> For the setup price alone of a low-end home lab1, you could buy several years of a paid subscription to one of the AI providers. The power costs would come out to around $50-$300 per month, depending on how much inference you’re running: again, the price of a couple more paid subscriptions.
Okay, but I already have multiple computers capable of running local models with acceptable performance, so for me the cost is $0. I suspect that's true for most people of sufficient technical inclination to be interested in and capable of running models on their local machines.
Also, no, the monthly electricity price of even my power-hungriest machines ain't anywhere close to that figure. Hell, at the high end that's more than my power bill for my whole household.
The mean price of electricity for me is about 40c per kWh: about as high as anywhere in the US.
There are about 730 hours in a month, and a consumer machine can sustain maybe 500W, so we can spend 0.4 * 0.5 * 730 = 146 dollars on compute power per box unless we start buying special stuff.
For my area (nw Iowa) where electricity is about 10¢ per kWh, that'd be like $43.8 ish? In reality much less for me personally with a Mac Studio, since it doesn't go anywhere near 500W
Win what? Money, fame? That's not the point of small models, freedom is the point. What's this occultist obsession with "There shall be only one" monopolies? Why only one? That's so irrational and childish.
Every so often someone hypes local models as "Don't pay for Claude check out this local modal" and the local modal isn't at all equivalent.
That doesn't mean that local models are going anytwhere (as the article motes they're useful) but if you want the best we're stuck with expensive remote models right now.
> Or maybe models get so good that a 30B model is genuinely smart enough to do everything, so nobody really needs a model like Opus or Sol unless they’re trying to solve the Reimann Hypothesis. I don’t really buy this. Models can do frontier mathematical work today while still being not smart enough to refactor large codebases as well as me, so it’s hard to imagine a world where I don’t just want to use the smartest model available.
Idk, I already don’t bother with Opus and stick with sonnet med. I really care more about speed. I use qwen3.6 27b for personal projects and I think it works pretty great.
So like the article mentions, if scaling stalls and small models get better it’s not impossible to imagine a convergence and hardware costs drop.
Having said that, self hosting will be a niche thing like it is today for other services.
"In a year you might be able to run something about as strong as GPT-5.6-Sol on your laptop. But by then, you’ll think of GPT-5.6-Sol as too weak to be useful." This assumption is unfounded and invalidates almost the entire premise of this piece
I think local models will not be “niche” in the future in the same way personal computers are not “niche” just because most computing happens in the cloud. They have entirely different uses
There is one more possibility. Mainboards all get shared memory and memory costs go down below $1/GiB. That makes GPUs cheap, as they do not come with memory. Then local models become the norm for many, because you can buy a 10 TiB machine for 10k, and upgrade the GPU when needed. I guess that models will not grow endlessly, ones the growth in size flattens, the demand for more and more memory flattens. So, IMHO, long term local models will win. However, short to medium term it may not be. My 2 cent.
> For the setup price alone of a low-end home lab1, you could buy several years of a paid subscription to one of the AI providers. The power costs would come out to around $50-$300 per month, depending on how much inference you’re running: again, the price of a couple more paid subscriptions.
Okay, but I already have multiple computers capable of running local models with acceptable performance, so for me the cost is $0. I suspect that's true for most people of sufficient technical inclination to be interested in and capable of running models on their local machines.
Also, no, the monthly electricity price of even my power-hungriest machines ain't anywhere close to that figure. Hell, at the high end that's more than my power bill for my whole household.
Let’s put numbers on it.
The mean price of electricity for me is about 40c per kWh: about as high as anywhere in the US.
There are about 730 hours in a month, and a consumer machine can sustain maybe 500W, so we can spend 0.4 * 0.5 * 730 = 146 dollars on compute power per box unless we start buying special stuff.
For my area (nw Iowa) where electricity is about 10¢ per kWh, that'd be like $43.8 ish? In reality much less for me personally with a Mac Studio, since it doesn't go anywhere near 500W
> No, local models will not win
Win what? Money, fame? That's not the point of small models, freedom is the point. What's this occultist obsession with "There shall be only one" monopolies? Why only one? That's so irrational and childish.
Win as the dominant for of LLM interaction.
Every so often someone hypes local models as "Don't pay for Claude check out this local modal" and the local modal isn't at all equivalent.
That doesn't mean that local models are going anytwhere (as the article motes they're useful) but if you want the best we're stuck with expensive remote models right now.
The fallacy is that being the best in everything isn't that relevant for most folks, in many scenarios being good enough is sufficient.
> Or maybe models get so good that a 30B model is genuinely smart enough to do everything, so nobody really needs a model like Opus or Sol unless they’re trying to solve the Reimann Hypothesis. I don’t really buy this. Models can do frontier mathematical work today while still being not smart enough to refactor large codebases as well as me, so it’s hard to imagine a world where I don’t just want to use the smartest model available.
Idk, I already don’t bother with Opus and stick with sonnet med. I really care more about speed. I use qwen3.6 27b for personal projects and I think it works pretty great.
So like the article mentions, if scaling stalls and small models get better it’s not impossible to imagine a convergence and hardware costs drop.
Having said that, self hosting will be a niche thing like it is today for other services.
"In a year you might be able to run something about as strong as GPT-5.6-Sol on your laptop. But by then, you’ll think of GPT-5.6-Sol as too weak to be useful." This assumption is unfounded and invalidates almost the entire premise of this piece
They only need to be good enough to solve specific problems, that is already a victory on my book.
Is this the old mini vs micro argument again?
I think local models will not be “niche” in the future in the same way personal computers are not “niche” just because most computing happens in the cloud. They have entirely different uses
There is one more possibility. Mainboards all get shared memory and memory costs go down below $1/GiB. That makes GPUs cheap, as they do not come with memory. Then local models become the norm for many, because you can buy a 10 TiB machine for 10k, and upgrade the GPU when needed. I guess that models will not grow endlessly, ones the growth in size flattens, the demand for more and more memory flattens. So, IMHO, long term local models will win. However, short to medium term it may not be. My 2 cent.
Most of my inference already happens locally, actually.
- not if apple m6 mac studio comes with 1TB of RAM and 144 core cpu / gpu
What about that edge computing. Those cheap drones
If they dont the gap between rich and poor will be unsustainable.
Isn’t it already?
I guess we just have different timelines
[dead]
[dead]
[dead]