The carbon footprint of AI

Episode Summary

Executive Summary: This Tech Life episode examines AI’s growing energy demands, including a forecast that AI servers could use electricity on the scale of the Netherlands by 2027, while noting key uncertainties around demand, chip supply and efficiency. It also explores African opportunities and barriers in AI, the need to prepare workers for technological change, and a home-built AI system that identifies and safely deters foxes and badgers.

Main Topics: AI’s electricity demand and data-centre growth: Researcher Alex de Vries estimates that, if demand stays high and chip supply constraints ease, AI servers could require electricity on the scale of the Netherlands by 2027. BBC technology editor Zoe Kleinman stresses the forecast’s assumptions, including continued growth, full utilisation and no major efficiency gains. Data-centre hardware, heat and cooling: A visit to a data centre illustrates the physical costs of AI computing: AI racks can use far more power than conventional racks and generate heat requiring additional cooling. Scottish operator DataVita uses outdoor air in its free-air cooling system to reduce cooling energy use. AI opportunities and infrastructure needs in Africa: Entrepreneur John Kamara argues that African countries should become AI producers as well as consumers. He identifies young talent and applications in agriculture, healthcare and climate change, while pointing to gaps in skills, university programmes and computing infrastructure. Preparing workers for AI-driven change: Kamara says technological change can displace people who lack relevant skills, but argues that training can help workers use AI as an enabler. He frames preparation around people, data and infrastructure, with human capability central to creating value. A home-built AI wildlife deterrent: Computer expert James Millward combined garden cameras, a custom-trained image-recognition model and an ultrasonic device to deter foxes and badgers. After an off-the-shelf model misidentified animals, he labelled his own images and trained a model that successfully triggered the deterrent.

Key Arguments: AI’s energy footprint may grow substantially if demand remains strong and chip supply catches up; the Netherlands-scale estimate is conditional, not a certainty. The projected energy burden depends on assumptions about industry growth, equipment efficiency and server utilisation, and the research focuses mainly on NVIDIA because of its dominance in AI hardware. AI data centres face both server-power and cooling demands. Free-air cooling can reduce cooling energy use where climate conditions allow, but it does not remove the servers’ underlying electricity consumption. African countries can gain economic and social value by developing local AI talent and infrastructure, rather than relying only on imported systems and acting as consumers. AI-related job disruption is not inevitable for every worker: education and reskilling can help people adapt, while failure to prepare may leave some workers redundant. Purpose-built AI applications can address practical problems with off-the-shelf components, but may require locally relevant training data and human tuning.

Data Points: Projected AI-server electricity consumption: Potentially equivalent to the electricity consumption of the Netherlands by 2027 - Alex de Vries’s estimate depends on high demand and the easing of AI-chip supply constraints. Share of global electricity consumption attributed to the projected servers: About 0.5% - De Vries describes this as the potential share associated with the AI servers in his projection. Netherlands population: About 17 million people - Used in the programme to give scale to the country-sized electricity comparison. Conventional data-centre rack power: About 4 kilowatts - DataVita managing director Danny Quinn compares a standard rack used by a bank or council to the power use of a family house. AI rack power: About 80 kilowatts, roughly 20 times a standard rack - Quinn’s comparison for an AI rack at the facility. Scale of some AI deployments: Thousands of racks - The programme notes that some deployments use thousands of AI racks. Potential reduction in cooling power from free-air cooling: Up to 80% - Quinn says DataVita uses Scotland’s cool outside air to reduce the power used for cooling. Objects in the off-the-shelf model’s training set: 1.5 million common objects - The model James tried initially did not recognise foxes or badgers and misidentified a badger as an umbrella. Custom wildlife model accuracy: About 69% - Millward’s first trained model’s reported accuracy identifying foxes or badgers. Time spent training the custom model: About two and a half hours - Millward’s description of the training time for his initial model. Images labelled for wildlife training: A few hundred - Millward labelled foxes and badgers in nighttime camera images to train the model.

Pivotal Quotes: "You're talking about half a percent of our total global electricity consumption that would be related to those servers." — Alex de Vries: Explaining the possible global scale of electricity use by AI servers in his 2027 projection. "So we have to be a combination of both production and consumption." — John Kamara: Making the case that African countries should build capacity to create AI systems, not only use systems produced elsewhere. "And at my first stab of training my own model, it came back with about a 69% accuracy of identifying foxes or badgers, which seemed like quite good progress for what was really just two and a half hours of training a model." — James Millward: Describing the initial results of his custom-trained wildlife-recognition model.

Implications: AI’s benefits will depend on managing its electricity and cooling costs, improving hardware efficiency and preparing workers. African participation will require investment in skills, research and computing infrastructure. Small-scale AI can also solve practical problems, but often needs locally relevant data and testing.