The number Google could not engineer away
On 30 June, Google published its eleventh annual environmental report, and by the numbers it is an engineering triumph. A ninth consecutive year of matching 100 percent of electricity with renewable purchases. Twelve gigawatts of new clean-energy deals signed in 2025 alone, part of roughly 35 gigawatts contracted since 2010 across more than 240 agreements. A fleet-wide power usage effectiveness of 1.09, meaning its data centres burn 83 percent less overhead energy than the industry average. Operational emissions actually fell 2 percent.
And then the number none of that engineering could touch: electricity consumption rose 37 percent in 2025, the largest single-year increase in the company's history, bringing the rise since 2019 to more than 250 percent. Supply-chain emissions climbed 25 percent, with data-centre construction alone adding about 2.3 million tonnes of CO2 equivalent from semiconductor suppliers on carbon-heavy grids in Taiwan, Japan, Vietnam and India. Google's own wording concedes the race: its AI infrastructure buildout is, in the report's words, accelerating faster than the grid is decarbonizing.
Why it matters: the efficiency trap in AI budgeting
Why it matters: this report is the cleanest live demonstration of a decision error that is currently sitting inside thousands of AI budgets: converting a per-unit trend into a total-cost forecast. When a vendor shows you cost per token falling or a team promises the rollout pays for itself as models get cheaper, the intuitive conclusion is that spend will fall. The base rate points the other way. Efficiency lowers the price of each unit, demand for units explodes, and the total rises. The most efficiency-obsessed operator on the planet just posted a 37 percent increase in a single year while running at 9 percent overhead.
Economists have a name for this: Jevons paradox, observed in 1865 when more efficient steam engines increased, rather than reduced, Britain's coal consumption. The pattern survives because it is not a technical failure but a market response. Applied to your own planning, the rule is short: treat every per-unit efficiency claim as a forecast of more usage, never as a forecast of lower spend.
The bottom line: demand efficiency, forecast consumption
Yes, but: efficiency is still worth demanding. Google reports that its hardware and software interventions avoided 58 million tonnes of CO2 equivalent in 2025 and that its footprint would otherwise be roughly five times larger. The lesson is not that efficiency fails; it is that efficiency is a brake on unit cost, never on appetite. The two claims live on different lines of the budget.
The bottom line: three moves follow for an owner. Cap AI spend with hard budget lines and alerts, not with efficiency assumptions. Write consumption-growth clauses into cloud renewals, because committed-use discounts quietly assume your usage keeps climbing. And for European companies, put the grid gap into planning: electricity prices near data-centre hubs such as Frankfurt, Dublin and Amsterdam, and the scope 3 lines in CSRD reporting, inherit exactly the curve Google just published.
Read next: Cybercrime Borrows Your Home Internet · Europe Just Made Its Google Case Final
