Nvidia Throws Gamers Under the AI Bus

S

sram

[H]ard|Gawd
2FA
Joined
Jul 30, 2007
Messages
1,854
https://www.theinformation.com/arti...der_2.6.26&utm_term=daily-tech-insider-active


Sorry gamers, the data center ate your GPU.
A report from The Information says that Nvidia won't ship a new gaming graphics card in 2026. This ends Nvidia’s 30-year streak of releasing new chips every year. The reason? Memory chip supplies are so tight that the company is prioritizing its wildly profitable AI business over the gamers who built its brand.
Nvidia had already finished designing "Kicker," an incremental upgrade to last year's RTX 50 lineup, before managers pulled the plug in December. The culprit is a worsening global memory crunch driven by AI's insatiable appetite for the same raw materials that power gaming GPUs. Samsung, SK Hynix, and Micron—the only three major memory suppliers—can't spin up new fabs fast enough to meet demand from both sides.

The decision isn't just about skipping a refresh. Production of existing RTX 50 cards is getting slashed too, pushing retail prices even higher than the scarcity-driven spikes we've already seen over the past year. A spokesperson from Nvidia confirmed that there is indeed a memory supply constraint in the market but gave no confirmation about whether this was the cause of the delay.
The delay also pushes Nvidia's next-gen RTX 60 lineup—originally slated for late 2027 mass production—further into the future with mass production only starting by 2028. With AMD focused on budget-friendly midrange cards, and no direct high-end competition on the horizon, Nvidia can afford to let gamers wait.
 
sram said:
The delay also pushes Nvidia's next-gen RTX 60 lineup—originally slated for late 2027 mass production—further into the future with mass production only starting by 2028. With AMD focused on budget-friendly midrange cards, and no direct high-end competition on the horizon, Nvidia can afford to let gamers wait.
Click to expand...

looks like upcoming AAA games are also lowering their specs which is a bad thing in terms of innovation
 
We need to invest in AI dudes! Although I'm not an expert on that, I wish I bought some shares from AI startups worth billions now. I'm not lucky with stocks/shares though. I was there in the US when google first started. I didn't put any money into it. Now google is an empire or a country by itself. Nobody knows the unseen. A company making hardware for gamers suddenly becomes the maker of hardware every nation in the world is fighting for. Ironic.
 
polonyc2 said:
looks like upcoming AAA games are also lowering their specs which is a bad thing in terms of innovation
Click to expand...
Is it though, or are they being forced to not suck?

Too often, developers are taking shortcuts, designing for the top end, then trying to scale down, because it's "cheaper" to develop, assuming you have no restrictions.
But it never works out that way because scaling down is the hard (aka expensive) part, so they leave it last, rush it, and the result is a "terribly optimized" mess that takes 4+ months to correct, which then detracts from bug fixes and content updates, assuming they do it at all.

Telling a team to go hog wild with virtually no restrictions leads to a mess because it always results in a "well, I'll fix that later" approach. But when you are busy, there is absolutely no time to go back and fix it, and by the time you do have time, you now have 30 things built on that thing, which all need to be modified to accommodate the fix on the one thing.
Now, if this were a one-off item you needed to "fix later," that would be one thing, but it's never one thing, it's hundreds of things made by hundreds of people over 3+ years of development, leading to 10's of thousands of things you need to fix that realistically there is no time or budget to ever fix.

Some of the best games made were done so under restraints, because it leads to smaller, more specialized teams, which makes communication and management easier, and it greatly reduces the ability to say "this is good enough" because if you do that too many times, you quickly encounter the bounds the project needs to operate within.

They say necessity is the mother of invention for a reason. If you don't need to actively improve things, there is no reason to try to improve those things.

What I am most interested in right now is this: How will Epic respond to this with UE5? The low VRAM GPUs are essentially here to stay, and with upgrade costs what they are, 8GB will be the norm for at least another 5 years.
What will they do to make Unreal Engine work smoothly and reliably within that 8GB limit, because if the answer is nothing, then we are in for 5+ years of "This game runs like shit!", because the argument of "well you shouldn't have bought a card with only 8GB" won't fly if those cards make up 80% of the market.
 
4saken said:
Or perhaps spend time optimizing. Wishful thinking, I know.
Click to expand...
It's not even about optimizing; it's about changing the entire game development cycle.
I feel like I could write an entire book on why designing for the 0.1% of the market, then trying to scale down for the 60%, will always result in a sub-optimal experience for everybody.

I've been doing this for too long, and the era of "just buy more RAM" has clearly passed.
 
If the Engineering costs are high enough, it makes no sense to make new models when there isn't enough parts (memory) to manufacture enough to make it worthwhile.
If the development costs have already been expended, could simply be delaying production/release as there is insufficient memory supply. Memory costs what they are, the cards prices would simply reach completely unreasonable levels, especially at the low to mid range, which is the only place they refreshed in the 4000 series.. there was no 4090 Super. So this doesn't exactly affect the high-end anyway, the one card where they can't keep it in stock. Nvidia has to decide if the cost of a 5090 and using some of the limited memory supply on it is worth it to make use of the chips that have a few defects and cannot be used in the Datacenter parts. i.e. the production cost of those wafers. Selling consumer GPU's helps them recoup wafer costs. If memory gets too high, might be cheaper to just not utilize that silicon. My guess would be that the 5090 could be the only card to have any production at all in 2026. They will likely sell the lower end GPU chips to AIB's, and leave them to their own devices to secure memory supply. It's not a good situation at all, some AIB's could fold or stop selling GPU's.

I doubt even lack of competition at the high end has much to do with it.

They can delay the manufacture of those video cards, build up a huge supply of chips, while waiting for memory to become available in 2028 to then sell those cards. What sucks is that the 6090 might be delayed while they sell the large supply of 5000 series chips.
 
Last edited:
Hardly surprising. This is the way it has been going for some time.

We are simply no longer profitable enough. Why spend limited silicon manufacturing capacity (they can only get so much latest node manufacturing capacity out of the likes of TSMC and Samsung) on consumer products which are the lowest margin?

This is an over-simplified view as there are many more parts to a video card product than just the GPU chip, but it helps tell the story:


CategoryProductDie Size
(mm^2)		Price ($)Price per area of Manufactured silicon
($/mm^2)		Multiple of
consumer revenue
ConsumerNvidia GeForce RTX 50601812991.651.00x
Enterprise (non-AI specific)RTX 6000 (Ada)6086,80011.186.78x
AI SpecializedBlackwell B2001,60040,000+25.00+15.15x +

So, they can get a fixed amount of silicon manufacturing capacity. If they wanted more, they'd have to outbid every other player on the market, who will go down swinging because they want that capacity too, so it is unlikely to happen.

I mean, if you look at the numbers, It's surprising these businesses haven't all dropped out of consumer GPU's a long time ago. Why would you consider making consumer parts, when for every mm^2 of silicon you can make almost 7x more for an Enterprise product, and over 15x more for an AI product?

Staying in the consumer GPU business in the current market conditions makes no business sense at all.

And that's before you even consider how much more expensive it is to sell to consumers. You have to warehouse large amounts of product in pretty boxes that get shipped out in onesy twosies. Yopu have marketing budgets to advertise your products. You have mass customer service, returns and RMA's to deal with. Whereas for a datacenter AI product you ship pallets of the stuff, packed in space efficient brown/white boxes, and the customers keep coming. You never need to advertise.

In this market we as consumers don't have a chance.

There are many efforts to bring more silicon fab capacity online to help deal with demand, but these are large and extremely complicated plants operating at the borders of what is scientifically possible. It takes many years and billions of dollars to bring new plants online, and even then, success is not guaranteed (as Intel learned with their 10nm process)

This is still the biggest example of just how difficult this industry is. Intel went from being the undisputed king of silicon fab manufacturing, the company all others measured themselves against when it came to process excellence, to being in ruins in just one process node generation. This is not a tech you can scale up quickly.

Some will presumably be successful and eventually come online, but the question is, when they do will it actually help the situation, or will AI's seemingly insatiable appetite continue to gobble up all of the capacity. Who knows? Maybe there will be some crumbs of supply left for the rest of us, but will there even be a consumer PC market left to save at that point, or will years of overpriced RAM, storage and GPU's have completely killed it?

I don't have the answer to that question.

I suspect the best case situation for our hobby is that eventually new process nodes will launch, instantly making current nodes unprofitable for AI data centers due to the cost of electricity, meaning all of the big players will shift their acquisition to the latest gen products, freeing up more current gen nodes for consumer products. Hopefully then we will get more supply. We will be a full generation or more behind datacenter AI, but at least it will be something.
 
Last edited:
OK, nvidia, you greedy twats, but don't come crawling back to gamers when that idiotic AI bubble pops all over your pants like you do when you see a schoolbus go by.
 
bugsy siegel said:
OK, nvidia, you greedy twats, but don't come crawling back to gamers when that idiotic AI bubble pops all over your pants like you do when you see a schoolbus go by.
Click to expand...

It is my theory this is the only reason they have stayed in the consumer GPU market at all. They want to have something to fall back on if the AI gold rush slows.

Also, I'm not convinced accusing them of greed is the right term.

It's just reality.

The reason any business exists (Nvidia included) is to maximize return for its shareholders. That's where people stick their retirement savings, etc.

A business has to go where the profit is. There is no concept of loyalty or nostalgia for a certain customer base. A business cannot choose to voluntarily make less money. That's how businesses fail or become irrelevant. The market evolves, and they have to follow it. That's the way it always has been, and the way it always will be.

It is a crying shame that this is the way the market has turned out, but I don't blame Nvidia or AMD for that. I blame the lack of will in our political system to regulate this destructive technology.in order to prevent all of the harm it is doing.
 
Lakados said:
Too often, developers are taking shortcuts, designing for the top end, then trying to scale down, because it's "cheaper" to develop, assuming you have no restrictions.
But it never works out that way because scaling down is the hard (aka expensive) part, so they leave it last, rush it, and the result is a "terribly optimized" mess that takes 4+ months to correct, which then detracts from bug fixes and content updates, assuming they do it at all
Click to expand...

designing for the low end causes just as many issues and is basically the same as porting a game from consoles to PC without any work optimizing for those with higher end systems...we'll be stuck with mediocre graphics for years
 
polonyc2 said:
designing for the low end causes just as many issues and is basically the same as porting a game from consoles to PC without any work optimizing for those with higher end systems...we'll be stuck with mediocre graphics for years
Click to expand...
Not necessarily low graphics, DX12, EPIC, AMD, and Nvidia all have some excellent compression algorithms that aren’t really in use which substantially reduce memory usage in GPUs. Heavily dependent on matrix multiplication so not overly usable on RDNA 3 and below, but tech options exist that developers can take advantage of them.
 
You must log in or register to reply here.
Back
Top