NVDIA may have just changed the game...

[Otto Von Herunterhängen]

NVIDIA RTX Spark​

Get RTX Spark in a small, ultra-efficient desktop. Built to run personal AI agents 24/7 right at your desk plus game and create with the full power of RTX graphics.

Hype aside, this may well make the huge data center buildup obsolete - along with Intel and AMD. Check out the link above.

Also here's a WIRED article on it.

I sold a bunch of data center building stock earlier this month, (MU, WMD, etc) because they are cyclical and I thought things were getting overheated.

What NVIDIA just came out with confirms it.

Things are moving right along...

Up to 1 Petaflop
FP4 AI Performance

Yeah, now we're talkin'

And yeah, I bought some NVDA

 

[Otto Von Herunterhängen]

Hey @TailsWin I was interested in getting your view on this one. It looks like Nvidia is moving in the direction you were suggesting and creating a laptop that can move some of the load off the server farms.

 

[TailsWin]

I'm skeptical of this particular product. IMHO nvidia is doing this mostly to stay ahead of other ARM providers and Chinese companies before they flood the market with their chips, because that's absolutely coming. Nvidia is making all the money in the world from datacenters already so they don't really need another market, but yea it's pretty clear that locally running AIs are very important. So nvidia has to play ball.

So, I see this more as a stopgap and market manipulation rather than serious effort. Nvidia gets all the hype and drowns out anyone else who might want to announce something similar in the next couple months. Which is exactly what a bunch of other ARM companies are rumoured to do, including Qualcomm and probably Samsung too. Because the local/mobile chips are an open and nimble market with a bunch of players, unlike datacenters.

Ironically enough, DGX Spark so far haven't been that great for AI, and stupidly expensive. Anyone with some sense buys a Mac instead, which does basically the same thing (ARM, unified memory and FP4/MX4), is easier to get and you can also use it as a normal Mac. One thing DGX Sparks have going for them is you can use them to train AI models, which is blocked on Macs. But well nvidia can block it too on consumer products, and we're back where we started. A de facto monopoly isn't breaking into new markets out of goodness of their heart, and nvidia is among the last companies to be trusted.

So I say, don't buy into the hype and rather watch what else is happening, besides nvidia. But we'll see I guess.

 

[TailsWin]

I also want to point out that nvidia doesn't have some sort of magic for AI. They made a great bet with CUDA 20+ years ago, which became a de facto standard, that's where everyone is still behind. But for AI from the hardware perspective, you need just dumb cores that do dumb math, but you mostly need fast RAM. That's why Macs can do AI just as well without nvidia - they have fast RAM and some dumb cores that do math. Nvidia doesn't make memory, they don't even design it. Samsung and Hynix make RAM, and China is now starting to make RAM. Once there's enough Chinese RAM on the market, anyone can add their dumb math cores and get some AI running nicely wiithout nvidia. No wonder nvidia is trying to stay ahead of the flood.

Ed: btw this is the 10000th post on this forum

 

[Nyghtfall3D]

But for AI from the hardware perspective, you need just dumb cores that do dumb math

Something like this?

 

[TailsWin]

Something like this?

You have to draw a bunch of faces in there, since there's couple thousands individual cores in there that do this kind of math

 

[Otto Von Herunterhängen]

I'm skeptical of this particular product. IMHO nvidia is doing this mostly to stay ahead of other ARM providers and Chinese companies before they flood the market with their chips, because that's absolutely coming. Nvidia is making all the money in the world from datacenters already so they don't really need another market, but yea it's pretty clear that locally running AIs are very important. So nvidia has to play ball.

So, I see this more as a stopgap and market manipulation rather than serious effort. Nvidia gets all the hype and drowns out anyone else who might want to announce something similar in the next couple months. Which is exactly what a bunch of other ARM companies are rumoured to do, including Qualcomm and probably Samsung too. Because the local/mobile chips are an open and nimble market with a bunch of players, unlike datacenters.

Ironically enough, DGX Spark so far haven't been that great for AI, and stupidly expensive. Anyone with some sense buys a Mac instead, which does basically the same thing (ARM, unified memory and FP4/MX4), is easier to get and you can also use it as a normal Mac. One thing DGX Sparks have going for them is you can use them to train AI models, which is blocked on Macs. But well nvidia can block it too on consumer products, and we're back where we started. A de facto monopoly isn't breaking into new markets out of goodness of their heart, and nvidia is among the last companies to be trusted.

So I say, don't buy into the hype and rather watch what else is happening, besides nvidia. But we'll see I guess.

Thanks, that's what I was looking for. I largely sold my AI manufacturing stocks and took profits - lots of it. I figured this stuff was coming along and was going to knock the crap out of it. I'll sell the NVDA and sit on the sidelines for a while. Yeah, with that much profit to be made I figured the Chinese were going to come roaring in at some point, and I don't see the Server Farms being the answer in the long run, technology is going to pass them by rather quickly.

Technology stocks are still the way to go but valuations are just insane at the moment. The crowd never learns, particularly when they mistake cyclical stocks for growth stocks in super cycles.

 

[TailsWin]

Technology stocks are still the way to go but valuations are just insane at the moment. The crowd never learns, particularly when they mistake cyclical stocks for growth stocks in super cycles.

Obv warning, don't take my speculation for financial advice lol. I don't know shit about markets and their psychology. While computing technology generally pans out the way I expect in the long run, it tends to take longer than I expect (which then makes me frustrated when I'm like "we should've had ARM desktops 15 years ago!"). And markets/financials go way over my head, otherwise I'd have some stocks too, or bitcoins.

Jensen is a smart cookie and it's actually pretty crazy how consistently good nvidia's long-term technological decisions have been over time. They're never afraid to overhaul everything, which is unusual for a company like that.

So I guess they have a plan for the post-datacenter-boom world, this is probably part of it. Only just like with GeForce 3, GeForce FX, Quadro and CUDA, the initial product may be something else than what people expect. On the other hand, nvidia has never been in such a crazy market position as they are now, with so many looming competitors, so monopolistic logic pretty much dictates they have to go FUD mode to stay ahead. How well are they gonna surf this wave, I have no idea.

 

[Otto Von Herunterhängen]

I wasn't asking for financial advice What I was looking for was your view on the technology involved. Bubbles are bubbles though, the dot com bubble of course did not change the fact that everything moved to the Internet, it still did, it's just that valuations became insane and the market had to reset.

The same thing I believe is happening now. I bought Seagate at $60 a share in April of 2025. It's a good company, it had been traveling along at between $40-80 a share about forever, and was paying a great dividend at the time. I sold half of it at $450, and the rest of it at $800 because, much as I love the company, it ain't worth $800 a share even with HAMR. I expect when the AI Hysteria dies down it will be back under $200 and I expect to own it again.

---

There was a time real early on when CPU's only did integer math and they needed a co-processor for Floating point calculations, that was about around the 386, which used a 387. They were finally put in the same chip with the 486

I suspect the same thing is going to happen with GPU's, and it looks like that's where NVDIA is heading. It doesn't make sense to put that across a bus, when it was primarily handing video chores that's one thing, AI training is another. I will be keeping a close eye on who is moving in that direction, because I think that's going to become the next level of "Standard" technology.

Anyway, that's the part I wanted your opinion on, if a bunch of others are on NVDIA's heals on that I need to take that into account because NVDIA has an insane valuation also.

Stock advisors always tell you not to time the market, and they are right in general. But if you notice hysteria buying, which is what we have now in AI relates cyclical stocks, investors damned well better take note. Because cycles happen.

 

[TailsWin]

There was a time real early on when CPU's only did integer math and they needed a co-processor for Floating point calculations, that was about around the 386, which used a 387. They were finally put in the same chip with the 486

I suspect the same thing is going to happen with GPU's, and it looks like that's where NVDIA is heading. It doesn't make sense to put that across a bus, when it was primarily handing video chores that's one thing, AI training is another.

Well yes, but from the exact opposite direction. GeForce 256 was the first (consumer) GPU with hardware vertex shaders, GF3 with fragment shaders, GFFX unified them and then CUDA solidified the concept into a general computing device, but suitable for fast FP calculations like in scientific fields. From that point onwards, the whole architecture had little to do with video, a GPU was a supercomputer with a video output.

This concept coincidentally makes it useful for AI training (and nice graphics), but an absolutely ridiculous overkill for inference, which is what normal people want. To actually run AI, you mostly need memory bandwidth, which is always lacking, so 95% of time, all those expensive CUDA cores end up doing nothing. When you run an LLM, that 5090 is some fast memory chips with an expensive paperweight.

For running AI assistants locally, it literally makes more sense to make RAM first and bolt some simple AI core right into it. ARM especially has being going in the direction of having CPU and RAM on one die, lately so adding some extra math shader cores isn't that big of a deal, in concept. RAM chip companies probably could do that too if they wanted, sell AI RAM to run models directly on the sticks. Now there's an idea...

It's just that historically, CPU, GPU/FPU/TPU and RAM have been different universes of know-how. Only mobile chips eventually went in the direction of complete unification, which is why AI cores started popping up on ARM processors very quickly. Those matrix cores are really dumb, which is why GPUs have tens of thousands of them. But you can fit a couple hundred on a tiny mobile chip and still get something useful.

And that's still universal computing. If you want to design a chip to run some specific model architecture, you can go even simpler and cheaper. I wouldn't be surprised if eventually, every computing hardware will just naturally come with some AI or assistant stock right in the chip, just like today everything comes with a screen.