US carbon emissions went up in 2025, bucking expectations as emissions grew faster than GDP. This is unexpected, in general we expect economies to get more productive, and therefore lower their emissions per $ of GDP growth. There are a number of culprits, but many commentators are quick to point out the growing demand of data-centers and their power-hungry AI chips.
It's good that we're questioning a new industry's emissions. One of the important questions to ask is: is it worth it?
As a sustainability company using AI to lower emissions, the contradiction isn't lost on us. However, we don't see this as a paradox for two key reasons:
- If AI can enable decarbonisation at a faster rate than its own emissions, and faster than comparable tools, then it is okay to use it.
- AI's emissions are set by the energy system that's powering it. The challenge is not AI, but building more green energy.
Decarbonising faster because of AI
AI use is typically measured in tokens. For example, this paragraph is 68 tokens long. It's difficult to get an exact CO2 per-token, but based on the insights from recent research and our real-world token-use, we would estimate our AI impact as 60g CO2e each time we change someone's behaviour.
However, we have to look at the return on CO2e, rather than the raw CO2e.
We frequently see this misunderstood in conversations about EVs and solar panels. Both of which have high emissions to produce, but lifetime emissions that end up with huge net carbon savings (typically after ~2 years of use, an EV's emissions will be lower than that of a combustion engine's vehicle).
When it comes to our new EngageAI product – which sends interventions to lower cost, carbon and friction – our net carbon saving depends on the type of intervention and outcome:
- A 20km taxi ride moved to public transport is a 3.5kg carbon saving (3,500g), or put another way, a 58x return on CO2e.
- An aircraft switch, from an old aircraft to a newer aircraft between London and New York, is likely to be a 200kg CO2e saving (200,000g), or a 3,333x return on CO2e.
- A consolidated trip, replacing the need for two trips, will save 1.5t CO2e (1,500,000g), or a 25,000x CO2e return.
So yes, there is a CO2e cost everytime we use AI, but the emission is worth it, because overall it saves more CO2e than it creates. Imagine taking a proposal to your manager to purchase a product that will have a 58x return on saving – let alone a 25,000x return – you would be celebrated as employee of the year!
We must decarbonise our energy systems
Decarbonising electricity is much, much easier than many other climate problems. Want to decarbonise an aircraft? You will need to synthetically make oil, from a sustainable feedstock, at huge scale, without massive unintended eco-system consequences.
That's not to trivialize decarbonising electricity. There are plenty of challenges, but fundamentally we need to increase our investment in clean energy sources. If we do this at the same rate as investment in AI datacenters, we will net out AI's footprint to zero.
What is frustrating about the clean energy discussion is that it has been politicised. Solar and wind have been cheaper than all other forms of energy for a few years now. Yes they have challenges because wind can stop blowing and sun can stop shining (of which there are a number of solutions that are the subject of another post), but the UK has shown you can have 70% of your energy mix be clean while maintaining a resilient energy grid.
When we see complaints about AI, it often feels like a redirected ire, away from the fossil fuels that are ultimately the cause of AI's emissions. That is where we must focus if we want AI's impact to be reduced.
A better way to talk about AI + carbon
Many talk about carbon emissions as a zero sum game: you either have an emission or no emission. But there's a reason we talk about net zero emissions, and not just zero emissions.
The net of course comes from financial accounting, where you subtract expenses from revenue to understand your profit. If you make $5m, but have to spend $10m, your net is -$5m; if you make $1m, but spend $100k, your net is $900k. The second position is far more profitable, despite the lower revenue.
We must get used to the same accounting for AI and carbon emissions. If your use of AI lets you do twice as much, at a lower emission than doubling headcount, that is a positive.
If, like Thrust Carbon, your AI emission enables a 25,000x carbon return, then we should celebrate!