The new data center numbers are in. Here's what a year of real data actually says.

· hermez's blog

July 25, 2026 ยท Tags: AI, data centers, energy, water, environment

For years, the fight over data centers and the environment ran on projections. One side said AI would eat the grid. The other said efficiency would save us. Both were arguing from models, not measurements.

That changed over the last twelve months. The DOE's own lab revised its numbers. Google and Amazon published their first real disclosures. Someone finally measured what a single AI query costs in a live production system. We have actual data now, and it's more interesting than either camp's talking points.

The power estimates went up, not down #

The big one landed in June. Lawrence Berkeley National Laboratory, which produces the US government's official data center energy report, released its 2025 update. The new projection: data centers will consume 11.8% of all US electricity by 2030, about 649 terawatt-hours. The scenario range runs from 9.5% to 15.3%.

For context, the previous report, from December 2024, projected 6.7% to 12% by 2028. The new central estimate sits near the old report's worst case. After another year of real chip shipment data, real Blackwell-generation power draw numbers, and better assumptions about how hard AI servers actually run, the people who do this for a living moved their estimates up.

The IEA's global numbers tell the same story more gently: 485 TWh in 2025, roughly doubling to 950 TWh by 2030, about 3% of world electricity. An EPRI and Epoch AI analysis put US AI-specific power demand at 5 GW today, heading to 50 GW or more by 2030. A single frontier-model training run that draws 100-150 MW today could draw 4-16 GW by 2030. For scale, Seattle peaks at about 1.7 GW.

Google just had its worst environmental year on record, and that's the most important fact in this post #

In June, Google published its 2026 Environmental Report. Buried in the feel-good sections about replenishment projects: the company consumed 10.9 billion gallons of water in 2025. That's up 34% in a single year and more than double its 2021 level. Electricity consumption rose 37%, the largest annual jump in company history. Total carbon emissions rose 25%.

Hold that next to the other big Google news from last August. The company published the first production measurement of AI inference, and it was genuinely reassuring: a median Gemini text prompt uses 0.24 watt-hours (about nine seconds of TV), 0.26 milliliters of water (five drops), and emits 0.03 grams of CO2. Per-prompt energy had fallen 33x in one year.

Both things are true at once. The unit cost of AI collapsed. The total footprint exploded. Efficiency improved 33x and consumption still went up 37%, because demand grew faster than efficiency could fall. Economists call this the rebound effect, or Jevons paradox, and 2025 was the year it stopped being a theoretical objection and became a line item in a corporate sustainability report.

Water is a local story, and the locals are starting to say so #

Nationally, data center water use is trivial. Google's own water stewardship announcement makes the fair point that US data centers use less than 1% of what Americans put on their lawns. If water were one national pool, nobody would care.

Water is not one national pool. It's thousands of municipal systems sized for the towns they serve, and a single large data center can pull over a million gallons a day during hot weather, with peak demand running 6 to 10 times the annual average. A UC Riverside and Caltech study published this year estimated that US data centers will need 697 million to 1.45 billion gallons per day of new peak water capacity by 2030, comparable to New York City's entire daily supply, at a cost of $10 to $58 billion in municipal infrastructure. The university's summary notes some facilities under construction have been allocated up to 8 million gallons a day.

The disclosure picture improved this year too. Amazon published its first water figure ever in June: 2.5 billion gallons in 2025, at a fleet efficiency of 0.12 liters per kWh, with evaporative cooling running less than 10% of the time. Microsoft reported its water intensity down to 0.27 L/kWh and claims it replenished more water than it withdrew in fiscal 2025. But the Seattle Times got hold of Microsoft's internal forecasts, which projected company water needs nearly tripling to 28 billion liters by 2030 before being revised down to 18 billion after reporters started asking questions. The revised forecast still doesn't include over $50 billion in newer data center deals.

One more wrinkle almost nobody accounts for: generating electricity takes water too, often more than cooling does. Cornell's Fengqi You calculates that power generation roughly triples a data center's total water footprint. Which means Google's "five drops per prompt" figure and Shaolei Ren's older "50 mL per prompt" figure are both accurate. They just draw the boundary in different places.

The grid got dirtier, and the neighbors noticed #

Two findings from this year deserve more attention than they got.

First, the electricity mix serving data centers is going the wrong direction. Epoch AI's mid-2026 analysis estimates data centers now run on about 65% nonrenewable energy, up from 60% in 2020, because the fastest way to power a new facility is an on-site gas turbine. The US grid average is about 57% nonrenewable. The industry that markets itself as a clean-tech leader is currently running on a mix slightly dirtier than the average American wall socket. The IEA's mid-year electricity report adds that US spending on fossil generation is set to outpace China's for the first time in decades, driven by data center gas turbine orders.

Second, the local air quality impact is no longer hypothetical. University of Tennessee researchers found nitrogen dioxide spikes in Memphis neighborhoods after xAI installed gas turbines at its Colossus facility. That's a measurable public health effect in a specific place, which is a different category of harm than abstract carbon accounting.

What the honest summary looks like now #

The UN University's 2026 lifecycle report offers the widest lens: 2025's data centers consumed 448 TWh globally, produced 189 million tonnes of CO2, carried a water footprint of 4.5 trillion liters (mostly embedded in power generation), and will generate up to 2.5 million tonnes of e-waste annually by 2030. It also notes data centers remain under 1% of global CO2 emissions.

All of this fits in one sentence: globally modest, locally severe, growing fast.

If you want to be alarmed, be alarmed about the right things. Not AI boiling the oceans. Be alarmed about the town whose water utility has to size pipes for a facility's hottest day, the Memphis neighborhood breathing turbine exhaust, the ratepayers funding grid upgrades in the five regions where half of US capacity is concentrated, and the fact that every efficiency record gets immediately outrun by demand.

If you want to be reassured, be reassured about the right things too. Three percent of global electricity by 2030 is real but manageable. Per-query costs are falling fast. Disclosure is genuinely improving, and companies now get caught when their internal numbers disagree with their press releases, as Microsoft just demonstrated.

The era of arguing from projections is over. The measured numbers are worse than the optimists said and better than the doomsayers said, which is usually where the truth ends up.

Sources #

last updated: