Can India Power the AI Boom Without Burning the Planet?

Every time someone marvels at what a new AI model can do, a quieter fact goes unmentioned: it runs on electricity, and a great deal of it. AI energy consumption is not a footnote to the technology story — it is part of the story. Training a large model burns power for weeks across thousands of processors, and then every answer it gives afterwards draws power again, one query at a time, millions of times a day. Data centres, the warehouses where all this computation lives, have quietly become one of the fastest-growing sources of electricity demand in the world. For India, a country still working to keep the lights on reliably for everyone, that is not an abstract concern. It is a question about who gets the power, who pays for it, and who decides.

Why AI is an electricity story

We tend to picture AI as something weightless — a chat window, a clever reply, a picture conjured from a sentence. But behind the screen sits physical infrastructure: racks of specialised chips running hot, cooled around the clock, fed by transformers and backup generators. The intelligence may be artificial, but the electricity bill is real.

There are two distinct appetites here. The first is training — the enormous, one-time-per-model effort of teaching a system by pushing vast datasets through it repeatedly. The second, and increasingly the larger over a model's life, is inference — the electricity spent every time the finished model actually answers a question, generates an image, or drafts a paragraph. As AI features get bolted onto search engines, phones, office software and customer service lines, inference happens billions of times over. The per-query cost is tiny; the aggregate is not.

How large the total becomes is genuinely uncertain, and anyone quoting a precise figure years out is guessing. Estimates of data-centre electricity demand vary widely depending on how fast AI is adopted, how efficient the chips get, and how much of today's excitement translates into durable use. What the credible projections agree on is direction, not magnitude: the arrow points sharply upward. Reportedly, some grid operators and utilities are already revising their long-term forecasts to account for clusters of demand they did not expect a few years ago. The honest summary is that AI is adding a new, concentrated, and fast-rising load to electricity systems that were planned for a gentler curve.

India's particular bind

India sits at an awkward intersection of three pressures that pull in different directions.

The first is ambition. India has actively courted data-centre investment, and companies are building. Cheap land in some regions, a large domestic market, data-localisation rules that require certain information to be stored within the country, and the global scramble for AI capacity have all made India an attractive place to put servers. That expansion is real and, in economic terms, welcome — it brings investment and a claim on the industries of the future.

The second is the grid itself. India's electricity system has improved dramatically, but it is still strained in ways that richer countries' grids are not. Demand peaks — especially in brutal summers, when air-conditioning and irrigation pumps run hard — already test supply. Distribution utilities in many states are financially fragile. Reliable, round-the-clock power is something millions of households and small businesses still cannot fully take for granted. A data centre, by contrast, wants exactly that: uninterrupted, high-quality power, every hour of every day, without fail.

The third is climate. India has made serious commitments to expand clean generation and to bring down the carbon intensity of its economy. But a large share of the country's electricity still comes from coal, and when new demand appears faster than clean supply can be built, the gap tends to be filled by whatever is already running. That is the quiet risk with AI load: a surge in round-the-clock demand can end up leaning on fossil generation, not because anyone chose coal, but because it was there and the clean alternative was not ready in time.

The intelligence may be artificial, but the electricity bill is real — and in India, the grid that pays it is already stretched.

Can clean power be added fast enough?

The optimistic case is straightforward, and worth taking seriously. India has built solar and wind capacity at remarkable speed and at some of the lowest costs in the world. In principle, new data-centre demand could be met by new clean generation, so that AI's growth and the energy transition move together rather than against each other. Some operators talk about siting facilities near renewable resources, signing long-term clean-power contracts, and investing in the efficiency of their cooling and hardware. None of that is fantasy.

But two hard problems sit underneath the optimism. The first is timing. Solar produces during the day, wind when it blows; a data centre needs power at three in the morning in the monsoon just as much as at noon in April. Matching a steady, around-the-clock load to intermittent clean sources requires storage, grid upgrades, and firm backup — and those take years and money to build. The second is honesty about accounting. It is easy for a company to buy clean-energy certificates on paper while the actual electrons feeding its servers come from the grid's fossil-heavy average. A green claim and a green kilowatt-hour are not the same thing.

So the real question is not whether India can power AI with clean energy in the abstract — it can, eventually. It is whether the clean generation, storage and grid capacity get built at the same pace as the demand, and whether the heavy new users are sited and priced in a way that helps rather than hurts. That is a policy and accountability question far more than a technical one.

The capture pattern, applied to the grid

This is where I find it useful to step back and look at the shape of the thing, because it is a shape I keep seeing across technologies. New tools arrive wrapped in the language of progress and inevitability. The benefits are concentrated — they accrue to whoever owns the infrastructure and the models. The costs are diffuse, spread thinly across everyone, and often not counted at all until much later. I've written before about how technology gets captured: the pattern where a genuinely useful innovation ends up structured so that a few capture the gains while the many absorb the risks.

Electricity makes this legible. A data centre serving global AI demand can consume power on the scale of a small town, drawing on a grid that ordinary Indians and small businesses also depend on and pay to maintain. If that demand pushes up peak strain, or nudges the system toward more fossil generation to stay reliable, or absorbs grid investment that might have gone elsewhere, then the costs — higher strain, dirtier air, a heavier climate burden — are socialised across the whole population. The profits, meanwhile, concentrate with the firms that own the servers and the models. Costs socialised, profits privatised. It is the oldest move there is, and it works precisely because each individual share of the cost is too small to notice while the aggregate is enormous.

Water follows the same logic, which is why I've treated it separately in the hidden water cost of AI — those same cooling systems drink heavily from local supplies, often in places that can least spare it. Electricity and water are two faces of the same physical footprint, and both tend to land on communities that had no say in the siting decision. And the question of who owns AI is not separate from any of this: ownership of the models is also, downstream, a claim on land, power and water in places far from where the profits are booked.

Costs socialised, profits privatised — it works precisely because each individual share is too small to notice while the aggregate is enormous.

Who takes, who pays, who could fight back

None of this is an argument against AI, or against data centres, or against India competing for a share of a growing industry. It is an argument against letting the terms be set entirely by the people who benefit most. The difference between AI load that strengthens India's grid and AI load that quietly degrades it comes down to the conditions attached — and right now, those conditions are thin.

Consider who is at the table. The companies building data centres are sophisticated, well-capitalised, and able to negotiate directly with state governments hungry for investment. The households whose grid they share, the farmers whose peak-hour power competes with server cooling, the communities near the sites — they are rarely part of the conversation. When one side can negotiate and the other can only absorb the outcome, the outcome is predictable. Fighting back, in this context, does not mean opposing the technology. It means insisting that the people who bear the costs have a say in the deal.

What accountability could actually look like

The encouraging thing is that the fixes here are not exotic. They are ordinary tools of good governance, and other countries are beginning to reach for them. A few would go a long way.

• Clean-energy requirements with teeth. Large new data centres could be required to bring genuinely additional clean generation online to match their load — not just buy certificates against the existing pool, but add new capacity — and ideally to match it hour-by-hour rather than as an annual average. That turns a green claim into a green kilowatt-hour.
• Transparent siting. Where a large facility goes should be a public decision, weighing the local grid's headroom, the water available, and the strain on the surrounding community — not a quiet deal between a company and a state agency. Communities have a right to know what is being plugged in next to them and what it will draw.
• Demand disclosure. Operators above a certain size could be required to publicly report how much electricity and water they consume, from what sources, and how that changes over time. You cannot govern what you cannot measure, and right now the numbers are mostly hidden behind commercial confidentiality. Disclosure is the precondition for every other kind of accountability.
• Honest pricing. If heavy, round-the-clock users impose real costs on the grid — for peaking capacity, for reinforcement, for backup — those costs should be reflected in what they pay, rather than smeared across every other consumer's tariff.

None of these is anti-investment. A serious operator planning to run on clean power and site responsibly should welcome rules that reward exactly that and stop less scrupulous competitors from cutting corners at the public's expense. Clear conditions are how you separate the investment that strengthens India from the investment that merely extracts from it.

The choice in front of us

India is going to build data centres, and India is going to use AI. Those are not the questions. The question is whether the power to run it all gets added cleanly and fairly, or whether the strain and the emissions get quietly loaded onto a grid — and a public — that never agreed to carry them. The exact numbers are uncertain and will stay uncertain for a while; the direction is not. Demand is rising fast, and the window to shape how it is met is open now, while the infrastructure is still being planned rather than already poured in concrete.

The technology is genuinely useful. That has never been the point of contention. The point is that useful technologies are exactly the ones most worth watching, because their usefulness is what makes it easy to wave away the question of who pays. On the grid, as everywhere else, the honest test is simple: when the benefits and the costs are finally tallied, do they land on the same people? If we get the accountability right — clean-energy requirements, transparent siting, real disclosure, honest pricing — then AI's hunger for power can be met without burning the planet or the public. If we don't, we will have built something remarkable on a foundation that quietly bills everyone but the owners. India still has the chance to choose the first path. The choosing has to happen now.