So image generation is where it's at?

I'm not sure I follow.

What Koboldcpp does is set up to call out to an external generator for the images. It itself isn't providing the image generation computation.

Like, you create a prompt and hand it to koboldcpp. It then computes a textual response. Part of that response is a prompt intended for an image generative AI. It feeds the prompt to something like Stable Diffusion or ComfyUI, and that does the image generation. It takes the output and displays it inline with the text it's generated for you in the KoboldCPP web page. You run both the image generator and koboldcpp side-by-side.

What OP is complaining about is that he feels that consumer hardware --- by which you're probably talking GPUs with up to about 24 GB of VRAM --- don't have enough memory to run large LLM models, to have a chatbot on par with what typical cloud-based services are running. He is okay with the image generation side.

Llama.cpp can split a model across multiple GPUs. In theory, you can run quad 4090s or quad XT 7900 XTXs. Each of those has 24GB. Each of those is maybe $1k for the XT 7900 XTXs. I'm pretty sure that the 4090s used to go for $1.5k-$2k, but it looks like they're currently about $3k on Amazon. So $4k for the AMD route, and $12k for the Nvidia route. For the 7900s, that's about 1420 watts, disregarding the rest of the system. For the 4090s, 1800W. A standard US household circuit is 15A or 20A at 120V so 1800W or 2400W, so in the US, you're probably running close to circuit limitations. There are apparently some computer systems that use dual PSUs that can feed off multiple circuits. You'd need a power supply capable of feeding that, and given that this is considerably more heat than a lot of space heaters, probably cooling. That'll get you 96GB of VRAM (assuming, as is possible with llama.cpp, that your problem is one that can be split across multiple GPUs). Whether-or-not that's reasonable consumer hardware may be debatable, but unless you start going to dedicated AI compute hardware, which costs more, that's about what you have to work with.

There's also the approach that some people have used of using machines with unified memory to get more memory for the GPU. OP mentioned Apple hardware (like a Mac Studio, at up to 192GB) , and I mentioned the AMD AI Max machines (at up to 128GB) that Framework is selling. Those probably aren't going to be able to crunch as quickly as dedicated hardware for your given problem, but they're a way to get parallel compute hardware with a lot of memory for less money.

Running cloud-based will save money if whatever you're doing doesn't have your hardware constantly in use, since otherwise you're paying for idle hardware. That is, it's definitely going to be cheaper to run things sharing hardware if your use case is intermittent use, like a chatbot.

Llama.cpp does support clustering multiple machines (dunno about for training). I have not done this myself, and if you're thinking about buying hardware to do that, I'd probably look into whether what software you want to run can actually run in that kind of environment, and what kind of performance penalties you're looking at, but it's one possibility.

But I think that the short answer to "can I locally run hardware that can do what the bleeding edge of cloud-based services are doing", the answer is "yes, but not inexpensively". You are going to need to accept paying more for hardware than you would if you were sharing the cost with other people, performance tradeoffs if you're going to have less-beefy hardware doing the crunching, or quality-of-output tradeoffs if you want to use smaller models or otherwise limit memory usage.

I personally find running smaller models than GPT-5 locally to have value. But...OP might not; for him, the quality tradeoff might not be acceptable. I also don't mind things running more slowly, but that might not be an acceptable tradeoff for OP. I am not willing to pay for the kind of hardware that cloud-based commercial AI services are using to do AI compute, but it is possible to get that (well, barring any availability issues) if one throws enough money at it.

It's also possible to use something like vast.ai to rent remote cloud-based parallel compute hardware, if one is comfortable with trusting remote hardware and hoping that whoever is running that hardware isn't snorfling up data from one's compute job (I'd guess probably not, but one never knows, once one's data goes out into the broader world). That's not local, but it might be preferable to using ChatGPT or whatever service, which will be logging user chats.