Training GPT-4 Used Enough Energy to Power 33,000 Indian Households for a Year

Training GPT-4 Used Enough Energy to Power 33,000 Indian Households for a Year

Part of “The AI You Don’t See” series by Akshay A. Walimbe

Let me tell you about a number that nobody in the AI industry likes to talk about.

Training GPT-4  the model that powers the ChatGPT most of us use daily  consumed an estimated 50 gigawatt hours of electricity, according to multiple independent analyses of OpenAI’s compute requirements. To put that in perspective, the average Indian household consumes roughly 1,500 kilowatthours of electricity per year. Fifty gigawatt hours is enough to power over 33,000 average Indian households for an entire year. Not a month. A year.

And that was just the training. Every time you ask ChatGPT a question, every time someone runs a query, every time an AI agent processes a task  that is inference. According to industry estimates, inference now accounts for 80 to 90 per cent of all AI computing power, making it the larger cost at scale when you add it up across hundreds of millions of users, day after day.

Let me put that in terms you can feel. If you live in a Tier-2 Indian city say, Nagpur or Indore imagine taking an entire mid sized township, tens of thousands of homes, and telling them: sorry, no electricity for a year. We needed it to teach a machine to write poetry and summarise emails.

That is the trade off. Except nobody asked those families.

The Numbers Nobody Wants You to See

Here is what we know about the energy appetite of large AI models, and I want to be precise because the AI industry has been remarkably imprecise about this on purpose.

A widely cited Goldman Sachs estimate from 2024 claimed that a single ChatGPT query consumed roughly ten times the energy of a standard Google search. More recent analyses, including data from Epoch AI and Google’s own disclosures, suggest the gap is narrower for simple queries  closer to comparable, around 0.3 watt hours each  though complex reasoning queries and long context prompts can consume substantially more, up to 40 watt hours per query. The point stands, even if the exact multiple is debated: AI is an energy hungry technology.

Google confirmed in its 2024 environmental report that the company’s total greenhouse gas emissions had risen 48 per cent since 2019, with data centre electricity consumption growing 17 per cent in 2023 alone. To its credit, Google achieved a 12 per cent reduction in data centre energy emissions in 2024 through clean energy projects, even as electricity demand grew 27 per cent. But the overall trajectory remains upward.

Microsoft told the world it would be carbon negative by 2030. Then its sustainability report showed emissions had increased by 23.4 per cent since 2020. The primary driver? Massive data centre expansion to support AI workloads. Microsoft itself called the 2030 goal a “moonshot” and acknowledged “the moon has gotten further away.” The very company promising to save the planet is building infrastructure that consumes more of it.

And it is not just electricity. Training a large language model requires enormous amounts of water for cooling the data centres that run the computations. Researchers at the University of California, Riverside estimated that every 100 word AI prompt uses roughly half a litre of water. At the model training level, one widely cited study estimated that training GPT-3 alone consumed approximately 700,000 litres of fresh water. GPT-4, with its significantly larger architecture, would have consumed substantially more. Microsoft’s own reporting showed its water consumption rose 34 per cent year over year by 2023, reaching approximately 6.4 million cubic metres.

In a country like India, where entire villages queue for water tankers in the summer, the idea that thousands of litres of fresh water are being evaporated to cool servers that help someone in San Francisco generate a marketing email should make you pause.

The Geography of the Burden

Here is what bothers me most. The environmental cost of AI is not distributed evenly. It never is.

The companies training these models are headquartered in the United States. Their data centres are spread across the globe increasingly in regions where electricity is cheap, land is available, and environmental regulations are lighter. Parts of Latin America, Southeast Asia, and yes, India.

India is aggressively courting data centre investment. We have announced data centre parks, special economic zones, fast tracked land approvals. The pitch is straightforward: cheap power, cheap land, growing demand.

But here is the question nobody asks during those investor presentations: whose power are we giving away?

India added approximately 18 gigawatts of new renewable energy capacity in the financial year 2023-24. That sounds impressive until you realise that a single hyperscale data centre can consume 100 megawatts or more. Build a dozen of those  and the plans call for far more  and you are diverting a meaningful slice of India’s hard won clean energy capacity to cooling servers that primarily benefit companies and users in the West.

This is not to say India should reject data centre investment  the jobs, the technological capability, and the sovereign infrastructure are real benefits. Companies like AdaniConneX, Nxtra, and Reliance Jio have pledged to power their facilities with renewable energy by 2030. India’s data centre operators are achieving power usage effectiveness ratios as low as 1.3, better than the global average of 1.5 to 1.8.

But the trade offs deserve honest acknowledgement. According to S&P Global, 60 to 80 per cent of India’s data centres could face high water stress within this decade. The electricity that powers an AI data centre in Maharashtra competes with the grid that serves rural Vidarbha. The water that cools those servers draws from the same supply that irrigates fields. The environmental cost of AI is not abstract. And the communities living near these facilities deal with the local costs  noise, heat, water consumption  while much of the value flows to global companies and their users abroad.

But It Gets Worse

The energy problem is not static. It is accelerating.

Each generation of AI models is larger than the last. GPT-3 was trained on roughly 300 billion tokens of text data. GPT-4 reportedly used around 13 trillion tokens. Meta’s Llama 4 Scout was trained on 40 trillion tokens. According to Epoch AI, the compute required for frontier training runs has been growing at 2.4 times per year since 2016. If that pace continues, the largest training runs will cost more than one billion dollars by 2027.

The International Energy Agency’s April 2025 report projected that global data centre electricity consumption could more than double by 2030, reaching approximately 945 terawatt hours  equivalent to Japan’s entire national electricity consumption. AI workloads are the primary driver. Goldman Sachs estimated that data centre power demand could grow 165 to 175 per cent by 2030, with AI responsible for the majority of that increase.

We are building an industry whose energy appetite doubles every few years, in a world that is trying desperately to cut emissions.

And here is the part that should alarm you: the efficiency gains are not keeping pace with the growth. Yes, models like DeepSeek R1 have shown genuine efficiency innovations  its mixture of experts architecture activates only 5.5 per cent of its total parameters per token. The company claimed a training cost of under six million dollars, though independent analyses by SemiAnalysis estimate the true all in cost at closer to 1.6 billion dollars when you include R&D, infrastructure, and failed experiments. Google’s own data shows that energy per Gemini query dropped 33 times between May 2024 and May 2025.

But the frontier labs are not using efficiency gains to consume less. They are using them to train bigger models more often. The efficiency savings get reinvested into scale, not sustainability.

This is not a technology problem. It is a priorities problem.

The Invisible Cost, the Indian Paradox

When you open an app on your phone and ask an AI assistant a question, you see the answer. You do not see the data centre humming with thousands of GPUs. You do not see the cooling towers consuming thousands of litres of water per hour. The AI industry has gotten very good at making the cost invisible.

Compare this to any other industry. When you buy a car, you know it burns petrol. When you fly, you know the plane emits carbon. But when you use AI? Nothing. No carbon label. No energy disclosure. No environmental impact statement.

And India finds itself in an impossible position. On one side, India has committed to reaching net zero emissions by 2070, investing massively in renewable energy. On the other side, India wants to be a global AI hub. More data centres mean more jobs, more investment, more technological sovereignty.

But the question nobody is answering publicly is: what happens when those two ambitions collide? India’s data centre power demand is expected to grow roughly five times by 2030, from 13 terawatt hours to 57 terawatt hours, according to the Institute for Energy Economics and Financial Analysis. That means an additional 15 to 30 gigawatts of renewable energy capacity will be needed just for data centres. If India succeeds in attracting the investment it is courting, the energy demand could consume a significant share of the country’s new renewable capacity. The environmental costs stay local. The profits often go global.

When We Talk About Ethical AI

We spend a lot of time talking about AI ethics in terms of bias, fairness, and transparency. These are important conversations. I am writing an entire book about them.

But there is a question we have collectively decided to ignore, and I think it is time we stopped.

When we talk about ethical AI, we forget to ask: ethical for the planet?

Is it ethical to train a model that consumes as much energy as tens of thousands of homes use in a year, so that someone can generate a birthday poem? Is it ethical to build data centres that drain aquifers, so that a company can shave three seconds off a customer service interaction? Is it ethical to pour billions into infrastructure whose carbon footprint is growing faster than any other sector in tech, while simultaneously pledging net zero commitments?

These are not anti technology questions. I build AI systems for a living. I am not suggesting we stop. I am suggesting that we stop pretending the cost is zero.

Every AI query has a price. Not just in compute dollars. In carbon. In water. In electricity that could have powered a home or a hospital or a school.

I’m have written a book about exactly this  how AI and automated systems make decisions about your life, where accountability disappears, and what we can do about it. If you want to know morea about this book or order a copy, you can do it here: https://akshaywalimbe.com/beyond-bias/