Truace
Subscribe
52/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Scenario study projects AI data centers using up to 1,719 TWh by 2035

Study title: Beyond PUE: An integrated scenario assessment of electricity, cooling, polymer–elastomer demand, carbon, and circularity in AI data centers

Evidence type
Benchmark or lab evaluation
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2498cite or check it

Study title: Beyond PUE: An integrated scenario assessment of electricity, cooling, polymer–elastomer demand, carbon, and circularity in AI data centers

A scenario-based screening study covering 2024–2035 projects that global AI data-center electricity demand will climb from 416 TWh in 2024 to between 707 and 1,719 TWh in 2035. Over the same period, power usage effectiveness is projected to improve to 1.13–1.22, yet the authors conclude that infrastructure growth can dominate facility-efficiency gains.

Read the study abstract

Artificial intelligence is transforming data centers into coupled energy–cooling–material systems that cannot be assessed adequately using electricity demand or power usage effectiveness (PUE) alone. This study develops a scenario-based screening framework for 2024–2035 that links institutional electricity pathways, H100 node-level calibration, facility-load decomposition, air/direct-to-chip/immersion cooling transitions, polymer–elastomer requirements, operational carbon, circularity, and uncertainty. Global el…

The problem

A scenario study projects global AI data-center electricity demand rising from 416 TWh in 2024 to 707–1,719 TWh in 2035, with efficiency gains (PUE of 1.13–1.22) outweighed by infrastructure growth.

In brief

The work also looks at cooling and materials. In the Base pathway, cooling electricity reaches 123.1 TWh in 2035 and the central plastic requirement proxy is 344.8 kt; the Lift-Off pathway reaches 174.4 TWh and 525.3 kt. Stronger liquid-cooling adoption lowers cooling electricity but raises material intensity.

These are modelled projections, not measured outcomes. Material-coefficient bounds widen the plastic estimates to 235.1–470.2 kt (Base) and 358.2–716.3 kt (Lift-Off), and a 5,000-sample Latin Hypercube test gives a 2030 global range of 169.6–482.8 kt.

Main points

  1. Projected global AI data-center electricity demand rises from 416 TWh in 2024 to 707–1,719 TWh in 2035 depending on pathway.
  2. PUE is projected to improve to 1.13–1.22, but infrastructure growth can dominate facility-efficiency gains.
  3. Base pathway: cooling electricity of 123.1 TWh and a central plastic requirement proxy of 344.8 kt in 2035.
  4. Lift-Off pathway: cooling electricity of 174.4 TWh and 525.3 kt of plastic in 2035.
  5. Material-coefficient bounds widen plastic estimates to 235.1–470.2 kt (Base) and 358.2–716.3 kt (Lift-Off).
  6. A 5,000-sample Latin Hypercube analysis gives a P5–P95 plastic range of 169.6–482.8 kt for 2030 globally.
  7. Stronger liquid-cooling adoption cuts cooling electricity but increases technology-weighted material intensity.

The full story

The framework links institutional electricity pathways, H100 node-level calibration, facility-load decomposition, and transitions between air, direct-to-chip and immersion cooling. It also covers polymer–elastomer requirements, operational carbon, circularity and uncertainty, rather than relying on electricity demand or PUE alone.

Results are given for Base and Lift-Off pathways. Cooling electricity reaches 123.1 TWh in the Base case and 174.4 TWh in Lift-Off by 2035, while the plastic requirement proxy reaches 344.8 kt and 525.3 kt respectively.

The authors identify a cooling energy–material trade-off: liquid cooling reduces cooling electricity while increasing material intensity. They conclude that sustainable AI infrastructure requires joint management of compute growth, cooling architecture, electricity decarbonization, material demand and circular supply.

Sources

1 on file
  1. Results in Engineeringdoi:10.1016/j.rineng.2026.113927
    Peer-reviewed2026-11-02

Talk it over

Four questions to start onWho carries this cost, and what would have to change to prevent it?

Start with a question

Scenario study projects AI data centers using up to 1,719 TWh by 2035https://truvace.com/item/beyond-pue-an-integrated-scenario-assessment-of-electricity-cooling-polymer-elas

  1. Who carries this cost, and what would have to change to prevent it?

    The problem A scenario study projects global AI data-center electricity demand rising from 416 TWh in 2024 to 707–1,719 TWh in 2035, with efficiency gains (PUE of 1.13–1.22) outweighed by infrastructure growth.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a benchmark or lab test, from Results in Engineering.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat The figures are scenario-based screening estimates for 2024–2035, and uncertainty in material coefficients widens the plastic projections considerably.

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Climate