1. AI‑driven data‑center electricity is outpacing all other sectors

The International Energy Agency (IEA) estimates that global data‑center electricity consumption reached about 415 TWh in 2024—roughly 1.5 % of total electricity use—and is growing at 12 % per year【1†L81-L85】. More strikingly, the agency projects that consumption could double to 945 TWh by 2030, representing almost 3 % of global electricity demand【1†L107-L110】. The acceleration is tied directly to the rollout of high‑performance accelerators (GPUs, TPUs) needed for AI model training, which now account for a significant share of server power draw【1†L115-L117】.

These numbers are not just abstract forecasts; they translate into real‑world climate pressure because electricity generation remains heavily fossil‑fuel dependent in many regions.

2. Accelerated servers are the engine of growth

Accelerated servers—machines equipped with GPUs or specialized AI chips—are responsible for almost half of the net increase in data‑center electricity use, even though they represent a smaller fraction of total hardware stock【1†L115-L119】. The IEA’s “Lift‑Off” scenario, which assumes faster AI adoption, predicts a 45 % higher electricity demand for data centres by 2035, pushing global consumption toward 1 700 TWh【1†L219-L223】.

This rapid scaling underscores a structural shift: AI is no longer a peripheral workload but a primary driver of data‑center design, location, and energy procurement strategies.

3. Training giant models burns massive amounts of energy

Training large language models (LLMs) is extraordinarily power‑intensive. According to a 2023 study in Joule, training the 176‑billion‑parameter BLOOM model required 433 MWh, while GPT‑3, Gopher, and OPT consumed 1 287 MWh, 1 066 MWh, and 324 MWh, respectively【4†L78-L82】. To put this in perspective, a single GPT‑3 training run uses as much electricity as a small town does in a month.

When multiplied by the growing number of models being trained annually, the cumulative carbon footprint becomes substantial. If each MWh of electricity is produced with an average emission factor of 0.4 kg CO₂ (global average in 2024)【1†L83-L84】, a single GPT‑3 training run emits roughly 500 t CO₂, equivalent to the yearly emissions of about 100 average‑size passenger cars.

4. Inference— the hidden, ongoing cost

While training is a one‑off event, inference runs 24/7, turning AI into a perpetual energy drain. The same Joule article notes that OpenAI’s ChatGPT infrastructure consumes 564 MWh per day during inference, translating to 2 9 Wh per request【4†L94-L100】. Extrapolated to billions of daily queries, the annual electricity demand can reach 29 TWh, comparable to the total consumption of a small European country such as Ireland【4†L150-L153】.

Google’s own analysis estimates that AI‑related electricity accounted for 10‑15 % of its total data‑center power in 2021, a share that could rise dramatically as generative AI becomes embedded in everyday services【4†L149-L152】.

5. Industry and policy responses are still nascent

Tech giants have begun to acknowledge the issue. Microsoft, Google, and Meta have pledged to shift AI workloads to renewable‑powered data centres and to publish environmental impact statements for major model releases. However, the scale of the problem outpaces current commitments. The IEA stresses that grid integration challenges will intensify because AI data centres are highly concentrated geographically, making reliable renewable supply harder to guarantee【1†L211-L214】.

Regulators in the EU are discussing mandatory AI carbon‑footprint disclosures, mirroring existing requirements for large‑scale ICT equipment. Such policies could create market incentives for more efficient hardware and for software‑level optimizations that reduce unnecessary inference.

6. A path forward: sustainable AI by design

To keep AI’s climate impact in check, three actions are essential:

1. Efficiency‑first model development – Researchers should prioritize architectures that achieve similar performance with fewer parameters, as demonstrated by Google’s GLaM model, which required 2.8 × less energy than GPT‑3 for comparable training【4†L200-L204】.
2. Dynamic inference scaling – Deploying models only when demand justifies it, and using techniques such as model distillation and quantization, can cut inference energy by up to 70 %.
3. Renewable‑backed data‑centre expansion – New AI‑focused facilities must be co‑located with abundant renewable resources (e.g., solar farms in the American Southwest or wind corridors in the North Sea) and must incorporate advanced cooling methods that minimize auxiliary power use.

By treating energy use as a first‑class metric, the AI community can avoid the trap of Jevons’ paradox—where efficiency gains paradoxically lead to higher overall consumption—highlighted in the Joule commentary【4†L205-L208】.