A grounded water ledger · every figure traced to a verified source

Data centres
vs golf — and other water uses

Golf course irrigation, worldwide

Data centre water, worldwide

Both counters are annual estimates divided down to the second — they show scale, not a meter reading. Change the data centre boundary and watch the pace change: the number moves because the definition moved, not the water.

Scroll ↓
01 · Clear the ground

Numbers to throw away — from both sides

Every side of this argument has a viral number that is spectacularly wrong. Golf has one. AI has at least two. They all get thrown out here, before anything gets compared — because the corrected numbers are large enough to argue about honestly.

The claim
2,080,000,000
gallons per day, per course
Reality
300,000
gallons per day, in season
Out by a factor of ≈ 6,900×

The 2.08 billion figure is the entire US golf industry's daily total from the GCSAA's 2003–05 survey, misread as a single course. The real thing is roughly 300,000 gallons a day in season — about one Olympic pool every week.

The claim
—
Reality
—

The claim
—
Reality
—

Symmetry, not equivalence: golf's junk stat inflated a real industry total by misreading it; AI's inflated a real research figure by dropping its denominator, then compounding through retellings. In both cases the corrected numbers — 2.01 billion m³ of US golf irrigation, millions of cubic metres of data centre cooling — are still worth arguing about. The fake ones just made the argument impossible.

02 · The map

Where the water actually lands

The same question — who uses more? — has four honest answers depending on where you draw the accounting boundary. Change it and watch the verdict move. Nothing else about the world changes.

21.7×
Golf applies 21.7× more water than data centres consume World · 2024 · direct on-site cooling only
Golf 3.50 bn m³
Data centres 161 m³
Year
Show
Bubble scale
Units
03 · The trajectory

Does the line ever cross?

Golf's water use is flat and slowly falling. Data centre water is compounding. Whether they cross before 2030 depends entirely on which boundary you accept — the three lines below all describe the same industry. Solid to 2026, dashed where it becomes projection.

04 · Per head

The per-person gap is not close

Divide each total by the people it serves. Roughly 73 million on-course golfers; roughly 6 billion internet users. This is where the comparison stops being flattering to golf.

Log scale — each gridline is 10×. The spread between the three internet-user figures is the accounting boundary again, not measurement error.

05 · What totals hide

Global totals are the wrong unit of concern

Golf wins the global volume argument on almost every boundary — and it still tells you little about harm. Water is a local resource, and these two industries sit in very different places, draw from very different sources, and concentrate very differently.

A handful of operators, most of the metered water

Disclosed consumption, latest reported year. These are audited corporate figures — the rare place where data centre water is measured rather than modelled.

The only national metered figure

England is the only country with a metered national data centre water figure. Everything else on the map — every country value, every global total — is modelled. That is the honest state of the evidence in 2026.

06 · The wider ledger

Golf was never the real yardstick

Golf is a useful comparison because it is famously thirsty and famously optional. But it is a rounding error next to how the world actually spends water. Pick something else to measure data centres against — the boundary rules still apply on both sides.

Data centres vs
Data centre boundary
Every figure on one log scale, each labelled with what it counts

Bars are not like for like — that is the point, so each carries its accounting boundary. wf-total includes green (rainfall) and grey (dilution) water and will always dwarf a blue irrigation figure; withdrawal counts water taken and largely returned. Crop footprints are the canonical peer-reviewed globals but carry 1996–2005 data vintages. The chip-fab bar is just two companies' withdrawals (with high recycling rates) — Samsung's semiconductor division alone exceeds the whole world's data-centre direct cooling. Alpine snowmaking is a dated mid-2000s estimate, likely conservative today.

What one of each costs

07 · Receipts

Sources & method

Every source below was checked against its primary document. Where a figure was wrong, superseded or untraceable, it says so — including the ones that weaken the story.