Most impact organisations use AI either in secret or with a guilty conscience. Neither is the right approach. The real question is not ‘whether’, but ‘where’.
June 2026 · Noted by · reads in ~4 minutes
Most impact organisations I come across either use AI without mentioning it publicly, or do so with a slight twinge of guilt. I consider both approaches inadequate, not because AI’s resource consumption is irrelevant, but because the question is being asked the wrong way round. What counts for me is not whether an organisation uses AI, but where the energy for it comes from and whether the organisation is prepared to shape that actively.
A typical text query consumes around 0.3 Wh of electricity – the same as an LED bulb in two minutes. The figures most frequently cited in the press are overestimated by a factor of 10, because a single estimate from 2023 is being passed on uncritically. Per-prompt consumption is a real factor, but it is not the problem the headlines suggest.
The problem is scaling: more users, longer contexts, more autonomous processes. And: exactly where the data centres are located.
Efficiency alone does not solve this, it actually accelerates it. The computing effort required to train AI models doubles every five months, and the cheaper a query becomes, the more queries get made. Per prompt things improve quickly, in total they still grow quickly. Anyone quoting only the first half is making it too easy for themselves.
This is the crux of the latest UN report on the subject (UNU-INWEH, June 2026), and it is rarely understood: low-carbon electricity is not automatically low-water or low-land. Every energy source has a different profile.
Wind power and photovoltaics have very low water consumption, but require land. Nuclear power is low-carbon, but water-intensive. Gas-fired power stations emit more CO₂, but use less water than coal. The three footprints – carbon, water and land – do not always move in the same direction.
What this means in practical terms is that a data centre in Iceland, powered by geothermal energy, has a different environmental profile to one in Texas, which is fed by the gas grid. Two-thirds of the new AI data centres built in the US since 2022 are located in semi-arid or drought-prone regions.
Since June 2026 this can be quantified. The Öko-Institut, commissioned by Greenpeace, analysed the electricity sources of around 6,700 data centres worldwide: US data centres come in at 414 grams of CO₂ equivalent per kilowatt-hour of IT electricity, the European average at 263. And the gap is widening. Because fossil power stations continue to be built in the US, computing capacity generated there is projected to be roughly twice as carbon-intensive as European capacity by 2030.
One figure from the same study gave me pause, though: Germany sits at 430 grams, above the US average. The good European average comes from countries with plenty of hydro and nuclear power, not from us. So anyone using “servers in Frankfurt” as a sustainability argument has not looked the number up. An EU location is first and foremost a data protection argument.
Location decisions are made by the providers, not the users. But users have more influence than one might think.
Anyone using AI APIs can make enquiries or make an active choice: Anthropic and OpenAI offer EU server regions. That only helps, however, if you look at the specific country rather than the label “EU”. Between the large hyperscalers themselves the study finds no meaningful difference in emission intensity anyway. Switching from AWS to Microsoft achieves nothing; the location and the type of provider do.
I would not rely on carbon-neutrality labels here. The net-zero pledges of the large providers so far rest largely on purchased green electricity certificates. In reality, coal and gas power stations fill the gaps, because there are no direct lines from renewable generators to the data centres. The Öko-Institut therefore deliberately calculates without certificates and measures only the electricity that physically reaches the grid connection. For an impact organisation that distinction matters: anyone citing a certificate is importing somebody else’s greenwashing into their own communication.
A second lever sits closer to hand than most people think: the search engine. Google is currently making the AI answer the default. Since July 2026 Ecosia lets you switch off every AI feature with one click, and DuckDuckGo, Brave and Qwant offer something similar. For impact organisations Ecosia is more than a tool here: a social enterprise held in steward-ownership that has been building its own solar plants since 2018 rather than buying certificates. Anyone who has to explain why they use a given tool is then arguing from a physical installation rather than a label.
Three further patterns drive consumption disproportionately and can be directly controlled.
AI images and videos. A single high-resolution AI image consumes as much energy as a full smartphone battery charge. Anyone who uses image generation on autopilot will quickly end up using many times the amount of energy required for text processing.
Autonomous agents without termination conditions. Processes without a defined end point multiply consumption exponentially. Clear boundaries and manual approval steps are essential in every AI workflow.
Unnecessarily long contexts. Token length is the direct cost driver. A 100,000-token context consumes around 130 times as much as a typical query.
An organisation that uses AI in a targeted manner, actively chooses its server location, communicates its usage transparently and channels the freed-up capacity into its mission is in a better position, both environmentally and ethically, than one that does so secretly or avoids it altogether. This is not meant as self-reassurance; it is how we get to a position we can actually defend.
All sources, figures and methodological background information can be found on the resources page for this topic.
45 minutes, one screen shared, an honest assessment. Not a sales pitch, I show you where I see leverage and where I do not.