When the AI bubble pops, what will remain? Cheap GPUs at firesale prices, skilled applied statisticians looking for work, and open source models that already do impressive things, but will grow far more impressive after being optimized:
There were supposed to be cheap GPUs after the crypto bubble burst
Also the AI GPUS probably won’t be great for gaming. And cheap could mean anything when they go for 20k a piece.
There were when the first Ethereum bubble burst. That was one easier for the average person to get into with gamer GPUs, and they flooded the market on eBay as soon as it was no longer profitable.
Bitcoin won’t do that, because it hasn’t been based on GPUs for a long, long time. Ethereum doesn’t even work like that anymore.
The AI bubble popping will only flood the market with GPUs that are useful for running AI models. The GPUs in AI datacenters often don’t even have a display output connector. I think Corey is overstating his case on that one. Most likely, those GPUs are headed to the landfill.
The AI bubble doesn’t mean AI/LLMs aren’t useful. It means datacenter speculation can’t make money.
those GPUs are headed to the landfill.
They’ll just have a similar discount to the Ethereum switch.
You can still use such GPU as an accelerator either for running AI, or for gaming. In either case, given that you workload is Vulkan-based on Linux, you can use vkdevicechooser.
Of course, you will need a second GPU (even the CPU’s integrated one) to connect your display(s).
That has never worked well. It might give high average framerates on paper, but it introduces jitter that produces a worse overall experience. In fact, Gamers Nexus just came out with a video on a better way to measure this, and it touches on showing the problem with multi-GPU setups:
I think that you misunderstood my comment.
The video shows how SLI makes the frame pacing more inconsistent, which is a known issue when multiple GPUs work together to solve the same problem.
What I am talking about is more like Nvidia Optimus. This is a common technology on laptops, where the display is connected to the low power iGPU, while games can use the dedicated Nvidia chipset.
I don’t know about potential frame pacing issues on these technologies, and it seems like it was not addressed in the video either. However, I know that newer laptops have a switching chip that connects the display to the dedicated GPU, which, I think, aims on lowering the latency.
I can strongly recommend the arricle from the OP blog post about marker dynamics and use of what is essentially accounting fraud by major companies involved in AI:
Lifespan of AI Chips: The $300 Billion Question
I am looking forward to reading the research paper they are working on.
While the author takes a relatively neutral tone, the analysis is brutal in its portrayal of major market players (Nvidia, Microsoft, Amazon, Google); they come of more as oligopolists who are happy to engage in what is de facto in an attempt to undermine true market competition.
OP’s post is largely right, but it doesn’t require that link to be true. Also, whether these $3m+ systems are warrantied is a relevant question. It’s hard to know exact lifespan from one person saying their gpu failed quickly. Paper still stands well.
Because of power constraints, I’d expect they replace GPUs every 2 years with new generations, and so there will be big write offs.
A 6 year depreciation schedule seems unrealistically long for a GPU.
Even in gaming terms (I know this is completely different use case), a 2080S from 2019, a high end SKU, would struggle with many modern games at 1440p and higher. A profession streamer would be unlikely to use a 2080S.
Then there is the question of incentives. An objective look at American technology and VC suggests they are far closer to criminal organizations than their treatment by media and US institutions would imply. They very much can be expected to engage in what is essentially accounting fraud.
a 2080S from 2019, a high end SKU, would struggle with many modern games at 1440p and higher. A profession streamer would be unlikely to use a 2080S.
On one hand a 2080s would still be good at doing what it was doing 6 years ago. If there are new needs, and unlimited power availability, then a new card in addition to whatever AI workload the 6 year old GPU can do in addition to the new card makes sense… if that card still works. Selling your 2080s or whatever old card, does mean a fairly steep loss compared to original price, but 6 year depreciation schedule is ok… IF the cards are still working 6 years later.
$3m NVL72 systems are a bit different, as one out of 72 cards burning out can screw up whole system, and datacenter power structure and expertise requirements, would have low resale value, though I assume the cards can be ripped out and sold individually.
They very much can be expected to engage in what is essentially accounting fraud.
Oracle this week “proudly boasted” that they get 30% margins on their datacenter, and stock went up. This is not enough, as it is just 30% over electricity costs. Maintenance/supervision, and gpu costs/rentals don’t count, and it is unlikely that they are profitable, though it’s not so much accounting fraud as it is accounting PR.
One thing I worry about is that there‘s going to be a fire sale on the polluting crap that are powering these GPU farms. It’ll likely end up in poorer countries because it’ll be cheaper than new renewables.
The Internet has already been mostly destroyed, drowned in AI slop. Is all that shit gonna be taken down? Are search engines going to go back to working again?
They will ruin it, and then move onto the next thing to subsume and destroy forever.
Dude this. Looking up how to pull off pci passthrough on an SBC I have as well as answer a few lingering filesystem questions I get nothing but slop. The useful shit isn’t even visible anymore. And if I ask chatGPT to sift through it all, it can’t do it either, instead regurgitating all the slop it can’t make sense of either.
We are looking at the destruction of the greatest library in mankind’s history. Because NVIDIA’s line must go up.
I was going to reply “at least the burning of Alexandria was an accident,” and then I thought to look that up. Seems egotists destroying public collections of knowledge is just baked into humanity. We’ll never be free of its scourge.
Local archives of Wikipedia and Project Gutenberg have never been a better idea.
I have a conspiracy theory this was intentional, about the greatest library. The system that happened was intentionally built to end like this.
The mechanisms that led it to this point were simple and easily predictable (which is why they were not present in previous concepts, like NLS and Xanadu and such dream systems), and follow from the fundamental structure. Mechanisms solving these flaws also emerged early enough - siloed services and search engines. Companies providing the main offerings to solve these flaws followed in a few years. And they are still here.
See, this has been built up in like 30 years, right? Then surely much of the material will enter a better built system (reverse links, transparent identities, delay-tolerant, versioned, global and message-oriented, not connection-oriented, and so on)?
Yes, but less because of reliance on the wrong system. It was a trap to siphon the most effort. There will be no similar explosion of enthusiasm in a global networked system. Just like there will be no similar explosion of enthusiasm in communism as there was during 20s’ USSR.
It’s what secret services do, among other things - encourage underground activities to control those and detect participants and make the effort go in a predictable direction. Similar to preventive activation of mines.
Some things’ only good media would be preserved in the main medium used for their reading and reproduction, but the main medium became, say, some scan stored in some Internet service not caring about it. So they will vanish.
At the same time - the road from first printed Bibles to the French revolution to printing millions of student books on math, for example, took a long time. But also, with the speed information travels, we might expect new wonderful things in the following decades which won’t all be dystopian. Of course we should also prepare for bloodshed.
nope, its useful for propaganda still, especially dimwits like conservatives that cant tell the difference.
I suspect a lot of the slop will disappear. Why host it when nobody is interacting with it?
Please let the pop take the tech bros with it.
They’ll move on to the next big thing, just like they did after bitcoin.
And after NFTs, blockchain, the metaverse and so on
The metaverse was/is the stupidest thing ever. They couldn’t even sell it to the average person. I’m into tech but even I can barely describe what its supposed to be. Is it just VR? VR office work? 🤷
I’d love a gang of nerds doing the real metaverse, like Gibson showed us.
Yea but it comes with all the other stuff in Gibsons books
All that other stuff comes already, but I do not see it create the metaverse.
Let’s stop giving meta credit. Metaverse is a shit name anyway. Cyberspace is cooler and it’s still not that cool.
The metaverse is awesome, invented in the book Snowcrash from 1992. Its by Neal Stephenson, not Gibson, my mistake.
Facebook just stole the name and are making some crap.
VRchat, the OpenXR format, and custom-built/3d printed VR headsets come to mind.
What if we were pro-active, and trapped them by creating some bullshit “next big thing” and then just pocket all of their cash when they invest.
“You thought your money was going towards a space elevator, but little did you know, you’ve been single-handedly housing and feeding the population of six US cities!! Muahahahahahah!”
How dare those awful people keep trying new things. The Internet was perfect in the 1950s and we should stick with it being exactly the same way forever.
Their “new things” have created a bubble that will chop the head off the US economy, and at the same time is directly propping up a fascist in the White House. Fuck them. I’d rather the Internet be frozen the way it was in 2006 than deal with this shit.
I wish. I suspect the mega-billionaires will be absolutely fine.
I’m visualising that scene from IT Crowd with the boss stepping out the window…
Someone think of the “prompt engineers”!
Hmm and what about the everyday user who needs to ask AI how long to cook potatoes? What will they do after the pop?
completely uncharted territory. No one tried to cook a potato until Sam Altman graced us plebs with ChatGPT
Local models are actually pretty great! They are not great at everything… but for what most people are using LLMs for they do a fine job.
Thats from llama3.1:8b and the answer is decent, it took about 20seconds to generate my answer and used no more power than if I were to play a video game for the same amount of time.
Boiling time isn’t related to original potato size, it’s related to the size of pieces you cut. So the first half is irrelevant and the second half is overly verbose.
They’ll ask their parents, or look up cooking instructions on actual websites.
Cue 8 paragraphs of “I first learned to cook potatoes with my grandfather back in…” and every reader screaming “Oh my god just get to the recipe…”
Another advantage of cookbooks. (There are good ones, but a lot are junk.)
Holy shit yes, do yourself a favor and buy a copy of The New Best Recipe cookbook by Cooks Illustrated or How To Cook Everything by Mark Bittman, or even the old Betty Crocker cookbook is good, they’re packed with every basic recipe you need. I don’t even fuck with the internet for recipes anymore unless it’s something more uncommon I can’t find or something very specific.
the same way they did in 2021
How will I remember to put avocado in my BBQ sauce, without AI?
Oh yeah, maybe using indexing website with fraction of the compute power needed where the answer will be first result anyway.
Oh, fuck þat noise. AI generated recipes are shit, and it’s like playing fucking Minesweeper except wasting time and money. I’m about to go back to buying cookbooks. My culinary bullshit detector has gotten much better, but I’m no chef. It can be hard to tell if some AI bullshit hallucinates e.g. too much liquid-to-dry ingredients and you get a slushy or inedible mess.
Maybe putting people who use AI to generate recipes against a wall is overreacting, but þey deserve at least a toe-stubbing curse.
It’s a waste of perfectly good food, and a legitimate crime.
Oh, there are still thorn-users trying to “poison” AI? It doesn’t work and only annoys human readers.
It’s just this person from what I can tell.
He don’t care. People have tried to engage him in conversation on this and he just ducks any attempt to get him to explain his rationale for being annoying.
He’s just a snowflake trying to validate himself by forcing a social media trend. It’s pretty cringe.
Oh he knows. But this lets him feel edgy and different
What does it take to pop this bubble? So many people are calling it a bubble but what actually makes it pop?
As stupid as it is: Faith is what keeps bubbles afloat. Faith can go a long way towards forcing reality to what you want it to be, and if you have the wealth, you can play nearly endless money-games to make it seem like you’re ahead when you’re actually losing your shorts.
The reason there is so much faith is because this is a make it or break it moment for late stage capitalism. The businesses (including non-AI businesses) viscerally need it to work so they can get rid of human workers. If they can’t make humans slaves, they will make digital slaves. This may be a last gasp for the old order if it fails because so many entrenched companies from automobile makers like GM and Ford to airframe makers like Boeing to general electronics like General Electric have finances that are literally upside down because they have been using stock buybacks to fake growth for the better part of two decades now absolutely need it to happen to stay afloat.
It is difficult to get a man to understand something, when his salary depends on his not understanding it.
This quote by Upton Sinclair is usually used to describe lower level employees who don’t understand how unionization could be good for them, but it applies here as well. The faith persists because this is their “salaries” that depend on this working so the bottom doesn’t fall out from under them. They have to believe it will work and as such will keep dumping money into it as long as humanly possible.
AI is like Theranos but bigger and affecting numerous industries who are all betting the future of their companies on this all working out. For their livelihoods and their plan to continue ignoring all the little people in the world, there is no losing state they can or will accept until they are on the edge and about to leap from the top of their buildings to avoid the consequences.
Once the faith breaks, it will be like a dam breaking and flooding out too fast to escape.
Venture capital drying up.
Here’s the thing… No LLM provider’s business is making a profit. None of them. Not OpenAI. Not Anthropic. Not even Google (they’re profitable in other areas, obviously). OpenAI optimistically believes it might start being profitable in 2029.
What’s keeping them afloat? Venture capital. And what happens when those investors decide to stop throwing good money after bad?
BOOM.
OpenAI optimistically believes it might start being profitable in 2029.
Which is absolutely buck wild when you consider they’ve already signed contacts to spend another trillion dollars over the next five years.
How the fuck is a company that has $5 billion in revenue today going to grow that revenue by at minimum $995 billion by 2029? There’s just no fucking way, man…
On top of that, there’s so much AI slop all over the internet now that the training for their models is going to get worse, not better.
To an extent, I think that’s already happening. ChatGPT5 released with a huge amount of hype, but when users started playing with it, it was incredibly underwhelming — and flat-out worse than 4 in many cases… all while burning though even more tokens than ever. Definitely seems like that capabilities of this technology have hit a plateau that won’t be solved with more training.
I’m a software developer and my company is piloting the use of LLMs via Copilot right now. All of them suck to varying degrees, but everyone’s consensus is that GPT5 is the worst of them. (To be fair, no one has tested Grok, but that’s because no one in the company wants to.)
I really have no idea how anyone is supposed to take this world seriously anymore.
🎼You can’t stop, techno pop 🎷🎹
I have been wondering the same thing. Since we have a bubble called TSLA that still hasn’t popped, that I know of. And “AI bubble” seems a much more diffuse and hard to define concept, in comparison.
This is a great question. Eventually the companies must run out of gullible investors, right?
The sad thing is those GPUs are in specialized boards with specialized servers and cooling, and not really good at the kind of work consumers would want it for, assuming you could even get drivers. So most all of those GPUs will realistically get scrapped if the bubble pops.
Maybe we’ll see the EPYC CPUs get sold secondhand at least, those are socketed.
I understand the failure rates for the GPUs is huge. The duty cycles tend to be high, with power and cooling issues.
(Actually the power issues are wild and can destroy power distribution and generation equipment. “Power Stabilization for AI Training Datacenters” 21 Aug 2025 https://arxiv.org/pdf/2508.14318)
There are plenty of these, yes. But there are also plenty of consumer-grade GPUs being used for it too.
Datacenters and GPUs will be sold for pennies on the dollar, both due to an incalculable used supply, as well as plummeting demand due to the inevitable economic depression.
will NVIDIA suffer from the pop, if they invested 100bn they expect ROI TWICE AT LEAST BACK.
Only in such a way that their profits eventually decrease back to “normal” levels.
I mean AI has been doing what its been doing without much porn so far. Im pretty sure it will shift to that. Its what has pushed every technology before forward.
You’re right about commercial ai.
Open weight visuals models ? Wan, Qwen, Flux and Stable Diffusion before, most of their usages is for NSFW.
Admission time: I actually used a porn AI chatbot.
To defend myself, what I was trying to do is experiment with high-protocol BDSM ideas in an environment where cost of failure was low. This is when there’s a more formal hierarchy and rules. Often, the rules are deliberately unfair to the subs so that they can be “punished” for breaking them.
Problem was that chatbots would straight up hallucinate one rule or another all the time. The idea isn’t for the subs to fuck up every time. Then it’s nothing but punishment, and that doesn’t work. There needs to be times when they’re successful so they can be “praised”, and the chatbots simply couldn’t do that.
The future is AI powered uncensoring of JAV. It’s already started anyway.
You just aren’t going to the right websites.
oh I know there is some but many of the official ones don’t do it but they will. oh yes they will. thats right. you know what I like. you need the money.
I’m all fine already with my RX 6400.
