GPU prices — is the worst behind us ?

I think we can safely say the worst is still to come....mid 2024 to 2025 was the lull before the storm hit again this time with unprecedented ferocity.

At least until 2030 based on current VRAM supply agreements. The antidote to this will be more companies bringing capacity online (and not being banned by the US government).
 
I think we can safely say the worst is still to come....mid 2024 to 2025 was the lull before the storm hit again this time with unprecedented ferocity.
Even if the worst passed us, prices are never going to return to what they were, and that's true even if the AI bubble popped tomorrow. These prices are the new normal for top end consumer cards. They'll go down in price through inflation devaluing worth, but not due to prices being lowered.
 
You've been able to run local LLMs with some good models on a 16GB card for a while now, there are threads about it.

Alibaba's Qwen in particular has been cited for driving the huge spike we've seen recently. Alex Karp's speech about owning your data by localizing AI has also been a big contributor.
 
its too bad Intel essentially killed off Arc....they would be the one saving grace right now. They made such good strides between A and B.
I wouldn't be surprised, Arc just was not selling against Nvidia, even AMD is struggling, but that is mostly of their own doing since they invested a whole lot more R&D money into AI related stuff and the CPU side of their business.
 
I wouldn't be surprised, Arc just was not selling against Nvidia, even AMD is struggling, but that is mostly of their own doing since they invested a whole lot more R&D money into AI related stuff and the CPU side of their business.

i think the major problem was RT and the idea that people need more than 1080p.......

outside of a few games RT is meh still and you do not need 4k t ohave good quality visuals and good framerates.
 
I wouldn't be surprised, Arc just was not selling against Nvidia, even AMD is struggling, but that is mostly of their own doing since they invested a whole lot more R&D money into AI related stuff and the CPU side of their business.
The ARC Battlemage gaming cards seem to move when NV gets out of the way. They were in and out of stock all the time in early 2025 after NV discontinued the 4060 ahead of the launch of the 5060 and ran the channel completely dry. Judging from the stock history on nowinstock.net, it looks like the same thing is happening now. Drops every couple of days, sells out in a few hours or the next day. Not a great show for Intel when they don't sell much when they have competition, though given prices on everything else I can see why they're moving at $330. $400 5050s, $450+ for a 5060, $500 9060XT 8GB, etc.

I don't think ARC is dead. It's just shifted to AI and iGPUs. The ARC G3 got a lot of good reviews in the handheld space, and Panther Lake is their best gaming iGPU yet. They're sort of ahead of AMD, especially in the handheld space. The big Strix/Gorgon Halo APUs are much more powerful, but also use a lot more power and their performance/watt is lower. So basically they're making a play for gaming on battery. Word is the next dGPU was cancelled months ago, but they're still working on drivers, did a multi-frame-gen update a few months back, and don't really seem to be slowing down on the mobile iGPU side. If they get some traction maybe we'll see another batch of dGPUs someday.

I suspect what happened to the next batch of ARC cards is they made a big shift towards AI and did a radical design change that probably makes them unsuitable for a gaming card. The ARC B-series pro cards are mostly AI cards. Or at least that's why you'd buy them. They're pretty "meh" for traditional workstation tasks, but have solid price/perf vs. AMD and NV. Or did. Prices are way up lately. The Intel GPU D270 uses DDR5, 120-480GB on a 1280-bit bus with 1.5TB/s of memory bandwidth. 350W, fits in a standard 16X PCI-e slot. The chip is 32 Xe3p cores, so the same core count as an ARC A770 or ARC Pro B70. They call it a "datacenter" card but the specs look like they tried to make the cheapest local AI card they could for running good size models. So lots of less expensive DRAM instead of HBM, pack a ton of it onto one card, and make the whole thing in a package that easily fits in a normal computer. I won't be a bit surprised if some of their board partners slap fans on these and put them in retail boxes. The problem with using DRAM on a gaming card is you need a lot more of it for the same bandwidth. If 120GB is a hard minimum, even if they cut the bus width in half it'd still be a 60GB card.
 
i think the major problem was RT and the idea that people need more than 1080p.......

outside of a few games RT is meh still and you do not need 4k t ohave good quality visuals and good framerates.
I've never played at 4k. I rocked a CRT well into the LCD era. However, the 1080 LCD monitor that I did get was not high end, but it looked so much better than the CRT monitor. Now, I have a 1440 monitor that makes the 1080 monitor look worse than it made the CRT look! 1440 is good enough for me, and not so hard to run that I need a high-end card to do it, but I'm not going to drop back to 1080p unless I cannot afford 1440 anymore. It could be 2035 or later before I even consider moving up to 4k. I like solid visuals, but I also like really fun games. I don't want to play a game because it looks gorgeous and has little else to offer.

RT has been a tough road for GPUs. I like it, but when it actually adds to the game and not added as a tacky add on because a game dev wants to keep up with every other game that has RT. RT can add that touch of realism that makes the environment more immersive.
 
The ARC Battlemage gaming cards seem to move when NV gets out of the way. They were in and out of stock all the time in early 2025 after NV discontinued the 4060 ahead of the launch of the 5060 and ran the channel completely dry. Judging from the stock history on nowinstock.net, it looks like the same thing is happening now. Drops every couple of days, sells out in a few hours or the next day. Not a great show for Intel when they don't sell much when they have competition, though given prices on everything else I can see why they're moving at $330. $400 5050s, $450+ for a 5060, $500 9060XT 8GB, etc.

I don't think ARC is dead. It's just shifted to AI and iGPUs. The ARC G3 got a lot of good reviews in the handheld space, and Panther Lake is their best gaming iGPU yet. They're sort of ahead of AMD, especially in the handheld space. The big Strix/Gorgon Halo APUs are much more powerful, but also use a lot more power and their performance/watt is lower. So basically they're making a play for gaming on battery. Word is the next dGPU was cancelled months ago, but they're still working on drivers, did a multi-frame-gen update a few months back, and don't really seem to be slowing down on the mobile iGPU side. If they get some traction maybe we'll see another batch of dGPUs someday.

I suspect what happened to the next batch of ARC cards is they made a big shift towards AI and did a radical design change that probably makes them unsuitable for a gaming card. The ARC B-series pro cards are mostly AI cards. Or at least that's why you'd buy them. They're pretty "meh" for traditional workstation tasks, but have solid price/perf vs. AMD and NV. Or did. Prices are way up lately. The Intel GPU D270 uses DDR5, 120-480GB on a 1280-bit bus with 1.5TB/s of memory bandwidth. 350W, fits in a standard 16X PCI-e slot. The chip is 32 Xe3p cores, so the same core count as an ARC A770 or ARC Pro B70. They call it a "datacenter" card but the specs look like they tried to make the cheapest local AI card they could for running good size models. So lots of less expensive DRAM instead of HBM, pack a ton of it onto one card, and make the whole thing in a package that easily fits in a normal computer. I won't be a bit surprised if some of their board partners slap fans on these and put them in retail boxes. The problem with using DRAM on a gaming card is you need a lot more of it for the same bandwidth. If 120GB is a hard minimum, even if they cut the bus width in half it'd still be a 60GB card.
Intel may have an angle with laptop GPU aspect because that has been a sore spot for laptops for a long time. Sure, some laptops have discreet mobile versions of GPUs added, but a GPU integrated to keep costs down is likely preferred by many over having discreet added. I bought my wife a laptop earlier this year that beats the pants of the other laptop I bought her years ago with a discreet GTX 1050 in it. The integrated AMD GPU can play Hogwarts! I wouldn't want to attempt that on the GTX 1050.

Making cheap AI cards definitely has a place at the AI table. People are paying ridiculous prices for 3090 cards because of memory capacity for AI stuff. I think Raja underestimated how hard it would be to grab marketshare from Nvidia when he set out to make the Arc series.
 
1793273083085.png

“GPU prices are getting worse. In this video we look at how much worse they've gotten in the last month, and explore the impacts of the RTX 5090 being discontinued, specifically on RTX 5080 pricing.”
 
I have yet to find a reason to upgrade my 3080 12GB card. It works fine for the limited gaming I do.

For AI workloads I use hardware designed for it.
You lucky you have the 12GB version. All I could score was a 10GB one. It's sitting in my office collecting dust and I'm still on a 3090 for AI work.
 
If the AI bubble burts prices are not going to come down. Maybe temporarily because the market might be flooded with used GPUs but once next gen cards come out they will be priced high because people have proven to Nvidia they will pay those prices.
 
There are still large amounts of society with stupid money and that "money is no object" mindset that will continue to support "jacket man' for years to come
I would say not a "die off"... but taking a step back to what it was like in the 80's-to-mid-90's.. Having a PC was not even a thought for households without too much disposable income... owning your own PC was a niche group...

(Source February '95 Computer Shopper for example)
View attachment 833342

Yep. There isn't a shortage of people that can afford spending that much if they really want it. Also plenty of people that really shouldn't spend that much but still do.

18% of households in the USA have a net worth over a million dollars, and about 1/15 individuals in the USA have a net worth over a million dollars.
 
CME/CBOE should introduce options on GPU pricing by class...would have a big market. Surprised one of the big 'speculation' platforms like Kalshi or Poly have not hopped on to this...they have basically no regulation and can bet on anything.
 
Yep. There isn't a shortage of people that can afford spending that much if they really want it. Also plenty of people that really shouldn't spend that much but still do.

18% of households in the USA have a net worth over a million dollars, and about 1/15 individuals in the USA have a net worth over a million dollars.

Net worth is a useless number when you're talking about purchasing power. It only represents a theoretical amount of money you do have if you sold all your assets. It is hardly cash in the bank. My net worth is probably in that range and, because of my situation, I have the cash where I COULD afford a 5090 but I'm not stupid enough to blow that kind of cash on it. My BMW motorcycle cost me less than a 5090 does right now.
 
Net worth is a useless number when you're talking about purchasing power. It only represents a theoretical amount of money you do have if you sold all your assets. It is hardly cash in the bank. My net worth is probably in that range and, because of my situation, I have the cash where I COULD afford a 5090 but I'm not stupid enough to blow that kind of cash on it. My BMW motorcycle cost me less than a 5090 does right now.
A lot of that money is also retirement savings, it is not money to burn, it is money to source income from to keep one from being a Wal-Mart greeter in their 70s.
 
Yep. There isn't a shortage of people that can afford spending that much if they really want it. Also plenty of people that really shouldn't spend that much but still do.

18% of households in the USA have a net worth over a million dollars, and about 1/15 individuals in the USA have a net worth over a million dollars.

Yes, but to be fair, "net worth over a million dollars" and "having a million dollars you can spend" are not necessarily the same thing. Depends a lot on where the money is. I live in the Toronto area in Canada, and many people here have a net worth over a million dollars simply because they bought a house 15 years ago and then the market exploded for a variety of reasons, many of which had nothing to do with local incomes (criminals using homes and condos for money laundering, immigration far exceeding the build rate, investors stuffing cash in condos but not renting them out, etc), but it's also trapped wealth, because all of it is trapped in the place where they currently reside, and unless you're borrowing against it, you can't spend it. To put it in American terms, we have San Francisco housing prices with Mississippi incomes and Wisconsin weather, so a lot of people have "high net worth", but are still struggling to keep up with costs of living because they can't spend their house. I can assure you that we have many millionaires that can't afford an RTX 5090 here. The net worth numbers don't tell the whole story.

In any case, the consumer market doesn't matter. GPU prices aren't going haywire because a bunch of wealthy gamers wanted to upgrade their PCs, they're going haywire because there's an AI datacenter buildout arms race going on, and the lenders are happy to continue backstopping loans because they know Meta, or whoever, is good for the money. Gaming used to be the bulk of Nvidia's revenue. Now, it makes up $16 billion out of $216 billion as of the last earnings report. Gamer net worth doesn't matter.
 
The last gaming generation that followed the traditional hw depreciation cycle was Pascal, somewhere around 2019-20 GPUs finally became the widely used GP parallel processing compute devices that LJM had been positioning for, also the start of the mega run in NVDA stock.

In a sane/traditional market, 5080 supers would be going for 900 right now with 5090s regularly dropping to 1500, given Blackwell is almost 2 years old now.
 
Last edited:
From my short research it seems there isn't much you can do with 5080s for AI when compared to 5090s.
Possibly, but what it does do is move silicon from producing 5080s to whatever AI is using. So if a company makes a 5080 it costs them more for the silicon and you feel it... plus I'm sure they're also profiting more too
 
In any case, the consumer market doesn't matter. GPU prices aren't going haywire because a bunch of wealthy gamers wanted to upgrade their PCs, they're going haywire because there's an AI datacenter buildout arms race going on, and the lenders are happy to continue backstopping loans because they know Meta, or whoever, is good for the money. Gaming used to be the bulk of Nvidia's revenue. Now, it makes up $16 billion out of $216 billion as of the last earnings report. Gamer net worth doesn't matter.
Maybe NVidia should spin off consumer GPUs as a separate company. Only half-joking. Jensen is too greedy to do that.
 
Last edited:
Maybe NVidia should spin off consumker GPUs as a separate company. Only half-joking. Jensen is too greedy to do that.

It wouldn’t work. A GPU is now just a general compute device and consumers will treat it as such. Nvidia has been advertising and developing parallel computing via GPUs for years, but along the way, gaming remained the primary application. It no longer is.

You also need the technical development pipeline. A lot of the gaming technologies being pushed by Nvidia were a result of other endeavours that made them possible. For example, DLSS is an AI technology. A separate company would need access to this to remain relevant.

You generally would only split a company like this if you think you could maximize profitability and growth in one or both divisions by doing so. That isn’t the case here.
 
"AI demand triggers massive PCB shortage for PC makers as order books full until 2030 — lead times stretch to 30 months as material costs surge 12x"

https://www.tomshardware.com/pc-com...etch-to-30-months-as-material-costs-surge-12x

"Motherboard and GPU Prices Could Rise Further as PCB Shortages Extend to 2030​

by AleksandarK Today, 15:59 Discuss (7 Comments)
Printed circuit boards (PCBs) are the foundation of every PC component, as they mount everything and connect the entire system using electrical components, wiring, and more. However, we are about to enter another phase of consumer electronics price hikes, as these PCBs are now experiencing severe component shortages projected to last until 2030. According to Nikkei Asia, an unnamed executive from an American PC maker told the publication that PCB manufacturers are confirming this situation, with some predicting that orders placed today may not be fulfilled until late 2028.
American PC Maker for Nikkei AsiaEven though PC shipments could decline again next year, we are still struggling to secure enough PCB supplies to meet our needs."
 

"Motherboard and GPU Prices Could Rise Further as PCB Shortages Extend to 2030​

by AleksandarK Today, 15:59 Discuss (7 Comments)
Printed circuit boards (PCBs) are the foundation of every PC component, as they mount everything and connect the entire system using electrical components, wiring, and more. However, we are about to enter another phase of consumer electronics price hikes, as these PCBs are now experiencing severe component shortages projected to last until 2030. According to Nikkei Asia, an unnamed executive from an American PC maker told the publication that PCB manufacturers are confirming this situation, with some predicting that orders placed today may not be fulfilled until late 2028.
Upstream suppliers realizing they can also get a much bigger piece of the pie, I suppose.
 
Back
Top