AI sweat: Chris Preist on data centre emissions

12 June 2026

Notes from a Sustainable AI seminar by Professor Chris Preist on 10 June 2026, part of a series of events on responsible AI by ClimateAction.tech with ustwo. My thoughts in [square brackets] - any errors will be mine.

• Direct emissions: digital’s own footprint. Enabling: good and bad things the AI lets/makes happen. Systemic: induced larger social change.
• Digital tech has roughly same emissions as aviation. But because it is much more widely diffused through global population than aviation (~20% vs ~1% of pop), different emissions reduction strategies will be needed.
• Data centre electricity use is exceptional, water use is normal. A large data centre uses the same electricity as a large city, and the same water as a small housing estate.
• Measuring AI emissions [inside the DC]:
  • Functional unit of measurement: one typical chatbot interaction. Measurement is inherently imprecise and ballpark, though numbers often present many digital places. [precision without certainty.]
  • Inference: ~0.5Wh Measuring the environmental impact of delivering AI at Google Scale, 2025. But closer to 5 Wh for complex reasoning.
  • Training: Grok 4, 310GWh, ~154,000 tCO2e. Assume model used for 1 year, allocate per query. (eg ChatGPT 19 B/week in July 2025 ~0.3 - 3 Wh/query)
  • Embodied: GPU 0.16 tCO2e More than Carbon: Cradle-to-Grave environmental impacts of GenAI training on the Nvidia A100 GPU, Assume GPU lifetime of 3 years [I think this is increasing while the economics favour it - previous gen GPUs are still getting deals] and 50% CPU load [conservative?]. Allocate embodied per Wh of use = 0.02 g, and double for the other equipment. Embodied = 0.04 g total.
  • Electricity mix matters: UK 0.14 - 1.0 gCO2e, US 0.35 - 3.0 gCO2e.

Screenshot of Preist’s deck: training and embodied allocation to use

• An extra computer screen (600 kg CO2e at Bristol) is equivalent to 200k - 4,200k chatbot requests in 2025.
• A 1 GW DC could offer every person in the world a prompt every 4 hours using 2025 technology.
• Modern data centres use adiabatic cooling - mainly aircooled, also spraying water onto the aircooled coils to speed up cooling through evaporation. AI sweating. More efficient than older methods. Many historic emissions models collapse the implementation detail and assume all compute or all DC is built the same way.
  • [Related: the way that AI tries everything look effortless, fast and convenient. The lack of transparency, the chat ui, the forced coyness around disclosing emissions. All of it making the numbers pretty opaque and the infrastructural hard work invisible. This approach feels like a relic of the 20th Century, when we enjoyed seamless supply chains and next-day magic and didn’t think about how it happened. AI really could do better.]
• Manufacturing of data centres, GPUs and equipment and model training are sunk costs, and usage costs are ongoing. [As GPU longevity increases, inference (usage) accounts for a larger and increasing proportion of costs and emissions.]
  • [A sense of carbon capex vs carbon opex.]
• AI gets the headlines, but “data centres” are much wider than ai. High performance compute isn’t just accelerated ai. and AI isn’t just chatbots.
• Efficiency improvements in eg GPUs -> more use [yeah Jevons] -> emissions going up. Drivers of increased usage: more users, more use-per-user (embedded, agentic), heavier workload (increased reasoning and media)

- Hannah Ritchie: How much electricity does AI consume? 2025 summary

• implied: sufficiency vs profligacy - some uses are more wasteful or justifiable than others?
• A “socio-technical imaginary” is being used to sell the story - the story/idea of a general and super intelligence. These are driving the investment case and the emissions impact. But maybe we won’t need as much data centre as everyone fears.
• Conclusions
  • each person’s footprint is small but billions of users = large total
  • wider R&D/experimentation increases impacts
  • electrical power is exceptional, water is not
  • hyperscale general intelligence [lower emissions?] or lighter specialist intelligence [higher?]
  • “good enough” AI at the edge/device
  • DC infrastructure risks overbuild, redundancy, resource waste
• Good question in the chat on whether AI will ever be profitable. [No doubt heavily subsidised currently esp for consumer, but the large players are clearly feeling their way from flat rate subscription to usage-based/metered pricing models. Eg Fable 5 moves to usage-based pricing soon.]

Good talk.