The discussion around AI and climate usually boils down to its consumption, which we documented in a dedicated article. That's half the story. The other half, less covered, is what these tools help avoid. Neither half is enough to draw a conclusion, and claiming otherwise is a matter of conviction rather than calculation.
We've covered the cost extensively. Let's look honestly at the other column, without turning it into a pitch.
What AI actually optimises
Power grids. This is probably the most significant lever. A grid with a lot of wind and solar has to constantly balance variable generation against variable demand. Better forecasting of output, anticipating peaks, managing storage: these are prediction problems where machine learning excels. Every point gained on balancing reduces reliance on backup plants, which are the most polluting.
Agriculture. We detailed this in our article on AI in the field: targeted weeding sharply cuts inputs, and modulating nitrogen application reduces particularly potent emissions.
Materials. Searching for new materials for batteries, storage, or low-carbon cement is a problem of exploring an immense space. It's one of the areas where AI has produced tangible scientific results.
Buildings and logistics. Optimising heating, delivery routes, truck loading. Modest percentage gains, applied to enormous volumes.
Three reasons make the calculation impossible today. The rebound effect: increased efficiency often drives costs down, which boosts usage and can cancel out the gain. Attribution: when a grid cuts its emissions, how much of that is down to AI rather than renewable investments? Scope: counting a data center's consumption is doable; counting the emissions avoided by a model deployed across a thousand uses is not. Anyone who gives you a quantified balance is selling you a conviction.
The asymmetry that should guide us
There is, however, one thing we can say with certainty, and it's useful.
Consumption is certain, immediate, and concentrated. The benefit is probable, deferred, and diffuse. This asymmetry doesn't mean the benefit is illusory, but it does impose a requirement: you can't justify a certain expense with a hypothetical gain without measuring it.
In practice, this calls for distinguishing between uses. A model that optimises a power grid has a measurable, verifiable benefit. A model that generates variations of ad images has a real cost and zero climate benefit. Both consume. Treating them as equivalent in the debate makes no sense.
What could improve the balance
Three levers, two of which are already at work.
Model efficiency. Progress is real and fast: compression, expert architectures, and execution optimisations that have driven an 80% price drop mechanically reduce energy per request.
Where you host. A facility powered by nuclear or hydro doesn't have the same footprint as one on a carbon-heavy grid.
Frugality of use. This is the least discussed lever and perhaps the most accessible. An oversized model for a simple task, a massive context sent without caching, an agent idling in a loop: most of today's waste is software-level.
What to take away
The question isn't whether AI is good or bad for the climate. It's what we make it do, and at what cost.
A technology that consumes a lot and enables significant gains is acceptable if the gains are real and measured. It isn't if most of its consumption goes toward producing content nobody needed. This isn't a technological question; it's a question of allocation, and it plays out at the level of collective choices far more than at the level of your individual requests.