The carbon cost everyone is ignoring in AI

I have been talking with people who build in cloud, data centers and hardware, and the same pattern keeps coming up: the sustainability conversation is missing from product design. We have standards for physical products, such as LCAs, energy ratings and certifications. Software, and AI in particular, is still working out how to account for itself.

The efficiency plateau

Data center power usage effectiveness (PUE) fell quickly, from about 2.5 in 2007 to around 1.65 by 2013. Since then it has been roughly flat, near 1.58, according to the Uptime Institute's annual surveys. The easy wins are done.

Demand has not stopped. The IEA estimated data centers used about 460 TWh in 2022, close to 2% of global electricity, with AI a large part of the growth to come.

What we talk about, and what we should

Typical prompt engineering advice is about cost: structure your prompts, shrink your context, save money per day. That is good advice. But the same optimization, multiplied across 10,000 tokens per request and 100,000 daily users, is also an energy and emissions decision.

What teams usually track:

• Token count
• Response quality
• Latency
• API spend

What we rarely measure:

• Carbon per query
• Energy per output
• Infrastructure load
• Grid impact

Building differently

This is not about stopping innovation. It is about climate-aware engineering from day one.

• Add energy impact to acceptance criteria.
• Set carbon budgets next to performance budgets.
• Treat prompt optimization as a sustainability practice, not only a cost one.
• Pick cleaner regions, schedule for renewable peaks, cache aggressively.
• Design graceful degradation for lower-power modes.

We are embedding computation in everything, from sensors to agents. Efficiency gains are real, but they are not keeping pace with demand. The answer is not to stop building. It is to build with the footprint in view. I tried to make that concrete in measuring software carbon intensity.