insight report · September 2026
All insight reportsFour doubts come up in almost every first conversation about AI: the environment, client data, the effect on people, and whether casual use is a waste. This report tests all four against the original figures.
Executive summary
The usual worry runs like this. AI needs enormous amounts of electricity, the world cannot build clean power fast enough to keep up, and so every question typed into a chatbot is another small push towards coal and gas. Very little of it holds up. What does hold up sits somewhere most coverage never looks.
Clean electricity is being built faster than data centres can use it. In 2025 new wind, solar and other clean generation outgrew new data-centre demand by roughly twelve to one, and global power-sector emissions stayed flat. The difficulty is what happened to that clean power: nearly all of it was absorbed by other new demand, cars and cooling and industry, before data centres were counted at all, and the grids that carry it cannot connect new supply or new load anywhere near quickly enough.
Meanwhile the way the largest companies count their own emissions has drifted a long way from what their power stations actually emit. Google publishes two audited figures for the same year and one is more than five times the other. For an individual, none of this turns on how often they use ChatGPT. For a business it does, because a business chooses at volume, and the difference between the cheapest and the dearest thing a model can be asked to produce is around sixty thousand to one.
The figures behind this report
- 12 to 1
- New clean generation against new data-centre demand across 2025
- 38 TWh
- The entire global fall in fossil electricity that year. Data centres alone added 70 TWh of new demand
- 5.4x
- The gap between the two audited emissions figures Google publishes for 2025
- 61 months
- The median wait to connect a new American generation project, against 22 in 2008
What the report covers
What a data centre is for
Most data-centre electricity serves streaming, social feeds, storage and company systems. AI is the fastest-growing slice rather than the whole, and no source publishes a clean split.
Where the new clean power went
The twelve-to-one figure is real. The condition underneath it is that clean generation has to outgrow all new demand, not data centres alone, and through July 2026 it did not.
Two ways of counting emissions
Location-based and market-based reporting are both standard and both audited. The distance between them is where a great deal of corporate climate progress now sits.
What this means for a business
Routing work between models, which account client data sits on, what can actually be owned, and the evidence on who gains and who loses.
How it was researched
Every figure is traced to a primary source rather than to coverage of it, and nothing about AI itself is older than 2025, because the technology changed too much for earlier figures to hold. Claims were independently fact-checked by a second pass whose job was to refute them. The report names what could not be established in a section of its own.
Questions for your team
- Which model is your routine work actually running on, and does it need to be?
- Is client work being done on a personal login or a business account?
- If you asked your cloud supplier for location-based emissions, would they answer?
