Beyond AI's Energy Footprint: What The Sustainability Conversation Misses
Mikko Kärkkäinen, CEO, RELEX Solutions.
gettyThe sustainability debate around AI has largely focused on energy consumption. Data centers draw measurable power, model training has a calculable carbon cost and inference requests can be metered. These impacts are visible and auditable, often attracting attention from boards, regulators and ESG teams.
The response is predictable. Companies publish AI energy commitments, invest in offsets and treat AI as a sustainability liability to be managed. That framing isn’t wrong, but it’s incomplete.
Some of the highest environmental costs in retail and manufacturing never appear on the same ledger as energy use. Overproduction, excess inventory, food spoilage and half-empty trucks generate significant emissions, land-use pressure and water consumption that don’t show up in a sustainability report.
The International Energy Agency (IEA) estimates that data centers account for roughly 1% of global energy-related carbon dioxide emissions. That footprint is real and important to manage. But energy is only part of the picture. The UN Environment Programme (UNEP) estimates food loss and waste at around 8% to 10% of global greenhouse gas emissions, making the climate cost of food waste roughly 10 times that of all data centers worldwide.
That ratio reframes where sustainability attention should be directed. When AI is applied to the planning decisions around demand, replenishment and network coordination, the waste avoided dwarfs the energy spent running the systems.
Consider what this looks like in practice. A grocery retailer applies AI to fresh food ordering and materially reduces shrink. The carbon equivalent of that avoided waste, accounting for land use, water, transport, refrigeration and decomposition, runs to multiples of the total energy consumed by the AI system generating those recommendations. The calculation isn’t complicated, but it isn’t being made consistently because the two sides sit in different parts of the business and are rarely brought into the same view.
There’s a second issue the sustainability conversation misses. Not all AI is the same, and the choice of tool for each task carries its own environmental weight. LLMs are suited to language, unstructured content and open-ended reasoning. Demand forecasting, replenishment and inventory optimization are a different kind of problem. They need to be reliable across millions of decisions, explainable to a planner or auditor, fast and cheap to run at scale and able to improve from their own outcomes. Purpose-built models trained on operational data meet those tests at a fraction of the compute cost of a general-purpose LLM, which means more waste avoided per unit of energy spent.
The sustainability implication is direct. Deploying an LLM for a problem a lighter, purpose-built model could handle well adds to the very footprint sustainability teams are working to reduce. As the range of available AI tools grows, organizations need a deliberate framework for matching computational approach to task type, treating energy efficiency as a selection criterion alongside accuracy and cost.
For leaders, this is a structural problem as much as a technical one. In most organizations, sustainability and supply chain teams aren’t having the same conversation about AI. That means the waste-reduction value of planning AI is underestimated, energy use is scrutinized in isolation and tool-selection decisions are made with no sustainability lens at all.
The environmental case for AI can’t be made from one side of the ledger. Energy consumption gets measured. The waste that better planning prevents is harder to quantify, so it rarely is. That asymmetry skews how leaders evaluate the trade-off. Directing AI toward the decisions with the highest avoidable waste and matching the tools used to the actual demands of each problem is a more useful sustainability frame than energy consumption alone.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
