Price of MSI GB300 workstation (DGX Station) appeared online ~ $97k
Posted by fairydreaming@reddit | LocalLLaMA | View on Reddit | 59 comments
Posted by fairydreaming@reddit | LocalLLaMA | View on Reddit | 59 comments
lucasxp32@reddit
I'd rather just keep paying my tokens like a regular person. It doesn't even make sense unless I had a net worth of over 5 million and I'm feeling spendy. How many billions of tokens somehow has to spend for it to be worth the price, or perhaps the privacy?
I'd totally run a quantized GLM 5.1 on that thing if I had it... But also I like using Gemini Pro 3.1, which isn't open source. Also I don't even think I do a personal use that justifies it. Just the interests alone I'd earn on that money would make me more money per year than my use for it.
ImportancePitiful795@reddit
And many got downvoted in here last year when said it will cost over $56K ($56K was the price of it's predecessor).
At least has full Blackwell SM 10.3 support, unlike the DGX Spark.
Do I want one? YES. Who doesn't.
Can afford one? Nope, without getting a loan against my home. 😥
Mindless_Pain1860@reddit
Even if you gave me one for free, I couldn’t afford to run it in Europe, electricity is too expensive here.
jjshab@reddit
If you guys will finally admit the climate stuff is a hoax then you should be in full-blown revolution over there. Your governments obviously hate you and want you to fail.
They prefer virtue-signaling to Islamists who want you all decapitated and humiliated. Why is the Left so braindead about this. The facts are undeniable at this point.
Avoiding bringing fossil fuels while importing so-called "green" tech from the most despicable regime in China who pollutes 10000x worse than any Western nation is the definition of stupidity.
Oh wait we'll turn off our insanely clean natural gas and clean goal energy plants so we can import solar cells which were built by poisoning the ground for thousands of years and required 109x more carbon footprint to create, because we're GREEN!
How is this all not the most obvious thing in the world? Are you all paid off by Chinese actors to knowingly destroy yourselves or what? It doesn't make any sense whatsoever.
Facts are objective and inarguable. Truth exists, but "your" truth is subjective bullshit that is by definition NOT TRUE. so wake the F up and I just these sociopaths who govern you.
sHuRuLuNi@reddit
Eat a bag of dicks.
powsta@reddit
Also so annoying when they’re closing every nuclear power plant left. Like I actually don’t want to hear it anymore about the environmental compromises in peoples’ lives. Sorry, nope we actually solved that one in 1945 and it was a pretty deal so I’m not going to listen for one second about anything I need to do to reduce emissions when the solution to that problem has been around for 80 years practically.
SnooCheesecakes2821@reddit
Climate change isn`t a hoax u idiot. our prices are high becouse russia started a war to imflate prices for OPEC+. Then america started a war in iran to inflate prices even higher so they can sell their oil at extremely high prices.
"insanely clean natural gass and clean coal energy plants" thats a verry complicated way of saying "i am severly mentally handicapped"
TaylorHu@reddit
Sir this is a Wendy's
No_Swimming6548@reddit
I dunno man I kinda don't want it. What am i supposed to do with a 20 TB GPU?
ImportancePitiful795@reddit
Where you saw 20TB GPU? It has 775GB combined which is good to run 700B+ model like Deepseek R1 at very good speeds and quantization, or 1T model at Q4. 🤔
No_Swimming6548@reddit
As I understand, VRAM capacity is 20.7 TB (288x72) https://developer.nvidia.com/blog/nvidia-blackwell-ultra-for-the-era-of-ai-reasoning/?utm_source=chatgpt.com
CKtalon@reddit
That’s a full rack costing ~10m
No_Swimming6548@reddit
What's the vram of $96K device? Damn these things expensive.
CKtalon@reddit
784GB unified ram
Miserable-Dare5090@reddit
It says 121k on the link not 97k.
I was going to say that 97 was fair when it is like 8 RTX 6000s with the computer built around that kind of monster. But 127 is too high by comparison.
No_Swimming6548@reddit
OK. I would love that too. Though, I don't think my kidneys worth that much
mastercoder123@reddit
Lol that rack is like 10 million for one and nobody is buying 1 rack. They sell those by the datacenter and probably limit orders to like 30 as the minimum which is just fucking insane
ImportancePitiful795@reddit
The article is about the workstation :)
No_Swimming6548@reddit
As I understand, VRAM capacity is 20.7 TB (288x72). Perhaps I am too poor to understand these things...
ItzDaReaper@reddit
Pretty cool you have a home though ngl.
mxforest@reddit
You can build an 8xRTX PRO 6000 system with 768 of VRAM. At this price.
Capable_Site_2891@reddit
That’s also so much slower than this is. For training, you’re talking about 20% the speed. Plus you have to waste time making it parallelise, which is very hard.
GifCo_2@reddit
RTX 6000 pro is not real Blackwell. So there is a huge difference
jhov94@reddit
You could, but then you'd be bottlenecked by the PCIe5 interface.
bigh-aus@reddit
The other option would be a quad H200NVL system with nvlink bridge - that'd probably cost around $130k though.
Honestly showing a list of the price of setups that could run say kimi k2.5 at what speeds would be interesting.
The only downside i can see of the multiple gpu systems is the power requirements.
mxforest@reddit
I have a Sharpie to make 6 out of that 5.
AideRemarkable5875@reddit
https://www.techradar.com/pro/msi-re-launches-usd85-000-nvidia-dgx-station-workstation-with-the-nvidia-gb300-ultra-a-pair-of-400gbe-lan-ports-and-768gb-of-ram
jduartedj@reddit
$97k is wild. For context, you can build a dual RTX 4090 setup for around $5-6k, or even go with 4x used 3090s for about $4k total. You won't get the unified memory or NVLink bandwidth, but for most local LLM use cases — inference, fine-tuning smaller models, RAG — it's more than enough.
I run a single RTX 4080 Super (16GB) and can comfortably do Qwen3 30B at Q4 with decent speeds. If I had $97k to blow, I'd rather build 15+ of those machines and run distributed inference, or just fill a rack with 3090s.
The GB300 makes sense for enterprise/research where you need 288GB unified memory for massive models, but for the LocalLLaMA crowd it's kind of a flex purchase. The consumer GPU path keeps getting better every generation.
mastercoder123@reddit
Why are you even comparing them? You literally CANNOT buy one of these from anyone without purchasing through a company. I tried emailing MSI, supermicro, ASUS, dell and the others and YOU HAVE to have a company email and 'credit card' to buy them, they literally wont sell to you any other way because they arent selling these 1 at a time.
ImportancePitiful795@reddit
Dual 4090/Dual 5090 at $5000-$6000 makes no sense when at $7500 can get a single RTX6000 96GB. 🤔
AdorableAd1817@reddit
What about the Chinese 4090s with 48GB wouldn’t two of those be cheaper than RTX 6000
ImportancePitiful795@reddit
Did they fix the BAR been just 24GB? 🤔
Imho if desperately need CUDA maybe, but at that money we can get 2xR9700 and have €1000 spare.
jduartedj@reddit
$7500? Not in these parts of the world, at least! Running GB300 *at home*.... I don't even know how to finish this sentence. It's way past overkill! It's like buying a Porsche GT3 as a dining table.
ImportancePitiful795@reddit
You do not buy a car to eat dinner. That was daft analogy.
Enough-Cartoonist-56@reddit
You’d be amazed at how many interior design challenges can be solved with a spare supercar:
https://robbreport.com/shelter/home-design/pagani-zonda-racecar-installed-inside-this-miami-condo-2807745-2807745/
ImportancePitiful795@reddit
Anyone having a Pagani Zonda in their living room and not on the street, needs to be arrested for crimes against cars and give the Zonda to me. 😂
Enough-Cartoonist-56@reddit
I agree. I think it’s a terrible use of a Zonda (R?). Though if you’re hanging one between your bedroom and lounge - you probably have one parked downstairs.Â
jduartedj@reddit
Ha fair enough, bad analogy on my part. The EU pricing with VAT reclaim is actually really reasonable at that point, €5600 is firmly in "serious hobby" territory rather than "remortgage your house" territory. I'm in Portugal so I know the VAT game well... still waiting for GPU prices to come down here tho, everything is like 15-20% more expensive than DE or NL
ImportancePitiful795@reddit
That's why buying from other countries eg Poland, Germany and not home country.
jduartedj@reddit
Yeah exactly, cross-border shopping within the EU is the move. I actually ordered my 4080 Super from a German shop because Portugal had like a 20% markup for no reason. Shipping was like 15 euros and still saved a hundred bucks. The VAT is the same within EU so its purely retailer margin differences.
Poland has gotten really competitive for hardware prices too from what Ive seen. PCKomponenty and similar shops are surprisingly cheap.
fairydreaming@reddit (OP)
Where? When I look at RTX PRO 6000 prices (geizhals.eu) I see plots like this:
ImportancePitiful795@reddit
The prices went up this month.
If you look at it, up to end of January the prices with VAT were hovering across the EU at €7500-€7800 with VAT.
During that period €7650 with VAT from a Polish website and £7500 from SCAN UK. Both with full VAT.
Hedede@reddit
Where did you find such prices? Where I live €7700 is the price excluding VAT.
ImportancePitiful795@reddit
Last month was €7650 with VAT from a Polish site.
In UK were £7500 with VAT at SCAN also.
Enough-Cartoonist-56@reddit
Personally, I like your choice of rhetorical device - using a Porsche as a dining table. It effectively communicates your point. Ignore the naysayers.
prusswan@reddit
Might even pay for itself if you apply it to the right business (i.e. trading)
jduartedj@reddit
Thats actually a really good point about the RTX6000, I keep forgetting those exist at that price point now. 96GB unified on a single card is way cleaner than juggling multi-GPU setups with tensor parallelism. The bandwidth alone makes it worth it for inference.
And yeah fair enough on the GB300 use case.. for a small office scenario with actual concurrency needs thats a completely different calcuation than what most of us hobbyists are doing. I was thinking purely from a "tinkering at home" perspective which is probably not the target market lol
prusswan@reddit
They come in sets of 8 too: https://servers.asus.com/products/detail/overview/XA-NB3I-E12
If these become standard issue in data centers over the next few years, I know I won't be able to match that so I will focus on getting the most out of the Pro 6000s (all things considered I see these as the sweet spot for serious individual users, for now at least)
mastercoder123@reddit
Those are HGX systems and cost like $400,000 each. They also are impossible to get for anywhere near MSRP because of ram shortages and the lead time on someone like supermicro is months at least. Those arent sold to us plebs, you cant even email supermicro support without a work email anyways
TechnicalGeologist99@reddit
This isn't DGX Station. It's a server rack with B300s
DGX Station is a standalone product with 2 Grace CPUs and 1 Blackwell 300
MayhemPayload@reddit
NGL, the idea of having 748gb of VRAM makes me drool a little.
fairydreaming@reddit (OP)
Found the same item here for $85k: https://www.centralcomputer.com/msi-ct60-s8060-nvidia-dgx-station-cpu-memory-up-to-496gb-lpddr5x-nvidia-blackwell-ultra-gpu-1x-10-gbe-2x-400-gbe.html
sputnyek@reddit
This may be more attractive than buying a threadripper with 4xblackwell rtx 6000 maxq (around $80K). For inference it may not matter as much but for training or building small models from scratch I wonder how they compare for that use case?
128G@reddit
$97,000 is a lot of money.
Direct_Turn_1484@reddit
You could probably get that easy by just selling three kidneys.
mxforest@reddit
What is the combined bandwidth of the RAM modules?
fairydreaming@reddit (OP)
CPU Memory 496GB LPDDR5X | 396 GB/s
mxforest@reddit
When using vllm, does it see 288 or 784? Will it be able to serve DeepSeek? If this gives a decent throughput with batch, this will be an easy buy for my team.
Investolas@reddit
I see this as for serving inference for a small tech business. Though at what point does it make more sense to use cloud conpute idk.