Energy and Sustainability Limits

Training a single frontier AI model can consume as much electricity as hundreds of American households use in a year. Running a large language model at scale — answering billions of queries per day — draws power continuously at the level of a small city. These are not hypothetical projections. They are measurements from the data centers that run today's systems. The rapid scaling of AI capability has a physical price paid in watts, liters of water, and tons of carbon dioxide. Understanding this price is not peripheral to understanding AI — it is central to evaluating whether the current development trajectory is sustainable.

Training Cost: The One-Time Expense

Training a frontier model is a one-time but enormous energy event. A rough estimate for training GPT-3 (released 2020, 175 billion parameters) was approximately 1,287 megawatt-hours of electricity — comparable to the annual electricity consumption of over 100 American homes. GPT-4 and comparable 2023-vintage models are believed to require substantially more, though exact figures are not publicly disclosed by most labs. This training energy has a carbon footprint that depends critically on where the data center is located and what its power source is. Training in a region powered primarily by coal generates roughly 500 grams of CO2 per kWh; training in a region with substantial renewables may generate under 50 grams per kWh. The same training run can have a carbon footprint 10 times larger depending on geography and energy source — a fact that has made data center location and power sourcing strategically important decisions. Beyond carbon, training data centers consume large quantities of water for cooling. Cooling systems — which dissipate the heat generated by dense GPU clusters running at full power — use water evaporation as a heat-rejection mechanism. Estimates for large training runs reach tens of millions of liters of water. In water-stressed regions, this creates competition with agricultural and municipal uses. The training cost is a one-time investment per model version. What matters for ongoing sustainability is inference cost.

Inference Cost: The Ongoing Expense

Inference is the process of running a trained model to produce a response to a user query. Unlike training, which happens once, inference happens millions to billions of times per day for widely-deployed systems. A single query to a large language model — generating a few hundred tokens — uses roughly the equivalent of running a search engine query times 10 in energy terms. Queries that require long outputs, multi-turn conversations, or multiple model calls (as in agentic systems) can cost substantially more per query. Microsoft's own sustainability reports noted that its emissions increased by nearly 30% between 2020 and 2023, partially attributable to the expansion of AI infrastructure. Google's 2024 environmental report similarly disclosed a significant increase in emissions driven by data center growth for AI. These disclosures, while incomplete, confirm that at-scale AI inference is a major and growing contributor to corporate carbon footprints. The inference cost creates a tension at the heart of AI deployment: the most capable models tend to be the largest, and the largest models are the most expensive to run per query. Deploying a frontier model for every query is not economically or environmentally sustainable at web scale. This drives the development of model compression techniques — distillation, quantization, pruning — that attempt to achieve near-frontier performance at dramatically lower inference cost. Smaller, specialized models (often called edge models) that run on device — on a smartphone, a laptop, an IoT sensor — consume orders of magnitude less energy than cloud-based frontier models. This architectural alternative trades some capability for dramatic efficiency gains. The tradeoff is not neutral: edge models cannot match frontier models on the most demanding tasks, but for many common tasks they may be adequate while consuming a fraction of the energy.

Renewable energy is frequently cited as the solution to AI's carbon footprint. In principle, powering data centers with wind and solar eliminates operational carbon emissions. In practice, three complications arise. First, renewable supply is intermittent — the sun does not always shine, the wind does not always blow. Data centers require 24/7 power. Without massive grid-scale battery storage (which does not yet exist at the needed scale), data centers cannot run solely on renewable supply without fossil backup. Second, large-scale AI infrastructure expansion is outpacing renewable capacity additions in many regions. Building and powering new data centers faster than the renewable energy grid grows means that marginal electricity consumption draws on fossil sources even if average consumption is offset by renewable certificates. Third, the physical infrastructure of AI — the GPUs, servers, networking hardware, and building materials of data centers — has a manufacturing footprint that is separate from operational electricity. Producing a high-end GPU involves rare materials, energy-intensive semiconductor fabrication, and global supply chains with their own emissions. None of this means AI development should stop. It means the sustainability calculus is genuinely complex, and claims of 'carbon-neutral AI' require careful scrutiny of what exactly is being offset and how.