Agreed but I want to see how it plays out. Historically a good Windows computer cost $1000 and it was all it took to start programming. How much does it cost a computer with enough resources to run a good enough AI model for agentic workflows and a reasonable time to first token? Can "most of the world" afford buying one?
Qwen 3.6 27B is quite good for agentic coding, and practical to run on consumer hardware. You need a system with either 32+ GB VRAM, or a unified memory system with 48+ GB VRAM and a decent integrated GPU. While not cheap, such a setup is still attainable for much of the world, and will eventually get cheaper over time. Open models hosted on non-American clouds also remain an option with a much lower barrier to entry, for cases where privacy is less critical.
There was an article on HN a few weeks ago where someone detailed how they managed to get an old datacenter GPU to run in their consumer PC, getting decent performance with qwen. He spent something like $200 on the GPU (second hand of course).
So yeah, I think models on local hardware will be quite common soon among the tech savvy (such as people creating software).
Especially considering the millions of 2026-class data center GPUs that massively overinvested companies are currently buying, which will be obsolete in a few years.
I think those are going to be run until they die. The capex vs opex is too high to obsolete them in a few years. They'll keep serving current gen LLMs for as long as they keep running.
It won't make sense to run them after two years. The vendors will be limited on datacenter space, power and cooling, and there will be new hardware available that will run the same models at a fraction of the power.
A100 -> H100 was >3x tokens per joule, H100 -> B200 >10x. There are significant low-hanging fruit still available in architectural efficiency, and the vendors are chasing them.
This is the big risk for AI companies that I feel is not being sufficiently priced in. Almost none of the investments they are making are durable, the depreciation schedules for everything but the real estate should be less than 24 months. Until the hardware is stable enough that you only get double-digit % improvements per generation, it should almost be counted as opex.
The annual operating cost of these is <10% purchase price annually. Even if the B200 is 10x more efficient in practice you can still operate the H100 at profit.
As it stands there's way more demand than supply. The new GPUs are going to run frontier models while the older ones serve smaller ones.
That said some of these are running in tents hooked up to mobile turbines. I can see some of those going away but generally I think you'll see them used until they start to fail in 5-10 years.
They can also be used for other things than running the main frontier whatever model as well.
E.g. grok isn't truly multi-modal, it has a callable tool that is a separate VLM it invokes on image URLs or files (for a long time it was grok-1.5v, but I think they have upgraded now, it was pretty bad).
And then you have the small summarizer models for the CoT/thought traces, the guidable summarizer models for the standard browse tools, etc.
I think you are right when you factor in the much more efficient newer high end GPU. That is what may make the current GPu investments obsolete in 2-4 years.
I do hope you're right that it will get cheaper over time (it should), but right now 32GB of VRAM is not affordable to a lot of people. You're talking ~$4500 just for the GPU, or $800 ish used if you can find one.
A Mac laptop can be had with 32GB of RAM for far less than $4500. Not sure if they actually need 32GB of discreet GPU RAM. My Mac laptop does run Qwen at a reasonable speed.
I've seen folks make it work with a 3090 on 4 bit quant using turboquant for KV cache. That's key because 3090s remain the most cost effective gpu metal for enthusiasts (albeit 24g) and the jump to 5090 (32g) is quite expensive and not always worth the LLM specific performance; sadly, good 32g metal is somewhat lacking in the price point at or above the 3090.
> Open weights/source doesn't necessarily mean running on local hardware, though.
And we've barely started to scratch the surface on helping open-weight models "be the best they can be", with cloud burst parallel sampling and prompt mutation. Looking for best probabilistic results for a prompt, and looking for best prompt variants for a task. Adaptively scaling computes at generation, not just training.
And speculatively, if agentic coding is naturally a multiplicity, what UX might enable human devs to dance with that quantum superposition? Rather than quickly collapsing to one monkey and its keyboard.
How far back are we talking? Historically there were just a few computers in the world, personal computing revolution happened much later. And a private inference rig costs roughly the same in 2026 as a personal computer did in the early 80's, if you want the performance comparable with a personal workstation of that time (vs a mainframe/supercomputer).
> Historically a good Windows computer cost $1000 and it was all it took to start programming.
Gotta remember inflation here.
$1K in 1995 was roughly equivalent to $2K now and wouldn't have been a particularly "good" machine then.
In 1982 the Commodore 64 started at about $600 bucks, also roughly around $2K today.
If you outgrew that, beefier machines back then were A LOT. It was easy to find $2k+ towers and (especially) laptops even into the 2000s, and a lot of those would be $5K+ equivalent today.
I don't understand the justification for local hardware with cost as the motivation. The same (or bigger/better) open weights models can served by third parties at much higher resource utilisation, and will therefore be much cheaper!?
Especially because the world is likely to persist, at least for a while, in state where computing hardware demand drastically exceeds supply resulting in high prices for hardware. So why wouldn't you want to max out utilisation and amortize costs, at least for typical (non sensitive) use cases.
IMO the more useful distinction is in analogy to VPS versus SaaS/PaaS. Open models allow you to use any inference provider you like, including local ones, similar to running open-source software using VPS providers. You’re not bound to a particular SaaS/PaaS as you are with closed model providers. That same freedom also allows you to self-host when you care about that.
Before the AI "crisis" it used to take about $3500 to get a prebuilt with a 5090 which can run good enough LLMs. I run reasonable LLMs on just 16GB of VRAM on my Mac, and the 5090 has double that.
Software models and hardware are getting better all the time—and that’s where some big companies spending billions might stumble! In fact, Microsoft recently announced that they’re scaling back a bit on their AI investments.
Moore's law or one of its generalizations still holds, so it will only be a short matter of time before a $1k computer will be able to train and run a powerful enough model.
I thought Moore's Law came to an end in the last decade?
Certainly the transistors/chip or transistors/$ or flops/$ have not been progressing at the same exponential rate as during 1970-2010. There is still progress, but it's rather slower.
I was careful to say "Moore's law-like". Moore's law is stated as "transistor count per area doubles about every 2 years." [0] As stated, then yes, this might be true but, while important, that's not really the quantity we care about.
As you point out it's really cost per transistor or cost per flop that we mostly care about. I'm finding it hard to find a succinct and clear plot, but I believe one is provided by Our World In Data on "GPU computational performance per dollar" [1] which, to my eyes, clearly shows exponential growth in computational power per dollar.
The picture for storage is a little more muddied but if you squint just right you might still be able to recover an uninterrupted exponential growth [2].
In my view, it's pretty clear that advances in AI have progressed so quickly because GPUs have been keeping up with the exponential growth of computational power (per unit cost).
Exponential growth in this area is usually characterized by "S-curves", where one technology gets saturated but the exponential increase in power or decrease in cost is picked up by another, adjacent, technology, that allows the growth to continue. For compute it's CPUs to GPUs. For storage it's platter drives that are now being overtaken by SSDs.
The more general phenomena is called Wright's law, or experience curve effects [3].
I think AI is still in it's "mainframe era". Once upon a time, computers that ran useful workloads filled a whole room. And now that's true of computers that run (state of the art) AI models. My guess would be that this won't always be the case, although I certainly don't claim to have a clear idea of how things will play out.
Roughly about Eur 3-4K right this minute I think? The graphics card, ram and storage are punishing. Under more normal circumstances (hopefully late 2027) it'd be 1500-2500 depending on what you think is realistically useful.
Possibly it's the same price range, allowing for inflation.
> It was only in 2025, as memory prices began an unprecedented surge, that the memory makers started to build new fabs targeted at HBM, all slated to start producing chips in 2027 or 2028.
The RAM shortage is predicated on both the huge datacenter buildout (many of which are already mired in delays, with a few even cancelled outright), and the massive memory purchase commitments various hyperscalers have made - hyperscalers who seem to be running short on cash lately...
History? This isn't the first RAM shortage. When one happens, producers build more fabs. The fabs come online, the availability of memory shoots up, and the shortage goes away, usually replaced by a glut.
If you want to argue that this is different from all previous RAM shortages, you can, but the burden of proof is on you to show the difference.
There is certainly economic pressure to create an exponential demand for tokens, but we've already seen a pullback from the costly "token maxing" companies were pushing last year.
It's also pretty unclear to what degree the RAM shortage is driven by inference (versus by training). We're rapidly approaching the point where frontier models are "good enough" for everyday use, are at some point we're going to hit diminishing returns on training new trillion-parameter models...
Historically the cost of compute has also gone down. Like just look at it as compared to a year ago. We have amazing open source models that can run on consumer hardware and if we go away from our obsession of using opus 4.8 or mythos for everything then it actually is super amazing to see what these open source models could do. I use qwen3.6:27b as a daily driver and I am heavily impressed with it.
> Historically a good Windows computer cost $1000 and it was all it took to start programming
Started with computers around 2009 and later bought an oldish computer (a pentium 4 PC) for the equivalent of 50 usd. Codeblocks and Python Idle were free at the time (C and Python were the first languages I learned). The barrier to programming has always been low as the only thing you needed was books (the internet made things easier) and access to a PC (I had friends with laptop and my school lab).
In 1998 I had a second PC for the first time in my life. It was an old i486DX2-66 that my dad gave to me. He said it cost him $50 secondhand and was bought so I could put Linux on it.
Hence why brute force needs to be replaced with examples such as neuromorphic methods. It could realistically could be combined with mesh networking as well to utilise the capabilities of all computers locally.
