If you had $2000-$2500, what would you build or buy?

[Greg_E]

If you had $2000 to $2500 usd, what would you build or buy?

Needs, general light coding : make me a web app that does this and that.

Diffusion type work: make me a still image of this or that at 1920x1080, maybe make me a short video of this or that

Make me a sound effect of this, or maybe expand to make me background music that sounds like this

What's important in a build? Lots of CPU cores and system ram or would 24 cores and 32-64GB of system ram work?

In the build side of this I might have some old Supermicro X10 series with dual processor and 24 cores total, 32 or more DDR4 ecc, and dual 16x slots (one per cpu) that I might be able to use for a build. I'd need something like AMD V620 and probably want two of them, probably want to add more system ram too.

Obviously, with only this much money, performance will be modest and really only support one or two users. We would have the ability to just send the job and let it churn overnight or days at a time if that's what it needed, it will not be getting new tasks every day.

 

[nexox]

I kinda went through this exercise a couple months ago, minus the diffusion requirement, I ended up deciding to throw $200 at anthropic and use claude for a year to tackle my project backlog, then see what GPU/DDR5 prices look like when that subscription is up for renewal. My read of the various benchmarks is that the hardware to do reasonable LLM coding just costs too much, I don't have the patience to wait overnight (or 20 minutes) for every iteration.

 

[louie1961]

I don't really do any coding, so take this with a grain of salt, but I don't run a lot of AI models locally. I have Ryzen 7 5800XT, 128gb of RAM and a Nvidia RTX 2000 Ada 16gb GPU. I run small embedding models, a RAG model and will (in the near future) run Frigate models locally. For heavier stuff, mostly deeper RAG and data pipeline processing, I use my local instance of Ollama, but with cloud models instead of local models. I use that for things like Project NOMAD, the Jarvis plugin for Joplin, and such. And finally I save the frontier models for deep work. My usage of the cloud models from Ollama have stayed with in the free tier so far. And you can connect Ollama to other containers the same way you would local models. So I guess what I am saying is use that money for subscriptions.

 

[Stephan]

Prices are so spoiled at the moment, 2500 budget makes no sense for anything other than pure basics ("computer to type on", "display", etc.)? If you wait a little longer that will be the price of one RPi6. Maybe DDR5 from CXMT will be of help to break Samsung/SK Hynix/Micron monopoly. What is needed is 16x B300 for below 2 grand.

For agentic tasks dense 27B is bare minimum. Like the Qwens we all like. I mistrust 4-bit quantizations, so I always go for Q8_0 with MTP. Means 48 GB VRAM and that is already expensive. Newer Qwens have nerfed world knowledge to make room for working with tools. Gut says true universal models will need 200B parameters, maybe more like 500B or 1T.

On Reddit people use 2x 3090 with 24 GB each to run 27B models. Maybe that? But beware of half broken cards, test thoroughly or make sure there is a refund.

Speaking about "building for 2.5k". Do you have a finnish jacuzzi? With a little work that should get you one.

 

[WANg]

If you had $2000 to $2500 usd, what would you build or buy?

Needs, general light coding : make me a web app that does this and that.

Diffusion type work: make me a still image of this or that at 1920x1080, maybe make me a short video of this or that

Make me a sound effect of this, or maybe expand to make me background music that sounds like this

What's important in a build? Lots of CPU cores and system ram or would 24 cores and 32-64GB of system ram work?

In the build side of this I might have some old Supermicro X10 series with dual processor and 24 cores total, 32 or more DDR4 ecc, and dual 16x slots (one per cpu) that I might be able to use for a build. I'd need something like AMD V620 and probably want two of them, probably want to add more system ram too.

Obviously, with only this much money, performance will be modest and really only support one or two users. We would have the ability to just send the job and let it churn overnight or days at a time if that's what it needed, it will not be getting new tasks every day.

...if it's me and I am starting things clean without repurposing things in mind? Something like this Macbook Pro M2 Max/64/512GB - although I'll check the usual secondary marketplace to scour for a deal (that one is viable but not great). It has the 38 core GPU to give you decent (not leading edge but solid) prompt rates, it has the 400GB/sec memory bandwidth to do tokenwalking well, 64GB is a good number for loading large-ish LLMs / run VMs on UTM/Parallels, it's already past its prime so it won't be stupid expensive (that's a 4500 USD machine back in 2023). At the end of the day it's still a Macbook pro so it holds value, and you can use it as a workhorse machine. I'll recommend getting some kind of 3rd party warranty for it, and see if Apple has a support program that can cover for issues with LCDs or the battery. Mac studios aren't a bad choice but the secondary markets for them are a bit limited...but I would cap the hardware spend at 2k (including taxes shipping and warranty) and then spend the rest on a cloud LLM provider like Anthropic.

What I won't do is to go out to buy GPUs unless it's stupid cheap and have a use for it after the LLM fever dies down. The Radeon Pro V620 (or the Instinct MI50, or the Tesla V100 or the T4) is...kinda okay for this, but they are not really boards you can use in a quiet office unless you mod it extensively, and most have not much practical use minus using it for inferencing (the V620 require a custom firmware to run it as a W6800 but one Displayport along with the compute unit gimping doesn't thrill me, either). In the caee of the V620 you can wait for the SRIOV GIM drivers for GPU partitioning/virtualization, but it might not (ever) show up. I park mine in an Oculink dock with a remote power toggle just to keep the power/noise in control, and the N5 Air has an Oculink port already. I bought the N5 Air because I needed a 5 bay NAS with more oomph (which I ended up not needing) and I have extra RAM/SSD sitting around. If I haven't had the extra RAM/SSD that the N5 Air can use, the extra NAS chassis or the extra Oculink port, I wouldn't have spent the 600 on the V620+PSU+dock to work with the 400 I spent on the N5 Air.

The BC250 Cyan Skillfish APU that was recently hyped as sleeper eTrash? If you got it cheap (less than 135) and is willing to overlook every single flaw with it (RDNA2-ish, need for PSU/custom cooling/non-standard chassis, limited unified RAM, and fundamentally hacky as hell), you might be able to do something fun with it, but unless you can cram a model and its underlying software and OS image into its 16GB of unified RAM, it’s not worth it.

I also won't buy a DGX Spark or Strix Halo - both of them are stupidly overpriced as of Q2 '26...
The DGX is pricey but then nvidia almost always overcharge for their hardware SDKs (well, what else won't they stupidly overcharge?), and until recently, getting an ARM64 board like that to run 3rd party stuff like proxmox or TrueNAS scale is a pain even if it's UEFI.
The Strix Halo is definitely overpriced, just a gaming oriented APU that got dragged into inferencing tasks - the fact that networking them into nodes are like an afterthought, well, it’s not that easy to take it too seriously.

 

[neobenedict]

**** all with these DRAM prices

 

[Greg_E]

Hmmm... This was for work so it would go in the server racks. But I'm not getting much response to the thought of having our own local AI. So probably a dead project. The V620 have gone up a lot recently, if they were back down at $300-$350, I'd probably grab one and start fooling with it for personal use, but not at $600+.

At $600+, I'd be better off buying a newer mini PC with a decent igpu and a decent amount of ram, like an M3 or M4 mini, though I would prefer if it ran Linux natively (yes I know about Asahi). Probably be in for $1000 on a decent mini PC these days, and still probably slow. But for personal use, moving along at 10t/s seems like it would be fine, as long as I can load a big enough model to get my projects done accurately. Guess I'll stick with the T740 and see when it finally doesn't keep up with what I want it to do, and then decide on the next step.

 

[WANg]

Hmmm... This was for work so it would go in the server racks. But I'm not getting much response to the thought of having our own local AI. So probably a dead project. The V620 have gone up a lot recently, if they were back down at $300-$350, I'd probably grab one and start fooling with it for personal use, but not at $600+.

At $600+, I'd be better off buying a newer mini PC with a decent igpu and a decent amount of ram, like an M3 or M4 mini, though I would prefer if it ran Linux natively (yes I know about Asahi). Probably be in for $1000 on a decent mini PC these days, and still probably slow. But for personal use, moving along at 10t/s seems like it would be fine, as long as I can load a big enough model to get my projects done accurately. Guess I'll stick with the T740 and see when it finally doesn't keep up with what I want it to do, and then decide on the next step.

Eh....if GPU is your thing, the V620s are listing for 5-600, but in actuality some vendors will take offers for 380-420, depending on interest. That being said, smart money always look for a plan-B instead of going for the same exits as the others. Have you considered a V340 or a Tesla P40 instead? The good thing about the V340 is that it's a half-witted Instinct MI50, so ROCm support (gfx900) is still there but at a steep discount (50 bucks instead of the usual 300-400). It won't win any speed awards, 16GB is barely useful and you still need the blower setup of the V620, but if you want a quick and stupid cheap way to get a 16GB card working in ROCm, that's how you do it.
I am not sure how to get that t740 to work with the V340...yet.

As for that entire "600 bucks for a mini-PC" thing, you won't be buying too much modern with it. Why yes, you can get a Minisforum X1 or ThinkCentre 75q-5 MiniPC chassis then have the onboard Radon 780M to run your LLMs, and the chassis itself is not likely that expensive (probably around 250-350)...the issue is that the DDR5 SODIMM RAM that it needs is stupid expensive (2x16GB DDR5 SODIMMs sits at 450 as of mid '26...which is more than how much I paid in mid '25 for 2x64GB), and DDR5 really doesn't give you that much bandwidth to tokenwalk well after a certain point.

So if you got a machine with 16GB of DDR5 for 450 or so (the usual price) the most you can run is an 8-10GB LLM. For 600 USD and 32GB of DDR5 RAM you can run between a 18 to 24GB LLM but the performance will depend on whether it's MoE (mixture of experts), AxB (active with a subset of parameters, like a 35 A3B is 35 billion parameters with 3 billion active) or whether you are getting a dense, "compute everything" model, like the Gemma4-27b model that's well known for being slow-ish. Even for my N5 Air (with 128GB of DDR5) there comes a point where running a very large LLM ceased to make sense, and the only thing I can do is either deal with it, fire up the V620 and let it take some of the load/use some of its VRAM bandwidth and spill out the rest to main RAM via GTT, or just fire up Claude and hope that I am not helping Dario Amodei train up the next round of bullshit dispensers by sending him training data for free.

One major benefit of the t740/t755 is that they use DDR4 SODIMMs, which can be had relatively inexpensively. 2x32GB DDR4 (max capacity) is around 300-400, which is much cheaper than 900 for 2x32GB DDR5 (still quite elevated compared to the 120-150 I paid back in 2024). I don't even think most stores sell 2x64GB SODIMMs nowadays. A recurrent project of mine is to test machines with Oculink-to-PCIe adapters to see how they behave. I was honestly not surprised when it the Oculink adapter refused to work with the t740...but I am keen to see if the same also applies for the t755.

At some point you want to pay extra for capacity AND bandwidth, which is why I suggested Apple Pro silicon (since they have more memory channels and RAM bandwidth than the normal DDR5 stuff) - The Apple used hardware tax is often cheaper than the 4x markup on new DDR5 SODIMMs.. If you spot an M4 mini or an M5 MBP14 on sale, consider it...even if they have lower memory bandwidth their GPU compute in metal is still quite solid.

 

[Greg_E]

I was looking at the v340. If I spend any real money, it will be on a bigger platform. I have a Supermicro x10 coming out of service when the new servers arrive, I could set that up for an AI box. Kind of old, and maybe not worth the effort, but something I have available in the near future.

I saw some MI50 and marked that for research, I want to say those were relatively priced well for performance.

The big factor is my coworkers are plenty happy jumping through hoops for the free level with Claude or other free models on the web. Lead a horse to water... Kind of tired of trying to get them to open their eyes, just trying to stick out the next 7 years to retirement, and expand my skills in case I need to jump.

 

[caplam]

I run hermes agent as a vm in proxmox. The host is a minisforum ms02.
I tried some local llm to work on npu and igpu without great success. So far the only thing i had running correctly is a whisper container with openvino.
I took a subscription at nous portal and run DS4 flash as my main model but even with this model credits are out well before the end of the month. I also have a concern about confidentiality.
After many research on what gpu to run local llm, the recent announce of qwen3.8 27b and the first benchs i decided to pull the trigger and ordered a radeon ai pro r9700.
From what i have seen it will run the model with enough performance to use an agent (context long enough, tg, pp). The model itself seems to be great at tool calling.
at first i didn't want anything from amd (or intel) because of the lack of support but for the first time amd can run the model from day 0.
To be honest 3 or 4 months ago i was looking for a intel B70 but there is a great difference between price announced and the price in the shops. I also could see that finding informations about intel ecosystem on how to run llm with intel hardware is quite hard. Intel seems to be years behind the competition.

 

[BackupProphet]

I still use Claude 20x Max. I do planning with Fable, and code implementation with Opus5.

I have tried Grok 4.6, and Qwen3.8 27B.
By far I prefer the Anthrophic stack, but its also at a premium and I value get shit done properly and fast!

Qwen3.8 27B was suprisingly good. But slooooooooooow.

I think investing in hardware is still a bit early, but it depends, if you are a company with 10+ devs, having a local llm running on DGX Workstation may be a solid option. For a single developer, I think paying for Claude is the best option.

 

[caplam]

I have just seen apple announced the new macs.
not the same budget as my r9700 but you can now have a m5 ultra with 1.2Tb/s bandwith. that should be fast.
Perhaps i will regret having bought the amd wit its 640GB/s bandwith.

you tried qwen with alibaba provider or running locally ? i think that many people wanted to try so maybe it was a little bit overloaded.
I may try it in the next few days. But first i need to clean a little bit my model config (provider/model, fallback provider, delegation models,...)

 

[WANg]

I have just seen apple announced the new macs.
not the same budget as my r9700 but you can now have a m5 ultra with 1.2Tb/s bandwith. that should be fast.
Perhaps i will regret having bought the amd wit its 640GB/s bandwith.

you tried qwen with alibaba provider or running locally ? i think that many people wanted to try so maybe it was a little bit overloaded.
I may try it in the next few days. But first i need to clean a little bit my model config (provider/model, fallback provider, delegation models,...)

The Mac Studio M5 Ultra won't be out until late September 2026, and that's 5500 to start (30/64) with 96GB of RAM/1TB SSD, an extra 1300 for 36/80, and an extra 4000 for 256GB of RAM. Now imagine buying 3 of those 256GB nodes (or wait until 2027 for the 512GB nodes) and wire them up for RDMA over TB5. *shudder*. If your workplace have a need for something like this, the purchase requisition practically writes itself.
That's a lot of money thrown towards Apple, but at least Mac studios tend to retain their value.

I would hold off on the M6 Mini until someone does a teardown. One of the hidden features of the M4 is their swappable storage socket. If they kept it in the M6 it'll make the machine more desirable, especially for those of us who upgraded their base model M4 with a 3rd party storage module and want an upgrade path (i.e. ME, although migrating machines from Time Machine backups is always a leap of faith).

 

[marcoi]

I just ordered a second intel b70 card to hopefully run on my deg2 dock with the other b70 card already setup in a deg1 dock. This is running on minisforum n5pro.
btw newegg has asrock b70 for 1299 and offers a free aio cooler worth 80 bucks.
The intel version was up to 1600 and the amd 9700 ai 32 gb was also higher 1700.

im hoping rebar works on the second dock else ill have to move both cards to old 5950 amd system current doing htpc duty.

ill update my ai thread once the second card is installed and tested.

 

[WANg]

I just ordered a second intel b70 card to hopefully run on my deg2 dock with the other b70 card already setup in a deg1 dock. This is running on minisforum n5pro.
btw newegg has asrock b70 for 1299 and offers a free aio cooler worth 80 bucks.
The intel version was up to 1600 and the amd 9700 ai 32 gb was also higher 1700.

im hoping rebar works on the second dock else ill have to move both cards to old 5950 amd system current doing htpc duty.

ill update my ai thread once the second card is installed and tested.

Well, I bought an Oculink-to-PCIe card for the N5 Air and had no issues with Re-Bar on that port with the V620 (same 16GT/sec bandwidth constraint), so I don't think the resource allocation issue is port specific per-se. That being said, I don't know if the ministry of finance will be okay with me getting a second dock. I'll need to upgrade the UPS from the relatively weedy BE600M1 to something stronger, drop 330 USD on the dock+PSU, and I'll still need a blower for the V340 (for 50 bucks it's at least good for shits and giggles) and an enclosure for my BC250 (yes I went there).

 

[marcoi]

Well, I bought an Oculink-to-PCIe card for the N5 Air and had no issues with Re-Bar on that port with the V620 (same 16GT/sec bandwidth constraint), so I don't think the resource allocation issue is port specific per-se. That being said, I don't know if the ministry of finance will be okay with me getting a second dock. I'll need to upgrade the UPS from the relatively weedy BE600M1 to something stronger, drop 330 USD on the dock+PSU, and I'll still need a blower for the V340 (for 50 bucks it's at least good for shits and giggles).

i think the same as the card im testing with is 2080 ti and it doesn't support rebar. So hoping the new intel card will be fine. It comes Monday.
Luckily i been getting psu off woot/amazon when they go on sale so i had an extra psu. The deg2 i think was like 200-250 range on sale from amazon. working as a self employed consultant, i can business expense the hardware, so that helps with MOF aspect lol.

 

[MBastian]

If someone put a gun to my head to spend that money I'd probably try to convince him to let me buy a Claude Pro MAX 5x plan for a year and spend the rest on tokens on sites like openrouter to fool around with other models. Mainly to get a feeling what model to run locally when prices come down, if ever. Imho prices are just absurd and $2500 will not get me anything worth my while.
Personally I think all the 32B, 35B and even 120B models (at least at Q4) are just too erratic and need too much hand holding to get complex stuff done.

If you insist on hardware: Two 3090s and 256GB of RAM might let you run ~120B MoE Q8 models with reasonable speed. But I would advise you to use some openrouter tokens first to see what (if any) models are up to the task.

 

[TrevorH]

OTOH, once you have spent the $2500 on Claude, it's gone and you have nothing to show for it. If you spend it on hardware then at least you still have hardware with some residual value left over at the end.

 

[MBastian]

OTOH, once you have spent the $2500 on Claude, it's gone and you have nothing to show for it. If you spend it on hardware then at least you still have hardware with some residual value left over at the end.

My example was a bit extreme, a regular Claude Pro/Standard seat account would probably be just fine. It's rare that I run out of my weekly token allotment at work. The 5 hours limit is a bit more aggravating at times. Image and Video creation is something you can't do with Claude, I dabbled around with Comfy a bit, creation times where acceptable on my 3090. Quality was all over the place, especially for videos. I did not follow up on it as I had no real need.
Bottom line for me: I'd rather get important stuff done fast than waste money and time with fine tuning and hand holding lesser AI models. Especially if my business would depend on it. Dabbling around with local AI models is fascinating and I do it with what limited resources I have at home but unless you throw a lot of money at it it's just not good enough for daily work. Edit: Yes I know, many people use local models for easy tasks to save on token costs for the big models.

 

[WANg]

OTOH, once you have spent the $2500 on Claude, it's gone and you have nothing to show for it. If you spend it on hardware then at least you still have hardware with some residual value left over at the end.

That burns both ways - either the stuff you bought were so cheap that you don't mind it being e-waste (because they essentially are), or you have secondary uses for them besides running unadulterated punchline spigots LLMs. If it's green field spending at 2500-3000, well, don't pay today's money on overinflated items, or rather, don’t buy anything that you don’t mind having a secondary purpose that you can use to recoup your value. For example, L1Tech did release an interesting talk on using the B60/B70 for license-free GPU virtualization...which might justify having one for more than just inferencing. That's the reason why I also mentioned having a Macbook pro 14 M2 Max as a inferencing machine...it's still a plenty powerful laptop for most users.