The CO₂ you never knew

When AI saves more than time

My first post in this substack looked at the future of software development. One question I didn’t address back then, but which I’ve subsequently been asked several times, is what was the CO₂ impact of using the AI.

It’s an interesting question. Recently, there have been many stories of AI companies getting involved in nuclear. Microsoft has signed a 20-year deal to get nuclear power from the Three Mile Island plant. Sam Altman is on the board of Oklo, a nuclear startup. AI clearly needs power. Lots of power.

But how much CO₂ did I actually generate?

I decided to compare the CO₂ that the AI generated against the CO₂ I would have produced doing the same work. Who would win?

Me

In three days I did what would have taken me 30-60 days in the old world. Let's be conservative and say just 10x faster. That’s 27 days quicker.

Humans produce, roughly, 1kg of CO₂ per day. So I’ve, err, saved 27kg of CO₂.

Of course it’s more complex than that. I don’t work on the weekend - and I take vacations. Correcting for those the 27 days turns into 42 days. 42kg of CO₂ then.

o1 and Claude

Working out how much CO₂ each LLM query generates is hard. But the general consensus seems to be somewhere between 2 and 10 grams of CO₂ per query. I ran ~200 queries. Of those half were o1 queries, which are ~10x more expensive. So that's 100x10 + 100 = 1100 queries. At 10g per query that's 11kg.

And the result?

On the face of it, I saved 31kg of CO₂. Really? Well…

The first problem is it wouldn’t take a big error to change the calculation. If a query produced 100g of CO₂ then the AI produces 70kg more than me. There are error bars.

On the other hand AI efficiency will continue to improve (it’s a key reason why the per-token costs of LLMs have dropped by ~80% over the past year). Sadly I’m stuck producing 1kg of CO₂ per day - and am unlikely to get significantly faster at coding.

While interesting, this simple calculation barely scratches the surface. There are many other factors in play. Nearly all of them make us humans look worse.

Take commuting. Back in my office days that was easily another 2-3kg of CO₂ per day. Even working hybrid that's still an extra kg per day. Then there's heating and cooling the office space. Making coffee. Buying lunch. All the equipment needed to make an office function. The list goes on.

But the real impact is harder to quantify. AI is accelerating human progress. We can do more in our finite lifetimes. Increasingly more.

Consider my parents’ generation. They spent decades writing assembler by hand. One painful line after another. It was slow. Very slow. Looking back it feels almost comical - like watching someone build a house with just a hammer and chisel.

Now we have power tools. We can build faster. But it's not just about speed. When you can iterate quickly you can try more approaches. Find better solutions. Instead of being constrained by time, we're now constrained by imagination.

And this compounds. Every improvement we make becomes a foundation for the next generation of tools. Each advancement accelerates the next. We're not just saving time - we're bending the curve of progress.

Take energy efficiency. In the past, optimizing code for energy usage was often an afterthought. Who had time? But now we can explore multiple approaches, measure their impact, and iterate until we find the most efficient solution. The CO₂ spent on AI development today could save orders of magnitude more in the future.

The irony isn't lost on me. AI needs enormous amounts of power to run. So much that tech companies are turning to nuclear power. But it's using that power to help us solve problems faster. Including, potentially, the problem of its own energy consumption.

In the end, measuring AI's environmental impact by counting query emissions is like measuring the industrial revolution's impact by counting lumps of coal. Interesting, but misguided. It misses the bigger picture. We're not just doing the same things faster - we're changing what's possible.

These days I spend most mornings walking and thinking while chatting with AI. Each conversation generates a few grams of CO₂. But those conversations have changed how I think about software, about work, about progress itself. I’m learning and developing faster than I ever have. That’s got to be worth more a few grams of CO₂.