Let me start with something awkward. Training a large AI model emits a lot of carbon. Running it emits more. The data centres that host them draw from grids that are, in most places, still not fully clean. If you want a tidy story about AI being a climate hero, you have to grapple with this first.
Now the good news: AI is also enabling climate work that would otherwise be impossible. The net balance depends on what we build and how we power it. That's the story I want to tell.
Where AI is genuinely helping the climate
Smarter grids. Electrical grids are increasingly hard to run. Intermittent renewables (wind, solar), distributed generation (rooftop solar, EVs feeding back), and changing demand patterns have made grid operation dramatically more complex than it was twenty years ago. AI forecasting — demand, supply, weather — is keeping the lights on while letting us integrate more renewables. This is unsexy. It's enormously important.
Materials discovery for clean tech. Better battery electrolytes, more efficient solar cells, cheaper catalysts for green hydrogen. These all depend on finding new materials in a design space that's too large to explore experimentally. AI-assisted materials discovery (like DeepMind's GNoME work) is accelerating this dramatically.
Optimising industrial processes. We looked at this in the manufacturing post. A 10% reduction in energy use at a medium-sized industrial facility is more carbon saved than a thousand individual lifestyle choices. AI-driven optimisation at industrial scale is where most of the climate impact of AI will come from over the next decade.
Monitoring deforestation and emissions. AI on satellite imagery can spot illegal logging in near-real-time, track methane plumes from specific facilities, and hold polluters accountable at a level of granularity that wasn't possible before. This feeds into policy and enforcement.
Climate modelling. Weather and climate models benefit enormously from AI — both as a replacement for parts of traditional physics-based models (faster and sometimes more accurate), and as a way to produce many more simulations to understand uncertainty. Google DeepMind's GraphCast and similar systems are genuinely important for operational weather forecasting.
Agricultural yield. We talked about this in the agriculture post. More food per hectare, less water, less fertiliser. The climate relevance is enormous — agriculture is a major source of emissions and a major driver of land-use change.
Transportation efficiency. Routing algorithms that cut delivery miles. Traffic management that reduces idling. Right-sizing of fleets. None of these feel like climate tech. They're moving serious carbon.
The shadows
Training costs. The big foundation models cost millions of dollars and thousands of tonnes of CO₂ to train. This is real and significant. Over time, these numbers are improving (more efficient architectures, more renewable-powered data centres), but anyone claiming "AI is carbon-negative" today is not being honest with you.
Inference at scale. Running a model once is cheap. Running it a billion times a day is not. As AI becomes part of everyday applications, the aggregate energy footprint matters.
Rebound effects. Sometimes efficiency gains get eaten up by increased use. A factory that's 10% more energy-efficient may choose to produce 20% more output. This isn't unique to AI, but it applies here too.
Greenwashing. Some companies are using AI-and-sustainability branding to distract from their actual impact. A fossil fuel company running an "AI for carbon capture" pilot while continuing to expand extraction is a net negative, no matter how clever the pilot.
What we can choose
The climate impact of AI isn't fixed. It's being decided now, by choices that individuals and organisations are making:
- Power AI with clean energy. The big cloud providers are making progress here, but faster is better. Choose providers that are transparent about their grid mix. Pressure them for more.
- Use efficient models where possible. Not every task needs a frontier model. A smaller, well-chosen model often does 90% of the work at 1% of the cost, and 1% of the emissions.
- Measure. You can't manage what you don't measure. Track the carbon cost of your AI workloads.
- Offset honestly. Credits from verified projects, not greenwashing scams.
- Focus on climate-relevant applications. Some AI work directly helps the climate. Some is neutral. Some is making things worse. Know which kind yours is.
Something I think about a lot
We're going to use AI for climate work because we have to. The scale of the problem is beyond what humans can address with spreadsheets and intuition. Every tonne of carbon avoided, every species preserved, every coastal city protected — the tools we deploy matter.
But we have to be honest about what those tools cost, and careful about how we use them. The best climate AI practitioners I know are the ones who track their own footprint, choose their projects carefully, and refuse to help fossil fuel companies extend their extraction curves.
If you're working on something in this space — we'd genuinely love to help. Book a chat. Climate projects are some we take on at reduced rates when the impact case is clear.
