Sunday, September 13th 2026

AI Firms Are Reportedly Buying NVIDIA RTX 5090 GPUs in Bulk for Server Use

New photos shared by HKEPC allegedly show pallets of brand-new, boxed GeForce RTX 5090 cards being unpacked at AI server assembly sites. Notably, these firms are not using the RTX PRO 6000 or dedicated data center accelerators NVIDIA actually markets for that work. The cards in question are standard retail RTX 5090s, the same cards gamers have struggled to find at anywhere close to MSRP, being pulled straight from their packaging and slotted into multi-GPU AI rigs. It must be mentioned that some elements of these photos appear to be altered, or at least edited with AI, so take these with a bit of caution.

Despite that, the current situation is not entirely new. During the crypto boom, miners bought up GeForce cards in bulk while gamers got squeezed out, and NVIDIA eventually responded with Lite Hash Rate models that throttled mining performance without touching gaming specs. That restriction got fully bypassed within about a year. This time, the workload is AI instead of mining, but the outcome for gamers looks familiar. As a result, retail RTX 5090 pricing has reportedly climbed past $5,000 on some listings, more than double its $1,999 launch price. Industry analysts even warned about this possibility months ago, and we now have a good idea why the RTX 5090 is at the price it's at.
Notably, the cards in these shared photos aren't even RTX 5090D or v2 variants, the China-specific SKUs NVIDIA already nerfed for AI workloads to comply with export rules. That means NVIDIA has already proven it knows how to limit a GeForce card's AI performance while keeping gaming performance intact, yet hasn't applied that same restriction to the standard RTX 5090 sold everywhere else. Given how central AI has become to NVIDIA's business, it seems unlikely the company will step in this time either.
Sources: HKEPC on X, Reddit, via Videocardz
Add your own comment

112 Comments on AI Firms Are Reportedly Buying NVIDIA RTX 5090 GPUs in Bulk for Server Use

#1
theglaze
5090 sold out from all (online) retail outlets in USA and Canada, so yeah, I believe it.
Posted on Reply
#2
Crackong
Is there a way to 'SLI' 5090 ?
Otherwise it is just 32GB of VRAM, and the 'AI firm' will get many timeouts trying to run anything beyond 30b ...
Posted on Reply
#5
AusWolf
Just release cards without Tensor cores. You could still run FSR 3, have the superior RT performance, and it would be cheaper, too.
Posted on Reply
#6
Visible Noise
Imagine, Reddit posts aren’t truthful. Who woulda thought?
Posted on Reply
#7
Fatloader
CrackongIs there a way to 'SLI' 5090 ?
Otherwise it is just 32GB of VRAM, and the 'AI firm' will get many timeouts trying to run anything beyond 30b ...
You can use multiple GPU's VRAM in more ways than just SLI.
Visible NoiseAI slop. Look at the model numbers.

I noticed that too. AI is getting scary good these days. The one that really gave it away for me was the upside-down "5" on one of them.
Posted on Reply
#8
AusWolf
Visible NoiseAI slop. Look at the model numbers.

AI creating images of AI farms. That's some Inception level stuff there. :laugh:

Now the question is, why?
Posted on Reply
#9
Chaitanya
Darmok N JaladAll roads lead to NVIDIA.
Not atleast according to Leather jacket wearing man who will deny everything.
Posted on Reply
#10
Hecate91
What a shocker, a gaming card priced out of reach of gaming consumers is no longer a gaming card when only AI companies can buy them.
It's the crypto scam all over again, except the bubble is much bigger this time.
Posted on Reply
#11
AusWolf
FatloaderI noticed that too. AI is getting scary good these days. The one that really gave it away for me was the upside-down "5" on one of them.
And the R instead of 9 on another one... Thinking about it, a GeForce SORO would sound cool. Maybe next gen will change the naming scheme? :D

Edit: Or maybe not. Looking at GPU prices is SORO-ful enough. :laugh:
Posted on Reply
#12
Bigshrimp
Lol... The images do look doctored as others have mentioned. Looks all over the place on those product box labels.
Posted on Reply
#13
wolf
Better Than Native
Notably, the cards in these shared photos aren't even RTX 5090D or v2 variants, the China-specific SKUs NVIDIA already nerfed for AI workloads to comply with export rules
More like notably, the cards in the shared photos aren't even real.
Posted on Reply
#14
wNotyarD
While for whatever my opinion's worth I don't discredit or disagree with what the news here is supposed to convey (after all, probably most 5090s are indeed being used for such ends), using that bad AI image is bad journalism.
The first and last pictures look more real, though.
Posted on Reply
#15
Rulumi
AusWolfJust release cards without Tensor cores. You could still run FSR 3, have the superior RT performance, and it would be cheaper, too.
That would of have worked until around 2019 or 2020 as an alternative model, but nowdays almost no company and OEM would be willing to purchase such GPU. Specially now that is being pushed in a small way across devices, NPUs and integrated graphics even in some cases for small models, across computers and phones.
Posted on Reply
#16
Dr. Dro
CrackongIs there a way to 'SLI' 5090 ?
Otherwise it is just 32GB of VRAM, and the 'AI firm' will get many timeouts trying to run anything beyond 30b ...
LLMs are generally not limited to a single GPU's memory pool, mirroring isn't necessary.
DrUsmanSaleemNotably, the cards in these shared photos aren't even RTX 5090D or v2 variants, the China-specific SKUs NVIDIA already nerfed for AI workloads to comply with export rules. That means NVIDIA has already proven it knows how to limit a GeForce card's AI performance while keeping gaming performance intact, yet hasn't applied that same restriction to the standard RTX 5090 sold everywhere else. Given how central AI has become to NVIDIA's business, it seems unlikely the company will step in this time either.
A note on this, the D v2 is significantly slower than the global and D variants owed to its reduced memory bandwidth. The reduction from 32 to 24 GB also reduced it from 512 to 384-bit bus, reducing memory bandwidth accordingly. Its memory is 448 GB/s slower in addition to the tensor performance limits imposed on the "regular" D variant, so it suffers accordingly, gaming included. This is how they got the card below the performance level allowed for export.
Posted on Reply
#17
zenlessyank
Phake phucking photos. I feel a different part of my brain getting angry.
Posted on Reply
#18
AusWolf
RulumiThat would of have worked until around 2019 or 2020 as an alternative model, but nowdays almost no company and OEM would be willing to purchase such GPU. Specially now that is being pushed in a small way across devices, NPUs and integrated graphics even in some cases for small models, across computers and phones.
Well, true. I did actually have a 1660 Ti and it was great.
Posted on Reply
#19
Tess
Visible NoiseAI slop. Look at the model numbers.

Good catch! I wonder how and why @DrUsmanSaleem missed that?!?
Posted on Reply
#20
AusWolf
TesselatorGood catch! I wonder how and why @DrUsmanSaleem missed that?!?
And I wonder why people would create such pictures.
Posted on Reply
#21
zenlessyank
AusWolfAnd I wonder why people would create such pictures.
Clicks and ego.
Posted on Reply
#22
matar
F this shit what happened to limit one per household days we don't want one per customer household is ok every household should be able to buy 1 gpu at MSRP. but its always the rich get what they want and they forget who supported them back in the days when they were no one i have been buying Nvidia since 1999 so since day 1,
Posted on Reply
#23
halcyon
I'm all for this. The situation is already fubar. Connector-burn RTX5090s now going for 3-4x the price doesn't change anything for the ordinary consumers
However, the more hw the final stages of AI bubble spends on, the bigger the crash and the bigger the discounts at the end. Bring it on!
Posted on Reply
#24
A Computer Guy
FatloaderYou can use multiple GPU's VRAM in more ways than just SLI.

I noticed that too. AI is getting scary good these days. The one that really gave it away for me was the upside-down "5" on one of them.
The shiny floor with no tile seams and the dudes that look like they are doing nothing also clues.
There is also the reflection of a stack of boxes on the floor has some inaccuracies.
Posted on Reply
#25
AusWolf
A Computer GuyThe shiny floor with no tile seams and the dudes that look like they are doing nothing also clues.
Of course they're doing something! They're installing GPUs into industrial printers, obviously! :rolleyes: And the green goo on the floor is glue to make sure they can't run away before their shift ends... Or rather, ever. :D
Posted on Reply
Add your own comment
Nov 13th, 2026 05:43 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts