The Quiet Cost of AI Images: Why Energy Use Is Becoming a Selection Criterion

There's a lot of talk about AI images – about image quality, about copyright, and lately about labeling requirements. There's noticeably less talk about the electricity behind them. Yet the number is striking: a single AI-generated image request consumes, by current estimates, roughly a hundred times the energy of a plain text query to a language model. Global energy demand for AI training and operation doubled in 2025 compared to the year before, now exceeding 150 terawatt-hours annually – more than countries like the Netherlands or Argentina consume in total.

For a studio like ours that offers CGI, photography, and generative AI side by side, that's not an abstract climate debate. It's a question that should directly feed into the choice of technique – alongside cost, time, and image quality.

Why images specifically cost so much energy

The difference from a text query lies in the computational effort per result. A language model generates character by character; an image model has to compute an entire pixel grid out of noise across multiple passes – for every single attempt, for every variant that gets discarded before one fits. That very iteration, which makes generative image AI so useful – quickly running through ten variants until one convinces – is also the reason for the high consumption. Every discarded variant carries the same energy cost as the one that ends up being used.

The point usually missing from calculations

Energy consumption is almost never factored into quotes today, because it doesn't show up as a direct cost to the end client. That's foreseeably going to change: larger companies are increasingly reporting on the carbon footprint of their supply chain, and digital services – content production included – are increasingly part of that. A studio with no answer to that loses credibility at exactly the point where it otherwise builds its reputation on quality and creativity.

That doesn't mean avoiding generative AI on principle. It means treating energy consumption as one of several variables – alongside time, cost, and image quality – when deciding how an image gets made.

Where classic CGI is the lower-energy choice

Here's an aspect that almost never comes up in the AI-image discussion: classic, hand-modeled CGI – not a generative model, but built 3D with classic rendering – has an entirely different energy path. The heavy cost sits in modeling and rendering the scene once; after that, the same scene can be reused for years with new products, new lighting, new perspectives, without a model having to run through a data center again for every variant. That exact difference between one-time effort and per-image model inference adds a third dimension to the comparison between CGI and photography that's been missing so far: not just cost and control, but energy per reused image.

For catalogs with hundreds of variants – precisely the case where CGI is already structurally superior – that means the ecological advantage reinforces an economic one that already exists.

Where generative AI still remains the right choice

For quick variant creation, concept drafts before the actual shoot, or retouching tasks – the use cases I described in AI-assisted content production – the time saved often clearly outweighs the energy cost, especially when only one or two images actually end up being used instead of hundreds of variants. The calculation isn't a blanket "CGI good, AI bad" – it depends, just as with cost and control, on repetition and the number of variants needed.

What a studio can concretely do about it

Three things can be implemented today without waiting for new standards. First, deliberately cap the number of generated variants per job instead of iterating endlessly just because it's technically possible – that saves energy and usually time too. Second, for foreseeably recurring visual worlds (catalogs, campaigns running for years), systematically check whether classic CGI is both the lower-energy AND the economically better choice. Third, include energy consumption as its own line item in quotes and sustainability reports before clients ask – that's cheaper than having no answer later.

Conclusion

The energy consumption of AI images is currently a footnote in a debate mostly about quality, law, and labeling. For studios offering CGI and generative AI side by side, though, it's a third, so far underrated selection criterion – one that reinforces the already-existing economic advantage of reusable CGI, without calling generative AI into question where it genuinely saves time.