AI’s hunger for power: Can India’s grid keep pace with compute boom?

A typical AI-focused data centre consumes as much electricity as 1,00,000 households, but the largest ones under construction today will consume 20 times as much, according to the International Energy Agency.

The world’s most powerful artificial intelligence (AI) models don’t just run on code, they run on megawatts. India’s grid is about to find out what that really means.

When Google announced a $15 billion investment to set up a large-scale artificial intelligence (AI) hub in Visakhapatnam, Andhra Pradesh, on October 14, its single largest outside the US, the move marked a turning point not just for Andhra Pradesh, but for India’s national AI ambitions too. The project will build the country’s first gigawatt-scale data centre campus tied to AI, alongside subsea cable gateways and renewable power infrastructure.

Union IT Minister Ashwini Vaishnaw called it a boost for the IndiaAI Mission, designed to provide shared compute infrastructure for sovereign AI development. Andhra Pradesh Chief Minister N Chandrababu Naidu termed it as “a new chapter in India’s digital transformation journey.”

But the real challenge lies in powering this chapter. AI demands electricity at a massive scale, raising a critical question: how much energy will India’s AI revolution consume and where will it come from?

IndiaAI Mission and GPU clusters

India’s AI strategy rests on the IndiaAI Mission, which has already deployed over 38,000 graphics processing unit (GPUs) across national compute clusters, far surpassing its original 10,000-GPU target. Yotta Data Services alone has committed 9,216 GPUs, including 8,912 Nvidia GPUs.

Sunil Gupta, co-founder and CEO of Yotta Data Services, said that each cluster is “capable of training large language models comparable in scale to GPT-4,” which itself consumed an estimated 50 GWh in a single training run. However, unlike OpenAI’s concentrated mega-training approach, India is distributing its compute.

“If GPT-4 training consumed ~50 GWh in one continuous run, then each India AI GPU cluster training a similar model may consume 30–70 GWh (including overheads). Ten such clusters would amount to ~300–700 GWh annually, but since they run episodically (weeks or months), the average load shall be much lower, and with intelligent scheduling tied to renewables, the energy burden shall be manageable,” Gupta told Moneycontrol.

Gupta added that “AI workloads still form a small share of India’s total data-centre power use, likely under 10 percent, but they’re doubling every 12–18 months. “This shift is driven by GPU-based training, fine-tuning, and inferencing, which consume 5–10× more power per rack than conventional CPU workloads,” he said.

Ganesh Gopalan, co-founder and CEO of Gnani.ai, one of the startups chosen to build foundation models under the mission, emphasised that India’s compute growth will be steady and broad-based.

“Power consumption will be continuous and operational instead of peaking in one large event. As clusters expand, demand could eventually match international benchmarks, highlighting the need for renewable integration and efficient cooling to ensure both cost control and sustainability,” Gopalan told Moneycontrol.

AI’s appetite for power

Training and deploying AI models takes place in large and power-hungry data centres. A typical AI-focused data centre consumes as much electricity as 1,00,000 households, but the largest ones under construction today will consume 20 times as much, according to the International Energy Agency.

India’s data centre load is projected to jump from 1.2 GW in 2024 to 4.5 GW by 2030, largely driven by AI and digital adoption, according to Unaise Urfi, Partner, KPMG India. AI-driven data centres alone could consume an additional 40–50 TWh annually by 2030, Urfi added, citing sector estimates.

Graphic credit: Upnesh Raval

This growth must be seen in the context of India’s broader electricity landscape. As of July 2025, the country’s total installed capacity stood at 490 GW, with fossil fuels accounting for 49.7 percent (243 GW) and non-fossil sources slightly higher at 50.3 percent (246 GW). Renewables, including hydro, made up 48.5 percent of the mix, with solar alone contributing 119 GW, according to data from the Central Electricity Authority (CEA) under the Ministry of Power.

According to S&P Global Commodity Insights estimates, India is expected to become the second-largest market for data centre electricity demand in Asia-Pacific over the next two years, surpassing Japan and Australia.

The government is already aware of the looming crunch. In December 2024, MeitY Secretary S. Krishnan said the ministry was working with the power and renewable energy ministries to ensure that India has enough electricity for data centres and AI compute facilities.

"Globally, all major technology companies are trying to figure out ways in which they can secure adequate power for AI applications. Related to that is data centres. How do you ensure that this power supply comes in a way that we are able to address this issue? This is something MeitY is looking into right now along with the Ministry of Power, Ministry of New and Renewable Energy, and other related agencies," he had said.

Graphic credit: Upnesh Raval

India’s data centre market is expanding rapidly, with 268 facilities now operational, placing the country eighth globally, just behind Australia (273) and Canada (286). The upcoming Google–AdaniConneX–Airtel campus in Visakhapatnam is part of this surge, but competition is intensifying. Mukesh Ambani-led Reliance Industries is building a massive AI-driven facility in Jamnagar, Gujarat, with a planned capacity of 3 gigawatts (GW), among the largest in the world. Meanwhile, Tata Consultancy Services (TCS) has announced plans to develop 1 GW of AI-grade capacity across India over the next five to seven years.

Even OpenAI, the developer of ChatGPT, is reportedly exploring partnerships with several Indian data centre firms as part of its plan to localise its $500 billion Stargate project.

In response to Moneycontrol’s queries, Tarun Pathak, Research Director at Counterpoint, said, “As data centres are projected to consume around 3 percent of India’s national power by 2030, India's compute ambitions will have to contend with its existing carbon output.” At the same time, he noted, India’s renewable growth provides a counterbalance: “India has already achieved a major milestone, 50 percent of its installed capacity now comes from non-fossil sources, meeting its national target five years ahead of schedule.”

The cooling crisis

Electricity is only part of the problem. Cooling and water use are quickly emerging as the biggest bottlenecks for India’s AI data centres.

“The biggest challenge is thermal management. AI workloads generate heat densities exceeding 70–150 kW per rack, making traditional data-centre cooling architectures inadequate. In a water-stressed country like India, the goal must be to achieve high cooling efficiency with minimal water use,” Yotta CEO Gupta told Moneycontrol.

Gupta claims that advanced technologies are already in play at Yotta’s campuses. For its H100 GPU clusters, the company deploys Rear Door Heat Exchangers (RDHx) to remove heat directly at the rack level. Its upcoming B200 clusters will employ Direct-to-Chip (DTC) liquid cooling with Coolant Distribution Units, allowing extremely dense configurations to run safely.

“These systems run within a closed-loop liquid cycle and air-cooled chillers with adiabatic assist, ensuring negligible incremental water consumption,” Gupta said. “By combining rack-level liquid cooling, efficient airflow management, and AI-driven controls to dynamically adjust cooling loads in real time, we can dramatically reduce both power and water usage.”

According to Gupta, these innovations have helped Yotta bring its Power Usage Effectiveness (PUE) close to 1.4, among the best in Asia, while keeping Water Usage Effectiveness (WUE) negligible. “These design choices reflect a core belief: as India’s AI compute capacity grows, engineering innovation, not just policy, will determine our sustainability trajectory,” he added.

Andhra Pradesh’s pitch

No state has leaned into this challenge more than Andhra Pradesh. State IT Minister Nara Lokesh said the government sees cheap renewable energy and surplus Godavari water as strategic assets.

“We've promoted renewable energies at a big scale. At a gigawatt scale, we've signed with Tatas, we've signed with Renew. GreenCo is a homegrown company and it's not just about solar and wind. We are also doing PSP pump storage projects. We are implementing battery energy storage systems, and there's always going to be a base load that has to come from thermal power,” Lokesh told Moneycontrol in an interview.

On water, he pointed to a river-linking project that will carry Godavari water from north to south Andhra.

While several countries have curbed new data centre developments due to resource constraints, Lokesh argued that India should take a different approach. “I believe that India should not do it, and India should split the grid into two, if I may suggest, and say, alright, data centres, this is the way we are going to do it, and this is the kind of energy that we're going to use for it from renewable sources,” he said.

Investors eye ‘green AI’

The financing side of AI is also turning green. Vasudha Madhavan, founder and CEO of Ostara Advisors, a climate-tech investment banking firm, said investors are beginning to back sustainable AI infrastructure. “We are beginning to see green capital flow into AI-linked infrastructure, though not yet at the same scale as EVs or batteries,” Madhavan told Moneycontrol.

“Investors increasingly recognise that AI’s significant energy and water demands create both a challenge and an opportunity for sustainable innovation. This is driving funding towards innovations in energy-efficient cooling of data centres, renewable energy-powered data centres, and energy-efficient chips and computing frameworks,” she said.

Madhavan argued that India’s sovereign AI mission could learn from EV policies like FAME-II by tying incentives directly to sustainability metrics. “AI policies could reward deployments that prioritise renewable-energy powered data centres, energy-efficient GPU clusters, and advanced cooling technologies,” she added.

The road ahead

India’s sovereign AI ambitions are colliding with the reality of energy, water, and climate limits. Policymakers, data centre operators, and startups agree that slowing down is not an option, but building responsibly is the only way forward.

As Gupta of Yotta put it: “India’s sovereign AI ambition is non-negotiable. But sovereignty must go hand-in-hand with sustainability.”