News Most RTX 50-series GPUs sold out in five minutes at Newegg — entire inventory evaporated in just 20 minutes

[abufrejoval]

While I do understand, that the main motivation of Nvidia is making money, I don't quite understand how this shortage increases their bottom line. Consumer dissatisfaction is less of a concern when those have no alternatives, but I can't see Nvidia intentionally aggravating their customers: they pay too much in marketing for that.

They properly wound down the last generation to avoid sitting on excess capacity, but the ramp-up for the next gen doesn't seem to have happend. Of course they'd always want to service the more lucrative DC market first, but fab capacity doesn't seem to be an issue for the Blackwell ASICs

For all I heard they have quite a stockpile of DC parts, because the production bottlenecks with the DC variants seem to pretty much tied to HBM assembly capacities and thus the shortage there. Of course in this generation they can't just switch those chips between consumer and DC use, that's the downside of their monolithic approach, but why not more wafer starts?

I haven't heard about TMSC being at capacity limits, no idea if 4NP FinFET lines are exlcusive or multiplexed with less custom variants...

Yes, ramp-up is long, fabs need months from wafer starts to a finished product, but a little more seems at play here.

Is it all because because of the tariffs threat that Nvidia tried to play it extra safe?

I sure hope that grocery retailers won't pick up on this sort of thing and starve us of what we really need...

 

[acadia11]

Hearsay, I'm going to go with hearsay.

I'm not saying the other comment is correct. I'm saying you exchanged hearsay you didn't like for hearsay you preferred. "My hearsay is better than your hearsay." <-- This is exactly what you said.

Hearsay, I'm going to go with hearsay.

I'm not saying the other comment is correct. I'm saying you exchanged hearsay you didn't like for hearsay you preferred. "My hearsay is better than your hearsay." <-- This is exactly what you said.

To split hairs , my comment would be an observation not hearsay, your comment implies third party source claiming jnsen said XYZ … therefore hearsay, I’m neither implying my opinion or observation or rather analysis is fact based or correct … it’s simply my view of the situation. Just saying … but yes I’ll take my B,S over yours … cause it’s mine to sum it up.

 

[rluker5]

What chips Intel 3 has mass produced and are commercially available?

It's kind of sad; people aren't getting it - what is needed here is a track record, not empty statements. That's the difference between TSMC and Intel on this front.

Granite Rapids and Sierra Forest Xeons. It is in the article. They are selling, just not as fast as Intel wants as there was a price cut on Granite Rapids. It seems like there aren't capacity issues if they want to sell more. As for track record that is short because Intel is advancing through nodes much faster than what is the norm.

Edit: Intel is putting out 18A this year btw.

 

[Gaidax]

Granite Rapids and Sierra Forest Xeons. It is in the article. They are selling, just not as fast as Intel wants as there was a price cut on Granite Rapids. It seems like there aren't capacity issues if they want to sell more. As for track record that is short because Intel is advancing through nodes much faster than what is the norm.

Edit: Intel is putting out 18A this year btw.

And that's it. It's not available outside Intel and likely because it's simply not competitive with yields.

As compared to TSMC delivering literally whole Snapdragon 8+ Gen 3 and Bionic 17 with literally millions of units made and shipped.

Again - it's incomparable. TSMC is so much ahead it's funny to even try to sell Intel in comparison.

---

If Nvidia could have used other foundry for their GPUs they would, they used Samsung for Series 30. They are not dumb, they would love better profit margins, but they also need a track record of delivery, not just promises.

You can bet Jensen is not happy being married to TSMC, just as people here not happy being hostages to Nvidia in the enthusiast high end segment. We just don't have a viable alternative and that's it.

 

[daftshadow]

Easy fix.

1. Make scalping illegal
2. One gpu purchase per customer
3. Block bots

But people lining up outside for hours to buy overly expensive pc hardware is ridiculous tbh.

 

[eisenb11]

I was manually monitoring Newegg for 5090s and if any showed up they were gone within a second. I, personally, didn’t see a single one switch to in stock. The stock checker site that I like wasn’t able to register a single 5090 coming available on Newegg, BH, and Amazon as well. I’d really like to know what AIB 5090 quantities they got since the FE here seems limited to BB.

 

[YSCCC]

I actually am wondering how many are legit buyers and how many are scalpers who hope to earn a big easy cash

 

[Loadedaxe]

I don’t know about anyone else but I never saw any stock listed on any of these sites, they all said sold out the minute you looked. if Intel had any since as a foundry they would have been priming a long time ago to support the likes of nvidia. What I don’t understand is why Nvidia is exclusively going to TSMC for foundry services , it’s got to be the Taiwan good old boys network. While there is a global chip demand crunch across the board … it seems to be in part due to contracts and loyalties.

Yeah, the ‘sold out instantly’ situation has been the norm for a while now. As for Nvidia and TSMC, it’s not just about loyalty. TSMC has had the most advanced nodes for years. Intel is trying to catch up, but foundry business isn’t just about capability, it’s also about reliability and relationships.

That said, politics and business ties probably do play a role in some of these decisions.

 

[Sluggotg]

I was manually monitoring Newegg for 5090s and if any showed up they were gone within a second. I, personally, didn’t see a single one switch to in stock. The stock checker site that I like wasn’t able to register a single 5090 coming available on Newegg, BH, and Amazon as well. I’d really like to know what AIB 5090 quantities they got since the FE here seems limited to BB.

I was refreshing Newegg/Best Buy and B&H Photo constantly starting a few minutes before the release. None of them showed anything for sale. They would go from "Coming Soon" to "Sold Out". I don't see where Newegg came up with "It took 20 minutes to sell out" because they instantly sold out between refreshes.

On the positive B&H photo did have an anti bot feature. If you refreshed to fast a thing would pop up and require you to click and hold it for like 5 seconds to help deter bots.

Not a problem. It will be 6-18 months before I will be able to buy one of the 5090 models I want. Certainly annoying but that is the modern business model for releasing popular products. (Like game consoles).

 

[ManDaddio]

Everyone keeps talking about quantity. It's known NVIDIA had to delay release because of a slight rework. So this should be expected.

But also, in fact, contrary to the hate talk people do want to buy these gpus.

NVIDIA gpus are still in demand. That's just the way it is.

 

[blppt]

It's already known at this time that 5080 and 5090 are a paper launch.

What's really sad about this is that the performance of the two are overall disappointing gains, AND they can't get them in stock, lol.

 

[Gaidax]

Everyone keeps talking about quantity. It's known NVIDIA had to delay release because of a slight rework. So this should be expected.

But also, in fact, contrary to the hate talk people do want to buy these gpus.

NVIDIA gpus are still in demand. That's just the way it is.

Ofc, we hatin' and we payin'.

It is what it is, all in all these are the best GPUs in existence for now, even if 16GB of 5080 makes my eye twitch. Let alone $3600 I paid for 5090 Astral.

 

[palladin9479]

While I do understand, that the main motivation of Nvidia is making money, I don't quite understand how this shortage increases their bottom line. Consumer dissatisfaction is less of a concern when those have no alternatives, but I can't see Nvidia intentionally aggravating their customers: they pay too much in marketing for that.

To understand why the paper launch and ridiculously low supply we have to stop viewing nVidia as a "gaming GPU company" and instead as a "datacenter GPU company".

The premier fabrication company that nVidia use's is TSMC and nVidia has paid TSMC money for a fixed amount of capacity. Nvidia can either chose to use that capacity to make consumer GPUs or to make datacenter GPUs. An H200 GPU from Nvidia is 814 mm2 per die and costs about $31k USD.

NVIDIA H200 Price Guide | TRG Datacenters

The NVIDIA H200 is the latest NVIDIA GPU made using Hopper architecture. It is a memory-upgraded version of the H100 and offers significant performance optimization with reduced power consumption and running costs. Despite offering 50% more performance improvement, the NVIDIA H200 is only...

www.trgdatacenters.com

A 5090 is 750 mm2 in size and costs $2k USD. MSRP That large size is why I call it a datacenter GPU lite.
A 5080 is 378 mm2 in size and costs $1K USD MSRP, far more modest and consumer GPU orientated.

So 1 H200 GPU is about the same silicon as 2.15 5080's or 1.085 5090, yet garners far more profit and is from the older architecture. The demand for "AI" datacenter GPU's is ridiculous with nVidia capacity being presold for a good twelve to eighteen months. From nVidia's point of view, any consumer gaming GPU's they make are a good will gesture and charity work. They would actually make more money by not selling a single consumer GPU and instead allocating that capacity to more datacenter GPUs.

This is why we need AMD and Intel to step up and make competitive products.

 

[abufrejoval]

To understand why the paper launch and ridiculously low supply we have to stop viewing nVidia as a "gaming GPU company" and instead as a "datacenter GPU company".

The premier fabrication company that nVidia use's is TSMC and nVidia has paid TSMC money for a fixed amount of capacity. Nvidia can either chose to use that capacity to make consumer GPUs or to make datacenter GPUs. An H200 GPU from Nvidia is 814 mm2 per die and costs about $31k USD.

NVIDIA H200 Price Guide | TRG Datacenters

The NVIDIA H200 is the latest NVIDIA GPU made using Hopper architecture. It is a memory-upgraded version of the H100 and offers significant performance optimization with reduced power consumption and running costs. Despite offering 50% more performance improvement, the NVIDIA H200 is only...

www.trgdatacenters.com

A 5090 is 750 mm2 in size and costs $2k USD. MSRP That large size is why I call it a datacenter GPU lite.
A 5080 is 378 mm2 in size and costs $1K USD MSRP, far more modest and consumer GPU orientated.

So 1 H200 GPU is about the same silicon as 2.15 5080's or 1.085 5090, yet garners far more profit and is from the older architecture. The demand for "AI" datacenter GPU's is ridiculous with nVidia capacity being presold for a good twelve to eighteen months. From nVidia's point of view, any consumer gaming GPU's they make are a good will gesture and charity work. They would actually make more money by not selling a single consumer GPU and instead allocating that capacity to more datacenter GPUs.

This is why we need AMD and Intel to step up and make competitive products.

Sure, I understand that they give DC variants priority. And those GB100 dies have had around a year of head-start, to a point where DC parts have long ceased being limited by fab capacity. From what I hear, it's HBM assembly capacity which limits output, and Nvidia is rumored to sit on stockpiles of GB100 chips waiting to get turned into sellable parts (which is why DeepSeek hurts them extra, with the value of all that existing but unpaid inventory in balance).

Beyond that, there is the issue of GB100 based aggregate products failing to heat issues, but again that's not adding stress on GB20x wafer starts, because it's a completely separate engineering domain, at least until they finish the Rubin design--if that is what it takes to fix that.

And in any case if Nvidia were to decide that consumers get the short straw, why even bother beating the marketing drum and doing all that software effort for MFG etc.?

Contrary to Covid times, fab capacity seems much less of an issue, currently. My impression is that everybody is most afraid of sitting on stockpiles of chips that will only move at clearout prices so they rather err on the side of having too few to sell.

That's a pretty fundamental turn in this industry, as far as I can tell, and I wonder how much of this is due to sanctions and tariffs killing the market.

 

[palladin9479]

And in any case if Nvidia were to decide that consumers get the short straw, why even bother beating the marketing drum and doing all that software effort for MFG etc.?

Because Jensen Huang has a serious hero complex and still wants the attention from gamers.

 

[YSCCC]

Everyone keeps talking about quantity. It's known NVIDIA had to delay release because of a slight rework. So this should be expected.

But also, in fact, contrary to the hate talk people do want to buy these gpus.

NVIDIA gpus are still in demand. That's just the way it is.

It was the same for each iPhone gen due to scalpers, and this gen Nvidia have some design rework to do to begin with so it's likely the yields arn't good anyway, but then I do hope (wishful thinking) not a lot will buy at those prices so that scalpers are sufferring soon, just by then the next gen probably we won't need to get along with these POS

 

[acadia11]

To understand why the paper launch and ridiculously low supply we have to stop viewing nVidia as a "gaming GPU company" and instead as a "datacenter GPU company".

The premier fabrication company that nVidia use's is TSMC and nVidia has paid TSMC money for a fixed amount of capacity. Nvidia can either chose to use that capacity to make consumer GPUs or to make datacenter GPUs. An H200 GPU from Nvidia is 814 mm2 per die and costs about $31k USD.

NVIDIA H200 Price Guide | TRG Datacenters

The NVIDIA H200 is the latest NVIDIA GPU made using Hopper architecture. It is a memory-upgraded version of the H100 and offers significant performance optimization with reduced power consumption and running costs. Despite offering 50% more performance improvement, the NVIDIA H200 is only...

www.trgdatacenters.com

A 5090 is 750 mm2 in size and costs $2k USD. MSRP That large size is why I call it a datacenter GPU lite.
A 5080 is 378 mm2 in size and costs $1K USD MSRP, far more modest and consumer GPU orientated.

So 1 H200 GPU is about the same silicon as 2.15 5080's or 1.085 5090, yet garners far more profit and is from the older architecture. The demand for "AI" datacenter GPU's is ridiculous with nVidia capacity being presold for a good twelve to eighteen months. From nVidia's point of view, any consumer gaming GPU's they make are a good will gesture and charity work. They would actually make more money by not selling a single consumer GPU and instead allocating that capacity to more datacenter GPUs.

This is why we need AMD and Intel to step up and make competitive products.

Nail meet head. The only caveat is AMD is a massive player in the data center space and is investing very heavily in AI centered processing power and outright is a monster in the CPU front and very competitive with Nvidia on AI GPU space with its instinct range of GPUs … so it seems to be following the same approach abandon the consumer space for their extremely profitable data center offerings. It may be AMD is diverting its resources hence the lack of competitiveness. And Intel well … ironically the grand daddy of them all is the one who could capitalize sliding under the radar.

 

[Peksha]

And that's it. It's not available outside Intel and likely

It seems like Arrow Lake-H(U?) might be partly on Intel 3, but it's not on sale yet.