Google's electricity use grew by 37 percent in 2025, the biggest annual increase in the company's history, and it was blunt about the cause: the rapid buildout of AI. The figure comes from Google's latest environmental report, the annual accounting in which the company lays out the resource cost of running its services. A 37 percent jump in a single year is enormous for a company already consuming electricity on the scale of a small country, and it lands the abstract idea of an AI boom in concrete, physical terms. The longer trend is starker still. Since 2019, Google says its total electricity demand has grown by more than 250 percent, most of it in the last few years as AI moved to the center of its business.
What makes the number interesting rather than simply alarming is what sits next to it. Over the same period, Google says it made the AI itself far more efficient. By its own accounting, the energy used to answer a median text prompt with its Gemini model fell roughly 33 fold over twelve months, a genuinely large improvement driven by better models, hardware, and software. Its data centers are also close to the practical limit of efficiency, with a fleet wide power usage effectiveness of 1.09, meaning almost all the electricity that goes in is used for computing rather than lost to cooling and overhead. On paper, Google is doing many of the things a company is supposed to do to shrink its footprint.
And yet the total still went up sharply, which is the whole point. This is the rebound effect at industrial scale. When each AI query becomes cheaper and more efficient, it does not reduce overall energy use if the number of queries grows even faster, and that is exactly what is happening as AI is folded into search, phones, workplaces, and products used by billions of people. Efficiency is winning per unit and losing in aggregate. It is a useful correction to two comfortable stories, the one that says AI will be fine because the models keep getting more efficient, and the one that says nothing can be done. Both miss the reality that a 33 fold efficiency gain and a 37 percent rise in total consumption are happening at the same time.
Google also points to the ways it is trying to cushion the impact, and they are real but partial. The company said it matched 100 percent of its annual electricity consumption with renewable energy purchases for the ninth consecutive year, and that it actually reduced its operational emissions by about 2 percent even as usage climbed. Matching consumption with purchases is not the same as running every data center on clean power at every hour, a gap the company itself acknowledges, but it is not nothing. The water numbers are harder to dress up. Google's water consumption rose 34 percent to 10.9 billion gallons, and while it says it replenished about 7.7 billion gallons, roughly 78 percent of its freshwater use, through stewardship projects, the underlying demand is still climbing in step with the compute.
The reason a single company's utility bill matters is that Google is one of the more efficient and more transparent operators in the industry, which makes its numbers a floor rather than a worst case. If the company doing many things right still sees electricity use jump 37 percent in a year, the aggregate demand from every company racing to build AI is going to place real strain on power grids and water systems that were not planned around it. That is already visible in the scramble for energy, from restarting old power plants to signing long term deals for future capacity. None of this means AI is not worth it, and the efficiency gains are genuine and fast. But the report is a reminder that intelligence delivered at planetary scale has a physical bill attached, that the bill is growing quickly, and that efficiency alone, however impressive, is not currently enough to bend the curve down.