Nvidia shelves 2026 new card releases

Torquinox

The good news keeps coming. Apparently NVidia shelved every new gpu card release for 2026. https://tech-insider.org/nvidia-skips-2026-gaming-gpus/ 
 

Okay, so it's September, but the 24Gb 5080 Super expected in December is apparently also shelved. IDK really what it means except that, coupled with spiraling hardware price increases, folks are going to have a harder time upgrading or replacing systems due to cost. 
 

I've read a lot of articles speculating about what will happen if NVidia abandons the consumer gpu card market because it represents a small portion of their total income. Some believe NVidia will provide GPU as a subscription service. I don't like that idea, but I could see that happening.

 

Nvidia Is not alone in pulling back from the consumer market. Micron also ended their Crucial line of consumer ram kits because data center build out is far more profitable. Interesting times!

Padone

The data centers are always payd by us, the consumers. It is us who have control of the market. If Micron abandons consumers for data centers it means consumers pay more for data centers than for new hardware. Personally I'm good with old hardware until I'm forced to change it by the applications I use. I can also stay outdated with applications until I'm forced to update them for some new features I need. Actually my old 1060 card works fine with DS4 and I see no reason to change that, especially with the actual market prices.

@3DIO Small correction, avoid -paying- for cloud services, we can use them for free there's no angle for them then.

p.s. However, AI services are truly impressive, apart too much censorship that is really blocking creativity. I wasn't able to get a girl wearing a bikini, or a night gown, really nothing NSFW but the AI decided so.

WendyLuvsCatz

while I use AI, it is locally

cloud based is a big no for me

I suspect that is actually the case with most people

and the other uses for data centres such as monitoring people are not popular

it will bite them in the bum, might take a while for the crash, but when it happens it will be spectacular 

Padone

As I understand it, using AI locally requires at least 16G vram and 64G ram, a powerful cpu, better if it's a nvidia gpu. Just to have basic results not comparable with those data centers can provide. Not something everyone can afford with the actual market prices.

Please correct me if I'm wrong, I'd like to.

WendyLuvsCatz

I use low VRAM versions in Pinokio

8GB VRAM and 32GB RAM versions 

it's not fast but faster than 3D rendering 

Padone

Thank you for letting me know how it goes on your side. That can be a reasonable upgrade for my hardware. But are the results good ? I hear local AI only provides basic results, not comparable with data centers and barely useful.

Torquinox

I believe the vast amounts of money going into the ai industry is short-circuiting how the market usually works. While the whole ai build out is probably a market bubble, it's not clear if or when that will pop.
 

Meanwhile, there is this wholesale shift going on in the marketplace. It seems like a brute force re-engineering of how the computer market works, where the biggest players move all the necessary hardware out of our hands and shunt it into paywall-regulated data centers.

I think people have already accepted the computing cloud as a fact of life. Those of us who want to keep everything local are outliers.

Valiska

Padone said:

As I understand it, using AI locally requires at least 16G vram and 64G ram, a powerful cpu, better if it's a nvidia gpu. Just to have basic results not comparable with those data centers can provide. Not something everyone can afford with the actual market prices.

Please correct me if I'm wrong, I'd like to.

I have 48G RAM and 8G VRAM. I can run moderately capable LLMs in LM Studio or Ollama.

Useful stuff so far: 

• They've managed to tell me about some recent science I wasn't aware of (for writing science fiction), and have managed to usefully summarize some stuff I fed them.
• They sometimes can partially make up for the growing gross deficits of Google search, in that they might be able to identify information on the basis of a description. That's helpful if you can describe what you want but don't know the right terms of art to call it up.
• I've managed to get one satisfactory science fictional image out of the image generators at Nightcafe (because I haven't gotten around to setting up any local image generation, and probably won't try that for a while yet).
• I got a handy summary of one batch of literary information I fed it.

Not so useful:

• I haven't yet managed to get one to take over some text transfer and editing (too bad, it's a nuisance job that mostly involves putting back together English-and-tables jigsaws from some badly constructed PDFs. Time-consuming but not very interesting).
• Neither LLMs or image generators actually understand how the world works, or what anything means. Online image generators, last time I tried, could not be made to make a successful image of science fictional objects rarely represented in existing art.
• Processing time and heat generation wasted flattering the user.

The LLMs' ability to write grammatical English is impressive.

The most capable LLMs I can run start losing track of the logic of a conversation at somewhere between 5-8 pages, give or take. That means LLMs often don't turn out to be a great timesaver, because if you're poking at something of much complexity, you have to keep stopping to make sure you have summaries that will allow the LLM to keep going a little farther. Comes a point where it just loses track of what it's supposed to be trying to do.

Sometimes it never lands in the right quadrant and then it doesn't do much good even in short form.

So far for me it's an interesting toy, but mostly not yet a particularly efficient tool for many purposes. Maybe it will soon become so; I don't think it would need too much improvement to get over the hump, for me. From what I've seen so far, I am expecting it to be pretty interesting for post-processing.

The only kind of question I'll ask a publically accessible AI is a question I don't care if everyone, including radically untrustworthy giant entities, knows I asked.

Valiska

Torquinox said:

I believe the vast amounts of money going into the ai industry is short-circuiting how the market usually works. While the whole ai build out is probably a market bubble, it's not clear if or when that will pop.
 

Meanwhile, there is this wholesale shift going on in the marketplace. It seems like a brute force re-engineering of how the computer market works, where the biggest players move all the necessary hardware out of our hands and shunt it into paywall-regulated data centers.

I think people have already accepted the computing cloud as a fact of life. Those of us who want to keep everything local are outliers.

If I can't afford the next gaming laptop for computer graphics, I'm buying something lower-end, switching to Linux, coding to amuse myself, and reverting to traditional artistic media. I can use the existing Daz/Poser library for quick line renderings, to save some time in preliminary sketches. I will largely drop offline, rather than move farther online.

jmucchiello

I recommend reading The Tech Report youtube channel. They constantly talk about the inevitable AI bubble and how we'll all end up paying for it. There are six or so companies trying to ship 10 tons of canaries in a 5 ton cargo hold by having half the canaries in flight at all times. Many of the RAM foundaries have commitments through 2028 that are exclusively AI data center commitments. And since they are so lucritive, they aren't spending on consumer grade anything in the near term. This is why NVIDIA isn't releasing newer video cards. They cost too much to manufacture (well, assemble technically) given the ton of cash being dumped into AI.

SnowSultan

I have what's probably considered a decent system for running AI locally (nvidia 4090, 128 GB RAM, but an older CPU). I can tell you a bit of my experiences with using generative/image-related AI (I don't run LLMs on mine, I use ChatGPT for that).

Useful stuff:

• If you can figure out ComfyUI - which is a whole other frustrating ordeal - it is excellent for making 3D renders look realistic. You can turn a test render into a photo with little trouble.
• Makes environments easily, no need for huge meshes, skyboxes, or scattering systems.
• You can replace clothing, hair, and other things, as well as making effects that would be difficult in 3D, like realistic water, smoke, clothing folds, and soft skin dynamics.
• Depending on the models used, you can restyle your image easily into cartoon or anime (with mixed results), or cinematic styles like 1950s, noir, vintage photos, vector, or pixel art.
• It's free if you do it at home and uncensored (which virtually no online service is nowadays).

 

What it doesn't do well:

• Generate highly accurate objects from scratch with no references. You'll often get very odd results when asking for objects that are not common or that are creative, like a demon tail. This is where using 3D in conjunction with AI is important; the 3D can be the reference to help guide the AI (it will still often mess it up, which is why it's good that you're not paying for each attempt).
• Consistency with characters or environments can be tricky to achieve, it can be done but it's a bit random. You can get much better consistency using a online service like OpenArt or Higgsfield.
• Local video generation doesn't come close to what can be done online. Almost every cool cinematic video with lots of action that you see done with AI are done online, likely using expensive video models like Seedance. You can make videos at home, but they can take a long time, run your graphics card hard, and aren't going to give you Hollywood results.

 

I have a lot of examples of both local render enhancements and Seedance fight scene videos on my deviantArt page, as well as ramblings in my journals about my daily tests if you want to see any examples. Hope that can be of some help.  :)

Torquinox

@3DIO I hope you're right. That's an area where I'm happy to be wrong.

Padone

@Valiska @SnowSultan Thank you for sharing your experience, at least this gives us some info about usage and limitations, even with high-end hardware. I think I'll try it myself starting with my actual limited hardware, just to see how it goes. I understand Comfy UI can also use optimized models that are good for low vram.

kyoto kid

...this is worse than the crypto craze of a decade ago.  Back then it was just GPUs which was bad enough, now it is pretty much everything,

The AM5 upgrade is now 3.400 USD. may as well buy a lotto ticket.

Even backstepping to an AM4 Ryzen 9 5900XT and 64 GB DDR4 3200 is ridiculous.

 

 

Torquinox

jmucchiello said:

I recommend reading The Tech Report youtube channel. They constantly talk about the inevitable AI bubble and how we'll all end up paying for it. There are six or so companies trying to ship 10 tons of canaries in a 5 ton cargo hold by having half the canaries in flight at all times. Many of the RAM foundaries have commitments through 2028 that are exclusively AI data center commitments. And since they are so lucritive, they aren't spending on consumer grade anything in the near term. This is why NVIDIA isn't releasing newer video cards. They cost too much to manufacture (well, assemble technically) given the ton of cash being dumped into AI.

The pattern of available information supports what you're saying, too. I'll look into the tech report. Thanks. 

donniekeidic

I'm actually kind of surprised I haven't heard of organized crime rings breaking into data centers. Or maybe they are all heavily guarded.

jmucchiello

donniekeidic said:

I'm actually kind of surprised I haven't heard of organized crime rings breaking into data centers. Or maybe they are all heavily guarded.

You would have to know what you're doing to rob a data center. As soon as you take the plug out of the wall, a dozen or more people probably are immediately alerted. And that assumes you got into the building undetected.

Next you have to move a bunch of rackmount equipment because a smash n grab approach can't possibly yield much.

It would have to be an inside job.

Maybe it could the next Ocean's movie.

Torquinox

kyoto kid said:

...this is worse than the crypto craze of a decade ago.  Back then it was just GPUs which was bad enough, now it is pretty much everything,

The AM5 upgrade is now 3.400 USD. may as well buy a lotto ticket.

Even backstepping to an AM4 Ryzen 9 5900XT and 64 GB DDR4 3200 is ridiculous.

I agree! Ridiculous!

Dareshiranu

They are going to discover that consumers will not be satisfied with thin clients and sub-standard services provided to them (for enormous recurring fees, of course) by the inevitable cloud services which themselves will demand ever increasing amounts of hardware and thus increasing fees irrespective of the greed of corporations. It will become the inverse of the economy of scale.

Padone

INCREDIBLE .. I just got ComfyUI and RevAnimated and everything is working flawless. I was expecting huge generation times with my old 1060, instead it just takes a few seconds. There's a couple of caveats as one has to get the portable cu126 for ComfyUI and the pruned version for RevAnimated. I'm reporting my experience here for anyone interested, with a old 1060 no new hardware needed at all.

The attached image can be dropped into ComfyUI as it stores the workflow parameters.

  AI FOR 1060 CARDS

  GET COMFY UI
  1. The ComfyUI installer doesn't work for 1060, get the portable version with cuda 12.6 aka cu126.
  2. Unzip the portable into "C:\", it will not work under "Program Files".
  3. To run use "run_nvidia_gpu.bat" adding --lowvram, do not use fast_fp16 as it doesn't work on 1060.

  GET REV ANIMATED
  1. To generate images get "Rev Animated V2 pruned" from CivitAI, the pruned version works fine with 6gb.
  3. Place "revAnimated_v2Pruned.safetensors" into "ComfyUI\models\checkpoints".
  4. Run ComfyUI and ask AI how to correctly set the KSampler for Rev Animated, as it wants its own parameters. Or drop the attached image.

 

Torquinox

Of course, applicable corporations are cashing in on the AI build out. And there is a noticeable shortage of computer components for the consumer market as a result. That means prices are high and may well go higher, at least for a time. But the build out is not an infinite event.

 

Eventually, there should be enough data centers to do whatever these companies are actually trying to do. Or something unexpected may happen... I'm not really sure what would happen then. Some of the wilder speculation here seems a bit unlikely - not impossible, but probably not the way things will actually happen. Actual events have a way of turning out different than expected.

WendyLuvsCatz

...or someone will discover a technology that will make it all obsolete 

Torquinox

@WendyLuvsCatz That would be unexpected!

Padone

Well if you want the unexpected, China is developing organic chips, using real neurons. That means ram vram will be obsolete soon enough, at least for data centers. And there's already a commercial line it's not a prototype.

https://corticallabs.com/cl1.html

kyoto kid

...Direct Neural Interface. 

Just plug your computer into a USB like jack in the back of your neck that links to a neural net implant and "think" modelling, art, music, prose, or whatever.

Even the "Make Art" button would be obsolete

Richard Haseltine

Torquinox said:

Of course, applicable corporations are cashing in on the AI build out. And there is a noticeable shortage of computer components for the consumer market as a result. That means prices are high and may well go higher, at least for a time. But the build out is not an infinite event.

 

Eventually, there should be enough data centers to do whatever these companies are actually trying to do. Or something unexpected may happen... I'm not really sure what would happen then. Some of the wilder speculation here seems a bit unlikely - not impossible, but probably not the way things will actually happen. Actual events have a way of turning out different than expected.

How long do the parts last, though? Replacing them could still put a strain on the suply chains for a while, though presumably companies will be looking at expanding production to meet the (currently) increased demand.

RAMWolff

I was experiencing black screens.  I finally zeroed in and realized that my memory modules were slowly dying out on me.  Never in all the decades using computers have had that happen.  Anyways, after allot of sleuthing, I found a "deal" on Newegg for 32 gigs of DDR5 RAM for just a little over $400.00, a year ago that same kit was about 100 bucks.  I just went out to check on priceing in hopes it was down to something more affordable to give my rig a full 64 gigs but NOPE it's up another100 bucks.  This is insanity!  

jmucchiello

Richard Haseltine said:

Torquinox said:

Of course, applicable corporations are cashing in on the AI build out. And there is a noticeable shortage of computer components for the consumer market as a result. That means prices are high and may well go higher, at least for a time. But the build out is not an infinite event.

 

Eventually, there should be enough data centers to do whatever these companies are actually trying to do. Or something unexpected may happen... I'm not really sure what would happen then. Some of the wilder speculation here seems a bit unlikely - not impossible, but probably not the way things will actually happen. Actual events have a way of turning out different than expected.

How long do the parts last, though? Replacing them could still put a strain on the suply chains for a while, though presumably companies will be looking at expanding production to meet the (currently) increased demand.

Three years tops. These are high end GPUs and the demand level is like pro-gamer level demands for high end graphics cards. A top of the line spec card today would be mid-tier after 3 years. So they dump billions of dollars on the equipment for the facility and then they do it again year after year keeping the facility state-of-the-art. It just isn't sustainable unless customers are pumping trillions of dollars into the AI companies.

Luckydraft

jmucchiello said:

Richard Haseltine said:

How long do the parts last, though? Replacing them could still put a strain on the suply chains for a while, though presumably companies will be looking at expanding production to meet the (currently) increased demand.

Three years tops. These are high end GPUs and the demand level is like pro-gamer level demands for high end graphics cards. A top of the line spec card today would be mid-tier after 3 years. So they dump billions of dollars on the equipment for the facility and then they do it again year after year keeping the facility state-of-the-art. It just isn't sustainable unless customers are pumping trillions of dollars into the AI companies.

Improvements in hardware, specially GPUs, have noticeably slowed down in the last few generations. Top of the line card 3 years ago is still top-tier today - the 4090 was released 4(!) years ago and it is only beat by a 5090, not even a 5080. And it has more RAM than the 5080 because Nvidia keeps delaying the launch of a 5080 with 24gb. In any case, all big cloud providers are working on their own chips that are more efficient and cheaper than buying from Nvidia.

They'll last less than 3 years, just not due to the cards specs... its because these are just devices running 24/7/365, and they've been rushed and cobbled together to proritize providing as much capacity as fast as possible to meet demand.

Something that's worth keeping in mind is that while training new models is not profitable, running them is very profitable. So it's very much worth it for them to keep replacing any hardware they need to as long as the models are solving any real-world need. And some of them are starting to. I would be surprised if anything improves throughout 2027 at all and I imagine it'll only get better once more suppliers enter the market.

csaa

It's interesting how the flow of discussion in this thread has bifurcated.

On one hand we have people trying to grasp the big picture: how the overarching narrative surrounding Big Tech has become disjoint, even sinister. In our modern age, the loss of personal agency is a frightening fate, be it to illness or to a soulless societal system. One could even venture to say that it ranks high up there along with cancer, death and taxes. On the other hand we have folks that are focused on what they can accomplish given the resources they have on hand. Practical and grounded -- problem-solving driven -- celebrating, it seems, the small personal wins.

First World Problems? Hmm.  Hope lies in the fact that the end to our Summer of Discontent 2026 may just be around the corner. Or not. 

Cheers!

garrett_3d

As long as people continue to use AI, in whatever form it takes, then this situation will continue. The only solution, bar an entire collapse of the internet, is to ban AI completely worldwide.

AI is not the answer - it is only as intelligent as the muffins that program it. Humans aren't that intelligent. We're screwed.