upvote
For those asking in threads. 4x r9700s will cost about 6-8k depending on how patient you are with ebay auctions.

From there any motherboard and CPU that can give you 4x slots of 16x Gen 5 PCIe slots positioned such that you can install 4x cards is probably fine. Ideally with a CPU capable of memory encryption.

My latest pairing is a ASUS Pro WS W890E-SAGE SE For $800 open box, and a Xeon 654 for $1300. My first build was an AMD EPYC rig for similar cost a few months ago using a Gigabyte ME03-PE0 and a AMD EPYC Siena 8224P for about 2k combined. I specifically wanted one amd and one intel machine for reproducible build diversity which most probably do not care about.

Do not get too attached to specific cpus and boards. Get what is available for cheap with the needed pcie bandwidth open box on ebay from reputable sellers to save thousands of dollars.

Be sure to get a power supply that can handle all 4 cards. I run Silverstone Hela 2500s at about $800. Do not go cheap on power supplies. It will end in pain. Can confirm.

Then 2-3k for 64gb in DDR5 RDIMMs. This is the most painful thing to purchase but it is what it is.

Not only do I recommend this setup, I plan to build more of them to share with my team.

If you are patient and buy used, a 10k budget can get you a very powerful local AI rig running 100% FOSS software/drivers even today.

reply
>I turned down the free OpenAI and Anthropic tokens in my YC batch because I have absolutely no interest in being reliant on third parties to do my job.

Which is exactly why said power and utility will soon be BANNED under a myriad of false aegis'.

An economy that almost exclusively extracts ECONOMIC RENT from the working class, won't permit any "leakage" or autonomy.

reply
Did you turn down this for internal use-case? OR for your core product? I am keen to learn how are people running stuff locally for their core product to be able to make it easy for their customers to use.
reply
I have no interest in having my business or my personal AI usage being reliant on any single third party. We currently have no customer facing LLM products, but if we did they would be only options we can run confidentially which no providers can do ATM to our standards. Current plans are just racking GPUs like I did at home.

Most of my peers do use OpenAI and Anthropic for development work, which pains me, but I will never tell someone else how to do their job if the end result meets my quality bar.

I lead core systems engineering building custom enclaves and operating systems and I do 100% of my work with hardware I own on my own property. When the internet goes down, I do not care and keep building.

reply
If you don't mind me asking, how do you drive the models? Right now it seems like everyone is feeling their way around in the dark about how best to steer the work. AWS AI-DLC, SpecKit, BMAD-METHOD, TwitTech Influencer Pile Of MD files...

It all seems hopelessly adhoc at the moment.

reply
My own fork of jcode stripped down to only my needs running in herdr. That is it. Just lean rust code and no javascript nonsense needed.
reply
I exclusively only use local models. I do not subscribe to anything, and I have my agents wrapped and inspected.

2xDGX Spark cluster is my current load, it is sufficient for a great deal many wonderful, wonderful things. It isn't fast as consumer AI, but I have nevertheless been able to do wonderful things with it for the past year'ish.. I'd get a couple more, but indeed they've tripled in price, so yeah .. things will just do for now, and it's fine. Pretty great, actually.

I think the way we control AI is we air gap it. My cluster is air-gapped.

reply
whats electricity bill and do you not fear that your hardware will be obsolete while others will have access to better models on faster hardware?

i just calculated that it should be Per month (30 days) 380euro in my region, which is a lot after an initial 20k.

I have to say, no limits stopping sounds nice though.

reply
> whats electricity bill

$300/mo in the bay area for 1kwh running 24/7, at least until I get solar installed, then free.

> do you not fear that your hardware will be obsolete

If it never gets better than it is right now, I cannot think of any problems in software I cannot solve locally, so I am all set for a while.

Though also, the capabilities on my -existing- cards have gone UP dramatically in the past 6 months since as the local LLM community keeps optimizing.

Sure in 3 years all these cards will be a lot cheaper than I paid, but having them for the next 3 years is absolutely worth it for me.

This costs me less than a used car. Most people chase car and phone upgrades every 2 years. I tend to get something good enough and upgrade once a decade. Depends on your goals really.

reply
Are you able to interface with it entirely through voice?
reply
I could easily, but not been a priority yet.
reply
why not? that's very easy to do with tts.
reply
pardon what is tts as im just beginning my journey into building my home local/private AI companion.
reply
text-to-speech
reply
How cost effective had it been? I don't want to rely on 3rd parties either, but my experience is it's super expensive to run locally versus using a provider (both in terms of hardware costs and ongoing electricity usage).
reply
No third party can ever provide the privacy and sovereignty of being able to run locally with abliterated models. I would do it this way even if the providers were much cheaper as I value those properties a -lot-.

In practice though, I can produce about 3B tokens/month of Qwen 3.8 Flash Next which is the model I crank almost 24/7. That would cost something like 1k/mo on OpenRouter. My setup costs about $300/mo in electricity, though getting solar soon to drop that to zero.

reply
It doesn’t really drop the cost of electricity to zero; you have to amortize the cost of the panels over their lifetime, plus the opportunity cost of the capital outlay (or interest on the loan)
reply
That is fair. Still, dramatically reduced cost and reliance on one less third party.
reply
Running locally is not a problem if you allow it to be slow. I get about 3.5 tps with Qwen3.8-27B:UD-Q4_K_XL on my 64 GB laptop and about the same tps with Qwen3.8-Flash-Next:UD-IQ4_XS (IIRC, or UD-Q4_K_XL) on my 128 GB desktop with a 16 GB GPU.
reply
Can you share more, or direct us to more resources around local setups? For starters what's the upfront cost, what are the space, weight, energy requirements, etc.
reply
Really depends on what GPUs you go with but I hate proprietary software so I went with AMD AI PRO r9700 cards which you can still pick up used on ebay for $1500 used on auctions. Started with 1 then expanded over time as I won auctions. Though on a lower budget my $400 16gb 6800 XT can run Qwen 3.8 27b which I could absolutely get by with at significantly reduced speed.

My entire 12u rack has roughly the same capability as an H200, though at maybe 15k all in with mostly used parts vs 50k for a used H200 checking todays ebay prices and that is -just- the card you still need to build a system for.

Consumer cards are a lot more power hungry than enterprise cards and they take up more space. I run my setup next to my dryer, running on 240v for 15% better power efficiency.

Also using transparent memory encryption on a Xeon and an EPYC for confidential and private inference guarantees no third party can offer me. Of course I can run abliterated models that do any security research I want without refusals.

Strata and Radiance are the state of the art ATM for running big MoE models on small consumer GPUs at high performance.

reply
Re: running qwen 3.8 27b on a 16gb card - that isn't a MoE model, was it just heavily quantized? How much context window were you getting?

Just starting to dive in on local AI with my own 16gb card, and the struggling with evaluating what tradeoffs are worth it.

reply
On a 16gb card check out Bonsai2. I run it on my 6800XT.
reply
There are tons of resources out there on how to run a local setup. Once you have something that's somewhat functional you can then have the system self optimize your software stack to suit your needs.

As for cost a 9700 is running at about 2k USD MSRP right now and the rated power is 300w.

reply
deleted
reply
deleted
reply
How does this make you less reliant on third parties?

It seems like it just shifts which third parties you rely on. I think local inference is cool and your setup sounds dope, but you still have third party hardware manufacturers, and third party models in the mix.

What is the advantage of this over OpenRouter, which lets me select for, jurisdiction, quantization, ZDR, etc.

note: My use case can't even come close to having this make sense financially, but I understand if yours does. Either way, that doesn't make the third party argument better.

reply
Not the OP but I would assume: - no data retention policies to worry about in the first place - no monthly billing outside of power bill and capex - no jurisdiction issues - full control over your data - no network round trips or even worry about network outages in the first place
reply
Exactly. I own the GPUs and bought them all used on ebay. One time cost.

When the internet goes down, my productivity is not impacted, and I have total privacy on my work at all times. Also no rate limiting etc, and I can burn tokens freely.

reply
Maybe a naive question, but how do you even run 8x GPUs? Threadripper Pro? Anything special about OS / software stack?

Sounds interesting, though a bit pricey (if I'm reading the prices correctly it's 20k just for GPUs). Fingers crossed that the prices will fall.

reply
Gave a full parts list as a reply to my original post
reply
it's the difference between a one-time transaction with a third party and an ongoing, perpetual third-party liability. you buy your inference hardware, you download your model, and if you wanted/needed to, you could run inference in perpetuity (or until your hardware broke down). compare that to anthropic/openai/etc being able to throttle intelligence, deprecate models, or raise prices on you whenever they felt like it.

sure, in a very general sense you're dealing with third parties in both scenarios, but if you actually own the hardware and model weights, you can control your own destiny for the lifespan of that hardware.

reply
> it's the difference between a one-time transaction with a third party and an ongoing, perpetual third-party liability.

no ongoing liabilities except the power bill and opportunity cost of dropping $20k on hardware.

He isn't using OAI or Anthropic models. He is using models that area available for absurdly cheap from completely independent providers in multiple jurisdictions.

I get the base level argument, but from a practical point of view there are so many different dependencies that shifting one isn't really doing much.

reply
Is this some sort of gotcha? You didn’t build your computer from scratch with sand, sticks and pebbles, so you’re just as reliant on third parties?

If you own your own hardware, you don’t have to depend on and trust third parties continually, just once when you bought it.

reply
This feels a bit pedantic, but my scope here is third party services, like Anthropic or OpenAI subscriptions.

If the power goes out, internet goes down, I fire up my generator and keep building.

We are talking about digital sovereignty and ownership.

I own the means of my production.

reply
Sounds like commie bunk to me! We can no longer sit back and allow Open Weight infiltration, Open Weight indoctrination, Open Weight subversion, and the international Open Weight conspiracy, to sap and impurify, all of our precious bodily fluids.
reply
> Is this some sort of gotcha? You didn’t build your computer from scratch with sand, sticks and pebbles, so you’re just as reliant on third parties?

I'm not the one claiming to be free of third party dependencies. Third parties built my hardware, they wrote the firmware in the hardware, they wrote the OS on the hardware, they designed the network stack that you might use to connect to a local server, they provide security updates to all of that software, they ensure that I have a stable power supply to run all this, etc.

I'm incredibly reliant on third parties, some of which happen to sell me tokens. I recognize that I live in a world where I will ALWAYS have a dependency on third parties, and rather enjoy the fact that I can use GLM flash on someone else's $20k computer for a few cents.

reply
Yes. They built the hardware. Now it’s mine, and they’re no longer involved, so I no longer depend on them. That’s how it works.
reply