Powering intelligence: Why India’s AI race will be won at the substation

A few months ago, in a meeting about a proposed data-centre campus in western India, I watched the conversation take a turn I hadn’t seen before. Nobody argued about land prices. Nobody argued about fibre. The entire discussion nearly two hours of it — was about power. Could the state grid deliver 300 MW to that site, at that quality, on that timeline? Nobody in the room was sure.

That meeting has stayed with me, because it captures where India’s AI infrastructure story is quietly heading. We spent years debating chips and cloud regions. The harder question turned out to be sitting in plain sight, humming away at the local substation.

The numbers explain why. India’s data-centre IT capacity crossed 1.8 GW in the first half of 2026, and most projections put it past 6.7 GW by 2030. Call it a near-quadrupling in under five years, landing on a power system that is itself in the middle of the biggest transition in its history. Something has to give — or something has to be built.

These are not your father’s data centres

There’s a lazy assumption floating around that an AI data centre is just a regular data centre, only bigger. It isn’t, and the difference matters enormously for anyone who has to keep the lights on.

Start with density. A rack of AI accelerators pulls several times the power of a conventional server rack, and every one of those watts eventually becomes heat that has to be pumped out of the building. So the electrical problem and the cooling problem compound each other in the same few square metres.
Then there’s the volatility, which frankly surprised even seasoned facility engineers when these workloads first showed up. A large training run doesn’t draw power like a factory. It ramps, checkpoints, stalls, resumes. Inference traffic surges with whatever users are doing at 9 pm on a Tuesday. The load can swing hard within seconds, and the electrical infrastructure has to ride those swings without blinking.

And the tolerance for failure is essentially zero. A voltage dip that a steel plant would never notice can wreck a training run that has been going for three weeks. Twenty-four-seven doesn’t quite capture it; what these facilities need is continuous, high-quality power with no asterisks.

Take those three things together and you arrive at an uncomfortable truth: no grid on earth, however well run, is a naturally good match for this kind of load. Something has to sit between the workload and the wire.

The battery, reconsidered

Which brings me to batteries — and to a mental model I’d like to retire.

Ask most operators where batteries fit and they’ll point you to the backup room, somewhere between the diesel gensets and the UPS units. Insurance, basically. Dead weight until the bad day arrives.

That view is a decade out of date. A well-integrated battery energy storage system is better understood as a buffer — a reservoir that fills when electricity is cheap, clean or simply available, and drains when it’s expensive, scarce or urgently needed. Once you see it that way, a whole set of uses opens up beyond keeping the servers alive during an outage.

The battery responds to a grid disturbance in milliseconds, long before a generator has even coughed to life. When compute demand spikes, it absorbs part of the surge locally instead of slamming the full jolt onto the network — a shock absorber, in effect, between the AI workload and the grid. During the facility’s most expensive hours it can discharge to shave the peak, which any operator who has stared at a demand charge will appreciate. It captures solar generation at noon and hands it back at midnight, which is exactly the mismatch that has bedevilled renewable procurement for data centres. And with the right controls and, crucially, the right regulation, it lets a gigawatt-class campus behave as a flexible consumer rather than an immovable one — something grid operators will increasingly reward, or eventually demand.

Five jobs, one asset. The interesting question for an operator is no longer whether to buy storage. It’s whether you’re set up to capture all five of those value streams or just the first one.

An opening India shouldn’t waste

Here the story gets bigger than any single facility. India’s own planners estimate the country could need something like 47 GW / 236 GWh of battery storage by FY2031-32, alongside pumped hydro, simply to absorb the renewable capacity being added to the grid. Data centres now add a demanding — and deep-pocketed — new customer to that build-out.

They also change where digital investment goes. Land and fibre used to decide site selection; they’re becoming table stakes. The next round of data-centre capital will flow to the states and corridors that can answer one question convincingly: can you power a very large campus, reliably and increasingly with clean energy? A few states will get organised around that question early. They’ll take a disproportionate share of the boom, and honestly, they’ll deserve it.

There’s a rarer opportunity buried in here too. Most countries built their digital and energy infrastructure separately and are now retrofitting the connections at great expense. India still has the chance to plan them together — data-centre parks co-located with renewable generation and grid-scale storage, transmission planning that actually looks at compute demand forecasts, tariff structures that pay data centres for flexibility instead of merely tolerating them. It would be a shame to let that chance slip out of bureaucratic inertia.

A warning about the hard part

I’ll end with the trap I expect many storage projects to fall into, because I’ve seen early versions of it already.

Buying a battery is easy. Operating one well is not. The value of storage isn’t fixed on the day of purchase; it’s earned hour by hour afterwards, through thousands of small decisions. Compute load just spiked — does the battery discharge now, or hold its charge against a possible grid event this evening? The right answer depends on renewable output, the current power price, grid conditions and what another deep cycle does to the battery’s lifespan. All at once. Every hour. Forever.

That’s a software and controls problem far more than a hardware one, and it’s where projects will separate. Get the orchestration right and the asset pays for itself several times over. Get it wrong and you’ve bought a very expensive UPS. My advice to operators evaluating vendors: scrutinise the energy-management software and the optimisation track record at least as hard as the cell chemistry. Probably harder.

The bottom line

Chips will stay at the heart of AI; no argument there. But chips at scale need power at scale, and power — reliable, flexible, increasingly renewable — is turning into a competitive weapon in its own right.
So the practical takeaways are these. If you operate data centres, treat energy strategy as a day-one design decision and architect storage in, not on. If you invest, notice that the picks and shovels of India’s AI boom include batteries, inverters and control software, not just servers. And if you make policy, accept that compute, grid, renewables and storage are now one infrastructure agenda, because functionally they already are.

The next chapter of India’s AI story won’t be written by compute alone. It will be written by how intelligently the country powers it. The substation, it turns out, is part of the stack now.