AI's Hidden Environmental Footprint Is Becoming a Board-Level Risk
Edition 17 — Why the board should treat AI's environmental cost as a test of capital discipline and disclosure credibility, not an ESG afterthought.
Most boards have now approved an AI investment of some size, and most boards have separately signed off on a climate commitment — a net-zero target, a transition plan, or a line in the annual report about reducing emissions. The two decisions are usually made in different meetings, owned by different executives, and supported by different papers, and almost no board treats them as connected. They have become a single decision governed in two separate rooms, and the gap between those rooms is where the risk sits. AI’s environmental cost is not, in itself, the thing that should command a board’s attention; in global terms, the emissions are real but modest. The problem is that scaling AI quietly enlarges a number the organisation has publicly committed to shrinking, and from 2027 that number will be reported, assured, and read by investors. This is a question of capital discipline and the credibility of what the board has already told the market.
A real cost, sitting in someone else’s operations
The cost is easy to miss because it does not appear on the organisation’s own meters. When a company buys AI and cloud services, the energy, carbon and water sit in the provider’s operations, and are recorded as a supplier’s activity rather than the buyer’s own. The scale is no longer trivial: global data-centre electricity use was around 415 TWh in 2024, and the International Energy Agency expects it to roughly double to about 945 TWh by 2030, with AI the main driver. What should concern a board is less the raw figure than how readily it is made to look smaller. Microsoft’s most recent environmental report is an instructive case. The company reported that its direct Scope 1 and 2 emissions had fallen by close to 30% against its 2020 baseline, which reads as progress — until you note that this is the market-based figure, based on renewable energy certificates and purchase agreements. On a location-based measure that reflects the actual carbon intensity of the grids its data centres draw from, Microsoft’s Scope 2 emissions more than doubled over the same period; its total emissions rose by more than 23%, and its supply chain now accounts for over 97% of the footprint. Microsoft is among the most sophisticated and best-resourced reporters in the world. If its headline can diverge this far from its physical reality, a board should assume its own providers’ assurances do the same.
In 2027, the externality becomes an audited disclosure
Until recently, that divergence carried little consequence for the buyer, because the footprint was an externality — unmeasured on the company’s books and absent from its accounts. That is the part that ends. On 25 February 2026, the Department for Business and Trade published the UK Sustainability Reporting Standards, and the Financial Conduct Authority has proposed, under consultation CP26/5, that in-scope listed companies report against the climate standard from 1 January 2027, with a policy statement expected in the autumn. The detail that matters most here is the treatment of Scope 3, the supplier emissions that include cloud and AI use. Scope 3 is to be reported on a comply-or-explain basis, with an optional deferral into 2028. The temptation that creates is plain: a board that cannot yet measure its AI footprint, or does not like what the measurement shows, can choose to explain rather than disclose. That choice is not the safe harbour it appears to be. A reasoned explanation that later proves to have understated a known and growing exposure is precisely the territory in which greenwashing challenges and assurance failures arise. The comply-or-explain window is where the credibility risk concentrates, not where it disappears.
One balance sheet, two committees
This is the heart of the problem, and it has a structure worth naming. The AI investment and the climate commitment are a single capital decision that most organisations split across two board committees, which never reconcile: the AI or technology programme approved in one room, the transition plan owned in another, each blind to what the other is signing. Call it the two-committee problem. Every material expansion of AI use scales the supplier emissions the company has pledged to reduce, so a board approving an aggressive AI programme alongside a validated decarbonisation target is, in effect, approving two trajectories that pull against each other while believing it has settled on one strategy and one commitment. The usual reassurance — that models keep getting more efficient, so the footprint will take care of itself — does not survive contact with how organisations actually behave: cheaper, faster AI invites far more use, the efficiency gain per query is swamped by the growth in queries, and the saving is spent on more consumption rather than banked. Researchers call this the rebound effect, and there is sufficient evidence that efficiency cannot be the board’s containment strategy.
The contradiction is easiest to see in a real case. Sainsbury’s has held targets validated by the Science Based Targets initiative since 2020, and it reports that Scope 3 — its value chain including upstream and downstream activities — accounts for around 98% of its total emissions, with a validated commitment to cut the energy, industry and transport portion of that by just over half by 2030. In 2024, it also signed a five-year partnership with Microsoft, in its own words, to become the UK’s leading AI-enabled grocer, with the work running on Microsoft’s Azure cloud. That cloud and AI consumption sits inside the same Scope 3 the company has committed to reduce, and its provider is among those whose location-based emissions have been climbing even as the headline market-based figures fall. Food and agriculture still dominate a grocer’s value chain, so the digital line is not the largest one; the point is that it is among the fastest-rising, it lands in the hardest scope to cut, and it is the line a board is least likely to have priced into a target it has already had validated. In more digitally intensive sectors — banking, insurance, professional services — that line is materially larger, and the question sharpens accordingly. Whose Scope 3 are you becoming is not a rhetorical flourish; it is what determines whether the two numbers in your own report can be reconciled when an auditor reads them together.
What disciplines this is treating compute as a costed input to the AI business case rather than a free good. A board that asks, for each significant use case, whether the value justifies the energy, carbon and water it consumes, and that prices that consumption into the return it expects, is applying the discipline it already brings to any other scarce and rising cost — and, in doing so, is governing the two committees as the single decision they have always been. The alternative is to meet the contradiction in the 2027 report, when both numbers are public, and the room for explanation has narrowed.
The constraint that can stop the roadmap
There is a harder constraint that turns this from a reporting question into an operational one, at least for anything built or hosted in the United Kingdom. The capacity to power AI at scale is not assured. The volume of demand seeking to connect to Great Britain’s grid rose from 41 GW in late 2024 to around 125 GW by mid-2025 — close to three times the country’s peak demand — and data centres are the single largest driver of that demand queue. New grid infrastructure takes the better part of a decade to build, against months for a data centre, so connection, rather than model availability, increasingly sets the pace. A board that has approved an AI roadmap on the assumption that compute will simply be there is carrying a continuity risk it has probably not named: the workloads it has funded may not be able to connect, or may connect later and at a higher cost than the business case assumed. Where operations sit in water-stressed regions, the same logic applies to water as a gating input. AI being environmentally harmful is beside the point here; the resources it depends on are constrained, and a plan that ignores that is a plan that may not run.
The line that breaks the target
It is worth being specific about where this lands first, because the early evidence already points to it. When the Science Based Targets initiative removed the net-zero commitments of 239 companies in 2024 — Microsoft, Unilever, Walmart and Marks & Spencer among them — the recurring reason was Scope 3: more than half of the companies surveyed said the supply-chain emissions were simply too hard to measure and reduce, and the standard now requires any company whose Scope 3 exceeds 40% of its footprint to cover at least two-thirds of it. Into that already-failing category, AI and cloud are pouring a fast-rising and structurally awkward new load, because the providers absorbing that demand are themselves reporting rising emissions that flow straight into their customers’ Scope 3. The forcing date is fixed: the first accounts under mandatory UK SRS cover the 2027 and 2028 financial years and will be published and assured across 2028 and 2029. My expectation is that in those first two reporting cycles, AI- and cloud-driven Scope 3 growth will become a named, recurring reason that large UK-listed companies revise net-zero targets, miss interim milestones, or choose to explain rather than disclose — moving digital emissions from an unremarked line in the accounts to one that investors and auditors interrogate directly. The boards that find themselves exposed will not be the negligent ones; they will be the ones that set a credible target when compute was cheap and invisible, and never reconvened the two committees when it stopped being either.
What changes in the boardroom
The corrective is to convert an invisible externality into a measured, owned and costed input, which is work the board can direct rather than admire. The starting move is to require the organisation’s AI and cloud providers to disclose carbon and water on a location-based as well as a market-based measure, and to write those reporting obligations into the contracts rather than accepting headline assurances. Accountability should sit with a named executive — a chief sustainability or operating officer — rather than being diffused across whoever happens to procure the technology. The board can set a threshold for when a frontier or general-purpose model is justified against a smaller, task-specific one, since the largest models can consume on the order of 30 times the energy of a fit-for-purpose alternative for tasks that do not require them. An internal carbon price on compute turns the rebound problem into a budgeting discipline, making teams weigh consumption instead of treating it as free. And the AI footprint belongs inside the transition plan and the AI business case, assessed together, so that the contradiction surfaces in the boardroom rather than in the annual report. None of this asks the board to slow its AI ambition; it asks the board to fund that ambition with the discipline it applies to every other commitment it has made to its investors.
The emissions, set against the global total, are modest. The governance gap is not. A board that leaves the two-committee problem unresolved — its AI ambition settled in one room, its climate commitment in another — is allowing a contradiction to compound quietly until a disclosure regime makes it visible to the investors and regulators whose trust the company depends on. The work between now and 2027 is to close that distance on purpose: to know the footprint, price it into the business case, assign an owner, and ensure the account the company gives of its emissions is one that its own numbers will support under assurance. Handled that way, the environmental cost of AI becomes another input the board governs with discipline, which is the standard the rest of its capital decisions already meet.
AI in the Boardroom is written for the directors, executives and advisers who have to make these trade-offs in practice rather than in principle. If governing AI as a capital and credibility decision, rather than a separate ESG exercise, is the kind of analysis you want more of, consider subscribing. Future editions will continue to examine the points where AI strategy, governance, and the numbers a board is accountable for intersect.