Amazon used more electricity than New Zealand last year. That single fact, buried in the company's 2025 sustainability report, captures a reckoning the whole industry is now facing. Amazon's total emissions rose 16 percent to 80.9 million tonnes, roughly what the entire nation of New Zealand emitted in 2024. Google, reporting in the same week, disclosed a 25 percent jump. Neither company names the cause outright, but as TechCrunch put it, "there's plenty of indirect evidence," and it all points the same way: the race to build AI is colliding with the climate pledges both firms made when AI was still a rounding error.
For years these companies had a workable answer to their carbon problem. Data centres for offices and search were modest enough that buying renewable power could cancel out most of the pollution. AI has broken that arithmetic. Amazon added more than 1.2 gigawatts of data centre capacity in the final quarter of 2025 alone, and to keep pace with demand it and Google have begun investing in natural gas plants, sliding back toward the fossil fuels they spent a decade trying to leave behind. Emissions from Amazon's purchased electricity jumped 34 percent in a single year.
The more revealing numbers are the ones the companies control least. Most of the growth in both footprints comes from what accountants call Scope 3 emissions, the pollution embedded in the things a company buys rather than the fuel it burns directly. For Amazon and Google that increasingly means the physical stuff of AI: the steel and cement poured into new data centres, and above all the chips. Semiconductor fabrication is enormously energy hungry, much of it done on Asian grids still running on coal, and some of the chemicals used in chip plants are greenhouse gases thousands of times more potent than carbon dioxide. Every training cluster is a carbon debt incurred long before the first token is served.
Both companies are quick to argue the other side, and it is not nothing. Amazon points out that its carbon intensity, the emissions generated per dollar of revenue, is still 38 percent below its 2019 level, and that it brought 80 new renewable projects online last year. The efficiency of the data centres themselves keeps improving; Amazon's power usage effectiveness has crept down to 1.14, though that still trails Google's 1.09. But efficiency per dollar is cold comfort when the absolute total, the number the atmosphere actually feels, is climbing at its fastest rate since Amazon first made its net zero pledge.
This is the uncomfortable shape of the AI boom's environmental cost. None of the problems are unsolvable. Cleaner steel and cement exist in prototype, carbon removal can be bought, renewables can be built. But all of it takes money and time that the current pace of expansion is not leaving much of. Both companies devote pages of their reports to how AI might one day help fight climate change, a case of protesting too much, as TechCrunch dryly noted. The technology may eventually earn back its carbon debt. For now it is running one up, quarter after quarter, faster than anyone's pledges assumed.