Mint Explainer: Why the surging use of AI may challenge India’s power sector

The peak demand exceeded the revised estimates of the government's power sector experts. The Central Electricity Authority (CEA) had estimated peak demand at 235 GW for May and to cross the 250 GW level in August and September. 

The power sector regulator had previously estimated peak demand for 2024-25 at about 245 GW and for 2025-06 at 260 GW in its electric power survey published in November 2022. This is forecast to climb to 328 GW by 2031-32 and 488 GW by 2041-42.

The CEA's power demand forecasts are based on demand from households, industry, commercial establishments, irrigation, railway traction, public water works, public lighting, and so on. 

Also read | Can India keep up with the ebbs and flows of power demand?

Additionally, the 2022 electric power survey considered emerging demand from electric vehicles—it assumed that 30% of all vehicles sold in 2030 would be battery electric vehicles and 100% by 2042. Solar power and hydrogen mission contributions that will ease the pressure on conventional energy sources were also considered.

However, the widespread use of artificial intelligence (AI), particularly generative AI, and the setting up of large data centres to support online AI interactions could lead to the overshooting of these demand forecasts. Mint explains how the rise of AI could lead to a sharp increase in power demand.

 

Why should power utilities worry about the rising use of AI?

Concerns about the energy impact of AI are on the rise in many parts of the world. The International Energy Agency, in its annual report on electricity market developments and policies published in January, estimated that data centres, cryptocurrencies and AI together accounted for 460 terawatt hours (TWh) of power or about 2% of global electricity demand in 2022. 

This demand could rise to 620-1,050 TWh by 2026; the agency’s base case for demand is at just over 800 TWh. The additional demand by 2026 from data centres, cryptos and AI is equivalent to the 2022 power consumption of Sweden at the lower end and Germany at the higher end.

An article published by the World Economic Forum noted that the energy requirement to run AI tasks was already accelerating with an annual growth rate of 26-36%. “This means by 2028, AI could be using more power than the entire country of Iceland used in 2021," the article read.

Also read | For AI, a few seconds of power becomes a booming business

A widely cited and peer-reviewed analysis by Alex de Vries, a doctoral candidate at the Vrije Universiteit Amsterdam, highlighted that AI servers are power-hungry devices. “A single NVIDIA DGX A100 server can consume as much electricity as a handful of US households combined. Because of this, the electricity consumption of hundreds of thousands of these devices will start to add up quickly," De Vries wrote on his blog. 

His paper was published in the scientific journal Joule in October 2023. AI systems require graphics processing units (GPUs) to run large numbers of calculations quickly, 95% of which are supplied by Nvidia.

He estimates that the worldwide AI-related electricity consumption could increase by 85.4-134.0 TWh by 2027. This figure represents half a percent of the worldwide electricity consumption and is comparable to the annual electricity consumption of countries such as the Netherlands, Argentina and Sweden.

What makes AI power-hungry?

According to New Scientist, De Vires in his paper noted that if Google were to switch its entire search business to AI, it would end up using 29.3 TWh per year–equivalent to the electricity consumption of Ireland, and almost double the company’s total energy consumption of 15.4 terawatt hours in 2020. 

“Making that sort of switch is practically impossible, as it would require more than four million powerful computer chips known as GPUs that are currently in huge demand, with limited supply. This would cost $100 billion, which even Google’s deep pockets would struggle to fund," he wrote.

The IEA estimates that a single Google search takes 0.3 watt-hours of electricity, while a ChatGPT request takes 2.9 watt-hours.

Also read | What’s behind the renewed interest in nuclear power?

Online interactions using AI such as chatbots need the support of large data centres equipped with hundreds of machines with high-performance chips. Energy is required not just to keep these data centres running 24x7 but also to keep the machines cool. That’s just one part of the energy requirement of AI.

The training and inference phases of AI models also consume a lot of electricity. During the training phase, models learn and develop by digesting vast amounts of data. In the inference phase, models go live and respond to specific real-world problems and queries. While both phases are energy intensive, the inference phase accounts for the bulk of the consumption. 

According to some estimates, training a large language model like OpenAI’s GPT-3 uses nearly 1,300 MWh of electricity, equivalent to the annual consumption of about 130 US homes.

Can India handle a surge in demand for power from a widespread deployment of AI?    

India is rapidly digitalising and that itself is set to increase the power demand. A joint study by Microsoft and LinkedIn found that over 90% of India’s knowledge workers use AI tools to improve productivity. 

Rising usage of generative AI models such as ChatGPT by students, professionals and industry will lead to the setting up of many large data centres. 

Also read | How India keeps up with its power needs, in 8 charts

India is already one of the larger data centre hubs in the world, and with increased data localisation the number of data centres will increase. These data centres need a reliable and stable flow of electricity.

Governments, regulators and electricity generating and distribution companies must consider this demand when forecasting to ensure adequate capacity is built.