I get asked this rhetorical question a lot, in various forms:
Sure, datacenters might use a lot of energy, but you don’t have to use a hosted frontier model to do software development. What if I just run a local open-weights model to do some coding, with an open-source coding agent? Video games also use my GPU. Is local model development any worse than playing a video game?
So I want to write down my comprehensive answer to this: Yes, using an LLM to write some code is worse than playing a video game, for a few reasons.
Video Games Are Interactive, LLMs Are Batch Jobs
Video games use compute to respond to human input. You are using your GPU while you are looking at a screen, displaying an image. When you are done playing, you shut off the game, and your computer goes back to idle. It’s much less energy. By contrast, agentic loops with evals (the only kind of “AI” that is meaningfully any good at coding) are running hot, for days. To use the most recent example of such a thing, a very rough first sketch of an implementation of a Windows graphics API backend to help port a paint program to other platforms, it took 3 weeks of Claude time, “day and night”. Do you play a lot of video games for 500 hours to make it past the tutorial level, while also using other computers for other things, as well as the rest of your carbon footprint?
Video Games Need Development, LLMs Need Training
Video games use compute to respond to human input during development, too. Your game has to be made, but your LLM has to be trained. LLMs use a historically extreme amount of power, probably using more than the entire Internet, but it’s kind of hard to say. Still, it seems a reasonable estimate to within several orders of magnitude that even over a multi-year project with hundreds of developers, the power used to develop an individual video game is nowhere close to training even a small LLM.
This is true even for local models. OpenAI has openly claimed that DeepSeek “stole its intellectual property”, and I have heard grumblings that none of the open-weights generalist models could realistically exist without the massive lift that the frontier labs are doing with their training, in various other ways too. Secrecy throughout the industry makes this kind of impossible to understand rigorously, but it seems fair to say that you are partially culpable for all that famously energy-intensive frontier lab training if you’re using a local model.
And They Keep Needing Training
You also can’t dismiss this as a sunk cost, because in order to stay current with industry developments, models need to be updated with new information from the rest of the world, which means that you need to keep training them. Beyond the energy for your own use, if you want a real-life agentic workflow that actually does useful stuff, practically speaking you would still need to update your local models over and over again, at least once every few months, which means you would be incentivizing continued energy consumption by whoever was doing that training for you, including the energy cost of scraping.
Let’s Be Real Here, You Aren’t Actually Using A Local Model
This question is a hypothetical thought experiment. Despite synthetic benchmarks that keep showing there isn’t much difference between open weight and frontier models, nobody’s actually using local models for much of anything beyond sharing those talking points. Depending on which benchmark you’re looking at, maybe it’s good enough or maybe it’s worse.
As an inveterate AI hater, all these systems seem pretty bad to me, but it seems that people who find them useful tend to subjectively believe the frontier models are worth the premium, and that’s what they’re actually using. Once you have accepted that it is OK to use LLMs for coding at all, it seems like a very quick slippery slope on down to “we’ll go ahead and use the frontier models for now anyway, but we could be ethically better in the future by switching to an open weights one, that option is always available”.
There’s A Reason We Have Data Centers
Devolving power usage to local LLMs might be good to make users responsible for their costs and decrease the impacts to communities that are physically next to huge concentrations of power utilization, not to mention generation. However, there’s a reason that it makes sense for the providers to build these giant facilities: economies of scale reduce total power consumption, they don’t increase it. If you do all the same stuff with a local model that they have to do in hosted environments, it will probably take more power, even though you will be incentivized to do different stuff. This incentive to “do different stuff” is why although local models can hypothetically hold their own against the frontier labs for some tasks, when people or businesses take their inference costs in-house they often find that it’s too painful and move back to hosted LLMs.
There Are Problems Other Than Power
These are subjects for a different post, but you have to consider a lot of other externalities: AI psychosis, de-skilling, comprehension debt, cultivating a dependency, introducing security defects, limiting your design space based on what LLMs can understand, context rot, wasting time on invalid solutions, introducing unpredictability into your workflows. You still have to consider the total cost benefit ratio.
To Sum Up
Local LLMs might alleviate some of the harms from using the hosted frontier providers. There are fewer privacy concerns, you can measure your power utilization and be more directly responsible for it, you can build interfaces with affordances that are less oriented towards addiction and dependency than the major frontier labs’ harnesses.
But they’re not automatically “the same as playing a video game” just because they can use the same GPU.
Acknowledgments
Thank you to my patrons who are supporting my writing on this blog. If you like what you’ve read here and you’d like to read more of it, or you’d like to support my various open-source endeavors, you can support my work as a sponsor!