AI's Carbon Footprint and Humanitarian Computing: The Sustainability Trade-Off

The footprint is no longer a research curiosity

By 2026, the energy and water demand of AI training and inference has become a measurable line item in national grid planning. The IEA Electricity 2024 report and its 2026 update project data-centre electricity demand could double between 2022 and 2026, with AI workloads driving a substantial share. The Stanford AI Index 2024 and Hugging Face's BLOOM training footprint disclosure provide the most rigorous bottom-up numbers.

What the numbers actually say

Three orders of magnitude are useful to hold in mind. Training a frontier model: low hundreds of tonnes of CO2-equivalent for a large training run, broadly comparable to a few hundred transatlantic flights, before considering the embodied carbon of the hardware. Daily inference at frontier scale: roughly comparable to training within a year for a heavily used model. A single chat query: on the order of a few watt-hours, dwarfed by the hardware lifecycle but non-trivial when multiplied across billions of queries per day. The Sasha Luccioni et al. published estimates are the most cited reference.

Water is the underdiscussed constraint

Direct water consumption for cooling in major data-centre clusters has become a politically salient issue in several jurisdictions. The Shaolei Ren et al. 2023 paper on AI water footprint put training of GPT-3 at roughly 700,000 litres of clean freshwater equivalent, and inference at meaningful fractions of a litre per interaction. In water-stressed regions, this is no longer a notional concern.

Why humanitarian organisations should care

Two pathways make this directly relevant. Operational dependency: humanitarian organisations increasingly rely on AI infrastructure whose carbon trajectory will affect both the climate they respond to and the cost of the services they use. Mission alignment: organisations responding to climate-displaced populations face an obvious coherence question about their own technology choices. Several major humanitarian organisations have published environmental policies committing to measurement and reduction of digital emissions; implementation is uneven.

What practical reductions look like in 2026

Five choices materially reduce the footprint of a humanitarian AI workload without sacrificing utility. Right-size the model: a 7B parameter model is enough for most summarisation and RAG workloads; a 70B model uses roughly 10x more energy per query for marginal gains on those tasks. Cache aggressively: most humanitarian queries are repeats with minor variation; caching at the application layer cuts inference calls by 30-60%. Pick the region: data-centre carbon intensity varies more than fivefold between regions; the Electricity Maps live data is the operational reference. Schedule training and batch jobs for low-carbon windows. Choose providers with credible disclosure: not all renewable-energy claims survive scrutiny, and the Greenpeace Clicking Clean and successor reports document the gaps.

What to ask a vendor

• Per-query and per-training-run carbon disclosure on actual workloads, not headline averages.
• Water-use disclosure for the regions you process in.
• Renewable-energy contracts that meet the Greenhouse Gas Protocol Scope 2 quality criteria, not unbundled certificates.
• Embodied-carbon disclosure for the hardware lifecycle.
• A roadmap that aligns with a 1.5C-compatible trajectory rather than offset-dependent net-zero claims.

Further reading and primary sources

• IEA Electricity 2024: https://www.iea.org/reports/electricity-2024
• Stanford AI Index: https://aiindex.stanford.edu/report/
• Hugging Face BLOOM carbon: https://huggingface.co/blog/carbon-emissions-on-the-hub
• Luccioni et al. on inference emissions: https://arxiv.org/abs/2311.16863
• Ren et al. on water: https://arxiv.org/abs/2304.03271
• Electricity Maps: https://app.electricitymaps.com/
• GHG Protocol: https://ghgprotocol.org/