In the past few years, Large Language Models (LLMs) have captured the world’s imagination. Services like ChatGPT, Google’s Gemini, and Claude can write poetry, debug code, and draft complex legal documents in seconds. They feel ethereal, like disembodied minds existing purely in the digital ether of “the cloud.” This perception of weightlessness, however, masks a very real and resource-intensive physical reality.
The magic of AI is not conjured from thin air. It is forged in the fires of immense computation, and that process comes with a staggering, and often hidden, environmental price tag. As we race to build ever more powerful models, it’s crucial to look behind the curtain and ask: What is the true environmental cost of training a single, state-of-the-art large language model? The answer lies in two critical resources: energy and water.
The Engine of AI: Energy-Hungry Data Centres
LLMs are not born in a typical computer. They are trained inside sprawling, factory-sized buildings called data centres. These facilities house tens of thousands of highly specialised, powerful computer chips known as GPUs (Graphics Processing Units) or TPUs (Tensor Processing Units), all wired together to function as a single supercomputer.
To train an LLM, this hardware runs at maximum capacity, 24/7, for weeks or even months. It performs a mind-boggling number of calculations (quintillions of operations) to learn patterns, grammar, and concepts from a vast repository of text and data scraped from the internet. This process consumes an astronomical amount of electricity.
How much? Precise, official figures are a closely guarded secret by the tech giants, but academic research provides some startling estimates.
A landmark 2019 study from the University of Massachusetts, Amherst, estimated that training one particular (now relatively small) AI model could emit over 284,000 kg (626,000 pounds) of carbon dioxide equivalent. That’s nearly five times the lifetime emissions of the average car, including its manufacturing.
That was years ago. Today’s flagship models are exponentially larger. While numbers for models like OpenAI’s GPT-4 are not public, researchers have estimated that its predecessor, GPT-3, consumed approximately 1,287 megawatt-hours (MWh) of electricity for its initial training run.
To put that into perspective:
- That’s enough electricity to power over 120 average American homes for an entire year.
- The associated carbon footprint was estimated to be over 550 tons of CO2, and this was accomplished using Google’s carbon-efficient data centres that utilize a mix of renewable energy. If the same training had been done using the average global energy grid, the emissions would have been significantly higher.
This immense energy draw is the direct carbon footprint of AI, contributing to greenhouse gas emissions and accelerating climate change. But it’s only half of the story. There is another, even more hidden cost.
The Unseen Cost: The Colossal Water Footprint
The thousands of processors running in a data centre generate an incredible amount of heat. To prevent them from overheating and shutting down, they must be constantly cooled. While some modern facilities use liquid cooling or are situated in cold climates to use outside air, a vast number of the world’s data centres rely on a simpler method: water-based cooling.
This system works much like a power plant’s cooling tower. Water is pumped through the facility, absorbs the heat from the hardware, and is then moved to a cooling tower where a portion of it is evaporated into the atmosphere to dissipate the heat. This process consumes massive quantities of fresh water.
Groundbreaking research from the University of California, Riverside, has begun to quantify this “water footprint.” Their study estimated that:
- Training GPT-3 in Microsoft’s state-of-the-art US data centres likely consumed 700,000 litres (about 185,000 gallons) of fresh water.
- For context, that’s enough water to manufacture over 300 cars.
This water is often drawn from local freshwater sources, putting a direct strain on community water supplies, particularly as many data centres are located in regions that already face water stress.
Beyond Training: The Never-Ending Cost of Inference
The massive environmental cost isn’t just a one-time expenditure during training. Once a model is trained, it is deployed for “inference”—the technical term for when users like us ask it questions. Every single query sent to an LLM requires a computational cycle, which consumes more energy and, in turn, more water for cooling.
While the cost of a single query is tiny, the scale is planetary. With hundreds of millions of people now using these services daily, the cumulative cost of inference is enormous and ongoing.
A report from Google itself revealed a shocking, tangible metric: a simple conversation of 20 to 50 prompts with an LLM can have a water footprint of about 500 ml (16 oz). Every time you ask a series of questions, you are essentially causing a bottle of water to be consumed for data centre cooling. Scaled across billions of daily interactions, this adds up to an unsustainable, continuous drain on global resources.
The Path to Sustainable AI: Mitigation and Responsibility
The goal is not to demonize AI or halt its progress. The benefits of these models in science, medicine, and education are undeniable. The challenge, therefore, is to pursue this progress responsibly. The industry is beginning to respond with a multi-pronged approach:
- Efficient Models: Researchers are developing new techniques like “model pruning” (trimming down less important parts of the model) and “quantization” (using less precise but more efficient numbers) to make LLMs smaller and less computationally expensive without sacrificing too much performance.
- Efficient Hardware: Companies like Google and NVIDIA are designing next-generation chips (TPUs and GPUs) that are specifically optimized for AI workloads, aiming to deliver more computational power for every watt of energy consumed.
- Smarter Data Centres: Tech giants are actively working to improve the efficiency of their facilities. This includes powering them with renewable energy (Google and Microsoft have pledged to match 100% of their energy use with renewable purchases), locating new data centres in colder climates to reduce cooling needs, and investing in advanced, water-free cooling technologies.
- The Need for Transparency: Perhaps the most critical step is for AI companies to be transparent. A future where major AI labs regularly publish an “AI Nutrition Label” for their models—detailing the training energy, carbon footprint, and water usage—would allow customers and researchers to make informed choices and drive the industry toward a more sustainable standard.
Conclusion: Intelligence Without Costing the Earth
Large Language Models are one of humanity’s most powerful new tools, but they are not magic. They are industrial products built on a foundation of immense physical resources. The ethereal cloud has a shadow on the ground, measured in megawatts of power and millions of litres of water.
As we stand at the dawn of the AI age, our greatest challenge is not just to build more intelligent systems, but to build them wisely. The true mark of genius in the coming decade will be found not just in the capabilities of our AI, but in our ability to achieve that intelligence without costing the Earth.
