Leaks Point to Return of 4GB GPUs as AI Flash Memory Crisis Worsens

It's now sheer desperation on the part of some of the (smaller) AIBs who are seeing their GPU buisness go down the drain.

Either they put something, anything, out for retail or they may as well shut up shop permanently.

I do think there is a market for this. Tens of thousands of games will run fine at 1080p, or below, for more casual gamers. It's mostly just AAA games from the last few years that struggle. Entry point GPUs are still needed.

Hopefully they will also concentrate on things like Neural Texture Compression that will lower the amount of RAM needed in future.

@EvilSpaceLasagna1 Sure if you are savvy consumer a second hand GPU may be a better option at the same price point, though most sales are new, not second hand. There’s also some inherent risk with second hand GPUs being overworked. Additionally most standard PSUs have an 8-pin connector, but you are right not all do.

This probably isn’t a card anyone coming to digital foundry is interested in, but I think there is a potential target market for it, many buy primarily on price.

@themightyant I understand your concerns. But from experience I disagree about used GPUs being a risk.

I have built multiple systems from used GPUs, usually bought a from CEX.

I’ve bought from them a GTX 970, GTX 1080ti, a RTX 2080 Super, a RTX 3070 and most recently an RTX 4070.

All of which came with 5 year warranties. All of which still work perfectly even to this day. The older PCs became hand me downs to various family members. None of them have ever had a problem. Not one.

I’ve bought a used RAM, used SSDs, used CPUs, motherboards, and entire gaming Laptops all used. Never had a single problem.

Id love to see digital foundry do an article on the cost savings made from buying used. As I find I can usually build a USD PC that achieves comparable FPS to a new PC for around 1/3rd to 1/2 the cost.

@Nick2310 I'm a big CEX fan, been using them for over 20 years. They actually test cards and offer a 5-year warranty which is great, but most second hand retailers sites do not and it's more of a lottery. CEX is also UK only.

I wouldn't mind seeing a second hand article, but the trouble with that is used pricing depends on availability which varies store to store, E.g. for CEX much of their stock isn't available online. Additionally second hand pricing varies massively country to country limiting scope for DFs international readership.

Used pricing would be a bit like a "paper launch", not really achievable by many.

If you need a graphics card to run a bunch of monitors for a security desk, as an example. Then 4GB is gonna be fine and its an upgrade over the old nvidia GT710 & GT1030

RAM is not apples to apples, and is only one element of a graphics card. Speeds and strategy for utilization is the key. Why is there a presumption that 'memory bloat' effected by programming must be accepted? Why not design to lessen the memory demands of the GPU via efficiency of programming and/or resource allocation? I think of the Volkswagen Beetle racing a Maserati - the former can win if the strategy is rethought. Time for a paradigm shift from developers, not a presumption that raw horsepower will be made available. Think smarter, not bigger.

@themightyant hi man. Thanks for the reply. You make some great points.

I have bought used parts on eBay in the past, a few PSUs, M.2s, a few CPUs, RAM sticks. Also bought a few refurbed laptops and utterly outstanding refurbed office class HP mini PC that I used for freelance work for years. I miss using that PC.

So it is possible. I’ve never been scammed, never had a dodgy part that didn’t work. And never had a part fail on me.

But, as you say this is not open to everybody. I’ve been building PCs since I upgrade my Pentium 75 from 8MB to 48MB of RAM. Yes that’s MB not GB. So I know my way around it. But for your average consumer it would be a nightmare. I once owned a gaming website where I wrote extensively on buying used components, do/don’ts etc. I forget experience has put me in a privileged position in what is a very hard time and expensive time for gamers.

@Bry159 The GTX 1080ti was an absolute beast of a card. It still runs non-ray traced games extremely well. It’s a GPU that just gives and gives and gives.

@ianbtv I’m a big fan of playing old games, and a friend of mine recently reminded me of how devs like Rockstar once managed to squash entire open worlds into the PS2’s tiny amount of RAM.

I’m also playing Fallout 3 on PC at the moment and that game requires extremely basic specs. Like recommends 2GB of system RAM and 512MB of Video RAM. From a graphics perspective it’s a good looking game, even by today’s standards. But from a gameplay point of view, the game is brilliant. For me games are about how different systems come together to create an experience that’s a delight to interact and experiment with. That provide hand/eye feedback loops that get the dopamine pumping. And you can do that without all the fancy graphics that’s making gaming so expensive. I’m fortunate that I only really notice the graphics/textures etc for maybe 2-3 minutes. Then I become immersed in the games world, in the story and characters. My mind and imagination fills in all the blanks.

@Nick2310 That PS2 example is exactly the point, and it's worth pushing further than nostalgia. Those studios didn't hit tight memory budgets by accident or through some lost art — they did it because the platform gave them no other option. There was no "add more RAM" escape valve, so engineering discipline became mandatory. Texture streaming, asset compression, clever reuse of buffers — all of that got solved because it had to be.

What we have now is almost the inverse incentive structure. If a game runs poorly on 8GB, the market's answer isn't "let's profile and fix our memory management" — it's "buy a 16GB card" or "wait for the 24GB refresh." The GPU makers, meanwhile, have every reason to let frame buffer size become the differentiator between SKUs rather than actual compute or efficiency, because it's cheap to sell you 8 more gigabytes and expensive to make a game engine that doesn't need them. So capacity keeps climbing on both sides — game memory footprints and GPU VRAM specs — in a kind of arms race that has nothing to do with whether the experience is actually better, just whether the ledger balances.

Your Fallout 3 point supports this too, but for a different reason than "old games are charming." It's evidence that immersion, feedback loops, and good design don't require the resource footprint the industry now treats as non-negotiable. That's not an argument for going back to 2008-era hardware ceilings — it's an argument that today's ceilings are much higher than they need to be, and the slack gets absorbed by unoptimized asset pipelines and lazy memory management because nothing forces otherwise. A 4GB card in 2026 isn't inherently a joke; what's a joke is a software ecosystem that's forgotten how to live within a budget, because it's never been made to.

I expect the reply to that will be about needing the headroom for 4K, high refresh, low latency — and sure, if we're talking a $6000 card driving an 80" 4K panel at 240Hz, that's a real conversation. But that's not the common case, and treating it as the default use case is itself the tell. Most players are on 1080p or 1440p at moderate refresh, on hardware nowhere near that tier. If the whole product stack — from the low end up — gets designed around the assumption that the high end is what matters, that's not a technical requirement, that's a market choice about who the "real" customer is.

And that's the same choice showing up on the supply side. The reason a 4GB card is even being discussed as a leak-worthy story right now is that DRAM and NAND capacity is being pulled toward AI datacenter buildouts, and consumer parts are getting the leftovers. That's not a coincidence sitting next to the GPU pricing story — it's the same allocation logic. Whoever can pay the most for a given wafer of memory or a given die gets served first, whether that's a hyperscaler buying HBM by the rack or a gamer buying a 5090. The consumer/enterprise balance that used to exist — where a company made a broad product stack because a broad market existed to serve — doesn't have the same pull anymore when a much smaller number of very well-funded buyers can absorb effectively unlimited supply at premium margins.

So yes, the 4GB card looks bad. But the more interesting question is why the industry's answer to scarcity is "sell what's left to the low end, cheap and thin," rather than "resource allocation is now a design constraint again, so let's engineer for it." Those are two very different responses to the same shortage, and the second one is exactly the discipline the PS2-era comparison upthread was actually about — just as it applies to hardware allocation as it does to software memory footprints.

@ianbtv Ian, that was an incredible reply mate. Honestly. Thank you so much for taking the time to write it. Damn, Digital foundry need to have a read and pull some ideas for content out of that.

I agree back in the day, Devs were forced to have tight memory management because they had no other choice. And we have kind have turned into the type of world where we like to throw more resources at the problem. I suppose the math has been done by senior management and the decision has been made that it’s more expensive these days to spend the people hours on optimising the game compared to the amount of money lost by small segment of people who won’t be able to buy the game due to a lack of RAM.

But aim just wondering if there is light at the end of the tunnel, concerning VRAM usage. There is neural texture compression Nvidia seems to be developing that can compress textures and other data down 7-8 times. Which from what I researched could lead to a 8GB ray traced frame buffer be reduced by about half. Could that lead to an age where large VRAM pools are no longer needed? Could this technology eventually be applied to any game lowering the amount of RAM needed? I’d love to hear your thoughts. And thanks again for the reply man. It’s appreciated.

@Nick2310 Thanks, Nick - that's kind of you to say, and this thread has actually been more useful than most of the ones I've engaged with in this realm.

On NTC: it's real, and the 7-8x compression figures are from Nvidia's own research, not just marketing claims. But it's worth being precise about what it actually fixes. Texture data is one consumer of VRAM, and a large one, but not the only one — render targets, geometry, ray-tracing acceleration structures, and frame buffers all still need their own room, and NTC doesn't touch those. It also isn't retroactive: it needs shader-level decompression support built into the rendering pipeline, so it's a "games built for it going forward" technology, not something that quietly lowers requirements on the back catalog. So yes, it could meaningfully raise the ceiling on what an 8GB card can do with ray-traced textures specifically. It's not a reprieve from the VRAM race in general, because the rest of the frame buffer keeps growing too.

But here's the part I think is being missed even in a generous read of the situation: this isn't really a story about optimization anymore, it's a story about who gets to be a participant. When you price the entry point high enough — whether that's the GPU, the console, or the game's minimum spec — you're not just making the experience worse for the low end, you're removing them from the audience entirely. And once that audience is gone, it doesn't come back, because a market decision has already been made that they weren't worth designing for.

That's the same mistake that happened with GPU and CPU enthusiasts who built the demand that justified the R&D in the first place, and then found themselves priced out once a bigger buyer showed up. Gamers built this industry. Now a chunk of them are being told the industry has moved on without them.

And it's not just the entry price, it's what happens to the thing after you own it. A phone can be carried for a decade if you accept it does 99% of what a current one does, and you can keep it alive by swapping a battery. A car is the same — parts, incremental maintenance, decades of life if you want it. That's a legitimate way to own something long-term. GPUs used to work that way too, more or less — a card held value, resale meant something, you could step down a tier and still play. Now the gap between something like a 4080 and a 5080 is wide enough, and the resale value of the older card weak enough, that "keep it running for years" isn't really an option anymore. You're not incrementally maintaining a system, you're on a forced replacement schedule, and the schedule keeps getting more expensive.

@Nick2310 I'll add a bit of where I'm coming from on this, not to posture, just context: I learned to program in BASIC on an Apple II in a school lab in the early '80s. That machine, and the philosophy behind it, was built on the idea that computing should be something an ordinary person could own, understand, and grow into — Woz's whole ethos was democratization, not gatekeeping. You can still see traces of that in Apple's products now, on and off, however imperfectly and however much they've fought right-to-repair along the way.

I don't see that same philosophy surviving at the companies actually driving the ingenuity now — Intel, Nvidia, and the rest of the silicon stack. Not because the engineers there don't want it. I'd bet plenty of them do. But the corporate incentive has shifted entirely toward whoever writes the biggest check, and right now that's data centers, not people. It's the same failure mode you see at Boeing — a company that used to be run by engineers making engineering decisions, now run by finance people making finance decisions, and the people who actually understand the product get overruled by people optimizing a balance sheet. I'm not equating the stakes, obviously nobody's harmed by a thin frame buffer the way they were by a MAX groundings. But it's the same structural failure: short-term capital allocation crowding out the long-term stewardship of the thing itself, and the people making the call aren't the people who'd have to live with what gets lost.

That's really what this whole thread has been circling. Whether it's memory bloat in software, VRAM tiering in hardware, or DRAM getting diverted to AI datacenters — it's all the same decision being made over and over: serve the biggest payer first, and let everyone else find their own way or drop out. That's not a technology problem. It's a stewardship problem. And unlike a lot of what's wrong in tech right now, this one didn't have to happen — there's no law of physics that says a broad consumer base and a lucrative enterprise tier can't both be served by the same industry. It's a choice being made, over and over, by people who've decided one of those markets doesn't matter as much as the other.

@Nick2310 One last thought, on its own: if Wozniak hadn't had the Homebrew Computer Club, or Gates and Allen hadn't been handed free time on machines they could never have afforded to own, there's no Apple, no Microsoft. Same story with chip design generally, and with the space race specifically — things got built because they needed to exist and someone believed they were worth building, not because a quarterly return demanded it. That's also how you get the next generation of engineers and builders: give a kid something to dream about and reach for. Price the dream out of reach early enough, and it just doesn't happen. Nobody dreams about what they've already been told they can't have.