Is AI a climate problem? How we think about it at GoClimate

As a climate company, we want to be clear about how we think about the climate impact of AI.

We have extra reason to care. We automate sustainability reporting by reading invoices and receipts, and making sense of large amounts of unstructured data is exactly where AI comes in.

Why it matters to us

If you use AI without thinking about which models you call and what for, it turns out to be a surprisingly carbon-heavy technology. As a climate company, we can't build AI into our product without keeping track of that trade-off.

We choose models that use little energy and run them on servers powered by green electricity. Sometimes we still end up on non-green electricity, when the servers are overloaded. One of the models we use is Gemini, where efficiency has improved quickly: according to Google, the energy per median text prompt in Gemini Apps fell 33-fold and the carbon footprint 44-fold between May 2024 and May 2025 [2][3]. We benefit from that improvement, even though it doesn't translate directly to our own use.

Our AI use in the tool comes to roughly 11.4 kg CO₂e per month [4]. Using Google's own figure of 0.03 g CO₂e per median Gemini prompt, that's about 12,500 AI prompts a day, every day, for a whole month [3]. But a comparison like that is a snapshot, not a universal rule, because there's no such thing as a standard "AI prompt". Different models are heavier or lighter and run on different infrastructure and electricity mixes. A small, fast model and a heavy reasoning model can differ in emissions by tens of times, and in some measurements by close to a hundred [5][6]. Lumping them into one emission factor makes the figure somewhat misleading, whichever way it leans.

At the same time, we help thousands of companies cut their emissions. At our current level of AI use, it would take more than 7 million years to emit as much [7]. To us, that's a calculation that adds up.

Want to see how it works in practice? Try GoClimate's sustainability reporting tool

How we choose tools and models

We have a clear strategy for what we use for what. We match the model to how complex the problem is, and only use the heaviest models when we really need them. Much of the foundation is also still rule-based logic that we built before AI was even around. We refine it with AI instead of rebuilding from scratch, and that affects how big a climate impact our AI use has.

The scale effect: our climate benefit weighs more

AI's big climate question is probably not about the single prompt but about scale. When billions of people use the technology every day, small footprints add up to big ones. But scale works both ways. A single company that gets its emissions under control can cut them by more than our AI emits over a very long time. The more people use GoClimate, the further the calculation tips in the right direction.

Every company that measures its emissions helps tip the scales. Create an account and see your company's carbon footprint.

We set requirements for our suppliers

We ask our AI suppliers to report their carbon footprint, and their answers count when we choose who to work with. It's the same thing we help our customers do: know what things actually cost in emissions, and choose based on that.

Want to know where your company's emissions come from? GoClimate calculates them from your invoices, so you can see which purchases weigh the most. Book a demo and we'll show you.

So, is AI a climate problem?

Yes, if it's used carelessly and at scale. But AI is a tool, and its climate benefit depends on what you use it for. We use it to help companies measure and reduce their emissions. We measure our own footprint, choose the model to fit the task and report the figures openly. That's how we think AI should be used.

What's your company's carbon footprint?

GoClimate reads your invoices and receipts and calculates your carbon footprint automatically. No manual calculations, no guesswork. Create an account to get started, or book a demo and we'll show you how it works.