AI’s
Dirty Secret: Accenture Warns Carbon Emissions Could Soar 11x by
2030
Artificial intelligence is being marketed as a solution to some of
the most pressing problems on the planet, including climate change. AI
systems are being deployed to optimise renewable energy grids, model climate
systems with greater accuracy, and identify efficiency improvements across
industrial processes. The irony that is becoming harder to ignore is that the
infrastructure powering these solutions is itself a significant and rapidly
growing source of the emissions AI is supposed to help
reduce.
A landmark report from Accenture, titled Powering
Sustainable AI, projects that without urgent action, AI’s carbon
footprint could increase elevenfold by 2030. That figure is not a worst-case
outlier from a fringe analysis. It is the central projection from one of the
world’s largest professional services firms, based on current enterprise AI
adoption trajectories and the energy intensity of existing model
architectures. It deserves to be treated as the serious warning it
is.
The Energy Arithmetic of AI
The energy demands of AI divide into two categories that each
carry significant footprint. Training large models, the computationally
intensive process of adjusting billions of parameters against enormous
datasets, consumes electricity in quantities that dwarf most other computing
workloads. A single training run for a frontier language model can consume
more electricity than a hundred US homes use in a year. As models grow larger
and are retrained more frequently, that figure compounds.
Inference, the process of running a trained model to generate
outputs for users, is individually less energy-intensive per query but occurs
at a scale that training does not. Every search query processed by an AI
mode, every customer service interaction handled by a chatbot, every
recommendation served by a content algorithm represents a small inference
computation. Aggregated across hundreds of millions of daily interactions,
inference energy consumption is substantial and growing rapidly with
adoption.
The arXiv
paper on sustainable AI scaling models multiple adoption scenarios
and finds that even under moderate growth assumptions, the energy consumption
of AI systems will increase by a factor that current grid infrastructure and
renewable energy deployment are not on track to absorb sustainably. Under
high-adoption scenarios, the paper projects a 24-fold energy increase by the
end of the decade.
The Corporate Blind Spot
One of the most striking findings in the Accenture report is the
gap between corporate sustainability commitments and AI strategy. The
majority of large organisations have made net-zero or carbon reduction
pledges. A much smaller proportion have any methodology for measuring the
carbon footprint of their AI workloads, let alone for incorporating those
footprints into their sustainability reporting.
As Axios
reported on the study’s release, corporate sustainability reports
consistently fail to factor in AI-related emissions. The omission is partly
methodological: measuring the carbon footprint of cloud-based AI workloads
requires granular data about the energy sources used by specific data centres
at specific times, which cloud providers have not historically made available
to customers. It is also partly a question of incentives: organisations that
have made ambitious sustainability pledges are not well served by discovering
that their accelerating AI adoption is materially undermining
them.
The result is a systematic blind spot in corporate sustainability
accounting. The AI transformation strategies being pursued by organisations
across every sector are generating carbon footprints that are real, growing,
and largely invisible in the metrics by which sustainability performance is
measured and disclosed.
The Hidden Infrastructure Problem
As explored in Invisible
Infrastructure, the systems that power AI operate largely out of
sight. Data centres processing AI workloads are typically located far from
the users they serve, in jurisdictions chosen for energy cost and land
availability rather than renewable energy availability. The carbon intensity
of the electricity powering any given computation depends on the grid mix at
the specific facility at the specific time, a figure that varies by location,
season, and time of day.
The hyperscaler infrastructure buildout described in The
Trillion-Dollar AI Arms Race is relevant here. The hundreds of
billions being invested in new data centre capacity will determine the energy
profile of AI workloads for decades. Facilities built now will operate for
fifteen to twenty years. Whether they are powered by renewable energy or by
whatever is cheapest on the local grid is a decision being made now, with
long-term consequences that current sustainability metrics are not
capturing.
Mitigation Pathways
Accenture’s report is not merely a warning. It outlines mitigation
pathways that, if adopted at scale, could substantially reduce AI’s projected
emissions trajectory. The most significant levers are architectural
efficiency, energy sourcing, and procurement standards.
Architectural efficiency means building models that achieve given
capability targets with less computation. Techniques including model
compression, knowledge distillation, and more efficient training algorithms
have demonstrated the ability to reduce energy consumption substantially
without proportionate capability degradation. The challenge is that the
commercial incentives for the largest model developers are currently aligned
with scale rather than efficiency: bigger models attract more attention, more
talent, and more investment, even when smaller models can achieve comparable
results on practical tasks.
Energy sourcing means powering AI workloads with genuinely
renewable electricity, not just purchasing renewable energy certificates that
may not correspond to actual clean power at the time and place of
consumption. Several major cloud providers have made commitments to real-time
renewable energy matching for specific data centres, but these commitments
cover a small fraction of total AI workload capacity.
The environmental concerns about AI’s energy footprint connect
directly to the food and resource security concerns examined in Algorithmic
Hunger. AI is simultaneously being deployed to address food
security challenges and, through its energy demands, contributing to the
climate disruption that makes those challenges more severe. The contradiction
demands honest accounting.
Regulatory and Standards Gaps
No regulatory framework currently requires organisations to
measure and disclose AI-related emissions. The EU’s Corporate Sustainability
Reporting Directive creates obligations for large companies to report on
their environmental footprint, but the methodology for AI workload
attribution is not standardised and the disclosure requirements for Scope 3
emissions, which is where cloud AI workloads typically sit, are subject to
ongoing debate.
Accenture calls for governments and enterprises to collaborate on
global sustainability standards specific to AI, including standardised
measurement methodologies, mandatory disclosure requirements, and procurement
standards that require sustainability performance from AI vendors. Without
those standards, the competitive dynamics of the AI industry will continue to
reward capability and cost performance over environmental stewardship.
The Question of Proportionality
The most difficult aspect of AI’s carbon footprint is the question
of proportionality. If AI genuinely delivers the productivity gains, health
improvements, scientific accelerations, and environmental optimisations its
advocates project, the energy it consumes may represent a worthwhile trade.
If many of the applications consuming that energy are delivering marginal
convenience improvements to already wealthy users, the calculus is very
different.
Making that assessment honestly requires the kind of rigorous
sustainability accounting that currently does not exist. The elevenfold
increase in carbon emissions that Accenture projects is not inevitable. It is
the consequence of a set of choices, architectural, commercial, and
political, that are being made right now, mostly without adequate information
about their environmental consequences. Providing that information is the prerequisite
for making better choices. The time to start is before the infrastructure is
built, not after.
About the Author
Stuart Kerr is the Technology Correspondent for LiveAIWire,
covering artificial intelligence, ethics, and the ways technology is
reshaping everyday life.