Software as heavy industry: AI has broken the carbon accounting model

One of my colleagues recently used our Ask Us Anything portal to pose the above anonymous question about how we intend to reconcile the agency’s rapidly increasing use of generative AI with our roadmap to net zero. It was a particularly astute question precisely because I’ve been thinking about this problem for more than a year, and while I agree that our use of AI should of course be factored in, as yet I’ve got no working methodology to do that in a useful or meaningful way. And I’m not alone.

That it comes with a substantial carbon footprint and uses an unholy amount of water is well understood, but it’s perhaps less obvious that in carbon measurement as in so much else Artificial Intelligence is doing more than accelerate change or challenge conventional ways of working. Rather it is breaking the model itself: we all know that our use of this technology carries environmental costs, but so far as the existing measurement techniques or attribution models are concerned those costs are almost entirely invisible. The problem we all face isn’t just that AI use risks an increase in emissions, it’s that it fractures the way that those emissions are measured and attributed.

Space & Time has pledged to achieve net zero by 2030 on a 90/10 basis against a 2019 base year: a 90% absolute reduction in our emissions, including our supply chain, with the remaining 10% removed through accredited schemes. While we have made some material progress on this path in the last few years, a great deal of the early work concerned simply measuring our emissions to an accurate standard. This involves surveying our team to understand how they get to the office; whether their car is combustion, full EV or hybrid; or whether their homeworking is powered by a renewable utilities provider or not; it means noting the consumption of electricity in each office and applying a standardised emissions metric to understand the CO2e produced per kWh consumed; trawling through expenses claims and clarifying the difference between a black cab and a minicab, or taking a view on at what length of journey the emissions value for “light rail and tram” should cede ground to the value for “national rail”. All slightly boring but all thoroughly doable and (an ironic observation if not an original one) all a good deal easier and faster with a decent LLM up your sleeve.

For our supply chain I use algebraic attribution to quantify our emissions. To the uninitiated this perhaps sounds recondite and complex and, as someone who is always eager to sound cleverer than I am, the opportunity to use terms like this brings me no small amount of glee. But in truth it’s just common sense: if Microsoft publish their global emissions figure and their global revenue, it’s a simple job to calculate the share of their revenue made up by our spend with them, and to take onto our carbon ledger a portion of their CO2e emissions proportional to that revenue share. This can be applied wherever the necessary upstream data is available, whether we’re buying media, using SaaS or restocking the office fruit bowl. It also has the added benefit of our vendors’ progress against their own net zero targets contributing to our own reductions: if Google achieve their stated ambition of hitting net zero by 2030 then their emissions reduction can be removed from our supply chain and will make a massive impact on the outcome of our own goal.

However the “if” in that last sentence is doing a lot more heavy lifting after the arrival of AI at scale than it was a few years ago. Alphabet have not yet published their emission figures for 2025 in full, but the 2024 value showed a 51% increase in carbon emissions vs their 2019 base figure. Turning that tanker around in six years seems challenging at best.

Ay, there’s the rub. The advent of AI is making material changes to the way businesses are operating across the supply chain of an agency like ours. Our suppliers’ emissions figures are headed in the wrong direction and consequently there’s a material risk that ours are too. That’s significantly unhelpful, but it is at least a change of direction within an existing model: it can be priced in and roadmaps can be recharted. Or at least they could, if the reporting and governance architecture needed to price that in actually existed. The real risk is that our emissions figures may be increasing to an extent which we have no capacity to measure. By way of an example, Space & Time uses Microsoft products extensively: we all use Office 365, we buy media from Microsoft Ads, our finance package is built on Business Central, we have ditched on-prem servers and embraced SharePoint. And over the last year or so we’ve been adopting Copilot at pace. Currently the only way I can measure the carbon resulting from our use of Copilot is the model identified above: (our Microsoft spend / Microsoft’s revenue) x Microsoft’s emissions. Writing emails in Outlook and building a complex model in Copilot are traded under one licence, but they are not remotely equivalent in terms of their computation needs, carbon consumption or water use. Consequently, attributing an arbitrary share of Microsoft’s carbon to our activities based solely on what they bill us each year becomes less and less helpful: it provides no incentive for me to use AI in a sustainability conscious manner, or not to use it at all, and it makes it impossible to measure the impact of our use of AI or model a roadmap to carbon neutrality. Indeed if data centres increase Microsoft’s emissions footprint such that our carbon emissions figures will be bloated by the global take up of Copliot whether we use it or not, our best intentions would see us picking up a carbon cost without seeing any benefit.

So, it’s clear that the algebraic attribution model doesn’t really work in the new AI-first operational context. What else might help? One extremely useful source of standardised data is DEFRA’s Greenhouse Gas Conversion Factors for Company Reporting, which are commonly used to inform UK companies’ reporting under the Streamlined Energy and Carbon Reporting framework. Predictably though these do not yet make reference to Artificial Intelligence, and are a slow-moving beast, produced only once per year and often unhelpfully behind the curve. For example, their measurement of the fuel used in transporting fuel to the user is a one area of continual development and by no means a settled truth (a shifting dynamic between “combustion only” measurement and “well to tank” metrics), and in 2025, five years after the pandemic, DEFRA made a huge change in the assumptions about how occupied planes are post-COVID, delivering reported emissions reductions of c.40% per passenger mile for some flight categories seemingly overnight. This is a complex and emerging science and to that extent incremental improvements should to a great extent be considered a feature not a bug, but in a global context where generative AI is becoming infrastructure just a few years after mass adoption began, annual updates and a multi-year lag in reporting indices simply aren’t sufficient. The truth is that the light-touch and quickly outdated requirements of SECR and the accompanying GHG indices are entrenching a model that is biased, unhelpful to a truly sustainable agenda and to an increasing extent miss the point: SECR incentivises optimisation against that limited set of variables which can be measured, while disincentivising businesses to consider water usage, land usage and the carbon that can’t be measured. The UNU estimates that by 2030 the water footprint of AI will match the basic water needs of 1.3 billion people, but meanwhile I’m required to worry over whether Uber journeys that people our team expense are more likely to have been in a “medium car” (0.2639kg CO2e per mile) or a “large car” (0.33808 kgCO2e per mile).

The challenge here is one of scale, and perhaps of a new politics: algebraic attribution functions well enough for SaaS or cloud computing. There is very definitely a carbon footprint to those products but it is modern and light touch, it’s an airy, ethereal phenomenon that happens “in the cloud” and it’s run by companies in California that “do the right thing” and have laudable ambitions. Their intentions, their transparency and the quantum of their emissions are pleasing, measurable, finite. By contrast AI may feel to the end user like SaaS, but in carbon terms it’s a heavy industry with an earthy physicality: stovepipe hats, iron bridges, smog and six billion gallons of water. And through a reassertion of a more overtly corporate agenda, through a more hostile geopolitics or perhaps just a perception that being nice is less commercially valuable than it used to be, those environmentally friendly tech companies are starting to look as though they are run by avaricious industrialists.

As consumers, agencies like ours might be tempted to think we’re buying “licences” and “cloud” from well-intentioned green companies, and we have sufficient tools to calculate our carbon footprint on that basis, but in practice we’re a low-carbon service industry which risks becoming addicted to leveraging heavy industry to deliver its capabilities: data centres already produce 0.5% of global combustion emissions and by 2030 are expected to require more electricity than all of Japan. AI is the driver of this change, and two years ago we didn’t need it at all.

I don’t really have any answers to offer and this piece is, on reflection, just a depressing list of problems. One piece of received wisdom is that AI will itself find the solution to global warming, so it’ll all be fine anyway. That would be nice, but I suspect we’d be foolish to bet on it. Similarly many people in my position have pointed to the benefits that AI can bring to measuring carbon; great but so what? There’s also the suggestion that the deleterious environmental impact of using AI will be offset by the improved efficiency of a workforce. I can see the logic of this to some extent: we have no plans at all to replace any staff with AI agents, but I can envisage a future scenario where fewer growth hires are made on account of the increased productivity of the current team, and our carbon emissions see a commensurate benefit in reduced commuting miles, smaller offices or lower air conditioning needs per £m of revenue. But will this be of a sufficient quantum to offset our upstream carbon emissions? Firstly, no I’m not convinced it will; secondly, crucially, there’s currently no means of finding out; and thing the third, is reducing the number of people on a train of any real benefit in an age when a paucity of water or land is the more pressing issue?

One hope we might have is that there will be a regression back towards the mean as the market catches up with this sea change. What I mean by that is that to a great extent the failure of the current measurement model lies in the disconnect between the emissions that LLMs produce and the financial, marginal cost to use them. If their cost to a business was more proportionate to their upstream emissions and to their value, realised or potential, to the business, then algebraic attribution might well be more fit for purpose. Our current location is at the bottom of a curve: a paradigm shift wherein the main innovators are still proving the model and aren’t yet making any money. Reports suggest that OpenAI lost nearly $39bn in 2025 (this includes one-off accounting changes; operating losses are suggested to be less, at $20.92, but the point stands). These AI companies all have massive start-up and R&D costs that must be recovered, and in-year operating losses that won’t be sustainable for long. The overall cost of AI to the end user will inevitably increase, and with increased costs will come greater measurement transparency on token usage, and a more direct relationship between spend and emissions. Potentially this increase in cost for generative AI will also see a retrenchment in its use or at least a flattening in the curve that in turn delivers a reduction in the more rampantly depressing predictions about where we might all be in five years’ time.

In the meantime at Space & Time we’re doing what we can: we will measure what can be measured, both in terms of carbon emissions and in terms of token usage. We have an AI policy which encourages colleagues to use the tool mindfully; to consider the environmental costs of their usage. It offers advice on lower-token prompt engineering, or manually looking for the last time you asked a similar question instead of just asking it again.

But for now all of this risks being so much greenwashing. We (all of us) are trying to manage industrial-scale environmental impact with a measurement framework better suited to software. Until there’s a more robust, more transparent model available, any apparent precision in carbon reporting is likely to be false comfort.

I’ve been at pains to avoid having AI actually write any of this piece for me. We’ve all read enough of those blogs already: “can LLMs really fool people into thinking its content was created by human? You tell me, I had GPT write this for me” etc etc, the ol’ bait and switch, lovely job. We thought it was a person, it was a robot all along, Turing would be amazed. (We thought he was a schoolboy but actually he was the Headteacher! Oh my sides).

But by way of transparency though, did I discuss this with an LLM and have it check my assumptions, search for sources and factoids? Absolutely I did, and our green bonafides notwithstanding I contend that I would be a fool not to have done so: this is a costly tool but of course also an extremely useful one.

Interestingly Copilot seems to have been concerned that I was too negative about AI, and has suggested I end on a more positive note, to wit: “This is not an argument against AI adoption. It is an argument against pretending that a tool with industrial inputs can be governed using software-era accounting.” (… and all watched over by machines of loving grace).