Hoarding, I mean, buying GPUs in 2026 is getting difficult. It feels a lot like piloting a large oil tanker through the Strait of Hormuz while
anti-ship missiles bounce off the boat and basically everything in your field of view. You have Sam Altman on one hand buying all the RAM and now all
the copper. Power is getting more expensive than ever because everyone seems to want to build data centers
in places where people live now. With that said, there is light at the end of the dark GPU tunnel and I think a lot of that might have to do with the
GeForce RTX 4090 and some really interesting developments with these GPUs being domestically modified in the US at the behest of Nvidia. Welcome to AI
Flux. Let's get into it. Really quick, I want to ask that you like, subscribe, and share. It helps
out the channel a ton. So, why the RTX 4090? The RTX 4090 sits in a really interesting place in terms of performance and capability. These GPUs are
getting less desirable in a lot of professional markets, which means they're spilling onto the secondary market and are actually starting to reduce in
price. There was a time when you couldn't find these for less than $2,400 and now they're finally
dipping under about $2,000 on eBay. And if you crawl around Facebook Marketplace long enough, you can also find some really good deals. Now, they're
not as cheap as a 3090 and they're not as capable as a 5090, but they present really interesting performance characteristics. And until I sold them
about 6 months ago, I had four of these, so I know a lot about what you can and can't run in them. And
yes, I do regret selling those GPUs. So, let's get into the rough performance characteristics of the official GPU from Nvidia. We'll get into the
spicy version in just a bit. So, if you look at specs here, they're Well, they don't share very much. You have 16,000 CUDA cores, really decent base
clock, and 24 gigs of GDDR6 memory. You also get the fourth generation Tensor cores with around 1,300 AI
tops, which again is an Nvidia metric, so this doesn't compare well to other GPUs. And what this GPU doesn't have in VRAM, it has the same 24 gigs as
the 3090, you get a nice bump in Tensor performance. It's roughly 15 to 26% faster in most cases. There are also cases where this GPU is almost twice
as fast. It just really matters what you're using for inference and what your system looks like. If
you're using proper risers and a proper kind of latest gen system running these all at full throughput, you're going to have a pretty great
experience. But unfortunately, this doesn't have NVLink like the 3090 and it lacks NVFP4 support, which is now one of the emerging reasons you'd want
to buy the 5090 or any 5000 series GPU. Hopping over to Local Maxing, which you should definitely check out,
there's a link below, there are some really interesting scores coming out with this GPU. So, curiously, some of the best scores are still held by
pairs of 3090 TIs and the Nvidia RTX Pro 6000 Black, but the 4090 as a singular GPU puts up really impressive performance numbers. This is impressive
because this is happening in Llama.cpp, not even one of the best local inference engines I prefer, which
are vLLM and SGang. So, we're seeing some really great performance here with Qwen 3.6 27B with a Q4KM GGUF model. And there's a lot of work that's
still going on to improve performance here, but the key here is these are really great GPUs in terms of speed. They're just a little bit harder to
group together to run on larger machines. But that's not a reason you shouldn't buy these. What's also
cool is people are starting to really understand how to use these with Deep
Flash and how to get really impressive performance out of Qwen out GPEX. That's really cool. And to look at pricing of just the normal 4090, let's see
what we can get here. So, like I said, these are getting quite a bit more affordable. Of course, there's a little bit of sticker shock with what's
being listed now, but if you look close enough, you can find some of these right at $2,000, which in a
day and age when a 5090 costs well above 3,000, really above 4,000 if you can even find them, there are some really pretty good deals to be had. And I
would not spend over 2,600. That's kind of crazy, but the $2,000 to $2,200 range is really pretty interesting. And what's nice is you can still find
some of these, like this one, which are not for parts only that are under $2,000 and are pretty good
deal. So, yeah, this is to say in 2026, I still think that the 4090 is a fantastic option, especially with models like when 3.627b. You'll be curious
to see if uh deep seek V4 gets some nice benchmarks with this model. And yeah, so I'm curious what you guys think. Are you going to go out and buy
kind of the standard 4090 or you going to stick to buying the 4090 48 gig for uh still $5,000 from the
US? Curious what you think. Uh as always, I hope you learned something from this video. Please like, subscribe, and share, and I'll see you in the next one.