Truace
Subscribe
52/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Modeling study projects a big company's AI power use could rise 24.4-fold by 2030

Study title: Exploring the sustainable scaling of AI dilemma: a projective study of corporations’ AI environmental impacts

Evidence type
Benchmark or lab evaluation
Sources
1 sourcePeer-reviewed
Published
3 min read
Truvace record
TRV-2026-2487cite or check it

Study title: Exploring the sustainable scaling of AI dilemma: a projective study of corporations’ AI environmental impacts

A modeling study in Scientific Reports estimates that a typical company among the world's 2,000 largest by revenue could see its AI electricity use rise 24.4-fold by 2030 in a high-adoption scenario. In that scenario, greenhouse gas emissions would rise 18.6-fold, driven by 47% yearly growth in generative AI use, 55% yearly growth in AI agent use cases and larger models with limited hardware improvement.

Read the study abstract

The rapid growth of artificial intelligence (AI), particularly Large Language Models (LLMs), raises environmental concerns extending beyond greenhouse gas emissions. Limited transparency from major providers hinders companies’ ability to evaluate their AI-related environmental impacts and achieve their net-zero targets. This paper proposes a methodology to estimate the environmental impact of a company’s AI portfolio, delivering actionable insights without requiring deep AI or Life-Cycle Assessment (LCA) experti…

The problem

A modeling study projects that, in a high-adoption scenario with little efficiency progress, a typical Global 2000 company's AI electricity use would reach 24.4 times its 2024 level by 2030, with greenhouse gas emissions up 18.6-fold.

In brief

The authors built a baseline for 2024: a company running 100 AI use cases uses 3.9 GWh of electricity a year, emitting 2,480,000 kg of CO2 equivalent and using 160,000 cubic metres of water. Generative AI makes up only 29% of those use cases but accounts for nearly all inference energy. Extrapolated to all Global 2000 companies, the figure comes to about 7.8 TWh in 2024.

Even the most favourable efficiency scenarios leave a gap. Matching a 90% cut in emissions from 2024 would require hardware efficiency gains of 175 to 565 times, according to the model. The projections rest on assumptions about adoption, model size and electricity mix that the authors describe as uncertain.

Main points

  1. In the high adoption scenario, energy use reaches 2,440 on an index where the 2024 portfolio equals 100, with GHG emissions at 1,862 and water consumption at 2,102.
  2. The intermediate scenario, with 40% yearly growth in generative AI and 45% in agent use cases, raises energy use 7.6-fold, GHG emissions 5.8-fold and water use 6.5-fold.
  3. A scenario pairing limited growth with efficiency breakthroughs cuts energy use by 70% and GHG emissions by 83% compared with 2024.
  4. Per chat inference, generative AI used 25 to 4,600 times more energy than conventional language-processing models; Llama 405B used 17 Wh and Llama 8B used 0.093 Wh.
  5. Energy per inference rises with workflow complexity: 1.55 Wh for chat, 2.64 Wh for retrieval-augmented generation and 8.54 Wh for agentic workflows.
  6. Embodied impacts matter more for water (28% for generative AI) and resource depletion (about 89%) than for GHG emissions (5%).
  7. The authors say that only 12% of companies measure AI's environmental impact and that 64% call AI energy use too complex to measure.
  8. The authors call for standardized assessment frameworks, more transparency from model providers and a 'Return on Environment' metric.

The full story

The method combines four models: a life-cycle assessment of compute, storage and network hardware; 192 clusters of AI use cases defined by AI type, use case, model size, daily users and usage frequency; a typical portfolio for a company with more than 20 billion in annual revenue; and 2030 scenarios. Impacts are measured across greenhouse gas emissions, resource depletion, primary energy, water deprivation and final energy.

Use cases scale sharply with model size. A chat inference on Llama 405B uses 186 times more energy than one on Llama 8B, because its throughput is 83% lower and it needs 35 times more virtual GPUs. The modeled portfolio is 29% generative AI and 71% traditional AI, yet generative AI accounts for 99.9% of use-case energy because of its user numbers and usage frequency. The 2024 estimate of 7.8 TWh across Global 2000 companies sits alongside International Energy Agency projections of AI data-centre electricity rising from 7.3 TWh in 2023 to 70 TWh by 2026.

Sensitivity tests show energy use moves one-for-one with changes in projected model size and output tokens: a 10% change shifts energy use by 10%. Energy use rises faster than linearly with agent adoption growth, from 512 on the index at 25% yearly growth to 1,237 at 65%. Hitting a 90% GHG cut would take hardware efficiency gains of 565 times in the high adoption scenario, even with a 1.04 data-centre power efficiency ratio and Paris-aligned grid decarbonization. The 'Technological breakthrough' scenario still lands at 402 on the index.

Sources

1 on file
  1. Scientific Reportsdoi:10.1038/s41598-026-77711-2
    Peer-reviewed2026-11-05

Talk it over

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

Start with a question

Modeling study projects a big company's AI power use could rise 24.4-fold by 2030https://truvace.com/item/exploring-the-sustainable-scaling-of-ai-dilemma-a-projective-study-of-corporatio

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

    The problem A modeling study projects that, in a high-adoption scenario with little efficiency progress, a typical Global 2000 company's AI electricity use would reach 24.4 times its 2024 level by 2030, with greenhouse gas emissions up 18.6-fold.

    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 Scientific Reports.

    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 2030 figures are scenario projections that depend on assumptions about adoption, model size and electricity mix, and the model leaves out pre-training and indirect effects.

    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