Could a company's AI electricity use rise 24-fold by 2030? Only in a high-adoption modelling scenario
A peer-reviewed Capgemini modelling study starts with a fictive 100-use-case company, then varies adoption, model size and efficiency. Its 24.4-fold energy result is a scenario—not an observed forecast—and the most optimistic technology pathway cuts energy below the modelled 2024 baseline.
The short version
- What happened
- The authors model a fictive large company with 100 AI use cases—29% generative and 71% traditional—rather than auditing a real company's meters, cloud bills or workloads.
- Who it affects
- Employees and customers may experience richer AI services while the environmental cost remains hidden in cloud contracts and repeated agent calls.
- What remains unproven
- The baseline is a fictive 100-use-case company, not an audited real portfolio.
Keep in mind
The baseline is a fictive 100-use-case company, not an audited real portfolio.
The Global 2000 extrapolation assumes diverse corporations share the same profile.
Growth rates and efficiency gains are scenario assumptions, not probabilities or observed trends.
Pretraining of public models is excluded from customer attribution, and life-cycle boundaries remain uncertain.
All authors are Capgemini Invent employees; the study was sponsored by Capgemini and authors disclose financial interests.
Source analysis · 2 linked sources. Primary analysed source date: 5 November 2026. Earliest linked supporting source: 24 January 2025.
Read the full report · 8 min The evidence, methods, implications and limitations
The result is a map of possibilities, not a corporate forecast
A new peer-reviewed study asks how the environmental footprint of corporate AI could change as companies add generative and agentic systems. Its most arresting result is a 24.4-fold increase in energy use by 2030. That number belongs to the authors' high-adoption scenario, which assumes rapid growth in generative AI and especially agentic workloads. It is not a measurement from Capgemini, a prediction for every company or a national electricity forecast. The same framework produces much smaller increases—and, under its most optimistic technology assumptions, large reductions.
That range is the study's real contribution. It connects portfolio choices to electricity, greenhouse-gas emissions, water consumption and mineral-resource depletion. It also exposes a governance problem: a company can make each inference more efficient while total impact still rises because it deploys far more systems and asks them to do more work. Conversely, slower adoption and dramatic efficiency gains can push modeled impacts below the starting point. Readers should evaluate the scenario assumptions before repeating the multiplier.[1][2]
The baseline company is synthetic
The model begins with a fictive large corporation running 100 AI use cases in 2024. Twenty-nine are generative-AI applications and 71 use traditional machine learning. From assumed use frequency, model type, inference energy and infrastructure factors, the authors estimate annual impacts of 3.9 gigawatt-hours of electricity, 2.48 million kilograms of carbon-dioxide equivalent, 160,000 cubic metres of water and 0.76 kilograms of antimony equivalent for mineral-resource depletion. These are model outputs, not readings from a named company's facilities or cloud invoices.
The paper extrapolates the same portfolio profile across the Forbes Global 2000 and obtains 7.8 terawatt-hours. That scaling exercise helps show order of magnitude, but it assumes that 2,000 diverse corporations resemble one constructed company. Banks, manufacturers, retailers and technology suppliers have different model mixes, user counts, data-centre locations and procurement arrangements. The extrapolation should not be added to national statistics as observed consumption. Its value is to ask what information companies need before making a defensible footprint claim.[1][2]
The distinction that matters
How the scenario model changes the answer
Each result combines assumed workload growth with assumed efficiency and infrastructure changes.
2024 model
Construct 100 corporate AI use cases
3.9 GWh synthetic baseline
High adoption
Grow generative and agentic workloads rapidly
24.4× energy; 18.6× GHG
Intermediate
Use mid-range growth and efficiency
7.6× energy; 6.5× water
Breakthrough
Limit growth and assume major efficiency gains
70% less energy than baseline
Why generative and agentic workloads dominate
The model assigns very different energy costs to different types of inference. A small chat request is estimated at 0.093 watt-hours, a medium one at 1.55 watt-hours and a large one at 17.3 watt-hours. A medium agentic task is set at 8.54 watt-hours because the agent may make multiple model calls and use tools. Depending on the comparison, the paper estimates generative or agentic workloads can require roughly 25 to 4,600 times as much energy per inference as a traditional natural-language-processing task. These figures are literature-based inputs, not direct measurements of every product.
Per-inference comparisons can still mislead when task boundaries differ. A single agent run may replace several human steps or many conventional model calls, while an apparently small chat can be repeated millions of times. Hardware utilization, batching, prompt length, generated tokens, model routing and cooling all affect the total. The portfolio view is preferable to one request estimate, but companies still need workload logs and supplier-specific data. Model size is varied by plus or minus 10% in sensitivity analysis and translated linearly, another simplifying assumption.[1][2]
Three 2030 pathways produce radically different outcomes
In the high-adoption scenario, the authors assume compound annual growth of 47% for generative AI and 55% for agentic AI. Modelled electricity rises 24.4 times from the 2024 baseline and greenhouse gases 18.6 times. In an intermediate pathway, electricity rises 7.6 times, greenhouse gases 5.8 times and water 6.5 times. Those multipliers follow from stipulated growth and efficiency settings. They show what happens if adoption accelerates faster than efficiency improves; they do not estimate the probability that those assumptions will occur.
The limited-growth plus technological-breakthrough scenario points the other way: energy falls 70% and greenhouse gases 83% below the 2024 modelled baseline. This optimistic path assumes strong efficiency gains and cleaner infrastructure alongside slower workload growth. It is a boundary case, not a promise that innovation will cancel demand. Efficiency improvements can be consumed by more use—the rebound effect—and clean electricity does not remove water, construction or mineral impacts. The wide spread between scenarios is evidence of uncertainty, not a licence to select a preferred number.[1][2]
Important impacts are simplified or excluded
For widely used public models, the study excludes pretraining from the corporate portfolio because an individual customer cannot readily attribute a share of a model's one-off training run. That avoids false precision but understates the full life-cycle demand created by the market for models. Fine-tuning, hardware manufacture, data-centre construction and network traffic also carry uncertain boundaries. Water estimates depend strongly on whether cooling water, electricity-generation water and local scarcity are counted. A cubic metre used in a water-stressed basin is not equivalent to one used where water is abundant.
Mineral-resource depletion is expressed in antimony equivalents, a life-cycle convention that is difficult to translate into an operational decision. The paper does not provide observed error bars for each workload because the company is hypothetical. Its authors all work for Capgemini Invent; the study was sponsored by Capgemini, and employment, shareholding or stock options are disclosed. Peer review adds scrutiny, but it does not turn a consultancy-authored scenario model into independent measurement. That commercial context belongs beside the results.[1][2]
What companies and policymakers can use now
The practical lesson is to measure portfolios rather than advertise one efficient model. Organisations should inventory use cases, distinguish training from inference, count model and tool calls, record token volumes and deployment regions, and obtain supplier data for electricity, carbon and water factors. Procurement teams can ask whether a smaller model, retrieval system or conventional algorithm meets the need. Product owners can set budgets for repeated agent steps and measure whether extra computation improves an outcome people value. Those controls work even when the exact 2030 scenario remains unknowable.
Evidence would strengthen with audited data from multiple companies, measured workloads across providers, uncertainty intervals for every impact category and location-aware water accounting. It would weaken if the portfolio assumptions fail to resemble real deployments or per-inference estimates change materially on current hardware. The paper is consequential because it makes competing futures explicit and shows why efficiency alone may not contain aggregate demand. It is not Breaking: it is a modelling framework, not an urgent outage, regulatory action or newly observed environmental incident.[1][2]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring ai energy: compare the finding and its evidence limits.
How much electricity does one AI task use? TokWatt exposes why published estimates disagree
The open database harmonises 134 estimates across 20 task types and shows a six-order-of-magnitude spread. Most reference values still rest on one source, and the boundary excludes idle capacity, hardware manufacture and user devices.
Read the next reportExplore the full source and evidence record
What this means for people
- Employees and customers may experience richer AI services while the environmental cost remains hidden in cloud contracts and repeated agent calls.
- Communities near data centres need location-specific electricity and water information; a global average cannot show local scarcity or grid stress.
- Procurement and product teams can reduce avoidable demand by choosing smaller models, limiting agent loops and measuring whether extra computation improves outcomes.
Global context
AI's environmental effects depend on where chips are made, where data centres operate, how electricity is generated and whether water is scarce locally. Corporate users may rely on suppliers that disclose only partial figures. A portfolio model can help organisations ask consistent questions, but results from a synthetic large company in a consultancy-led study cannot represent every region or firm. Transparent workload and location data are needed before scenario numbers become accountable targets.
What the evidence does not yet show
- The baseline is a fictive 100-use-case company, not an audited real portfolio.
- The Global 2000 extrapolation assumes diverse corporations share the same profile.
- Growth rates and efficiency gains are scenario assumptions, not probabilities or observed trends.
- Pretraining of public models is excluded from customer attribution, and life-cycle boundaries remain uncertain.
- All authors are Capgemini Invent employees; the study was sponsored by Capgemini and authors disclose financial interests.
What to watch next
- Audited corporate AI workload and supplier energy data
- Location-specific water withdrawal, consumption and scarcity measures
- Independent replication with uncertainty intervals and different company profiles
- Measured energy per agent task on current hardware
- Whether efficiency gains are outweighed by growth in use and longer agent chains
Living evidence record
Impact record IAI-0TJ5V1H
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
8 November 2026
Source trail
2 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 2 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 8 November 2026
Research question, topics and regions
How a modelled portfolio of 100 corporate AI use cases changes energy, greenhouse-gas, water and resource impacts under alternative adoption and efficiency scenarios
Topics: AI energy · Water use · Corporate emissions · Scenario modelling · Agentic AI
Regions: France · Europe · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.