The Price of Intelligence · a lens on the benchmark

The water lens

The third invoice: litres, not dollars

Data centre water use is a headline concern. AI needs water. So do humans.

—

loading…

The comparison

Water to finish each task

AI water is shown two ways: on-site data centre cooling (from Google's measured 0.26 mL per median prompt), and the full chain including the water consumed generating the electricity (~4.35 L/kWh US average — LBNL, via Andy Masley's analysis). Human water is the consumptive footprint of the food that powers the working hours — roughly one litre per kilocalorie, the FAO's rule of thumb — plus drinking water. Upstream compared with upstream.

Litres per completed task

Every gridline labelled, one order of magnitude past the data

Human (water embedded in food + drinking) AI, full chain (cooling + power-plant water) AI, on-site cooling only
One glass of water (250 mL) covers
—
—
Human water for the same task
—
—
Water gap
—
human footprint ÷ AI full-chain water
Honest accounting

What these numbers do and don't count

  • On-site vs full chain — we show both. Google's published 0.26 mL/prompt counts only data centre cooling (the hollow points). Most of AI's real water is upstream: generating a kWh of US electricity consumes ~4.35 L on average, mostly power-plant cooling (Lawrence Berkeley National Laboratory, via Andy Masley's analysis). The solid AI points add that in — full chain is roughly — the on-site figure, and it is the right comparison here, because the human food footprint is upstream water too.
  • Withdrawal is not consumption. Water taken from a basin and returned is not water evaporated away. Serious accounting reports the two separately and declares its boundaries — every figure on this page is consumptive. Blurring that distinction is how both the scary and the soothing headline numbers get made.
  • Closed-loop cooling changes the on-site story. Many new AI data centres use closed-loop liquid cooling: the loop is filled once and recirculated, so on-site consumption approaches zero — and liquid cooling cuts total facility power by ~10%, which also cuts the upstream water. The real engineering trade is water-for-electricity: evaporative designs save energy but consume water; dry cooling spends watts to save litres. This is why per-facility numbers vary wildly, and why viral per-prompt figures that mix cooling designs, climates, and accounting boundaries disagree by 100×.
  • Keep the denominator in view. As Masley documents, US data centres' water consumption is a rounding error next to agriculture — irrigated farmland dwarfs every data centre combined, and famous viral comparisons (a bottle per email!) collapse when you notice a single hamburger embeds ~2,400 L. Per-task water is micro; the meaningful questions are totals and siting.
  • The human figure is a food footprint, not a tap. Nobody pours a bathtub over a freelancer. The ~250 L per working hour is consumptive freshwater used to grow the food that fuels the person — the same accounting standard (upstream, consumptive) that makes the data centre number meaningful.
  • Diet moves the human number several-fold. The FAO's ~1 L/kcal is an average: a beef-heavy diet runs far above it, a plant-based diet well below. We use the average and badge it estimated.
  • Training has a water bill too. A frontier run's ~139 GWh at typical data centre water intensity (~1.1 L/kWh on-site) is roughly — — about — Olympic pools, once, amortised across ~1 trillion tasks.
  • Location matters more than totals. A litre evaporated in a water-stressed basin is not a litre in a rainy one — the same lesson as the carbon strip on the live page: where you compute is becoming as important as how much.

Sources