Lesson 10/12Intermediate12 min read·3 diagrams

Sustainability — Power, Water, Carbon

AI data centers are now the fastest-growing source of new electricity demand worldwide. This lesson covers PUE/WUE/CUE, hyperscaler carbon pledges, geothermal and nuclear deals, heat reuse, and the brutal honest math of where the watts come from.

1 · The three numbers — PUE, WUE, CUE

PUE
kWh total ÷ kWh IT
1.0 = perfect, 1.10 = hyperscale
WUE
L water ÷ kWh IT
0 = no evap, 1.8 = aggressive evap
CUE
kg CO2 ÷ kWh IT
Depends on local grid mix

All three are metrics defined by The Green Grid consortium. They're imperfect (operators choose boundaries cleverly) but they're the universal language for comparing facilities.

2 · Where the watts come from

A data center's CUE is dominated by the local grid mix. The same 100 MW facility produces:

Iceland (geothermal)
~5 g CO2/kWh
Effectively zero
France (nuclear)
~50 g CO2/kWh
Lowest in EU
Sweden / Norway
~30 g CO2/kWh
Hydro + nuclear
Texas (mixed)
~400 g CO2/kWh
Heavy gas + wind
Virginia (PJM grid)
~370 g CO2/kWh
Gas + nuclear + coal
India (coal-heavy)
~700 g CO2/kWh
Hardest to clean up

3 · The hyperscaler pledges

  • Microsoft — carbon negative, water positive, zero waste by 2030. Acknowledged in 2024 sustainability report that emissions actually rose ~30% since 2020 due to AI buildout.
  • Google — 24/7 carbon-free energy by 2030 (every hour, every grid). Currently ~64% averaged.
  • Meta — 100% renewable since 2020 (annual matched). Investing in long-duration storage (Form Energy) and geothermal.
  • Amazon — 100% renewable globally, claimed 2023 (matched annually). The world's largest corporate buyer of renewables.
  • Apple — already at net-zero on operations; targeting full supply chain by 2030.

The honest read: annual matching is much weaker than 24/7 matching. When solar dies at sunset and gas peakers ramp up, "100% renewable annually" companies are still emitting.

4 · Firm clean power — the new race

AI's 24/7 demand needs firm (always-on) clean power. The deals announced in 2024–2025 mark a new chapter:

  • Microsoft + Three Mile Island — 20-year PPA to restart Unit 1 reactor (~835 MW). Targeted online 2028.
  • AWS + Talen Energy — $650M acquisition of Cumulus Data adjacent to Susquehanna nuclear plant (~960 MW).
  • Google + Kairos Power — small modular reactor (SMR) deal for ~500 MW by 2030.
  • Google + Fervo Energy — 115 MW enhanced geothermal in Nevada (operational 2024).
  • Meta + Sage Geosystems — 150 MW geothermal pilot in Texas.
  • Oracle — announced gigawatt-scale SMR deployment (timeline unclear).

Source: Microsoft, AWS, Google, Meta press releases (2024-2025); IEA Electricity 2025 report; Lazard LCOE 2024.

5 · Water — the second front

Evaporative cooling consumes water — about 1.8 L per kWh of IT energy at typical sites. A 100 MW campus with PUE 1.20 and WUE 1.5 consumes ~1.5 million liters per day, comparable to a small town.

In water-stressed regions (Phoenix, Madrid, parts of Texas) operators are switching to closed-loop dry cooling — trading higher PUE for near-zero WUE. Microsoft committed in 2020 to a "water positive by 2030" pledge requiring replenishment of more water than they consume.

6 · Heat reuse

Heat doesn't have to go to a cooling tower. In cold climates with district heating networks, it's an asset:

  • Stockholm — Stockholm Exergi sells heat from multiple DCs to ~10,000 apartments.
  • Helsinki — Microsoft 2022 deal to provide up to 40% of city's district heat from a new DC.
  • Frankfurt — Mainova's network ingests waste heat from ~10 DCs.
  • Paris — Equinix PA10 supplies heat to the 2024 Olympic Village swimming pools.

Lesson 10 — TL;DR

  • • PUE measures energy efficiency, WUE water, CUE carbon.
  • • Local grid mix dominates carbon footprint. France < Iceland << Texas << India.
  • • Hyperscaler pledges target 2030 — but emissions rose with AI growth.
  • • Firm clean power (nuclear, geothermal) is the new battleground.
  • • Water is the second sustainability front; closed-loop cooling is the answer in dry regions.

Useful? Share so the next engineer learns this faster.

Share: