AI GPUs suck for gaming, I have seen a video from a guy playing Red Dead Redemption 2 on a H100 at a whooping 8 FPS! That's after some hacks, because otherwise it wouldn't run at all.

AI GPUs are stripped away of most things display-related to make room for more compute cores. So in theory, they could "work", but there are bottlenecks making that compute power irrelevant for gaming, even if they had a display output.

So there aren't actually GPUs, but more like some other architecture for CPUs.

I wouldn't mind my own offline Gemini or ChatGPT 5. But even if the hardware and model were free, I don't know how I'd afford the electricity.

If you can't afford the electricity to afford to run the model on free hardware, you'd certainly never be able to afford the subscription to the same product as a service!

But anyway, the trick is to run it in the winter and keep your house warm.

I think you're underestimating economies of scale, and today's willingness of large corporations to provide cutting-edge services at a loss.

I don't think I am. I don't think economies of scale on hardware will drive costs below free, and while subscription providers might be willing to offer services below the cost of the hardware that runs them, I don't think they'll offer services below the cost of the electricity that runs them.

And while data centers might sign favorable contracts, I don't think they are getting electricity that far below retail.

Here in California, as a residential customer, I pay close to 50 cents for even the very first kWh I consume each month, and there is no bulk discount. Meanwhile, Google et al. surely pay less than 10% that rate. A single Gemini query probably involves dozens of distributed machines that skilled SREs keep afloat, but let's say very optimistically that if I had a copy of the model, I could run it on a home workstation with six 80GB GPUs, consuming 4,200 watts when active, and 500 watts when idle. My utilization might be 10% on a day when I'm especially curious, whereas Google probably manages 95% utilization of its inference machines.

I could go on, but I think you get the point: the dollar cost for me to run a hypothetical version of Gemini at home far exceeds the price Google pays to deliver the same service to me. I'm effectively arbitraging the price of electricity by subscribing to an LLM service, because I could never run it for anywhere near the same price in my house -- and to provide the service at the speed and reliability of Gemini, I doubt I could even fit the equipment in my house-sized house and still have room for me to live in it.

Anyway, I stand by my original comment - it's neither easy nor cheap to run a frontier model at home.

A single machine for personal inference on models of this size isn't going to idle at some point so high that electricity becomes a problem and for personal use it's not like it would be under load often and if for some reason you are able to keep it under heavy load presumably it's doing something valuable enough to easily justify the electricity.