Behind the apparent weightlessness of AI is physical infrastructure – and associated emissions. Photo: halbergman/iStock
25 Sep 2026, 1:32 pm
The question is no longer simply whether AI has an environmental footprint, it is where those emissions arise, how they are allocated, who must report them and whether claims made about them can withstand audit and assurance.
Rapid AI adoption is accelerating demand for data centre capacity and electricity. This, in turn, is increasing the risk of higher absolute emissions even as major technology companies pursue ambitious decarbonisation pathways.
Responsible AI controls may initially be perceived as a drag on deployment, speed or innovation. In practice, however, proportionate controls over model choice, compute, workload timing and supplier transparency are likely to become part of mainstream governance, policy and the sector’s wider licence to operate.
The real challenge ahead, however, is how we can capture AI’s economic and operational benefits while ensuring that its energy demand, carbon consequences and environmental claims are visible, attributable and governed.
Generative AI is digital at the point of use, but industrial at the point of production.
For most organisations, AI appears deceptively simple. Employees interact with a chatbot, lawyers summarise documents, analysts interrogate datasets, finance teams automate reporting and product teams embed AI functionality into software.
Behind that apparently weightless digital service, however, is an increasingly material physical system of data centres, power generation, cooling equipment, servers, semiconductors, batteries, substations and electricity networks.
As we observed in a recent interview with FT Sustainable Views, the methodology for identifying all these different aspects – including where the emissions are originating from at the data centre level – has not yet been developed.
This creates an emerging challenge for businesses, regulators and technology providers. The question is no longer simply whether AI has an environmental footprint. It is where those emissions arise, how they are allocated, who must report them and whether claims made about them can withstand audit and assurance.
None of the greenhouse gas (GHG) Protocol, ISO standards nor the EU and UK sustainability-reporting frameworks create a standalone accounting category called ‘AI emissions’.
Instead, AI-related impacts must be located within the existing scope 1, scope 2 and scope 3 architecture.
Depending on the reporting and operational boundary:
For most ordinary corporate users, the largest immediate issue is likely to sit upstream in scope 3 because their organisation is purchasing AI capability rather than operating the underlying infrastructure.
The same physical emissions can therefore appear differently in the inventories of suppliers and customers.
The cloud is somebody else’s data centre, and somebody else’s scope 1 and scope 2 can become the corporate customer’s scope 3.
Microsoft’s Azure emissions calculation methodology, for example, allocates emissions associated with Azure and Microsoft 365 services to customers. It covers operational electricity and selected lifecycle impacts from data-centre hardware. It also illustrates the present limitations: some allocations use proxy usage measures, and the methodology does not currently include all critical data-centre facility infrastructure.
Google Cloud’s Carbon Footprint methodology similarly allocates emissions from computing infrastructure to products and then to customers according to their usage. Google expressly notes that its customer-specific data has not been third-party verified or assured and that methodological or data-source changes may result in material revisions to current and previous calculations.
The market is developing practical allocation approaches, but there is not yet a uniform, dedicated methodology specifically for AI services.
The existing GHG Protocol provides the accounting architecture within which AI-related emissions are considered. However, it does not currently prescribe a common “per prompt”, “per token” or “per model” methodology.
This matters because AI relies on highly shared infrastructure. A single data centre can support thousands of customers, numerous cloud products, multiple AI models and other storage and network services.
The methodological questions include whether energy and emissions should be allocated using:
Research published by Google describes a methodology for allocating energy across shared data-centre machines and software using resource reservations and hourly measured usage, combined with hourly, location-specific carbon-intensity estimates. The paper describes physical allocation factors, such as the approach preferred by the GHG Protocol’s scope 3 reporting standard.
That is a significant development, but it is not yet a universal standard followed consistently across all cloud and AI providers.
The result is that the industry has an established accounting framework, but not yet a settled method for translating millions of shared computing transactions into comparable, customer-level AI emissions information.
The environmental impact of one AI interaction may be small. The corporate significance lies in the cumulative effect of AI becoming embedded across thousands of employees, customer interactions, products and automated workflows.
ChatGPT is frequently being used in offices and households globally. It is estimated that approximately 0.15g CO2e (carbon dioxide equivalent) is emitted by a standard email drafted using ChatGPT-4o, while approximately 3g CO2e is emitted by ChatGPT-4o reviewing a 100-page report.
The important qualification is that small individual figures can become material when activity is multiplied across an enterprise and combined with company-wide AI platforms.
Prompt-level estimates can be useful illustrations, but they should not create a false sense of precision. The actual impact can vary according to:
The wider energy-system picture is also material. FT Sustainable Views reported, by reference to the International Energy Agency, that fossil fuels supply approximately 56% of the current global energy used for AI, although renewables are taking a growing share.
This means that a corporate user can acquire an AI service without seeing the fossil generation, grid constraints or infrastructure investment sitting behind it.
The developing legal and commercial challenge is not simply determining how much carbon was emitted.
Organisations are increasingly asking:
The future debate is therefore shifting from carbon measurement to carbon attribution.
This distinction is particularly important where data-centre power is supplied through different commercial structures, such as:
Contractual choices can alter reporting boundaries, data rights and the way emissions appear within the parties’ inventories, even where the underlying physical generation remains substantially unchanged.
Carbon accounting is therefore becoming both an engineering question and a contractual question.
However, risk allocation is not decarbonisation. Changing the ownership of a generator or the structure of a power contract may change how emissions are reported, but it does not remove them from the atmosphere.
AI growth is also drawing greater attention to the difference between physical power supply and contractual energy claims.
There are important distinctions between:
These instruments can perform legitimate functions within accepted carbon-accounting frameworks. They do not, however, answer every question about the physical generation supporting a particular data centre at a particular time.
A facility may purchase renewable-energy attributes on an annual basis while physically drawing electricity from a grid that includes gas, coal or other fossil generation. Corporate customers may receive contractual or market-based emissions information without seeing the underlying location-based position.
The legal and governance question is increasingly whether the renewable or emissions-reduction claim clearly communicates what has been achieved, the methodology used and any material limits.
This is especially important where companies incorporate the information into regulated disclosures, net zero progress statements, product claims or procurement decisions.
Claims should not be transferred automatically from supplier marketing material into corporate sustainability reporting without understanding first the methodology and its limitations.
One other significant aspect of AI growth is the time required to connect new data centres to the electricity network.
AI infrastructure can sometimes be developed faster than the grid reinforcements required to serve it. As a result, developers are considering:
In the UK, Ofgem has already consulted on proposed data-centre connection reforms, including commitment fees and queue-management milestones intended to prioritise viable projects. The proposed evidence requirements include a credible end-user, procurement of long-lead electrical equipment and financial and technical capability.
Temporary or dispatchable generation may have a legitimate role where network connections are delayed or where critical infrastructure requires resilience.
The issue is whether temporary power remains temporary.
A credible arrangement should establish:
These are governance recommendations rather than a single, prescribed regulatory test. They reflect the need to prevent a short-term infrastructure solution becoming a source of long-term carbon lock-in. Temporary power should be governed as a transition plan in miniature: with a defined purpose, transparent emissions, milestones and a credible exit.
Reliance on future carbon capture also requires caution. A future plan to abate gas-generation emissions should not be treated as evidence that existing operations are already low-carbon.
Companies need to understand when capture is expected to operate, the anticipated capture rate, the availability of transport and storage infrastructure, residual emissions and how unabated operation is disclosed before commissioning.
The EU’s Omnibus I reforms have significantly narrowed the scope of the Corporate Sustainability Reporting Directive (CSRD).
The final legislation raises the main EU-company thresholds to more than 1,000 employees and more than €450 million in net annual turnover. It also includes a transitional exemption for first-wave companies falling outside the revised scope for the 2025 and 2026 financial years.
This represents a substantial reduction in the number of companies required to report directly. It does not make AI-related impacts disappear.
Larger customers, banks and investors will continue to need value-chain information. The EU has also introduced a voluntary reporting standard and value-chain protections intended to give smaller businesses a proportionate basis for responding to sustainability data requests. Formal reporting obligations may start with the largest companies, but it is worth noting that information request travels through the supply chain.
The European Commission has also proposed a common rating scheme for data centres aimed at improving transparency in resources – energy use and water – are being used to support data centre operations across the EU. The plans are still being reviewed by the EU Parliament and the Council, but the Commission’s proposals indicate that sustainability will continue to be an important focal point in the ongoing discussions across Europe on AI and data centre development. If approved, the first sustainability labels for date centres are expected to be displayed next year.
In the UK, the Financial Conduct Authority’s (FCA) has proposed to introduce UK SRS-aligned reporting for specified listed-company categories for accounting periods beginning on or after 1 January 2027, subject to the final policy statement.
The FCA’s proposed approach provides additional treatment for scope 3, with comply-or-explain applying from 2028. The consultation has closed. The FCA plans to publish a policy statement in autumn 2026, with new rules expected to come into force from 1 January 2027.
Once the rules are finalised, it will be important for companies to distinguish between the current legal position and proposed future requirements. They should not, however, wait for final rules before creating the evidence needed to support future disclosures.
A smaller company may not be directly subject to the CSRD or proposed UK listed company rules. It may nevertheless face questions from corporate customers, financial institutions, investors, insurers, public sector procurers or supply chain partners.
Unnecessary model calls, oversized models and uncontrolled automated agents can increase cloud expenditure as well as energy consumption.
Responsible AI is likely to become part of ordinary cost and performance management. Companies of all sizes should map their AI use and agreements with AI providers.
For smaller businesses, the proportionate response is not to calculate the lifecycle impact of every employee prompt. It is to identify the most material services and decisions first.
That includes knowing:
Responsible AI should not be interpreted as discouraging valuable AI adoption. The objective is intentional, proportionate use.
Policies could encourage employees and system designers to consider:
Internal AI policies may increasingly ask if users need to run this through AI? AI work that is non-time-sensitive could also be carried out when the grid is using lower-carbon electricity.
It is even conceivable that responsible AI use could become part of ordinary operational governance alongside information security, privacy, accuracy, intellectual property and cost management.
The carbon intensity of a digital workload can vary by geography and time.
Research examining AI workloads on cloud infrastructure has found that the geographic region of a data centre can have a significant impact on operational carbon intensity and that the time of day can also be relevant.
As we are seeing in Australia, this is already laying the foundations for a new form of location strategy for data centres.
In the past, companies moved administrative and back-office processes between jurisdictions primarily for labour costs and access to skills. As AI automates more non-time sensitive activity, companies may increasingly consider:
Countries with abundant low carbon electricity and available network capacity may benefit from hosting automated back-office workloads. Equally, moving computing activity to another jurisdiction should not become a new form of carbon offshoring that ignores embodied infrastructure, water impacts or local system constraints.
The decision should be based on the full evidence rather than a single grid-emissions figure.
Organisations do not need to wait for a perfect AI emissions standard before acting. Prudent firms should act now by:
The EU Joint Research Centre’s Code of Conduct on Data Centre Energy Efficiency and its best practice guidance are relevant because they address operational energy performance rather than simply emissions classification.
The framework covers recognised practices for efficient data-centre design and operation and can be used by customers and suppliers as a procurement reference.
Its best practice guidance is also referenced in EU energy efficiency and taxonomy-related policy.
This provides an important distinction: the GHG Protocol and reporting standards help determine where emissions sit and how they are accounted for, while operational frameworks help determine how the physical infrastructure can use less energy. Both are needed.
The conversation about AI and sustainability is often framed as a debate about energy consumption or the emissions associated with an individual prompt. These questions matter, but they are incomplete. The more difficult challenge is connecting corporate AI use to cloud allocation, the physical data centre and the energy supply behind it.
AI also has an enabling role. It can support power forecasting – which predicts how much electricity will be needed in a particular location, area, or region – as well as renewable integration, predictive maintenance, demand flexibility and operational efficiency. Its impact should therefore not be assessed solely as a gross footprint.
The governance challenge is to demonstrate both sides of the equation: the resources AI consumes and the environmental benefit it claims to enable.
The next phase of AI sustainability is not merely disclosure. It is attribution, evidence, delivery and accountability.
Written by
OUT-LAW ANALYSIS
11 Aug 2026
