AI’s Energy Footprint — How Much Power Does AI Actually Use?

Status: 🟩 COMPLETE 🟦 LIVING Tags: energy, environment, sustainability, carbon, data-centres, AI-footprint

What it is

AI systems — particularly large language models and image/video generation — require enormous amounts of computing power, which means significant electricity consumption. This has become a major environmental, economic, and policy discussion as AI scales rapidly.

“AI’s energy footprint” refers to the electricity consumed by:

Training AI models (the one-time process of building the model from data)
Running AI inference (the ongoing process of answering queries — what happens every time you ask ChatGPT a question)
The infrastructure required (data centres, cooling, water cooling systems, hardware manufacturing)

Training vs inference: the two types of energy use

Training (one-time, enormous)

Training a large AI model requires running billions of mathematical calculations on thousands of specialised chips (GPUs or TPUs) for weeks or months.

Estimated energy costs for training major models:

GPT-3 (2020): ~1,300 MWh (enough to power ~130 average Australian homes for a year)
GPT-4 (estimated, 2023): Many times GPT-3; OpenAI hasn’t disclosed
Gemini Ultra (estimated): Similar scale to GPT-4
Llama 3 405B (Meta estimate): ~40 million GPU hours

These are large but one-time costs. Once trained, the model is “frozen” and inference is much cheaper.

Inference (ongoing, cumulative)

Every time someone sends a message to ChatGPT, the query is processed on data centre hardware. Individually, this is tiny — a single query uses a small fraction of a watt-hour. But at scale (ChatGPT has ~100 million daily active users), it adds up.

Key comparison (approximate):

A Google search uses ~0.0003 kWh
A simple ChatGPT text query uses ~0.001–0.01 kWh (3–30× a Google search)
An AI image generation request uses ~0.002–0.01 kWh
An AI video generation request uses ~0.05–0.5 kWh (much higher)

OpenAI’s estimated inference electricity (2024): ~500+ GWh/year — roughly equivalent to powering 45,000 average Australian homes.

Water use (often overlooked)

Data centres require enormous amounts of water for cooling. AI training and inference contribute to this:

Microsoft’s water consumption rose 34% between 2021–2022, largely attributed to AI workloads
A single training run for a large model may consume hundreds of thousands of litres of water

This matters particularly in water-scarce regions where data centres are located.

The nuclear energy connection

The energy demands of AI have led major tech companies to invest in nuclear power:

Microsoft: 20-year agreement with Constellation Energy to restart Three Mile Island (US) nuclear plant (2023)
Google: Investment in small modular reactors (SMRs) from Kairos Power (2024)
Amazon: Nuclear power agreements for data centre supply

AI’s energy appetite is one driver behind renewed interest in nuclear power globally — including the Australian debate about nuclear energy.

Carbon emissions and renewable energy

Whether AI’s energy use is carbon-intensive depends on where that energy comes from:

Data centres in regions with high renewable energy penetration (Nordic countries, Australia’s ACT with 100% renewable targets) have much lower carbon footprint per computation
Data centres in regions relying on coal or gas have higher footprint
Most major AI companies (Google, Microsoft, AWS, Anthropic) have 100% renewable energy commitments — though “renewable matching” (purchasing renewable energy credits) differs from actual on-site generation

Australian context:

Australia’s grid is increasingly renewable (45%+ in 2024)
But “renewable energy” on Australian servers depends on the specific region and time of day
Building new data centres for AI may accelerate demand for renewable energy — a mixed outcome

Is AI’s energy use justified?

This is a genuine debate:

Arguments that it’s worth it:

AI accelerates scientific research (protein folding, drug discovery) with high value multipliers
AI tools replace tasks that would require other energy-intensive activities (flights, printing, physical research)
Efficiency gains elsewhere (better logistics, optimised energy grids, materials science) may offset AI energy use
As hardware improves, energy per useful AI task decreases

Arguments for concern:

Current growth trajectory is unsustainable without massive new energy infrastructure
The benefits are unevenly distributed (wealthy countries benefit most; energy costs are global)
AI may be used for trivial purposes (generating marketing copy) not worth the energy
Rebound effects: as AI gets cheaper, people use more of it, increasing total energy use even as efficiency per query improves

How AI companies are reducing their footprint

More efficient models: Smaller, more efficient models (Mixtral’s sparse MoE, Llama 3.2’s smaller versions) achieve similar quality with less compute
Better hardware: Each generation of NVIDIA GPUs (H100 → B200 Blackwell) is more energy-efficient per operation
Renewable energy procurement: Signing long-term renewable energy contracts
Cooling innovation: Liquid cooling, underwater data centres (Microsoft’s Project Natick), AI-optimised cooling management
Inference optimisation: Techniques like quantisation and distillation reduce model size without major quality loss

What individual users can do

The truth: individual user choices have minimal impact on aggregate AI energy consumption. But for context:

Prefer shorter prompts that require less processing
Choose models appropriate to the task (a lightweight model for simple questions; frontier model for complex reasoning)
Be aware that AI image and video generation is substantially more energy-intensive than text

For organisations: measure and track AI energy use as part of sustainability reporting.

Australian relevance

Australia is a significant importer of AI services (using compute in US data centres); Australian carbon accounting for AI should consider offshore emissions
Australia’s National AI Strategy considers energy sustainability but detailed frameworks are still developing (2024–2026)
The Australian government’s data centre power agreements and location choices will affect AI energy footprint for government AI use

Gotchas

Energy numbers are estimates and vary widely. AI companies don’t uniformly disclose energy consumption. Most figures you see are estimates with significant uncertainty.
“100% renewable” claims should be scrutinised. Energy certificate purchases don’t necessarily mean the physical electrons running your computation are from renewable sources.
The comparison to Google search is often misused. “ChatGPT uses X times more energy than Google” is sometimes cited in misleading contexts. The relevant comparison is energy per unit of value provided, which is harder to measure.
Hardware manufacturing is often omitted. The energy cost of manufacturing the chips and servers themselves is often not included in AI energy footprint estimates.

Sources

Patterson et al., “Carbon and the Beast: Energy Use in Machine Learning” (2021, revised 2022)
Lannelongue et al., “Green Algorithms: Quantifying the Carbon Footprint of Computation” (2021)
Goldman Sachs AI energy report “AI is poised to drive 160% increase in data center power demand” (2024)
Microsoft Sustainability Report 2023 (water consumption data)
International Energy Agency — “Electricity 2024” (AI chapter)
Australian Energy Market Operator (AEMO) — renewable energy percentage 2024
OpenAI energy consumption estimates: various third-party analyses (2024)

Tech & AI, Explained

Explorer

ai-energy-footprint