Every company is racing to add AI to its products. Very few of them can tell you what that AI costs the planet. Ask how much carbon a single AI feature emits and you will get a shrug, a vague gesture at a data center, or a number that falls apart the moment you check it. The fastest-growing part of the world’s computing footprint is also the worst measured, and that gap is turning into a business problem.

The headline numbers about AI and energy are not really the news, because they are not in doubt. We know data centers are drawing more power every year, and we know AI is the reason. What almost no organization can do is the thing that actually matters to it: take that global trend and put a defensible figure on the carbon its own AI is responsible for. That is the blind spot, and it gets harder to live with as the numbers grow and the rules tighten.

The scale is no longer deniable

The growth is steep enough that it has stopped being a niche concern. Data centers used around 415 terawatt-hours of electricity in 2024, about 1.5% of global supply, and the IEA expects that to more than double to roughly 945 TWh by 2030, approaching 3% of all the electricity used on the planet (IEA, 2025). AI is the engine driving it: the GPU-heavy “accelerated servers” that run AI are growing at around 30% a year and account for close to half of that entire increase, and in advanced economies, where most of the new building is happening, data centers are on course to cause more than a fifth of all growth in electricity demand this decade (IEA, 2025).

It is tempting to assume the hyperscalers (Microsoft, Google and Amazon) have this under control, but their own reports say otherwise. Microsoft’s emissions are up 23.4% since 2020, almost entirely in the indirect, supply-chain emissions that now make up 97% of its footprint, even as it cut the emissions it directly controls by nearly a third (Microsoft, 2025). Google’s total is up 51% since 2019 (Google, 2025). Google’s report does contain a real success, with data center emissions falling 12% in 2024 even as the electricity those centers drew rose 27%, but it took some of the largest clean-energy contracts in corporate history to bend that one curve for one year, and the total still climbed. If the companies with the most money and the strongest reason to get this right cannot hold their emissions flat, the comfortable idea that AI’s carbon cost is being quietly handled upstream does not survive contact with the evidence.

The part you own is the part you can’t see

Most of the attention goes to the wrong side of the problem. Training a model is a single, energy-hungry event, and the carbon it produces belongs to the lab that built the model. Inference, which means actually running the model every time someone sends it a query, is the part that repeats, around the clock, for as long as the feature is live. Industry estimates put it at 80 to 90% of the compute a deployed AI system uses over its lifetime (Brookings, 2025), and that share belongs to you: the company that embedded the feature, with a footprint that grows in step with how well the feature does.

That is the carbon you are most directly responsible for, and for almost everyone it is invisible. The clearest sign of how invisible is that nobody can agree what a single query costs. For years the standard figure was about 3 watt-hours, roughly ten times a web search; more recent analysis revises a typical query down to around 0.3 watt-hours, about the same as a search, while long prompts and reasoning-heavy models climb into the tens of watt-hours (Epoch AI, 2025). The published range runs from a tenth of a watt-hour to forty, a several-hundredfold spread for the most common operation in modern computing. That is what happens when almost nobody measures the real thing.

The providers could close much of that gap, and mostly choose not to. Mistral, the French developer behind the Le Chat assistant and a family of widely used open models, is the welcome exception: in 2025 it published an audited life-cycle assessment of its Large 2 model with the consultancy Carbone 4 and the French environmental agency ADEME, putting a typical short response at about 1.14 grams of CO₂e (Mistral AI, 2025). One disclosure, from one provider, for one model, is not transparency, but it proves transparency is possible, and it leaves the rest of the field looking conspicuously quiet.

And it isn’t only carbon

Carbon is the headline, not the whole bill. The same Mistral assessment put the water behind a single short response at around 45 milliliters, and the cost of training the model at roughly 281,000 cubic meters of water and 660 kilograms of mineral depletion. Data center water use is barely counted in the first place (nobody really knows how much UK data centers consume), and behind all of it sits the embodied footprint of the hardware: the emissions and scarce minerals locked into GPUs that are replaced every few years, and the e-waste they leave behind. A measure that counts only operational carbon understates what AI really costs.

Why it lands on your desk now

None of this stays in the engineering department. Cloud and AI are Scope 3 emissions, and Scope 3 reporting is moving from voluntary to mandatory under rules such as the EU’s Corporate Sustainability Reporting Directive; auditors increasingly want figures they can check, and “we don’t know” will not hold for much longer. There is reputational exposure too, because the distance between “we are an AI company” and “we have no idea what our AI emits” is exactly where accusations of greenwashing take root. And there is plain cost, because energy that goes unmeasured goes unoptimized: nobody picks the cleaner region, right-sizes the model, or shifts heavy work to when the grid is greener, because nobody can see what it would save.

Measurement first

The good news is that this is fixable, and it starts with measurement rather than targets or offsets. A toolkit is emerging: open-source projects like EcoLogits, from the non-profit GenAI Impact, estimate the energy and embodied impact of individual API calls (GenAI Impact, 2025), CodeCarbon tracks training and compute jobs, and Hugging Face’s AI Energy Score rates models for efficiency. These are real progress, but they are mostly estimates, modeled from public assumptions rather than measured against the grid that actually served the request.

That distinction is where Tailpipe sits. Tailpipe measures emissions across cloud (AWS, Azure and GCP) and AI workloads, and it measures AI inference through a lightweight proxy, so you can see the carbon of your LLM calls across OpenAI, Anthropic, Gemini and Mistral by routing your requests through Tailpipe. The figure is real-time and location-based, tied to the specific grid behind each request and aligned to the GHG Protocol and the Software Carbon Intensity standard (ISO 21031) and its emerging SCI for AI extension (Green Software Foundation, 2024), rather than an estimate from a generic average. It is private by design, measuring a request without ever capturing the prompt or the response. Once that figure exists, the reductions follow: right-sizing models, moving inference to cleaner grids, and cutting the waste that was invisible before.

Everyone is shipping AI. The companies that come through it well will be the ones that can say, credibly and to an auditor, what theirs costs. The blind spot is going to close one way or another, and it is cheaper to close it deliberately than to explain it later.

If you want to see what measured AI emissions look like for your own workloads, get in touch.

Sources