I bought a datacenter GPU that doesn't fit in a normal motherboard, macgyvered the fan with jumper wires, and now I'm running a model that ties with Claude Sonnet 4.6 on benchmarks, all for £200.
Power consumption for AI is way higher than for games…your gpu is rarely at full usage when running a game, only really if you’re one of the few people that tries to get every single bit out of your game with the highest level graphics and even then you’re not using the full amount because the game would stutter. It doesn’t matter if an llm stutters. They use the full power of your gpu no matter what you’re doing.
And the models being trained on other models mean that local LLMs use more energy than a model like Claude. You have all the energy of multiple data centers training one model and then you go and train a model off of that.
You are mistaken. Power consumption is lower because the VRAM is what’s primarily being utilized, not the GPU. Not only that, but GPU & VRAM are used non-stop while gaming, but only intermittently while performing LLM inference. And it’s asynchronous, only processing when replying, while the rest of the time it’s idle, while waiting for the next prompt. Gaming uses a LOT more electricity than local LLM usage.
If you’re going to count the processes that had to occur prior to the current usage, then gaming uses a magnitude more electricity because of all the electricity of the developers who made the game, and the artists, and the sound designers, and the composers, and the compilers, etc.
Let’s just leave it at both use substantial energy, and both are difficult to justify. I still personally love and will use both, as a game developer with a focus on novel AI usage in games.
Power consumption for AI is way higher than for games…your gpu is rarely at full usage when running a game, only really if you’re one of the few people that tries to get every single bit out of your game with the highest level graphics and even then you’re not using the full amount because the game would stutter. It doesn’t matter if an llm stutters. They use the full power of your gpu no matter what you’re doing.
And the models being trained on other models mean that local LLMs use more energy than a model like Claude. You have all the energy of multiple data centers training one model and then you go and train a model off of that.
You are mistaken. Power consumption is lower because the VRAM is what’s primarily being utilized, not the GPU. Not only that, but GPU & VRAM are used non-stop while gaming, but only intermittently while performing LLM inference. And it’s asynchronous, only processing when replying, while the rest of the time it’s idle, while waiting for the next prompt. Gaming uses a LOT more electricity than local LLM usage.
If you’re going to count the processes that had to occur prior to the current usage, then gaming uses a magnitude more electricity because of all the electricity of the developers who made the game, and the artists, and the sound designers, and the composers, and the compilers, etc.
Let’s just leave it at both use substantial energy, and both are difficult to justify. I still personally love and will use both, as a game developer with a focus on novel AI usage in games.