India’s AI Moment Will Be Won on Power and Grid Capacity

The story of AI is often told in parameters and GPUs. The more important unit over the next decade is much simpler: kilowatt-hours. Data centres already consume a measurable slice of global electricity. As AI workloads scale, that share is rising fast enough that grid capacity, permitting, and power procurement are becoming binding constraints on model deployment. For investors, especially in India, this sits at the intersection of power markets, digital infrastructure, and long-duration capital.

How big is the load today?

Global data centres (including cloud, colocation and large enterprise sites) consumed an estimated 460 TWh of electricity in 2022, roughly 2% of worldwide demand. The International Energy Agency (IEA) projects that electricity use from data centres, AI and crypto combined could exceed 1,000 TWh by 2026 if current trends continue.

The US gives a sense of scale in a single market. Recent estimates suggest American data centres drew around 150 TWh in 2023, or roughly 3% of US electricity consumption, with that share expected to rise sharply as AI workloads ramp.

The important point is not the exact percentage; it’s the slope. In a world where overall electricity demand was almost flat for years in many advanced economies, data centres have become one of the few structurally growing loads.

Why AI changes the profile

Traditional data-centre demand was driven by a mix of storage, web serving, and enterprise applications. AI is different in three dimensions.

First, AI workloads are compute-dense. Training state-of-the-art models requires racks of high-end GPUs or specialised accelerators, each drawing several hundred watts under load. That pushes power density per rack and per square metre well beyond older facilities.

Second, utilisation is higher. Training runs and large inference workloads run close to full throttle for extended periods. The result is less variability and more “hard” baseload demand at each site.

Third, cooling requirements rise with power density. Even if a facility’s nameplate IT load is, say, 50 MW, the total site demand, including cooling and auxiliaries, will sit meaningfully higher unless the operator invests heavily in efficiency.

The result is that every major AI campus is effectively a new industrial-scale power customer. Location decisions are now being driven as much by available megawatts and grid connection timelines as by tax incentives or proximity to talent.

Grids, bottlenecks and who pays

This is where AI stops being a tech-sector curiosity and becomes an energy-system issue. On the supply side, utilities and independent power producers face a simple equation. Meeting AI-driven load growth means building generation, transmission and sometimes storage faster than originally planned. In markets with decarbonisation commitments, that effectively means more renewables, more grid reinforcement and, in some cases, more gas-fired capacity for flexibility.

On the grid side, many transmission networks were not designed for clusters of 100- to 500-MW loads popping up on the edge of major cities. Queue backlogs, interconnection delays and local congestion are becoming critical constraints. In some jurisdictions, hyperscalers are exploring behind-the-meter solutions or long-term power purchase agreements (PPAs) combined with dedicated grid upgrades to secure capacity.

For investors, this translates into a capital-formation problem. Somebody has to fund the incremental generation, the extra substation, the high-voltage line and the storage that trims volatility. That “somebody” increasingly includes private capital.

The India angle: capacity catching up with data

India is not yet at US-style levels of data-centre power load, but the trajectory is steep. Rating agency ICRA expects India’s operational data-centre capacity to rise from about 1,150 MW in December 2024 to 2,000–2,100 MW by March 2027, driven by data localisation, cloud adoption and digital-first business models. The associated investment over that period is estimated at roughly ₹2 lakh crore. Other forecasts suggest total data-centre capacity could reach around 3 GW by 2030, implying another step-up in power demand.

Overlay that on an already tight power system with summer peak-load stress, and you get a clear picture: AI and cloud demand are arriving just as India is trying to decarbonise, expand manufacturing and keep tariffs politically acceptable.

In practice, that means three things. First, new data-centre clusters will compete with other industrial and urban loads for firm capacity and grid headroom. Second, the marginal megawatt increasingly has to come from a combination of renewables, storage and flexible thermal rather than from cheap surplus baseload. Third, the planning horizon for both developers and regulators has to extend beyond the next electoral cycle.

Where the capital actually flows

From an alternatives/AIF perspective, the AI power story is to “follow the electrons”. At the core sit generation and grid projects. Utility-scale solar and wind assets with long-term offtake from data-centre anchors are likely to look different from generic merchant renewables. The counterparty risk, tenor and pricing structure are shaped by hyperscaler credit profiles and their own decarbonisation targets.

Around that core sits enabling infrastructure: grid-connected battery projects smoothing renewable intermittency, dedicated transmission links into new data-centre parks, and substation upgrades in peri-urban zones. These are capital-intensive, often regulated and, by design, long-duration.

Then there is the real-estate layer. Hyperscale campuses, edge facilities and specialised “AI zones” require land assembly, power-ready shells and, increasingly, integration with district-cooling or on-site energy systems. Whether that ultimately lives in REIT-like vehicles, private infra funds, or developer balance sheets is still evolving, but the capital requirement is not in doubt.

None of this is limited to India. Globally, the IEA’s projection that electricity use from data centres, AI and crypto could more than double to over 1,000 TWh by 2026 implies that a non-trivial slice of all new generation and grid investment over the next decade will be, directly or indirectly, in service of digital workloads.

Where are the risks

The obvious risk is over-building on over-optimistic AI demand assumptions. Forecasts for model size, usage, and monetisation are highly uncertain. If efficiency gains in hardware and software outrun workload growth, some planned capacity could be under-utilised.

Policy and social risk matter too. In markets where AI-driven power demand pushes up tariffs or crowds out decarbonisation budgets, you should expect political pushback. That can show up as windfall-style taxes, forced repricing of PPAs or permitting friction for new facilities.

There is also technology risk. If AI inference shifts heavily to the edge, or if new chip designs alter power density and cooling needs, some existing facilities may require more retrofit capex than current models assume.

Finally, the “green AI” narrative is not costless. Hyperscalers’ net-zero pledges can drive demand for renewables and storage, which is positive for certain asset classes. They can also compress returns if buyers use their bargaining power to push project IRRs down in exchange for scale and credit quality.

What actually endures

The first is that digital infrastructure and energy infrastructure are now joined at the hip. You cannot talk about AI strategy without talking about power procurement and grid capacity.

The second is that data-centre load is one of the few clearly growing, long-term electricity demands globally. Even if growth rates moderate, it will still be one of the main drivers of incremental generation and network investment in many markets.

The third is that capital formation for this build-out will not be solved by utility balance sheets alone. Project finance, infrastructure funds, speciality credit, and private market funds will all have a role.

Chief Brand & Investor Relations Officer

Sneha Dadrwal

Frequently Asked Questions

Q: How much electricity could data centres and AI use in the next few years?

A: IEA estimates that electricity use from data centres, AI and crypto combined could more than double from about 460 TWh in 2022 to over 1,000 TWh by 2026, or several percent of global demand.

Q: Why does AI make data-centre power demand different from older IT loads?

A: AI workloads are far more compute-intensive, run at higher utilisation for longer periods and require heavier cooling, which pushes up power density and makes each large campus look like an industrial-scale baseload consumer.

Q: How fast is India’s data-centre power demand likely to grow?

A: India’s operational data-centre capacity is expected to rise from about 1,150 MW in late 2024 to roughly 2,000–2,100 MW by March 2027, with some estimates pointing to around 3 GW by 2030.

Q: What types of assets benefit from AI-driven power demand?

A: Utility-scale renewables with long-term offtake, flexible generation, grid and substation upgrades, storage projects and power-ready data-centre real estate are all direct beneficiaries, provided contracts and regulation are sound.

Q: What are the main risks?

A: Key risks include over-building on aggressive AI demand forecasts, regulatory pushback if tariffs rise, technology shifts that change facility requirements, and return compression where hyperscalers use their buying power to drive down project IRRs.

Sources

1. EA
2. POWER Magazine
3. India Brand Equity Foundation
4. Business Standard