Google says it ran on 100% clean energy again last year. Its electricity use rose 37%, and the two facts don't touch.

Google's 2026 environmental report, its eleventh, contains one number the company wants you to read and one it would rather you skimmed. The one it wants you to read is on the cover in spirit: for the ninth year running, Google matched 100 percent of the electricity it consumed with renewable-energy purchases. The one it would rather you skimmed is a few pages in: its electricity demand rose 37 percent in a single year, the largest annual jump in its history. Both numbers are true. The useful thing to understand is that they are not measuring the same electricity, and the distance between them is the whole story of what AI is doing to the grid.

Start with the honest number, because it is the physical one. Thirty-seven percent more electricity, in one year, from a company that already ran on more power than some countries. That is not an accounting entry. That is real megawatt-hours pulled off real wires, generated by whatever happened to be running on those grids at the moment Google's data centers asked for it — and increasingly, what they are asking for is the steady, around-the-clock draw of AI training and inference, which does not politely confine itself to the hours when the sun is up and the wind is blowing. The report is candid about why, in a sentence that does more work than the executive summary: "Our AI infrastructure buildout is currently accelerating faster than the grid is decarbonizing." Read that twice. It is the company telling you, in its own document, that the megawatts are arriving faster than the clean electrons to fill them.

What "100% matched" actually buys

So what is the other number — the 100 percent — actually measuring? Not the electrons in Google's servers. It is measuring a year's worth of purchases across a year's worth of grids. Google adds up all the electricity it consumed over twelve months, then buys enough renewable generation somewhere on its grids over those same twelve months to match the total. It is annual, and it is regional. The solar farm it funds in one place and one season is counted against the coal plant that spun up for it in another place at 3 a.m. in December. On a spreadsheet, the two cancel. On the wire, they never met.

This is not fraud, and it is not nothing. Buying 12 gigawatts of new clean generation in a single year — Google's 2025 figure, part of more than 240 agreements totaling roughly 35 gigawatts since 2010 — genuinely puts clean capacity on the grid that would not otherwise be there. That matters. But it is a purchasing achievement, not a physics one, and the report's own finer-grained metric admits the gap. Measured the honest way — matching consumption with carbon-free energy on the same grid, in the same hour — Google runs at about 66 percent, not 100. And that global average hides the number that should interest anyone who cares about grids that can't be papered over: in the Asia-Pacific region, hour-for-hour carbon-free matching sits around 12 percent. The paperwork says 100. The clock says 66. The hardest grid says 12.

The solar farm it funds in one place and one season is counted against the coal plant that spun up for it at 3 a.m. in December. On a spreadsheet, the two cancel. On the wire, they never met.

I dwell on this because the distinction is the entire cost problem in miniature. A clean electron at noon is cheap and abundant and, increasingly, a solved problem — solar is the cheapest power humans have ever built. The expensive electron, the one nobody has a boring answer for, is the 3 a.m. megawatt: firm, dispatchable, carbon-free power available on demand in the dark and the still. Annual matching lets you buy a lot of the cheap daytime electron and count it against the expensive nighttime one you actually burned. It flatters the invoice by averaging across the hours where the real difficulty lives.

Watch what they buy, not what they say

You can ignore every sustainability claim a company makes and learn more by watching where it puts its money, and here Google's checkbook is more honest than its headline. Look at what it signed in 2025:

• A deal with NextEra to restart the 600-megawatt Duane Arnold nuclear plant in Iowa — reviving a reactor specifically because it delivers firm, around-the-clock output that solar and wind do not.
• A framework with Brookfield for up to 3 gigawatts of hydropower — again, dispatchable, storable, available at night.
• A power-purchase agreement with Commonwealth Fusion Systems, a bet on a plant that does not yet exist, for firm clean power sometime in the 2030s.

Notice the pattern. A company that had genuinely solved its clean-energy problem with annual matching would not be paying to restart a nuclear reactor and pre-ordering electricity from a fusion machine that has never sold a kilowatt-hour. It is buying firmness, because firmness is the thing the certificates don't deliver, and firmness is what a 24/7 AI load actually needs. The nuclear restart is the tell. You do not reopen a shut reactor to make a spreadsheet balance; you do it because you need real power on a real wire at 3 a.m., and you have concluded that buying more daytime solar will not get it to you. The checkbook is pricing the problem the headline is averaging away.

The number that touches the bill

Here is where the cost-first lens earns its keep, because the emissions accounting has its own version of the same trick. Google reports that it cut its operational emissions — the Scope 1 and 2 figures, the ones from the electricity it directly consumes — by 2 percent year over year, which sounds like progress until you see the number underneath it. Its supply-chain emissions, Scope 3, rose 25 percent, and the single biggest driver was building the data centers themselves: roughly 2.3 million tons of carbon-dioxide-equivalent from construction alone. The clean electricity is the part Google can buy its way clean. The concrete, the steel, the chips, the servers — the physical stuff you have to make and truck and assemble to house an AI build-out accelerating at 37 percent a year — is the part it can't. And that part is growing four times faster than the operational part is shrinking.

This is the number that will decide whether any of the pledges hold, and it is the number the coverage tends to skip, because it does not fit either the triumphant story or the outraged one. Google's data centers are, genuinely, extraordinarily efficient at the narrow thing efficiency measures: their overhead energy — everything spent on cooling and power delivery rather than computing — runs at about 0.09 against an industry average of 0.54, meaning they waste a fraction of what a typical facility does keeping the machines fed and cool. That is real engineering, and it is worth crediting plainly. But efficiency per unit of compute is a ratio, and the build-out is a volume. You can make each server 20 percent leaner and still burn 37 percent more power if you add enough servers, and Google added enough servers. Efficiency is losing the race to scale, and it is losing it by the report's own arithmetic.

The grid most of the world runs on

I keep coming back to that Asia-Pacific figure — 12 percent hour-for-hour carbon-free — because it is a preview, not an outlier. It is low not because Google tries less there but because the underlying grids are dirtier and thinner, dominated by coal, without the deep renewable capacity that lets you buy your way to a clean-looking hour. That is the condition of most of the grids the AI build-out is about to expand into: South and Southeast Asia, much of Africa, the places where the next data centers will land because that is where power and land are cheap and available. In those places you cannot average the problem away, because there is no surplus of cheap daytime clean generation sitting on the grid to buy. The physical grid is the ceiling, and the physical grid is the one most of humanity actually lives under.

What Google's report shows, stripped of the framing, is a company that has run out of grid faster than it has run out of money. It can sign 12 gigawatts of clean energy in a year and still watch its hour-for-hour clean share stall, because the constraint stopped being procurement and became delivery — the wire, the firmness, the 3 a.m. megawatt. Its answer is to buy nuclear restarts and fusion futures, which is the correct answer and an expensive, slow, physical one, measured in the years it takes to reopen a reactor rather than the quarters it takes to sign a certificate. That is the honest shape of the AI-power story: not a virtue and not a scandal, but a bill coming due in megawatt-hours that no accounting convention can net to zero.

So when the next hyperscaler tells you it ran on 100 percent clean energy, ask the boring follow-up. On which grids, in which hours, and what did it have to restart to get there? Google, to its credit, put the answer in the report. It is 66 percent on the clock, 12 percent where the grid is hard, and a reactor in Iowa spinning back up because the certificates couldn't cover the dark. The 37 percent is the number that touches the wire. Everything else is the number that touches the page.