Abundance · a design note
modeled, not measured — until someone builds the desk rig

Garbage heat is a water plant.

Every watt a datacenter spends on computing comes out as heat — about 415 TWh of electricity in 2024, essentially all of it ending up warm. Run the machines warm enough (liquid cooling, 40–60 °C) and that heat is exactly what membrane distillation needs to pull fresh water out of seawater or brackish wells. The machines get cheaper cooling; the coast gets drinking water. This page runs the arithmetic honestly: what the numbers say at full scale, what a $200 desk rig would prove first, and every reason the idea might die. Constants are cited; projections are labeled as models.

Back to the arithmetic · kinds of scarcity · bibliography

The whole trick in three sentences

Water evaporates at any temperature above freezing — cooler water just evaporates more slowly, and its vapor carries zero salts, zero minerals, zero microbes. A hydrophobic membrane with pores around 0.2 µm lets vapor through but refuses liquid water, so a warm saline stream on one side and a cool surface on the other produces distilled water at temperatures no ordinary engine would call useful. That process is called membrane distillation (MD), and its fuel is precisely the grade of heat a liquid-cooled computer throws away.

Latent heat of vaporization @ 50 °C
2,383 kJ/kg≈ 0.66 kWh per liter distilled — the thermodynamic bill every method must pay

NIST Chemistry WebBook, saturation properties of water

What MD actually rejects
>99.9% saltsnon-volatile solutes cannot cross the dry membrane; volatile contaminants partly carry over (see limits)

Alkhudhiri, Darwish & Hilal (2012), Desalination 287:2–18

How the built-properly version works

Click any part of the system. Ember is heat in motion, blue-green is clean water, grey dashed is brine. The two loops never mix — heat crosses metal walls, purity crosses membranes, and nothing crosses that shouldn't.

DATA CENTER warm-water liquid cooling PIT → almost all becomes heat HEAT EXCHANGER coolant heat → feed water 5 µm prefilter hot feed 50 °C cold permeate 22 °C hot feed 50 °C cold permeate 22 °C MD STACK vapor only crosses the membrane FRESH WATER <10 ppm TDS · mineralize · drink BRINE OUT saltier than the sea · diffuse or harvest COLD SINK sea / lake / night air DISTRICT HEAT optional winter mode: warm buildings instead 45–60 °C ~30 °C back heated feed brine chills permeate side seawater / brackish in heat crosses metal · vapor crosses membrane · the loops never mix
Architecture: warm-water liquid cooling coupled to direct-contact membrane distillation (DCMD) with counter-current heat recovery. Precedent pieces all exist separately — see sources below.

The machines

A modern GPU rack converts essentially 100% of its electrical draw into heat. Air cooling dumps it into the sky at ~25 °C — too cold to be good for anything. Liquid cooling carries it away at 40–60 °C instead, which costs the datacenter less chilling (often none, “free cooling”) and upgrades the heat from trash to fuel.

thermal output ≈ IT load · e.g. a 10 MW hall radiates ~10 MW continuously

Run your own numbers

Every figure below is computed live from four dials and two published constants (latent heat 2,383 kJ/kg at 50 °C from steam tables; the energy-recovery factor “GOR”, a standard MD performance metric — lab stacks run 1–6). Nothing here is measured at our scale yet; that is what the desk rig is for.

Datacenter IT load
10 MW
one mid-size AI cluster today: 10–150 MW
Heat actually captured
60%
losses to pumps, ducting, standby chillers
Coolant supply temperature
50 °C
higher ΔT ⇒ higher flux, more free-cooling hours
Stack efficiency (GOR)
3.0×
1 = no heat recovery · well-designed lab stacks: 3–6
6.0 MW
usable thermal power
652,700 L
fresh water per day
13,054
people at WHO 50 L/day basic need
0.26
olympic pools per day
litres/day = PIT × capture × 86,400 s × GOR ÷ 2,383,000 J/kg
specific energy = 2,383 kJ/kg ÷ GOR = 0.221 kWh/L (compare reverse osmosis ≈ 0.003–0.005 kWh/L of pure electricity — see limits)
per megawatt of compute = 65,270 L/day at current dials

And if the whole sector did this?

The world’s datacenters drew roughly 415 TWh in 2024 — an average continuous draw near 47 GW, essentially all leaving as heat. Same arithmetic, planetary scale (uses your capture and GOR dials):

3.09 B L
billion litres / day, whole sector
12.6%
of humanity’s direct drinking need (~24.6 B L/day)
0.75%
of WHO 50 L/person/day for all 8.2 B of us

Read that honestly: even perfect deployment covers a meaningful slice of drinking water — humanity’s total water withdrawal is ~4,000× larger — but a slice of drinking water is exactly the slice that is hardest to get where it matters. And it arrives attached to a business case (cooling), not charity. Source: IEA (2025), Energy and AI; population divisor UN WPP 2024; WHO minimum from Howard & Bartram (2003).

The desk rig — prove it for ~$200 before believing anything

The full-scale page you are reading is modeled. This rig replaces its central assumption (litres of water per kilowatt-hour of 50 °C heat) with our own measurement. It stands a mini PC in for the datacenter: same physics, same membrane, one ten-thousandth the stakes.

CPU + block pump +reservoir MINI PC stress test = the “datacenter” wall meter reads true watts hot feed salt water, 45–55 °C cold wall ice water, ~5–15 °C PTFE membrane 0.22 µm · 200–400 cm² ICE-WATER BATH the “ocean”, replaced as needed scale logs grams vs time DISTILLATE ~50 °C cooled back feed reservoir · 35 g/L salt T1 T2 T3 T4 ESP32 logs T1–T4 every second · scale + stopwatch give mass flow · bucket-and-stopwatch gives loop flow
#PartSpec that mattersEst. USD
Mini PC (or any spare PC)runs a stress test; wall-meter reads its true watts — this is the measured heat budget$0–150
CPU water block + 12 V pump/res combostandard PC liquid-cooling parts; any used AM4/LGA block works$40–65
PVDF/PTFE hydrophobic membranepore 0.22 µm, hydrophobic contact angle >90°; 47 mm lab discs ×25 or a cut sheet ≥200 cm²$40–90
Two-chamber acrylic celllaser-cut or hand-made; silicone gaskets; chambers clamp membrane between them$15–30
Ice-water bath + small pumpchills the permeate-side wall; the stand-in for sea/night air$15–25
4× DS18B20 sensors + ESP32T1 coolant out · T2 coolant return · T3 feed · T4 cold wall; 1 Hz CSV to serial$12–18
Kitchen scale (0.1 g) + TDS meterdistillate mass vs time = flux; TDS in/out = rejection$25–30
Tubing, fittings, epoxy, hose clampssaline-wetted parts plastic/epoxy only — hot saltwater eats metals$20–30
Total, newroughly half if any old gaming PC parts are in a drawer≈$170–440

Test protocol — and the pass gate

  1. Control run first: aquarium heater as heat source instead of the PC. Separates “MD works” from “PC heat works”.
  2. Salt run: feed at 35 g/L NaCl (seawater). Log scale mass every minute; TDS of feed and distillate hourly.
  3. PC run: swap in the mini-PC loop. Wall meter gives true input watts; T1−T2 and loop flow give recovered watts.
  4. 24 h soak: wetting failures show up as a sudden TDS jump in the distillate — that’s the killer mode, so watch for it.
  5. Compute: flux (L/m²·h), rejection %, GOR = (kg distillate × 2,383 kJ) ÷ recovered kJ.

Pass = all four

  • salt rejection > 99% sustained (distillate < 350 ppm from 35,000 ppm)
  • stable flux ≥ 1 L/m²·h at ≤ 55 °C feed
  • GOR ≥ 0.4 measured, not assumed
  • no wetting event across 24 h

Whatever GOR comes out goes straight into the calculator above as the measured preset — the model then stops being a model.

Do not drink the early batches. A TDS meter proves salt left, not that water is potable — microbes need sterilization, and MD partially passes volatile organics (solvents, some fuels): superb for seawater and brackish groundwater, unsafe next to industry without lab testing. Re-mineralize before anyone sips anything, ever. Membrane must stay dry-sided; one wetting event ends rejection instantly.

What would kill this — the steelman

The main thesis of this site survives on taking the strongest objection seriously. Here it is, unedited:

Reverse osmosis is ~50× more energy-efficient. Why bother?

True and decisive — where grid electricity is cheap and reliable. RO needs ~3–5 kWh of electricity per m³; MD needs ~200–600 kWh of low-grade heat per m³, but heat that would otherwise warm a river has zero opportunity cost. MD only wins where three conditions coincide: waste heat exists anyway (it does, inside every datacenter), water is scarce or expensive, and the heat would otherwise be thrown away. That intersection is real — Gulf coasts, arid coastlines with new AI buildouts — but it is a niche strategy, not a replacement for RO.

District heating often pays more for the same heat.

In Helsinki or Stockholm, a megawatt-hour of heat sold to buildings earns cash and displaces fossil boilers — usually a better deal than making water. But district heating only pencils in dense cold cities. In hot, dry, coastal regions — exactly where water stress bites and heating demand is nil — water is the highest-value use of the heat. The geography sorts itself.

Fouling, scaling, and wetting are brutal with real seawater.

Warm saline surfaces scale fast (carbonates drop out at elevated temperature), biofilms grow, and a single surfactant event wets the membrane and destroys rejection until it is dried and restarted. Pretreatment adds cost and complexity the tidy diagram hides. Lab reviews call this the central engineering problem of MD; the desk rig’s 24-hour soak test is a miniature rehearsal of it.

Nobody at the datacenter gets paid for making water.

The operator optimizes cooling cost, not community water output. Without a water utility partner, a water credit, or a heat-purchase contract, the water side stays a slide in a deck. This is a distribution problem — the site’s home turf — and it is political-economic, not thermodynamic.

All the numbers on this page are modeled.

Yes — flagged throughout. Latent heat is a physical constant; fluxes and GOR ranges come from peer-reviewed lab literature, not field deployments at datacenter scale. The honest sequence is: desk rig measures GOR → a single-rack pilot measures integration cost → only then does any regional number deserve belief. Until then this page is arithmetic, not evidence.

Sources

  • NIST Chemistry WebBook — saturation properties of water (hfg = 2,383 kJ/kg at 50 °C). webbook.nist.gov
  • Alkhudhiri, A., Darwish, N., Hilal, N. (2012). “Membrane distillation: A comprehensive review.” Desalination 287:2–18. doi:10.1016/j.desal.2011.11.033
  • International Energy Agency (2025). Energy and AI — datacentre electricity consumption ≈415 TWh in 2024. iea.org/reports/energy-and-ai
  • Howard, G. & Bartram, J. (2003). Domestic water quantity, service level and health. World Health Organization (50 L/person/day as the basic-access level).
  • United Nations (2024). World Population Prospects 2024 — 8.2 billion mid-2024 divisor.
  • Google Data Centers — cooling with reclaimed (treated wastewater) sources at Douglas County, GA. google.com/about/datacenters/sustainability
  • Stockholm Data Parks — datacentre heat recovery feeding city district heating. stockholmdataparks.com
  • Qarnot Computing — datacentre waste heat warming social housing, France. qarnot.com