How we know

Hum turns real, published numbers into a little world. This page is for grown-ups, teachers and anyone who wants to check our work: every number, where it comes from, how sure we are, and every time we changed it.

How your world's health is drawn

Your world's health is a scale we designed, not something we measured. We compare your weekly AI use with 7.24 times the average person's share of all data-centre electricity, which is about 971 Wh a week (415 TWh a year shared by 8.2 billion people). A world never drops below 15% health: it rests, it doesn't die.

We picked that multiplier so two promises hold whether we use our middle estimates or our high ones (mid and high). When we test "what if the high numbers are true?", we apply them to everyone, including the average person's share, so we compare like with like:

With our lowest estimates both promises can't hold at once (the lowest published video figure comes from an older, tiny model), so there we only keep the order: a video-heavy week is never shown healthier than a text-heavy one. The script that checks this runs on every build, and the build fails if a promise breaks.

Health is a designed scale: your weekly AI use compared with k x (all data-centre electricity per person). k sits in the middle of the range where the guardrails hold at the mid and high figures [5.37, 9.76). At the low figures only the order is kept: a video-heavy week is never healthier than a text-heavy one. Derived 2026-10-01.

The numbers

Nobody outside the AI companies can measure a single question exactly. So we start from what companies have published, add independent research, and always keep a low, middle and high estimate. Your world is drawn from the middle; the panels in your world show the whole range.

WhatLowMiddleHighUnit
Electricity per text answer0.240.341.86Wh
Electricity per AI picture0.631.222.9Wh
Electricity per short AI video (about 5 seconds)30944944Wh
Water used at the AI building (cooling) per kWh0.120.361.08L/kWh
Water used at power plants to make each kWh02.184.4L/kWh
CO2 per kWh of electricity125473473gCO2/kWh
Electricity used by all data centres worldwide per year415415415TWh/yr
People in the world8.2 billion8.2 billion8.2 billionpeople

The quiz turns answers into amounts: 1 / 6 / 30 / 80 questions a day, 0 / 3 / 20 pictures a week and 0 / 1 / 5 short videos a week. "A lighter AI" uses the lowest published figure for questions and pictures.

Constants version 3 · e1530fef40b95998 · built 2026-10-01

AI buildings on the globe

Each glowing block on the globe is a region with many data centres ("AI buildings"). Its size follows the region's power capacity on a gentle scale, so the biggest region doesn't swallow the map. Published figures measure different things, and we say which for each one.

RegionMegawattsWhat the figure counts
Northern Virginia, United States2,552–11,300mixed: Jan 2024 ranking (low), colocation inventory Q1 2025 (mid), operational incl. self-built 2026 (high)
Beijing, China1,799capacity ranking (Jan 2024, Cushman & Wakefield data)
London, United Kingdom1,189–1,380operational capacity, H1 2025 (low) to 2026 (high); mid is the geometric middle
Atlanta, United States1,279colocation inventory (Q1 2025)
Dallas-Fort Worth, United States654–1,000Jan 2024 ranking (low); colocation market size H2 2025 (mid, high)
Amsterdam, Netherlands927operational capacity (Q2 2026)
Johor, Malaysia897operational capacity (2025)
Singapore, Singapore876capacity ranking (Jan 2024, Cushman & Wakefield data)
Tokyo, Japan865capacity ranking (Jan 2024, Cushman & Wakefield data)
Paris, France391–862Jan 2024 ranking (low); listed facilities (mid, high)
Sydney, Australia786operational capacity (2025)
Frankfurt, Germany745live IT load (mid-2025)
Shanghai, China430–725listed facilities (low); Jan 2024 ranking (mid, high)
Seoul, South Korea698inventory
Mumbai, India670total capacity
Phoenix, United States617colocation inventory (Q1 2025)
Chicago, United States555–590Jan 2024 ranking (low); inventory H1 2024 (mid, high)
Hong Kong, China581operational capacity (2025)
São Paulo, Brazil493colocation inventory (Q1 2025)
Silicon Valley, United States489inventory (H2 2025)
Hillsboro, Oregon, United States475total inventory (end of 2025)
Greater Jakarta, Indonesia322operational capacity (2025)
Dublin, Ireland304capacity ranking (Jan 2024, Cushman & Wakefield data)
Johannesburg, South Africa161–278Jan 2024 ranking (low); listed facilities (mid, high)
UAE (Dubai and Abu Dhabi), United Arab Emirates250live capacity, whole UAE ("more than 250 MW")
Santiago, Chile148colocation inventory (Q1 2025)
Stockholm, Sweden74–132Jan 2024 ranking (low); listed facilities (mid, high)

We changed our numbers

When better information comes out, we update. Every change is listed here, newest first. Numbers are low / middle / high.

  1. 2026-10-01 · version 3

    • Electricity per AI picture: 0.63 / 1.22 / 2.91 → 0.63 / 1.22 / 2.9
    • Water used at the AI building (cooling) per kWh: 0.12 / 1.08 / 1.8 → 0.12 / 0.36 / 1.08
    • Water used at power plants to make each kWh: 0.49 / 2.18 / 3.94 → 0 / 2.18 / 4.4
    • People in the world: 8.16 billion / 8.16 billion / 8.16 billion → 8.2 billion / 8.2 billion / 8.2 billion
    • The health scale was re-derived.
  2. 2026-10-01 · version 2

    • Sources updated; no numbers changed.
  3. 2026-10-01 · version 1

    • Electricity per text answer: 0.24 / 0.34 / 1.86 (first version)
    • Electricity per AI picture: 0.63 / 1.22 / 2.91 (first version)
    • Electricity per short AI video (about 5 seconds): 30 / 944 / 944 (first version)
    • Water used at the AI building (cooling) per kWh: 0.12 / 1.08 / 1.8 (first version)
    • Water used at power plants to make each kWh: 0.49 / 2.18 / 3.94 (first version)
    • CO2 per kWh of electricity: 125 / 473 / 473 (first version)
    • Electricity used by all data centres worldwide per year: 415 / 415 / 415 (first version)
    • People in the world: 8.16 billion / 8.16 billion / 8.16 billion (first version)
    • The health scale was re-derived.

Your privacy