Are GPU prices hitting peak and falling?
Posted by DistanceSolar1449@reddit | LocalLLaMA | View on Reddit | 30 comments
I noticed GPU prices have gone up the past year, but recently it seems to have peaked and is falling again.
3090s seemed to have hit a peak and are now dropping in price.
I'm guessing the openclaw wave is dying out and supply/demand is now less on the demand side.
Important_Quote_1180@reddit
I purchased a 3090Ti last week after testing it for $800 in Northeast US. I’m building a workstation for building LoRa adapters and 3090 just sits in the best value areas right now unless you want Intel driver issues.
hurdurdur7@reddit
You are ignoring R9700 with a cold heart. While rocm isn't perfect it is far ahead Intel ...
Dave_from_the_navy@reddit
I have (probably misguided) faith that my Intel Arc Pro B70 is going to be good someday... It's miles ahead of what it was on launch. Still not great, but 32gb vram is 32gb vram.
ProfessionalSpend589@reddit
2 such cards would cost maybe a bit more than a strix halo last year, while being 3 times faster on bandwidth.
I’m not sure how low in power can the intel’s card go (Radeon AI Pro R9700 goes as low as 210W), but it may be competitive on that front too.
Not a bad choice at all for small dense models at around 30B with full context (may take up to 60GB VRAM).
Dave_from_the_navy@reddit
There are eco fanless models pushing as low as 160W tdp. Regarding my own testing (I've got a 290W tdp model), decoding speeds are largely flat at 170W and up. (Which makes sense, typically full memory bandwidth saturation across most GPUs happens at about 50% tdp.) Whereas prefill speeds (which I'm normally sitting at about 4000t/s at 290w) are mostly linear with about 3000t/s at 200w.
I'm very excited for the software stack to mature!
ProfessionalSpend589@reddit
I’m not sure if it’s because I’ve attached my two R9700 on PCIe 3 (old motherboard), but my power consumption yesterday was around 80W-90W during TG and 190W-200W during PP (for each card). Or maybe R9700 is just more efficient?
And with MTP my PP fell below 1000 on empty context, but it’s still faster than strix halo for Qwen 3.6 27B (quant 8). Maybe I’ll get a bit of speed up when I try ROCm instead of Vulkan.
Important_Quote_1180@reddit
I tried to work with my 9070 for a solid month and it just couldn’t get near the performance of CUDA. I really wanted it to work.
supracode@reddit
I have an R9700 and yes it will be slower than Nvidia, and ROCM is a mess. But the value for the 32gb of ram is what makes it attractive. Getting about 90 - 105tps output with Qwen 3.6 36B A4B Q5 with the full 256k context available on Vulkan. I would pay $2400 for a 64gb version if they offered it.
Miserable-Dare5090@reddit
the 9700 is going to be slower with less bandwidth. But more Vram so larger models. It’s not a bad tradeoff for \~400 more bucks than a 3090. You are lucky to find a 3090ti at 800. I bought one from microcenter at that price last October and then I chickened out and returned it. Bad idea but hindsight is 20-20
Charming-Author4877@reddit
The market is shifting fast, it's hard to predict.
NVIDIA has the largest stockpile in GPUs in it's corporate history, millions of highend GPUs are stockpiled and not being sold (their 2026 SEC filings show that). And they have no increased costs for RAM as well, so all their HBM/DDR stockpiles are in pre 2025 pricing still, they only stopped shipping RAM to their partners.
So the entire GPU pricing is manually modulated by NVIDIA, in Europe the big sellers are receiving 5090 GPUs in batches of 2-3, when they would normally want batches of 50-200. That's not because 5090ies are scarce, it's because NVIDIA is increasing their inventory as long as they keep a monopoly.
Intel and AMD are not really catching up, 11 year old NVIDIA GPUs are still beating them and China is entering the consumer GPU market now but with about the same performance gap.
As soon as competitors are able to sell true competitive GPUs NVIDIA will also empty their stockpiles, prices will drop rapidly. But for now that's not in sight.
Due-Function-4877@reddit
AMD's biggest problem has been software. Nvidia made big investments in building out their own proprietary software stack for inference. By comparison, AMD's efforts were half-hearted. They're still playing catch-up.
We need Intel to find their footing, but their (already tardy) attempt to get back in the discrete GPU card business was muted by panic. Short-sighted "activist investors" demanded austerity when the company needed investment and innovation the most. That further delayed and clouded the future of their GPU effort. Of course, the investors that cried about the deficits are blaming management now after Wall Street tried to kill their GPUs in the crib.
Charming-Author4877@reddit
It's more than software. The prefill speed of AMD GPUs is quite gruesome, that's not software it's lack of compute.
AMD has been very disappointing generally, just as Intel has been. Look at the line up of intel CPUs over the past years .. nothing has been released by that company that is worthwhile looking since their 13th generation CPUs.
For gaming GPUs AMD and Intel are doing not too bad, but for inference they are not on track.
China will do that for "us" - they were very quick in adapting to getting starved by a combination of monopolistic sales prices of NVIDIA and idiotic export limits by the US. Instead of falling for that, they specialized in efficient models and in their own hardware.
We all benefit from that.
Due-Function-4877@reddit
That's not fair.
You mentioned the products that AMD and Intel have right now. I noticed you didn't mention anything from this mysterious generalized "China" entity. Care to link the backfill speeds and other specs of these savior cards?... You can't because everything coming out of China right now is shit and the future is unwritten for everyone. 😏
Important_Quote_1180@reddit
One would have to think the AMD software should be improving rapidly alongside Intel and maybe it is but its not a first class citizen like nVidia.
MachineZer0@reddit
Not seeing a drop in 3090. Maybe potluck when a gamer upgrading has no awareness on the squeeze and posts a price below where they are changing hands. Any 3090 I see in Reddit HWS or HLS gets snapped up instantly if 800-850 with lots of replies in the thread. 900-1000 less responses, but usually marked sold same day.
The issue is model size. Most localllama folks are moving to multi-GPU setups. It’ll only get worse with Cursor, GLM and Copilot severely reducing subsidies to subscriptions. The only relief has been Qwen3.6-27b. I started reversing course on trying to run the latest GLM and Kimi locally after 27b’s capabilities were proven.
Important_Quote_1180@reddit
The 27B with 2x 3090's isnt that much better than the single card but its got more options like multi-user hosting and vLLM doesn't fail on huge context inputs. I also have a lot of DDR5 ram (192 GB) and I want the Qwen3.7 122B A10B to drop so bad it hurts.
HumanDrone8721@reddit
Sure, at least in EU the decline is visible, 6 weeks ago a RTX Pro 6000 could be found for 9.3-9.7K€, now the hoover around 13K€, same with the RTX 3090, they were 850€, now they sell around 1000-1100€, is clear there is an oversupply /s
LeatherRub7248@reddit
oversupply leads to higher prices?
HumanDrone8721@reddit
/s leads to to a bit of a different meaning.
LeatherRub7248@reddit
totally missed that.
tat_tvam_asshole@reddit
I just bought a rtx pro 6000 for $9k last week and wake up the next day they're at $11k. Felt like the last chopper out of saigon.
Dany0@reddit
Not really, sometimes regionally. What I DID notice was a decline in cloud gpu rental prices. Availability still sucks but y'know
Dany0@reddit
DRAM prices flatlined fwiw
nickless07@reddit
You mean after Nvidia just did that? Sure, you will get a small discount if you buy a couple thousand H200, but aside of that it will get worse.
AmusingVegetable@reddit
They will come begging to the home segment once the bubble bursts and the market crashes.
ImportancePitiful795@reddit
They have set up companies doing circular sales of GPUs. NVIDIA has 2 of them, funding then via the Huang foundation to buy GPUs to provide cloud gaming to NVIDIA.
This scam here is as big as the subprime ones which crashed the banking sector.
nickless07@reddit
No, they already got more then enough money to run the whole company for decades without selling a single product. Why should they 'beg' at all.
sleepingsysadmin@reddit
gpu prices probably keep climbing for the next 2 years or so.
Demand wont ease, but we have a number of new huge fabs coming online. We also have a new tech tier coming with ddr6. We also have denser intelligence so in 2 years 32gb of vram will be much more common but also sufficiently intelligent for many. Pressuring hardware lesser.
tat_tvam_asshole@reddit
Big 3 will just outsource DDR5 to the new fabs to service commitments, as they transition their fabs to next generation memory. And get ready for the real NAND inflation as they move to HBF.
Celestialien@reddit
The used 3090's a weird market since it's the cheapest 24GB card going, so it doesn't track the rest. But "prices falling" doesn't square with the memory crunch underneath everything: DRAM roughly doubled in Q1 and VRAM is like 80% of a card's BoM now, which is why Nvidia and AMD keep nudging prices up. And openclaw demand hasn't really gone anywhere, it passed 100k stars and most people run it on a cloud API anyway. I'd read any dip as people dumping stock before a new gen, not the start of a slide.