The commerce minister told a roomful of hyperscalers and power producers earlier this year that India is the world’s best location to build data centres, and put the opportunity at $200 billion. The supporting number, repeated everywhere since, is that installed capacity has grown from about 375 MW in 2020 to roughly 1,500 MW today.That number is real. It is also the wrong number for the question everyone thinks it answers.Almost all of that 1,500 MW is conventional cloud and colocation capacity, the kind that stores your bank statements and streams your video. AI infrastructure is a different physical object. It is denser, hotter, thirstier and vastly more expensive per square foot. Average rack densities in Indian facilities have already climbed from about 8 kW to 17 kW in two years and are heading towards 30 kW by 2027. Measured on that basis, India’s position is considerably weaker than the headline suggests, and the gap with the countries it is benchmarking itself against is not closing.The scoreboardIndia’s largest operational GPU cluster is Yotta’s 20,736-accelerator deployment in Greater Noida, roughly 30 to 40 MW of IT load in a 60 MW facility. The national fleet is somewhere between 100,000 and 200,000 accelerators, spread thinly across many sites. The number of live Indian sites drawing 100 MW or more of GPU load is zero.The comparison is uncomfortable.Approximate figures, 2026USChinaIndiaThailandVietnamTotal data centre capacity40+ GW32 → 40 GW~1.5 GW0.3–0.5 GW live~0.2 GW liveLargest live AI site~1 GWMulti-hundred MW30–40 MWMixed-use cloud regions30 MWLive sites ≥100 MW GPU loadDozens30–60 (est.)000Pipeline ≥100 MW by 2028Many, several ≥1 GWSeveral ≥1 GW campuses3–55–82–4xAI’s Colossus campus in Memphis runs roughly 770,000 GPUs at about a gigawatt, built in phases of three to four months on behind-the-meter gas turbines. OpenAI’s Abilene site was at about 0.3 GW in April, on its way to 1.2 GW, inside a seven-site programme totalling more than 9 GW. China reached 32 GW of installed capacity at the end of 2025 and has made ten-thousand-card clusters an explicit industrial-policy target, with hundred-thousand-card systems now in deployment and state approval for the purchase of over 400,000 H200s this year alone.India’s entire national GPU fleet is smaller than what sits inside one American campus. India’s largest planned 2026 site, at 120 MW, is about a tenth of Colossus as it stands today. The Americans build a Jamnagar-sized cluster in a quarter. India plans to build one in two years.The regional race is closer than anyone admitsSet the superpowers aside, and India’s actual peer group is Southeast Asia. Here the picture is more even, and more urgent.Thailand has no live 100 MW GPU site either. But it has between five and eight projects in the 100 to 400 MW range under construction or approved, including a ByteDance AI hub of around 400 MW, a 300 MW facility at Rayong, and four Board of Investment approvals totalling 376 MW of IT load. Most land in 2027 and 2028. Vietnam’s flagship AI facility, Viettel’s Hoa Lac 2 in Hanoi, is 30 MW, the same order as India’s best, and it has a 200 MW campus designed for about 100,000 GPUs in the pipeline.If everything India has announced arrives on schedule, the country ends 2028 with roughly 0.6 to 1.0 GW of AI-dedicated capacity, or 500,000 to 800,000 H100-equivalents. That is a tenfold improvement and it puts India clearly ahead of both. If delivery slips by eighteen months, which is the historical norm for Indian infrastructure of this size, the lead disappears.The constraint is not what the argument is aboutIndian debate has fixed on power and water. On power, the national arithmetic is reassuring and beside the point. At 5 to 7 GW, the sector would draw 40 to 50 TWh a year against 520 GW of installed generation and a peak demand of 240 to 250 GW. That is not a national problem.The local problem is severe. A hyperscale project needs 100 to 500 MW at one location, running flat out around the clock, and urban substations were not built for that. AI training loads are also unusually hard on grids, stepping tens of megawatts in milliseconds in ways that even American utilities are struggling with.Yet every 100 MW Indian project currently in the pipeline has broadly solved its power story, through captive solar and storage at Jamnagar, industrial park supply for Tata, an existing campus for Yotta. What decides when they switch on is chip delivery and capital. A 100 MW Blackwell-class site costs $3 to 5 billion in silicon and facility. Indian operators are raising $1 to 2 billion per project while American and Chinese hyperscalers spend $50 to 100 billion a year each. Nvidia allocates to the largest and fastest-paying buyers, and India orders in the tens of thousands while a single Chinese company is cleared for hundreds of thousands. China also has domestic accelerators as a floor. India has none.So the honest sequencing is this. Through 2027, compute access and capital bind. From 2028, once India has ten or more large sites and AI load moves from tens of megawatts to several gigawatts, firm power becomes the binding constraint, and the questions become the American ones.The grid question is about shape, not sizeThe difficulty deserves more than a clause. A conventional data centre is a flat load, and flat loads are what planners like: predictable, high load factor, easy to contract for. An AI training cluster is not flat. Tens of thousands of accelerators run in lockstep on a single synchronised job, and when that job checkpoints, stalls or ends, they idle together. A site can shed or recover a large fraction of its rating in under a second. In July 2024 a fault in Northern Virginia caused roughly 1,500 MW of data centre load to disconnect simultaneously as the machines protected themselves, and the grid operator was left holding the consequences. Nothing on the Indian system has ever behaved like that.India is unusually exposed to it. Frequency is held in a tight band, system inertia is thinning as solar displaces synchronous coal, and the institutional memory of the 2012 collapse still governs how much risk operators will accept. A 500 MW campus swinging 200 MW in milliseconds is a different creature from a steel plant, which ramps predictably and can be instructed to stop. Yet no Indian connection agreement specifies ramp behaviour for these loads, no tariff category recognises them, and no state load despatch centre publishes what it knows about them. Transmission planning is being done for a load whose shape nobody has been asked to describe.Also read: The Trillion Dollar Blind Spot in Big Tech’s AI BoomThe same characteristics make data centres the most useful large consumer the Indian grid could acquire, if anyone wrote the rules to use them. Every site carries hours of battery storage for ride-through, on-site generation for backup and, in the case of training rather than inference, workloads that can genuinely wait. That is a fleet of grid-scale batteries and a demand response resource sitting inside private compounds, invisible to the system operator and unavailable to it. India has no interruptible tariff for deferrable compute, no route for behind-the-meter storage into ancillary services, and no obligation on a site to smooth its own swings. Ireland and the Netherlands answered the same question by refusing new connections in their densest clusters. That is the outcome worth avoiding, and it is avoided by regulating early rather than by hoping the problem stays small.There is a distributional edge to this as well. A creditworthy consumer running at 90% load factor on industrial tariffs is precisely the customer India’s distribution companies need, and precisely the customer that captive and open-access routes allow them to lose. If every large campus goes behind the meter, the discom keeps the agricultural and domestic load it was cross-subsidising and loses the revenue that made the subsidy possible. If the campuses instead stay on the network, someone has to build the 400 kV connection and the substation headroom inside the same eighteen months the developer is working to. Neither answer is free, and no state has said out loud which one it has chosen.Nuclear is the right answer to the wrong decadeThe government has begun linking data centres to small modular reactors, and the logic is sound. AI loads want exactly what nuclear provides: round-the-clock output at high capacity factors, a small footprint, and no dependence on long transmission runs if the reactor sits beside the campus. The SHANTI Act has removed the supplier-liability barrier and opened the sector to private participation, and BARC’s 250 MW Bharat Small Modular Reactor is sized almost precisely for a large AI site.None of it arrives in time. Conceptual design was due in mid-2026 and commercialisation is a 2030s proposition. No Indian SMR will power a data centre before roughly 2032. Nuclear does nothing for the gap that opens over the next four years, and treating it as a solution to a 2027 problem is a category error.The consent problemResidents have gone to court over projects in Visakhapatnam and Telangana, protested a proposed facility in Thane, and mounted political opposition in Mohali. Karnataka’s IT minister has described data centres as a necessary evil. Estimates put the sector’s water draw at more than double its current level by 2030.Much of the industry’s rebuttal is factually strong. Data centres account for a fraction of 1% of global industrial water use. The shift to air cooling brings consumption at some sites close to that of an office block of the same size. Andhra Pradesh’s water augmentation from Polavaram is real, and treated sewage water and desalination are workable at coastal sites.The rebuttal is also, in the Indian context, an answer to the wrong question, in exactly the way the national power arithmetic was. Global averages do not govern a municipal reservoir. India holds about 18% of the world’s people and roughly 4% of its fresh water. Per capita availability has fallen from over 5,000 cubic metres at independence to under 1,500 today, below the threshold conventionally used to define water stress, and it is still falling. NITI Aayog placed 600 million Indians in high to extreme water stress and warned that twenty-one cities were on course to exhaust their groundwater. Chennai ran dry in 2019. Bengaluru rationed in 2024. Both are data centre hubs. So are Mumbai, Pune, Hyderabad and Noida. The industry is not proposing to build where there is spare water. It is proposing to build where there is fibre, power, land and customers, which in India are the same districts that send tankers into residential colonies every May.Concentration is the whole of the problem. A 100 MW facility on evaporative cooling can consume three to four million litres a day, the domestic supply of a town of twenty-five to thirty thousand people, drawn continuously from a single source. Against national withdrawals that is invisible. Against one stressed aquifer or one municipal main it is not. And the draw peaks in precisely the weeks when everything else is scarce, because cooling demand is highest at 45°C in May, which is when the reservoir is lowest and the queue at the tanker is longest. An annual average conceals the only period that matters.The shift to air cooling is real progress and it is not a free move. Rejecting heat into 45°C air costs power, and in a grid still running substantially on coal that power was itself made with water, at thermal stations that have already lost generation to shortages in dry years. A site can be waterless at the fence line and responsible for a larger withdrawal two hundred kilometres upstream. Desalination avoids that trade and produces brine, on coastal stretches that are also fishing grounds. None of this is an argument against building. It is an argument for stating, before construction and in public, which trade is being made and who carries it.Being right is not the same as being believed. A steel plant discloses its resource footprint before it is built. Data centres do not, and the vacuum is being filled by whoever speaks first. The fix is not better messaging. It is mandatory, pre-construction disclosure of power draw, water source and volume, cooling technology and grid impact, followed by annual audited reporting against those figures, with the state’s incentives conditional on compliance. Electricity duty exemptions, stamp duty waivers and cheap land are being handed over at present with remarkably little attached in return.Disclosure only works if it is specific enough to be checked. Litres per kilowatt hour, metered and audited rather than modelled. Withdrawal and consumption reported separately and monthly, so that the summer peak cannot be averaged into invisibility. The source split between municipal supply, groundwater, treated sewage and desalinated water, because those are four different political facts and only one of them competes directly with a household tap. Reporting aggregated to the basin or the city utility rather than to the company, so that a drawdown in Chennai cannot be offset against efficiency in Mumbai. On the electrical side, the connection point, the sanctioned load, the expected ramp behaviour, the backup fuel and the annual diesel generator hours, which in the Delhi capital region is an air quality question before it is an energy one. And one line nobody has yet asked for: accelerators are replaced every three to five years, so a gigawatt of AI capacity is also a very large electronic waste stream arriving on a schedule that is already known.Also read: Beyond Bigger Models: Why Smarter Scaling Will Define the Future of AIThe reason to do any of this is not virtue. It is that 1.4 billion people, most of whom have no stake in whether a frontier model is trained in Hyderabad or Johor, will not indefinitely accept scarce water and discounted power flowing to windowless buildings that employ a few dozen staff, unless they can see the numbers and check them against what was promised. Visakhapatnam, Thane and Mohali are not aberrations, they are the first instalment. Every serious infrastructure sector in India has eventually been forced into disclosure by opposition it did not anticipate, and almost always on worse terms than it could have negotiated for itself. The data centre industry has about two years to decide which of those it wants to be.What would actually helpPre-permitted compute parks: Jamnagar works because one company owns the land, the solar farm, the storage and the desalination plant. Replicate that near existing high-voltage substations and coastal renewable clusters, with 300 to 500 MW blocks cleared in advance and a target of six to nine months from approval to energisation.Behind-the-meter generation: The fastest American builds run on on-site turbines. India should permit captive gas and renewable-plus-storage at data centre sites now, with clear wheeling and backup rules, and SMRs later.Pooled chip procurement: A consortium buyer aggregating orders across IndiaAI, Reliance, Tata, Yotta and others would command better allocation than fragmented five-figure orders placed separately.The unglamorous supply chain: Transformers, switchgear, turbines and liquid-cooling equipment are globally back-ordered. Manufacturing incentives for these will matter more to 2028 delivery than incentives for chips.Domestic demand: The first Indian 100 MW sites are anchored by foreign tenants serving foreign models. Subsidised GPU-hours are useful, but two or three large sovereign training programmes would justify big sites on Indian demand.Treat data centres as grid assets: Ramp-rate obligations written into connection agreements, on-site storage sized to smooth the site’s own swings, an interruptible tariff for deferrable training load, and a route for behind-the-meter batteries into the ancillary services market. Done early, the sector becomes the flexibility an Indian grid short of it badly needs. Done late, the answer is the Irish one: no new connections.India is not failing at this. It is early, and it is slow, and it is measuring the wrong thing. The country counts announced billions. It should be counting energised megawatts, disclosed water, and months from clearance to first rack. On those numbers, the next 18 months decide whether India is Asia’s third compute centre or its fifth.Pravin Kaushal is an entrepreneur, public policy commentator and columnist, and tweets as @ipravinkaushal.