Written by Bijendra Shandilya & Ankit Kumar  final year students at Indian Institute of Management, Rohtak (IIM-R)

Each one of the large language models you ever used was trained within the shadows of a power station. For many weeks thousands of GPUs demanded energy which could illuminate an entire town, just to train a model that would be able to help you complete your sentences. This is no mere metaphor rather it resembles an accounting challenge that becomes legal too.

For the past three years the topic of AI regulation has been revolving around issues of bias misinformation copyright and existential risks. While these discussions were taking place, another question emerged quietly in the background who is accountable for the consumption of electricity, water, and carbon required for the existence of AI, and is it possible for our ESG reporting systems to even see it?

The simple answer for now would be partially. And the gap between what companies report about their AI training and what it really requires from our planet grows to be one of the most intriguing stories of 2026.

1. The Numbers That Should Worry a Sustainability Officer

According to the International Energy Agency, the amount of energy used by data centers is estimated to double from around 415 terawatt-hours (“TWh”) in 2024 to almost 950 TWh in 2030, which is slightly more than the total amount of electricity consumed in Japan. Notably one year later when the agency provided a follow-up analysis, the growth pace had only increased the electricity usage of data centers grew 17 per cent in 2025, with AI-oriented data centers increasing at an even faster pace and the capital expenditure of five largest technology firms exceeded 400 billion US dollars. 

Individual model estimates bring the abstract concept to life. It was calculated in a highly referenced 2021 study that it took 1,287 megawatt-hours of electricity and produced about 552 tonnes of CO2 equivalent in the process of training the GPT-3 model, which is the yearly emission of more than one hundred cars. A life-cycle assessment of BLOOM model, a 176 billion parameters model showed something even more revealing based on how the calculation boundary was defined, the footprint could be calculated either at 25 tonnes or more than 50 tonnes of CO2 equivalent in the exact same training period. One fact is important here if two conscientious researchers can get so far apart in their calculations of the exact same model then any regulator considering mandating “energy disclosure” on AI models faces a measurement nightmare before filling out a single form.

2. Big Tech’s Own Reports Are Starting to Say the Quiet Part Out Loud

The most striking evidence that something has shifted doesn’t come from activists or academics it comes from the sustainability reports of the AI companies themselves. Microsoft’s 2026 environmental sustainability report disclosed that its total emissions rose twenty-five per cent year-over-year, driven “primarily” by data-center expansion a jump reported in detail by ESG Dive. Its Scope 2 emissions the electricity-purchase category jumped from two per cent to thirteen per cent of its total footprint in a single year, largely because the company deliberately stopped counting certain renewable energy certificates that were not tied to genuinely new power generation. Google reported an eighteen per cent emissions increase for similar reasons Amazon, sixteen per cent. A United Nations telecommunications agency report found that the combined indirect emissions of four leading AI-focused tech companies rose by an average of one hundred and fifty per cent between 2020 and 2023.

Water tells a similarly uncomfortable story. Independent analysis of corporate disclosures found that Google’s data centers alone consumed an estimated twenty-seven billion liters of potable water in 2024 with more than a quarter of total withdrawals drawn from regions already under water stress even as the company simultaneously reported replenishing a rising share of what it used. None of this can be considered green washing in any crude sense. The way Microsoft puts this out is refreshingly honest it admits that there is an actual tension between the resource requirements of the AI on the one hand and its sustainability goals on the other and the solutions that it brings forth for the latter are not scaling fast enough to cope with the demand. It simply means that the numbers themselves in the industry are going the wrong way compared to the promises it makes about climate change.

3. Why Your Ordinary ESG Framework Was Not Built for This

Corporate ESG disclosure India’s Business Responsibility and Sustainability Reporting (“BRSR”) framework, the European Union’s Corporate Sustainability Reporting Directive (“CSRD”), the Greenhouse Gas Protocol’s Scope 1/2/3 architecture was designed to aggregate a company’s environmental footprint across its entire operations. That works reasonably well for a steel plant or a retail chain where environmental impact is spread fairly evenly across many activities.

AI training breaks that assumption. A single training run can concentrate more electricity demand into a few weeks at a handful of physical sites, than the rest of a company’s operations combined. Reported at the company level that spike simply disappears into an aggregate number indistinguishable from ordinary cloud-computing growth. India’s BRSR Core now requires the top listed companies to disclose assured data on energy, water, waste and emissions, and that is a genuinely significant step but it still cannot show what share of a company’s footprint came from training one specific foundation model as opposed to running email servers.

Ironically just as AI’s environmental footprint has become impossible to ignore, Europe’s flagship disclosure law is moving in the opposite direction. The “Omnibus” reforms finalised in February 2026 raised the CSRD’s mandatory-reporting threshold[1] to companies with more than one thousand employees and above €450 million in turnover removing a large number of mid-sized data-center operators and AI infrastructure suppliers from the regime entirely, in the name of competitiveness and reduced administrative burden.

4. Regulators Are Starting to Respond in Two Very Different Ways

Two regulatory experiments, running in parallel, show where this is heading.

The first is the European Union’s Artificial Intelligence Act, which took the unusual step of regulating the model rather than the company. Under Article 53[2], providers of general-purpose AI models must document the known or estimated energy consumption of the model itself, as part of their technical documentation the first binding, model-level energy disclosure rule anywhere in the world. It came into force for new models in August 2025. It is a genuine innovation, though it has a built-in weakness providers are permitted to estimate energy use from computational resources when direct measurement is unavailable, which reopens exactly the boundary problem the BLOOM study exposed above.

The second more dramatic experiment is happening at the level of the electricity grid itself in Ireland. Facing a real risk of blackouts as hyper scale data centers concentrated around Dublin which already consumes more electricity than all of the country’s urban households combined Ireland’s regulator effectively froze new data-center grid connections from 2021. In December 2025, it replaced that freeze with conditional access new large facilities must now bring their own dispatch able generation, site themselves away from constrained parts of the grid, and meet at least eighty per cent of their annual demand through genuinely new renewable generation within Ireland. Environmental groups have already sued, arguing the rules still lock the country into fossil-fuel-backed connections for years. It is messy contested and unresolved but it may prove a more consequential precedent than anything coming out of Brussels, precisely because it reaches the physical grid rather than a paperwork trail.

5. Where Does This Leave India?

India has not adopted an AI-specific energy-disclosure rule comparable to Article 53. Its response so far leans on BRSR the Bureau of Energy Efficiency’s data-center guidelines, and a policy preference set out in a recent Council on Energy, Environment and Water study for “frugal AI” smaller task-specific models suited to domestic infrastructure, rather than chasing frontier-scale training runs. This is a reasonable approach but as Indian companies start hosting fine-tuning or using large foundation models at scale the discrepancy between the disclosure of corporate BRSR and that of models will only increase, and it would certainly be more sensible to bridge the gap beforehand rather than be in the position in a few years’ time explaining why the figures simply do not match up.

6. The Uncomfortable Bottom Line

The environmental impact of AI is no longer a hidden secret if anything the corporate sustainability reports have become shockingly transparent on this issue. The real issue lies in its structure the metrics that were developed to monitor and control the environmental impact were meant for a world where the emissions accumulate gradually from diffused sources, but not for a technology where massive energy consumption takes place instantly within a few buildings.

This would require the completion of all three processes simultaneously disclosure that admits limitations of measurement grid regulation of physical access rather than paperwork and the ESG framework that recognizes the role of AI. Until then, the industry’s sustainability reports and its actual electricity bills will keep telling two different stories and regulators courts and investors are only just beginning to notice the difference.

---

[1]Directive (EU) 2026/470 of the European Parliament and of the Council amending Directive 2013/34/EU, Directive (EU) 2022/2464 and Directive (EU) 2024/1760 as regards sustainability reporting and due diligence requirements [2026] OJ L 2026/470.

[2]Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) [2024] OJ L 2024/1689, art 53 and Annex XI.