Corporate AI roadmaps are colliding with net-zero commitments. New compute demand creates long-lived power contracts and, in some cases, dedicated fossil generation. Even without building a data centre, AI use raises Scope 2 emissions through electricity. It also raises Scope 3 emissions through services, hardware, and construction.
For industrial organisations, the tension is specific. AI offers measurable efficiency gains in sectors like cement, steel, and energy management. The compute required to deploy it at scale carries an emissions footprint. Most governance frameworks are not yet equipped to measure or manage that footprint. The governance challenge is how to treat AI compute as the physical infrastructure it is.
The shift from cloud abstraction to dedicated power
Until recently, corporate AI compute sat behind a layer of cloud abstraction. Emissions were indirect, difficult to trace, and typically absorbed into broad Scope 3 estimates. That is changing.
A WIRED investigation published in April 2026 reviewed air permit documents for natural gas projects linked to just 11 US data centre campuses. Those campuses are connected to OpenAI, Meta, Microsoft, and xAI. The permits show potential annual greenhouse gas emissions of more than 129 million tons. That exceeds Morocco’s entire national output for 2024. Permit figures reflect maximum theoretical capacity. AI facilities can run close to steady load, and the direction of travel is unambiguous. Treat permitted levels as a material risk signal.
The IEA projects that natural gas and coal together will meet over 40 per cent of the additional electricity demand from data centres until 2030. Power use from AI-focused data centres is set to triple by 2030. For organisations with science-based targets, that trajectory sits directly in the procurement and finance risk register.
As capacity shifts from shared grid supply to dedicated assets, the emissions profile of AI workloads becomes shaped by siting, permitting, and fuel contracts. These are choices with long tails and they are hard to reverse once signed.
Why corporate targets get exposed
The headline risk is asset lock-in. Turbines, pipelines, manufacturing plants and long power purchase agreements can outlive interim carbon targets. They raise transition risk for buyers of AI services who have not built traceability into procurement.
Claims risk is equally practical. Many corporate targets rely on market-based accounting and certificates. Those instruments can leave a gap between contractual claims and real-world generation at the hour a model trains or a process runs. For company leaders, the question is whether public commitments match the physical system serving the compute.
The SBTi’s Corporate Net-Zero Standard requires Scope 1, 2, and 3 targets to be delivered through real emissions cuts. Where Scope 3 accounts for 40 per cent or more of total emissions, a Scope 3 target is mandatory. For most industrial organisations, purchased AI services sit in Scope 3 under purchased goods and services. Supplier disclosures linking workloads to electricity sources and locations are a governance requirement.
According to the 2025 ITU/WBA Greening Digital Companies report, the most recent edition available, operational emissions among digital companies heavily investing in AI had risen to 150 per cent of 2020 levels by 2023. None of the top ten emitters in the digital sector has an SBTi-validated target aligned with 1.5°C. Industrial buyers of AI services inherit a portion of that exposure.
Three questions that should be routine
Three questions should become standard in any steering committee or procurement process involving AI. Where does the workload run? What powers it, hour by hour where possible? What contract term fixes the emissions exposure?
These are the same questions that would apply to any supplier of an energy-intensive material input. In a materials contract, unknown fuel, unknown permits, and unknown lifetime emissions would be unacceptable. The same standard applies to compute.
A governance framework for AI emissions
AI compute should be treated as physical infrastructure with auditable energy and emissions supply chains. Governing it requires joint ownership across procurement, IT, finance, and sustainability, backed by a verifiable fact base.
For each major workload, the aim is to build an AI infrastructure file: a document that finance, internal audit, and sustainability teams can review. It turns compute from a vague line item into a governed input with defined exposure.
The file should address seven areas: the top AI applications by spend and compute demand; the regions and specific sites used for training and inference; the grid mix and power purchase agreements in use; whether any behind-the-meter generation is involved; how renewable claims are matched on an annual, monthly, or hourly basis; backup generation assumptions and actual run hours; and operational targets for facility energy and IT utilisation.
Interim power arrangements need end dates and decarbonisation gates. Time-bound commitment to lower-carbon supply separates credible disclosure from a holding position.
Procurement: the fastest lever available
Most industrial organisations already know how to govern carbon in procurement for steel, concrete, or logistics. The same discipline applies to AI.
Research consistently finds that the majority of companies fail to scale AI deployments beyond the pilot stage. If a workload may never move past proof-of-concept, locking in high-emissions infrastructure is difficult to justify. Contract for flexibility first. Expand only when the business case and the emissions boundary are both explicit.
Minimum procurement standards for AI should cover site transparency, power sourcing, and reporting cadence. They should also specify limits on fossil-based interim power, including backup generation assumptions. Enforcing these through contracts also improves internal decision-making. Teams can compare use cases on a consistent basis rather than approving new models on performance alone.
Operations: measure what matters
Some AI use cases reduce emissions in industrial operations. Process optimisation in cement and energy management in buildings are established examples. The IEA estimates that AI could cut energy costs in heavy industry by three to ten percentage points. Absolute compute demand can still rise even as individual use cases improve efficiency. Governance has to manage both effects through measurement and decision gates.
The most useful operational metrics are straightforward. They are kilowatt-hours per training run, carbon intensity of the serving region, and uptime requirements that drive redundancy. These are the constraints that push suppliers toward extra capacity, extra cooling, and diesel backup. Without specifying acceptable ranges, vendors will optimise for availability alone.
Finance and risk: treat compute as a carbon-exposed commodity
For finance and risk functions, the reframe is direct. Compute is a carbon-exposed commodity. It carries physical infrastructure risk, transition risk, and claims risk. The same applies to fuel, freight, and other materials.
Ask for an AI infrastructure file for each major workload. Require that finance, internal audit, and sustainability teams can review it. Set decision gates that tie compute expansion to verified value delivery and a defined emissions envelope. The aim is to ensure emissions exposure is known and bounded before contracts are signed, and compatible with the organisation’s carbon commitments.
The regulatory context is tightening. The EU AI Act already requires energy transparency for high-impact AI systems. Under CSRD, ESRS E1 requires companies to explain how their targets are compatible with limiting warming to 1.5°C. Investor scrutiny of AI-related emissions disclosures is increasing. Permit documents now make the gap between corporate commitments and physical infrastructure visible.
The industrial decarbonisation case for AI is real and significant. So is the governance obligation that comes with deploying it at scale. The two are fully compatible. They require the same rigour.
Written by Dominic Shales, Partner & CMO, Nexus Climate