AI's Environmental Cost: What Boards Must Ask and Do
- David Turner
- 3 days ago
- 4 min read
AI's Environmental Cost
Most organisations now have an AI strategy. Very few have asked what it is doing to the planet.
That is not an abstract concern. It is a capital allocation question with a measurable environmental consequence attached to it. Boards are declining to ask it. The omission will not age well.
The structural reason is straightforward: the cost does not appear on the organisation's books. A single large language model query consumes roughly ten times the energy of a standard web search. Training a frontier model like GPT-4 is estimated to have produced over 500 tonnes of CO2 equivalent. Data centres already account for around 1–2% of global electricity consumption, and that share is growing as AI workloads scale.
Microsoft reported a 34% rise in global water consumption in 2022, largely attributable to data centre cooling. Google's rose 22% in the same period. These are the operators' own published figures. The extraction cost of rare earth minerals - neodymium, lithium, cobalt - required for semiconductor production does not appear in corporate AI disclosures at all.
Companies and their auditors have not agreed on how to measure or report this load consistently. That is not an excuse for ignoring it. It is the mechanism by which the problem persists.
Executives who have signed net-zero pledges, obtained B-Corp certification, or published sustainability reports are, in a number of cases, actively undermining those commitments by scaling AI infrastructure without accounting for its draw. The carbon and water consumption does not disappear because it sits in a vendor's data centre rather than your own. It moves off the balance sheet. That is not sustainability. That is outsourced impact. Regulatory pressure is building. The EU's Corporate Sustainability Reporting Directive is tightening scope and enforcement. When the reporting line becomes mandatory, "we did not formally report it" will not be a defence.
Not all AI carries the same environmental cost or the same return. That is the distinction most AI strategies fail to draw.
Not All AI Carries the Same Cost
Generative AI - large language models, image synthesis, video generation - is computationally expensive by design. Every response requires inference across a model running billions of parameters. When that capability is deployed to draft internal communications, generate slide imagery, or produce copy variants for a campaign, the cost-to-value ratio is poor. The processing load is real and continuous. The output could, in most cases, be produced at a fraction of the cost by a person or a narrower tool.
Machine learning applied to scientific and engineering problems is a different matter entirely. DeepMind's GraphCast produces 10-day global weather forecasts in under a minute; conventional models require hours on supercomputers. AlphaFold effectively resolved a 50-year problem in structural biology. These are not marginal productivity gains. They are qualitative changes in what science can do. The same applies to AI used in grid optimisation, materials discovery, and climate modelling - domains where the compute expenditure is proportionate to the return.
The question for executives is not how much AI. It is which AI, for what purpose, consuming what resource. That is the same question applied to any other capital allocation. From my own experience, very few if any leaders are pushing back on vendors or analysing the environmental impact of their rush to adopt AI.
What Responsible AI Deployment Looks Like
Train your people to prompt precisely. A vague or exploratory prompt generates multiple inference passes. A specific one does not. At scale this is not a trivial saving, and it improves output quality in any case. Responsible prompting is not an environmental concession. It is operational discipline. Teams need to understand the impact, and keep it at the forefront of decision making and tool usage.
Interrogate your cloud vendors on energy provenance. Microsoft, Google, and AWS have made renewable energy commitments of varying credibility and specificity. Ask where your workloads run, what the energy mix is for those regions, and what the water draw looks like for your contract. Your sustainability reporting will require this eventually. Reconstructing it under audit is harder than collecting it now. Speak with your account managers and vendors, add the topic to RFIs and RFPs.
Extend the discipline to the broader digital estate. Storage draws power continuously. Email archives running to tens of millions of messages, redundant file versions, assets retained indefinitely in cloud storage - all of it consumes data centre capacity. Most organisations have never cleaned any of it. Starting is a signal to staff that the policy has teeth, not just language.
At its core, this is not a compliance question. The organisations best placed to use AI to accelerate climate science - modelling, grid management, physical system resilience - are deploying much of the same compute to generate commercial content. That is a choice. It is the kind of choice that looks obvious in retrospect, and inexplicable to the people who come after. For history to judge us favourably both as individuals and organisations, we have to learn from the mistakes of fossil fuels and tobacco, and act right away.
David Turner is the founder of Kói, an independent technology consultancy advising senior leaders and boards on strategic technology decisions, investment readiness, and effective AI deployment.
© Kói Holdings Ltd 2026. All Rights Reserved.
