At Rise of AI 2023 in Berlin, Susan Wegner, Managing Director at Deloitte Digital, presents an approach to measuring the ecological footprint of AI projects across their full lifecycle. She opens with a finding from MIT that digital solutions already produce more carbon emissions than aviation and cars, and cites studies showing GPT-3's training used energy equivalent to about 63 homes for a year, or 61 million phone charges, against roughly 3 homes and 3 million charges for the BLOOM model. She argues most attention goes to model training, while an 'iceberg' of other emissions, from employee devices and travel to data storage and ongoing monitoring, is usually ignored. Using a worked example from a small image-processing project with two employees, she shows how emissions accumulate across the lifecycle phases, with data processing and model development contributing the most. She reviews available measurement tools, including cloud-provider dashboards, a training footprint estimator and the open source tool CodeCarbon, and closes with practical levers for reducing emissions: model choice, hardware choice, cloud region and training timing.
Key takeaways
- Wegner cites an MIT-based finding that digital solutions already emit more carbon than aviation and cars combined.
- Training GPT-3 used energy equivalent to about 63 homes for a year, versus roughly 3 homes for the BLOOM model, by her cited studies.
- In her team's worked example, the data-processing and model-development phases produced far more emissions than deployment or monthly monitoring.
- CodeCarbon is an open source tool that can be integrated into Jupyter notebooks to track training energy use and emissions, though it excludes storage, travel and monitoring.
- Wegner points to Google's four levers for cutting AI emissions: model choice, machine type, cloud versus on-premise, and the region and timing of training.
Detailed notes
Why AI's carbon footprint now matters
Wegner traces her interest in the topic to an MIT finding from 2022 showing that digital solutions already produce more carbon emissions than aviation, and that emissions from digital products now exceed those of cars and aviation combined, a trend she says is growing exponentially as quantum computing and IoT expand. She connects this to the growing requirement for larger companies to report ESG figures, which pushed her team to ask what their AI solutions actually produce in terms of digital footprint, beyond the well-covered positive use cases such as optimizing flight rules or energy grids.
What training actually costs
She cites studies estimating that training GPT-3 used energy equivalent to about 63 homes for a full year, or the equivalent of charging 61 million phones, covering training only, not usage. The BLOOM model used roughly 3 homes' worth of energy and about 3 million phone charges. For comparison, she shows the emissions of a single flight from New York to San Francisco as a small mark on the same chart, far below either model's training footprint.
The lifecycle iceberg beyond training
Wegner argues most AI project discussions focus on model development and hardware choices, while a larger 'iceberg' below the surface includes employee devices such as laptops and phones, data storage, travel, and ongoing monitoring and fixes. Using a simplified version of a standard machine learning lifecycle covering business goal formulation, problem definition, data collection and quality, model development, deployment and monitoring, her team calculated energy consumption and, using Germany's public electricity map, the resulting carbon emissions for each phase, without accounting for retraining or upstream hardware manufacturing emissions.
A worked example and its results
In a small image-processing project with two employees on 50-watt laptops, a 50-gigabyte dataset and a 350-watt GPU for training, her team measured energy use per phase: laptop use throughout, added data storage during data processing, and GPU use added during model development, with deployment lower and monitoring counted only for one month since it recurs monthly. The results showed data processing producing emissions equivalent to charging 109 phones, and model development producing markedly higher emissions, equivalent to driving a very large number of kilometers in an average car (the transcript's exact figure is garbled and unclear), with monitoring much smaller per month but recurring.
Tools for measuring the footprint
Wegner says no single tool measures an AI project's full footprint automatically. She lists an internal travel carbon calculator built from CRM data, the Deutsche Bahn carbon calculator for comparing travel modes, hyperscaler dashboards that report cloud usage, a research-paper-based estimator that takes hardware type, usage hours, cloud provider and region to estimate training emissions, and CodeCarbon, an open source tool that integrates into Jupyter notebooks to track energy and emissions during training, though it does not cover data storage, travel or monitoring. For reducing emissions, she references Google's four levers of model choice, machine type, cloud versus on-premise, and region and timing of training, noting that emissions can vary significantly by country and season, plus Amazon's own environmental guidelines for its cloud.
Q&A highlights
Asked who should own the sustainability problem inside a company, Wegner says it increasingly appears in yearly ESG reporting and advisory questions, and compares it to the ethics discussion: teams need awareness of where they can reduce impact, such as choosing training regions, without expecting developers to stop working. On why emissions per region and season vary, she explains that countries further from the tropics see bigger swings because solar and renewable output rises in summer and falls in winter, which is why cloud providers let companies choose training regions such as South Africa or Namibia. Asked how to capture the full lifecycle for large language models, where training is costly but reused across many inference use cases, she agrees this is genuinely complex and says her simplified phase-by-phase approach is meant as a first, pragmatic view rather than a complete answer. On whether regulation should cap energy use, she expects continued ESG reporting requirements for companies in Europe but is skeptical governments will impose hard consumption limits, favoring education and awareness instead.
Quotes
You always have to look at it from the whole life cycle.
Susan Wegner
It's not that you can do nothing, but raise the awareness a little bit.
Susan Wegner
Questions answered in this talk
How much energy did training GPT-3 use, compared to households?
According to Susan Wegner at Rise of AI 2023, studies estimate that training GPT-3 used energy equivalent to about 63 average homes for a full year, or the equivalent of charging roughly 61 million phones, for training alone.
What tool can measure the carbon footprint of training a machine learning model?
According to Susan Wegner at Rise of AI 2023, CodeCarbon is an open source tool that integrates into Jupyter notebooks and calculates the energy consumption and emissions produced during model training on a laptop or in the cloud.
Why does the carbon footprint of AI training vary by country and time of year?
According to Susan Wegner at Rise of AI 2023, countries farther from the tropics see larger seasonal swings in green energy availability, with more solar power and lower emissions in summer and higher emissions in winter, which is why cloud providers let users pick training regions accordingly.