India Doubled Its Own AI Power Forecast in Five Months

The revision is the story

Governments revise forecasts. That is ordinary. What is not ordinary is the size and speed of this one.

A near-doubling of a five-year projection inside five months is not a modelling refinement. It is an admission that the demand signal moved faster than the planning process, and it is the single most useful data point in the release, because it tells you the current number is a snapshot rather than a settled estimate.

Set 26.3 GW against the Central Electricity Authority's overall outlook to see the scale. The CEA expects India's peak electricity demand to rise from 289 GW in FY27 to 388 GW by FY32 — an increase of 99 GW. If AI data centres account for 26.3 GW of that, they represent roughly a quarter of the entire projected growth in national peak demand, in a country simultaneously electrifying transport, expanding manufacturing and extending cooling to more households.

That is a significant claim on national infrastructure by one sector, and it is worth stating plainly rather than burying in enthusiasm about the AI build-out.

What has already been built

The capacity numbers show the trajectory is not speculative.

• ~375 MW installed data centre capacity in 2020.
• 1,575 MW as of the government's figures released in early August 2026 — more than four-fold growth.
• Over 1.6 GW operational capacity per Cushman & Wakefield's estimate, which would place India as the second-largest operational data centre market in Asia-Pacific.

The named projects give the abstraction some shape. Meta is leasing a 168 MW AI-ready facility being built by Reliance Industries in Jamnagar, Gujarat. Google has begun construction on an AI hub in Visakhapatnam as part of a $15 billion investment programme. Deloitte estimates the AI surge could require an additional 45 to 50 million square feet of real estate.

Note what those two projects have in common: both are hyperscaler capacity, sited where power and land are available rather than where the engineers are. Jamnagar is an industrial-energy location. Visakhapatnam is a coastal city with port and power advantages. Neither is Bengaluru. The AI build-out is landing where the electricity is, and that has consequences for where the jobs land too.

How the load is meant to be absorbed

The government's position is that the additional demand will be integrated into the national grid and served primarily by renewable capacity, with transmission infrastructure strengthened in phases. The named instruments are the Green Energy Open Access Rules and the Green Energy Corridor Scheme, alongside broader resource-efficiency measures.

Green Energy Open Access is the mechanism that matters most to anyone procuring capacity, because it is what allows a large consumer to buy renewable power directly rather than through a distribution utility's standard tariff. For a data centre operator, that is the difference between a power cost set by regulated tariffs and one set by a negotiated renewable contract.

Two honest caveats belong here. First, "primarily renewable" is a stated intention about a five-year horizon, not a delivered outcome, and renewable generation is intermittent while AI training load is not. The firming — storage, or thermal backup, or both — is where the real cost and the real emissions question sit, and the announcements are quieter on that. Second, phased transmission strengthening is a programme, and programmes run to their own timetables. The build-out is currently moving faster than the forecast, which was itself revised upward because it could not keep up.

The UK is hitting the same wall from the other side

This is where the story becomes genuinely dual-market, because Britain is dealing with an identical physical constraint using an opposite instrument.

UK grid connection applications surged from 41 GW to 125 GW in under a year, with at least 80 GW of that attributable to data centre projects. Ofgem's response, in a consultation opened on 29 July 2026 and running to 16 September, is a proposed Data Centre Commitment Fee of £237,500 to £712,500 per megawatt — roughly 2.5% to 7.5% of average project costs — refunded when a project energises and forfeited if it exits the queue early, alongside new milestones requiring developers to evidence financial capability, commercial maturity and procurement activity to hold their place.

Neither approach is obviously correct. India's bets that supply can be expanded to meet demand, which works if transmission keeps pace and carries real risk if it does not. The UK's accepts that supply is constrained and rations by willingness to commit capital, which clears speculative applications efficiently and also prices out smaller entrants who cannot post six-figure-per-megawatt deposits.

What they share is the thing builders should internalise: compute capacity is no longer primarily a procurement question, it is an electricity question. That is a structural change from even two years ago, when the binding constraint was GPU allocation rather than the power to run it — a shift visible in the way GPU lease prices have been rising even as model prices fall.

What this means for builders

For Indian teams, three practical consequences. First, capacity is arriving, which is genuinely good news for anyone who has fought for GPU allocation — and it complements the subsidised access routes we have covered through the IndiaAI Mission's GPU pool. Second, it is arriving in Gujarat and Andhra Pradesh rather than the established technology hubs, which means infrastructure and site-reliability roles will increasingly sit outside Bengaluru, Hyderabad and Pune. Third, power cost is becoming a differentiator between providers in a way it was not, so it is worth asking where a provider's electricity comes from and under what contract.

For UK teams, the commitment fee changes who can credibly promise you capacity. A provider who has posted deposits on its queue positions is making a materially stronger claim than one holding speculative applications, and it is a fair question to ask. Expect consolidation toward well-capitalised operators, and expect the practical answer for most British AI companies to remain buying capacity rather than building it.

For both, the architectural advice converges. Separate the workloads that must sit near your users and your data from the ones that can follow cheap, available power. Latency-sensitive inference and anything bound by residency obligations — and for Indian and UK teams that is a live constraint under DPDP and UK GDPR, as we set out in our guide to data residency and request routing — stays put. Batch training, evaluation runs and offline processing can be portable, and portability is worth engineering for when the constraint is this dynamic.

Three things to watch over the next twelve months

Forecasts of this volatility are not worth tracking closely, but a small number of concrete signals will tell you whether the 26.3 GW figure is holding, rising again, or quietly slipping.

Whether transmission keeps pace with connection

The government's stated approach is to strengthen transmission in phases. The number that matters is not how much generation capacity exists but how much of it can be delivered to the districts where data centres are actually being built. A facility in Jamnagar or Visakhapatnam draws on regional transmission, and a shortfall there produces the same outcome as a generation shortfall — an energised-on-paper connection that cannot carry full load. Watch state transmission utility announcements in Gujarat and Andhra Pradesh more closely than national generation totals.

Whether the renewable framing survives contact with load shape

AI training is a steady, high-utilisation load. Solar is not. The gap between those two profiles has to be filled by storage, by thermal generation, or by drawing on the grid at times when the marginal unit is not renewable. Statements that the additional demand will be met "primarily by renewable energy capacity" are about annual energy volumes rather than hour-by-hour matching, and those are very different claims. If firming arrangements start appearing in announcements — battery storage co-located with data centre projects, or long-term thermal contracts — that is the honest version of the picture becoming visible.

Whether a third revision arrives

The projection moved from 13.56 GW to 26.3 GW between March and July 2026. If a further upward revision lands within the next two budget cycles, the planning implication is not that the number is wrong but that the demand curve is steeper than the planning process can track — at which point capacity availability becomes the binding constraint on Indian AI deployment regardless of how much capital is committed. That is the scenario worth having a contingency for, and it is the reason the workload-portability engineering set out in our guide to planning AI capacity when the grid is the constraint is worth doing before you need it.

The short version

India's official AI power forecast went from 13.56 GW to 26.3 GW by FY32 in the space of five months, against a CEA outlook where national peak demand grows 99 GW over the same period. Installed capacity has more than quadrupled since 2020 to 1,575 MW, hyperscaler projects are landing in Jamnagar and Visakhapatnam, and the load is meant to be met primarily by renewables via green open access. Britain, facing the same ceiling, is rationing queue access with a per-megawatt deposit instead. Whichever market you build in, power availability has become an input to architecture, not a facilities footnote.