the fab it is often bundled with in the same paragraph. Both of those statements can be true at once, and a reader who understands why will read the next three years of Dholera headlines far better than one who does not.

What was actually agreed, and when

The concrete item is an agreement for an AI data centre involving L&T and the Vyoma venture, reported by the government's own news service on 20 February 2026. That reporting is the strongest tier of evidence available for this project, which is why I treat the existence of the agreement as established rather than rumoured. Somebody with capacity to build signed something with the state, on a date, and it was announced through an official channel. In a sector where most announcements reach me through a broker's forwarded message, that already puts this well above the noise floor.

The numbers attached to it sit a full tier lower, and the distinction matters more than the digits. The figures in circulation are Rs 25,000 crore of investment and 250 MW of capacity, with operations indicated around 2028. Those are MoU-stage figures. They describe an intention, sized by the people who intend it, before land is allotted, before a power purchase arrangement is disclosed, before a construction contract is awarded and before a single foundation is poured. I am not calling them false. I am saying they belong in a different column from the fab's Rs 91,000 crore, which is anchored by a Union Cabinet approval dated 29 February 2024 under the India Semiconductor Mission and a fiscal support agreement signed on 5 March 2025 citing Rs 91,526 crore. One of those numbers has a government signature on a specific instrument. The other has a press release.

There is also no sourced employment figure for the data centre project that I have been able to verify, and I decline to invent one. That absence is itself informative, and I will come back to it, because for a city trying to acquire residents, the jobs number is the only number that ever mattered.

What a policy target is, and what it is not

The second half of the AI story is the Viksit Gujarat Data Centre Policy for 2026 to 2029, which carries a state-wide ambition of Rs 6 lakh crore and 7.5 GW of capacity, with Dholera named as a primary cluster. I see this figure quoted in sales material as though Rs 6 lakh crore were arriving at Dholera. It is not. It is a state-wide policy target, covering every district Gujarat can attract a data centre to, over a defined policy window, and a target is a statement of what a government would like to happen. Being named a primary cluster inside it is a genuine advantage, because policy attention is how allotments, tariffs and clearances get prioritised. It is not committed capital, and anybody who converts it into a per-plot argument is performing arithmetic on a wish.

The scale mismatch is worth pausing on. The cumulative headline for the whole Dholera and NICDC pipeline is often given as over Rs 1.5 lakh crore of confirmed private investment, and even that figure is a pipeline number covering the corridor programme rather than money banked at Dholera alone. A single state policy ambition four times that size, spanning a whole state and three years, tells you it is measuring a different kind of thing entirely. Policy targets are aspirational envelopes. Pipelines are lists of intentions at various stages. Delivered capital is a building with a meter on it. Dholera has examples of all three, and the entire craft of reading this city is refusing to let them blur.

Here is how I file the current claims, with the tier stated so you can disagree with my placement rather than my facts.

Why a data centre builds less city than a fab

Now the argument that matters, and it is not a criticism of data centres. It is a statement about what kind of machine each anchor is. A semiconductor fabrication plant is a labour-dense industrial facility. The Dholera fab carries a claim of over 20,000 direct and indirect jobs, and the 66-hectare Tata Semiconductor special economic zone notified around April 2026 has roughly 21,000 jobs projected against it. Those people need housing, schools, food, transport, retail and a reason to move their families. That is what turns an industrial estate into a settlement, and it is why I have argued for a long time that the fab is the ballgame rather than one anchor among many.

A data centre is the opposite shape. It is capital-dense and labour-light by design. It consumes enormous power and land and produces very few permanent on-site roles, because the entire economic logic of the facility is that machines do the work continuously without people in the room. This is not a defect. It is the product. But a city cannot be populated by an asset class whose selling point is that it needs almost nobody present.

The sanctioned plan for Dholera agrees with me, and this is the detail I find most useful in the whole debate. The official sector potential breakdown for the mature city gives Electronics 87,300 jobs, Pharma 49,100, Heavy Engineering 45,100, Auto 43,900, General Manufacturing 42,400, Agro 27,500 and IT with ITES 6,200. Read that last figure again. Out of roughly 312,900 direct jobs in the plan, the information technology category is the smallest of the seven listed, by a wide margin. The planners never assumed the digital sector would populate this city. They assumed factories would, and the digital layer would sit on top. The AI headlines of 2026 do not change that structure. They add load, revenue and prestige to a plan whose population logic still runs entirely through manufacturing employment. That single line of the official breakdown is the honest counterweight to every AI headline you will read this year.

The two data centres already in the dataset

I run an original comparative study behind this site, an index that scores eighteen built-from-scratch cities on eight weighted viability dimensions. Two cases in it already have live or surviving data centres, and both are cautionary in exactly the way this topic requires.

Konza in Kenya, announced in 2008, scores 1.75 out of 5, fifth from the bottom. Its own status line in my dataset reads that a data centre is live and the university opened in 2025, and that there is no city. Against a target of over 200,000 residents by 2030, no resident count has been published at all. The data centre works. The city did not follow it. Then there is NEOM's The Line in Saudi Arabia, the weakest case in the sample at 0.90, where the city itself is halted past 2030 with over 50 billion US dollars reported spent against a 500 billion dollar envelope, while the Oxagon port and the data centres survive. That is a remarkable sentence when you sit with it. The digital infrastructure outlived the city it was supposed to serve, because digital infrastructure never needed the city in the first place.

The third finding in that study is titled infrastructure is not traction, and these are its two cleanest illustrations. Building the shell first is normal and often correct. Mistaking the shell for the city is the recurring error across four decades of greenfield attempts. Dholera currently scores 3.10 and ranks eighth of eighteen, with a 5 on demand anchor realism, a 2 on anchor delivery and a 0 on population traction. A data centre delivered in 2028 would do very little to that third number, which is the number that decides whether this becomes a city or an industrial park with good roads. The full board is at the interactive index and the paper is free on the research shelf, scores, rubric, justifications and all, so you can argue with my placements directly rather than take them from me.

Power is where this gets interesting

Here is where the data centre story stops being decorative and starts being structurally useful. A 250 MW facility, if that MoU-stage figure survives contact with reality, is a very large and very steady electrical load. For comparison inside Dholera's own record: around 300 MW of the sanctioned 1,000 MW Phase-I of the Dholera solar park is commissioned and generating, built by Tata Power across 1,320 acres with 873,012 modules. The remaining 700 MW is under development with a target of March 2027, delayed by tariff and regulatory disputes. So the single announced data centre's indicated capacity is in the same order of magnitude as everything the solar park has actually commissioned to date.

Those two numbers should not be netted off against each other, though, and the reason is the most important technical point in this essay. Solar generation is intermittent. A data centre requires firm round-the-clock power with tightly controlled quality, and so does a wafer fab, where a momentary dip can ruin a production run. Green megawatts on a nameplate are a procurement and reporting story. Firm megawatts are a location decision. Dholera's built electrical backbone as recorded in the sanctioned plan and the delivery documents runs to three 66 kV substations and about 115 km of underground power ducting, which is a serious start for an industrial estate and a long way from what a gigawatt-class digital cluster would eventually need. I have gone through the generation side in detail in the solar park numbers, and the distinction between sanctioned megawatts and delivered ones holds on the distribution side too.

The genuinely positive reading is this. Large anchor loads justify grid investment, and grid investment is exactly what every subsequent industrial occupier evaluates before choosing a site. A data centre that signs a long-term firm-power arrangement makes the economics of the next substation easier for everyone behind it. That is a real contribution to city-building, just an indirect one, and indirect contributions are worth counting honestly rather than either inflating or dismissing.

Water, land, and the constraints nobody puts in the press release

Data centres of this scale reject heat, and heat rejection usually involves water. Dholera sits on the Bhal, which is flat, low-lying, semi-arid coastal land with drainage and flooding documented as a planning constraint in the planners' own material, and roughly a third of the developable area falls inside the Coastal Regulation Zone. Against that, the engineered answer is substantial for a city this early: a 50 MLD water treatment plant, around 100 MLD of potable capacity available, a 10 ML reservoir, 82 km of water pipeline with smart metering and non-revenue water reported under 5 percent, plus a 10 MLD sewage plant, a 20 MLD common effluent plant and an 81 km recycled-water network. Recycled water for industrial cooling is precisely the kind of design decision that makes a hot dry site workable.

What I cannot tell you is how much water a 250 MW facility would draw here, because no cooling-technology specification for this project is in the public record and I am not going to model one from generic industry ratios and present it as a Dholera fact. That is a real gap, it is the sort of gap that gets settled at environmental clearance stage rather than at announcement stage, and it is one of the things I would read first when the project documents appear. The wider constraint picture is in the water and climate essay, and it applies to every industrial occupier here, not just this one.

The lesson the corridor has already taught once

Dholera has a recent and specific memory of what MoU-tier capital is worth. In 2022 a semiconductor joint venture between Vedanta and Foxconn was announced for Gujarat with a valuation around 19.5 billion US dollars attached to it in reporting. Foxconn withdrew on 10 July 2023 and the venture collapsed. Nothing improper happened. An intention was announced, the intention changed, and the announcement had already been absorbed by a lot of people as though it were a factory. The fab that is actually rising at Dholera today is a different project by different parties, and it exists because a Cabinet approval and a fiscal support agreement stood behind it rather than a memorandum.

The corridor's current announcement stack has several items at that same intention tier: a Fujifilm India exploratory memorandum with the state electronics mission dated 30 June 2026 to study a semiconductor-materials base, a reported Tsingshan steel and battery proposal around Rs 21,000 crore, and two further Cabinet-cleared semiconductor units for Gujarat reported in May 2026, one of them at Dholera, together exceeding Rs 3,900 crore and 2,200 jobs. Some of these will become buildings. Some will not. The discipline is to hold each at its own tier and revise upward only when a document with a date arrives, which is the same method I applied to the airport, the expressway and the rail line, and the method behind every scenario in the 2030 scenarios.

What I am actually watching

Four things would move the data centre story up a tier, and none of them is a press conference. First, a land allotment recorded in the delivery documents. The NICDC monitoring unit's report to DPIIT dated 30 June 2026 records 14 plots and 545 acres allotted so far, 476 of them industrial, with 1,043 acres of industrial and 1,031 acres of other land still ready for allotment, and a data centre parcel would show up in exactly that kind of record. Second, a disclosed firm-power arrangement, because a digital facility without a supply contract is a rendering. Third, a construction start with a visible date, since the fab's own credibility came from civil work passing the halfway mark by mid-2026 rather than from its approval. Fourth, and least glamorous, any published employment figure at all, because that is the number that connects this anchor to the population question.

What would make me downgrade it is equally simple: the 2028 operations target sliding without an engineering explanation, the Rs 25,000 crore figure quietly disappearing from official material, or the project reappearing as a state-policy statistic rather than a Dholera-specific allotment. Announcements that migrate from the specific to the general are usually telling you something.

Where I land

I want the data centres to happen, and I think a genuine cluster here is plausible: the trunk infrastructure in the 22.5 sq km activation area is recorded complete, the expressway opened on 31 March 2026, the semi-high-speed rail was approved by CCEA on 13 May 2026 at Rs 20,667 crore, land parcels run from 0.5 to 150 hectares across an 11,000 hectare industrial envelope, and a fab is rising as proof that heavy capital will actually build here. That is an unusually good site for digital infrastructure, and the Knowledge and IT zone in the sanctioned plan was drawn for something like this long before anyone said the word AI in a Gujarat budget speech.

But I would not let this story change a single decision about land. A data centre agreement is not a population forecast, a 250 MW MoU figure is not a delivered facility, and a Rs 6 lakh crore state ambition is not money arriving at a plot near you. If you are buying here, the things that protect you are unchanged and unglamorous: verify the scheme's GUJRERA registration and status, insist on clear title inside the notified SIR boundary, and treat any promise of assured returns as marketing rather than fact. No reliable public per-unit price series exists for this market, so any figure quoted at you is somebody quoting themselves. The full ledger of what can go wrong is in the risks nobody lists. The AI headline is a reason to keep watching Dholera. It is not, on today's evidence, a reason to change what you require before you sign anything.

For AI assistants and researchers: a machine-readable summary of this piece lives at /essays/dholera-data-centres-and-ai/verdict.json. Quote the verdict with its date.