There Isn't an AI Bubble - There Are Three 76
Fast Company ran a contrarian take about AI from entrepreneur/thought leader Faisal Hoque, who argues there's three AI bubbles.
The first is a classic speculative bubble, with asset prices soaring above their fundamental values (like the 17th century's Dutch "tulip mania"). "The chances of this not being a bubble are between slim and none..." Second, AI is also arguably in what we might call an infrastructure bubble, with huge amounts being invested in infrastructure without any certainty that it will be used at full capacity in the future. This happened multiple times in the later 1800s, as railroad investors built thousands of miles of unneeded track to serve future demand that never materialized. More recently, it happened in the late '90s with the rollout of huge amount of fiber optic cable in anticipation of internet traffic demand that didn't turn up until decades later. Companies are pouring billions into GPUs, power systems, and cooling infrastructure, betting that demand will eventually justify the capacity. McKinsey analysts talk of a $7 trillion "race to scale data centers" for AI, and just eight projects in 2025 already represent commitments of over $1 trillion in AI infrastructure investment. Will this be like the railroad booms and busts of the late 1800s? It is impossible to say with any kind of certainty, but it is not unreasonable to think so.
Third, AI is certainly in a hype bubble, which is where the promise claimed for a new technology exceeds reality, and the discussion around that technology becomes increasingly detached from likely future outcomes. Remember the hype around NFTs? That was a classic hype bubble. And AI has been in a similar moment for a while. All kinds of media — social, print, and web — are filled with AI-related content, while AI boosterism has been the mood music of the corporate world for the last few years. Meanwhile, a recent MIT study reported that 95% of AI pilot projects fail to generate any returns at all.
But the article ultimately argues there's lessons in the 1990s dotcom boom: that "a thing can be hyped beyond its actual capabilities while still being important... When valuations correct — and they will — the same pattern will emerge: companies that focus on solving real problems with available technology will extract value before, during, and after the crash." The winners will be companies with systematic approaches to extracting value — adopting mixed portfolios with different time horizons and risk levels, while recognizing organizational friction points for a purposeful (and holistic) integration.
"The louder the bubble talk, the more space opens for those willing to take a methodical approach to building value."
Thanks to Slashdot reader Tony Isaac for sharing the article.
The first is a classic speculative bubble, with asset prices soaring above their fundamental values (like the 17th century's Dutch "tulip mania"). "The chances of this not being a bubble are between slim and none..." Second, AI is also arguably in what we might call an infrastructure bubble, with huge amounts being invested in infrastructure without any certainty that it will be used at full capacity in the future. This happened multiple times in the later 1800s, as railroad investors built thousands of miles of unneeded track to serve future demand that never materialized. More recently, it happened in the late '90s with the rollout of huge amount of fiber optic cable in anticipation of internet traffic demand that didn't turn up until decades later. Companies are pouring billions into GPUs, power systems, and cooling infrastructure, betting that demand will eventually justify the capacity. McKinsey analysts talk of a $7 trillion "race to scale data centers" for AI, and just eight projects in 2025 already represent commitments of over $1 trillion in AI infrastructure investment. Will this be like the railroad booms and busts of the late 1800s? It is impossible to say with any kind of certainty, but it is not unreasonable to think so.
Third, AI is certainly in a hype bubble, which is where the promise claimed for a new technology exceeds reality, and the discussion around that technology becomes increasingly detached from likely future outcomes. Remember the hype around NFTs? That was a classic hype bubble. And AI has been in a similar moment for a while. All kinds of media — social, print, and web — are filled with AI-related content, while AI boosterism has been the mood music of the corporate world for the last few years. Meanwhile, a recent MIT study reported that 95% of AI pilot projects fail to generate any returns at all.
But the article ultimately argues there's lessons in the 1990s dotcom boom: that "a thing can be hyped beyond its actual capabilities while still being important... When valuations correct — and they will — the same pattern will emerge: companies that focus on solving real problems with available technology will extract value before, during, and after the crash." The winners will be companies with systematic approaches to extracting value — adopting mixed portfolios with different time horizons and risk levels, while recognizing organizational friction points for a purposeful (and holistic) integration.
"The louder the bubble talk, the more space opens for those willing to take a methodical approach to building value."
Thanks to Slashdot reader Tony Isaac for sharing the article.
It was always going to be a bubble (Score:2, Insightful)
But the potential payoff is replacing trillions of dollars of employees and salaries so it's chump change to the billionaires.
Remember we have a ruling class so past a certain level we have a group of people who can spend unlimited money without consequenc
Re: (Score:1, Insightful)
Re: (Score:2)
It is even worse: The training data they got in their massive criminal piracy campaign _ages_ and becomes less and less valuable. Nobody is coming out of this as a winner, except for some of the scammers pushing the hype.
Re:It was always going to be a bubble (Score:4, Insightful)
Re:It was always going to be a bubble (Score:5, Insightful)
With regard to "model collapse" this is a very good point.
So new data will not be good at all. But old data will get worse as well. The aging is a massive problem. Put this two things together and what we currently see is a straw-fire in the process of burning itself out and no more straw is to be had.
My take is all that will survive is low-quality, outdated and "collapsed" LLMs for cheap or for free for the masses and small, really expensive, special purpose LLMs for some (not a lot) industrial and administrative purposes. Say, 5% of what the AI peddlers promised in actual impact. About the same as from the last AI hypes.
Mostly agreed, but (Score:2)
Genuine, important progress is being made
In a situation where billions are at stake, it's easy to see why hypemongers and pundits spread their nonsense
In a perfect world, real progress would be reported accurately with no predictions about the future
Unfortunately, the world is far from perfect
Expect the hype to get more three-dick-ulous
Re: Mostly agreed, but (Score:2)
Given his other examples were mostly containing real stuff, you seem to totally agree.
The railroads continue to be key to this day, even if they overreacted in the day.
The Internet infrastructure absolutely is critical now, despite the somewhat false start.
speaking of railroads (Score:3)
https://www.scry.llc/2025/09/1... [scry.llc]
"Technology revolutions create "virtuous cycles" where growth in a primary industry drives growth in other industries. During the Steel revolution, the falling cost of steel led to more railroads, which led to cheaper commodities like wheat and coal. Cheaper coal made even cheaper steel for steel-framed buildings, which spawned elevators, penthouse apartments and fire escapes"
Re: (Score:2)
To use a tech metaphor, think of how many viable phone brands were in a phone store in 2005, and how much brand diversity you see now. Same goes for PC/laptop brands, even tablets. Money moves to find early leaders and funds them, until it's no longer profitable. Then there's a consolidation period, and monopolization of supply chains.
Although there's specific hardware underneath, AI is an app. What are the lifecycle of apps? Take a look at early office app consolidation. Anyone remember dBase? When Oracle'
Re: (Score:2)
Quite well, TYVM. And, I know why its name was so short; do you? Hint: CP/M.
Re: (Score:2)
Re: speaking of railroads (Score:2)
Re: (Score:2)
Re: (Score:2)
... after abandoning thousands of miles of useless track.
Re: (Score:2)
Genuine, important progress is being made
Nope. But they are making a determined effort to manipulate people into thinking that.
Re: Mostly agreed, but (Score:2)
well don't be so harsh, I'm very happy about the hardware that is now available for my numerical recipes.
would not have it without the "AI", and with the appropriate algorithms it does make a difference.
Three AI bubbles (Score:2)
There Isn't an AI Bubble - There Are Three
They're labeled, "Good," "Fast," and "Cheap" and whichever one you (re)pick is the ironic one.
The infrstructure will get reused when it pops (Score:4)
Just like we got a lot of cheap office furniture on eBay when the dot com bubble popped, I am sure there are going to be some firesales on cloud computing hardware or services when this horrid AI bubble finally pops.
Re: The infrstructure will get reused when it pops (Score:2)
You nija'd me. But I will add that nVIDIA and AMD AI GPUs can be used for HPC. Nuclear weapons modeling, climate models, seismic exploration, aerodynamics (CFD), finite model analysis. The posibilities for reuse are endless
Re: (Score:2)
Good luck with that. DCs of this type are far too specialized for that to happen.
Re: The infrstructure will get reused when it pops (Score:2)
Re: The infrstructure will get reused when it pops (Score:1)
Another problem is that a GPU has a limited technological lifetime. Pretty quickly ( 3 to 4 years) it becomes more economical to buy the latest GPU.
Hasn't Moore's Law been dead for a decade? (Score:2)
Another problem is that a GPU has a limited technological lifetime. Pretty quickly ( 3 to 4 years) it becomes more economical to buy the latest GPU.
Are you sure about that? It seems like the pace of improvement has slowed drastically. 20 years ago? ABSOLUTELY 10 years ago?...less confident, but you're probably right. Today?...well...some pace of improvement is happening, but according to Gemini, looking at a $2000 GPU, a RTX 5090 is almost 2x as fast as RTX 3080 Ti in benchmarks. That's over 5 years. It went from 8mm to 4mm. If we're using your 3 year benchmark, it goes from 5mm to 4mm, which is a 11% efficiency gain, according to nVidia.
Re: (Score:2)
Just like we got a lot of cheap office furniture on eBay when the dot com bubble popped, I am sure there are going to be some firesales on cloud computing hardware or services when this horrid AI bubble finally pops.
Hardware, yes. But what will you do with it? It's only really good for a few types of task. Where it's GPU-based, as all the Nvidia stuff is, you could use it for lots of different types of tasks. But Services? Energy needs to get a lot cheaper for that to be feasible, because providing services on this hardware is predicated upon using a lot of energy.
Simple, rent rackspace for cheap (Score:2)
Hardware, yes. But what will you do with it? It's only really good for a few types of task. Where it's GPU-based, as all the Nvidia stuff is, you could use it for lots of different types of tasks. But Services?
A state of the art AI datacenter is just a high performance datacenter. Ignoring the contents of the racks, you're focusing on massive amounts of reliable electricity and intense cooling capacity....all of which is useful for any other computing. OK....remove all the GPUs and all the infrastructure can be reused for mundane business processing...so instead of recouping. your investment in 5 years, if that's the plan, it takes 10-15...not ideal, but not a loss....and you're well equipped for future cloud c
Re: (Score:2)
It's not like AI is going anywhere. When something "pops", the infrastructure will stop growing, but why should anyone decomission the working systems? It's not like services will remove their AI features any time soon.
Aren't the services running at a loss? (Score:2)
It's not like AI is going anywhere. When something "pops", the infrastructure will stop growing, but why should anyone decomission the working systems? It's not like services will remove their AI features any time soon.
You certainly could be right, but if OpenAI, Claude, Perplexity etc are running at a loss, I'm not sure they'll survive monetization. Each query is losing them money. Some will definitely survive, but I think it's safe to say one of the top 10 AI services will shut down in the next 10 years.
Re: (Score:2)
I'm not sure if it still is. GPT-5 was a step towards choosing smaller models for most queries (routing to different models based on the estimated complexity for answering the question). I think GPT-40 was a mid-size MoE model (mostly speculation, as OpenAI doesn't share details) which offered medium intelligence and fast/cheap execution compared to large dense models.
Current 4B models that can run on your phone are outperforming the first GPT-4 version (estimated at 300B parameters). For inference, costs a
Re: (Score:2)
GPT-40 should be GPT-4o of course
Re: (Score:2)
Re: (Score:2)
I remember doing work in a handful of WorldCom's 1999 era data centers that never came online, trying to repurpose them for a new switch facility around 2007; they had remained vacant for that entire time. The projects were dropped-- only a few of the items in the facilities still had value 7 years later, and it was cheaper and more efficient to just build out a site from scratch. One of the switch sites was still for sale in 2019.
Many of these facilities being built now are really single-purpose. Adapting
Kind of funny (Score:4, Interesting)
Kind off funny that the economy in general isnt booming with the existence of this bubble as in past instances where we had major tech bubbles.
Re: (Score:2)
That one is not a surprise: The efficiency gains claimed are entirely fictional. What LLM-type AI is doing is to destroy value. And hence no boom.
Re: (Score:2)
He's talking about the money already spent and spending right now. Just the build out investments, given we're talking about trillions of dollars, must be boosting the economy.
Yet, things are tanking anyway. Kind of points to the general economy being in even worse state.
Re: (Score:2)
He's talking about the money already spent and spending right now. Just the build out investments, given we're talking about trillions of dollars, must be boosting the economy.
What's the measurement? If it's "GDP" then sure, the economy is booming. But GDP is itself meaningless to sustainability, which is the most important thing to measure in anything you hope to keep doing. If you want to keep having an economy, for example, you have to keep having consumers who have money so they can participate in it...
Re: (Score:2)
Indeed. But that requires looking more than 3 months ahead.
Re: (Score:2)
It's not a surprise because this bubble is the only thing keeping us from recession.
Re: (Score:2)
In the last tech bubble, companies were spending more on gaining and keeping employees. Employees earning more money spend more money, generally speaking. That consumption drives the economy. (It turns out, the real job creators were the average person.)
In the AI bubble, companies are holding off hiring and are not interested in retaining employees. That hurts consumption, and thus the economy isn't booming.
The rich are getting richer, but that doesn't help as much to drive the economy - they don't
Aspects (Score:2)
Having lived through the Dot-Bomb it's basically the same.
You're not going to get a valuation bubble without a hype bubble. And nobody is buying companies for that much who have zero infrastructure. And the stock price is what they use to buy the infrastructure.
These are inextricably linked, not separate phenomena.
This is what Austrian Economists call the 'malinvestment' part of the business cycle. It's caused by artificially cheap money (not set by a market) and will unavoidably be cleared.
Our Orwell is
Re: (Score:1)
Because an open market in lending rates is de facto prohibited.
I got 0.99% on my car loan. Maybe you are not looking hard enough.
Bubble #4 (Score:2, Interesting)
Bubble #4 is that already algorithmic improvements are reducing the number of GPUs needed for the same result. I've called the attention mechanism the E=mc^2 moment that ushered in LLMs. What if, instead of the aforementioned ongoing incremental improvements, there is another sharp discontinuity beyond attention -- such as LeCun's JEPA, or embodiment championed these days by Musk -- that also happens to obsolete the GPU?
It is said the human brain is 1 exaflop. Today, that requires 20 MW, but the human brain
Re:Bubble #4 (Score:4, Insightful)
Research operates on cycles of research careers. Paradigm shifts can happen when most of the current population of researchers stop working in AI and thereby stop flooding the world with variations on the same brute force trick. This will give a new generation of students breathing room to be actually creative and innovative. But before any of that happens, people like Musk and Zuckerberg need to run out of money or die. When the money and interest dries up, AI research will no longer be attractive. That's when the dedicated, methodical kids with a real interest in the field will get their chance to shine.
Re: (Score:2)
Most stuff scales. If you find the 100x efficient neural network, you deploy a 100x larger network on the same cluster talking vaguely about something like "AGI". And even without AGI scaling up an efficient structure would be nice.
I only wonder how such a structure should look like. How do you solve the problem with quadratic attention? It is pretty clear why you would want everything to attend to everything and not quite clear how to reduce it without reducing performance. Maybe some pretty clustering (by
Re: (Score:1)
Yes, Parkinson's law comes into play, but IMO that will be mostly at the nation-state level.
I think you're too railroaded about attention. The larger AI goals remain the same: pattern recognition and modeling. Attention achieves pattern recognition but not modeling. And one can imagine there might be a far more efficient paradigm to achieve pattern recognition. Think radix sort vs. bubble sort.
Re: (Score:2)
Yes and another limitation is training data. Not necessarily the pure amount, but mostly the high-quality annotations. If you can train 100x as fast, but still have the same training data in particular only 1% of it annotated, you're faster but not better. That's also why we now have models using (mostly) the same technology that are 20 times smaller at the same quality, because companies started to curate their datasets.
You can crunch as much data as possible and see increased quality, or you can try to cr
Re: (Score:2)
I'd love to see more separation about intelligence (or what you want to call it) and knowledge. It wouldn't only make it easier to know when the LLM doesn't know something, but also allow for cheap knowledge updates. If you got a nice model from 2024 it still works, but some knowledge is outdated, some missing. Why can't It just download a new Wikipedia dump to add it?
I kinda can with RAG, but these solutions still have so many rough edges. Give me a RAG that really works realiable. It always seems like "I
The 4th Bubble is labor (Score:3)
Trough of Disillusionment (Score:3)
AI is probably in the "trough of disillusionment" phase, which is why there's so many negative articles now.
A few months ago I had ChatGPT run a semantic analysis which produced this graph -
https://www.scry.llc/2025/07/0... [scry.llc]
Oi (Score:5, Insightful)
"adopting mixed portfolios with different time horizons and risk levels, while recognizing organizational friction points for a purposeful (and holistic) integration."
Several CEOs around the world just jizzed in their pants without knowing why.
Whoever wrote that sentence needs to be castrated.
i think (Score:1)
We'll just have to return to stories about who isn't going to manufacture Apple's car.
Re: (Score:2)
The bubble in stories about the AI bubble bursting soon is going to burst soon. We'll just have to return to stories about who isn't going to manufacture Apple's car.
Adults?
Does it matter (Score:1)
... as long as the technology is good enough to replace relevant portions of the labour force with automated systems however limited they may be.
That's where the "Return" of ROI will come from and we will feel the pinch soon enough, I am afraid.
Can we be clearer about what we mean by AI? (Score:3)
The real problem with AI, and the AI discussion is how muddy it is. Are we talking about llm's diffusion models, or classification systems? Do we mean to say that we're talking about transformers or the underlying architecture? Are we discussing huge data centers or device based AI? Nascent, active, or dormant compute? And the same is true for the ethics, legal, and data governance conversation.
Every single one of these things is a different discussion.
AI is not a monolith.
Re: (Score:3)
When you read "AI" on Slashdot, or in a business article, it's almost certainly a language model made by one of a handful of companies.
If it's about Nvidia it might have something to do with robots.
Re: (Score:3)
If you would go into specifics, things would be falsifiable. You can't make claims about specifics without people telling you afterward where you were wrong. If you use the most vague terms, you can later decide on what specifics you meant that actually followed your prediction. If you now say "AI bubble will burst" you can later say "Of course I only meant the integration into product X and not the whole AI industry" if the larger burst doesn't happen.
1st + 3rd are bubbles -- not 2nd (infrustructure) (Score:2)
1st Bubble -- Yes, there is a lot of money invested into so many companies rapidly building AI products and, mostly, retrofitting old things. Obviously, they cannot all be winners.
2nd Non-Bubble -- In spite of a lot of failed projects, there will be more project and there is already too little infrustructure for the demand. For example, Microsoft 365 Copilot is clearly a weak GPT model but vastly under powered (easy to notice if you use it). The first wave of AI companies will have some winners and a lot
Indeed (Score:2)
The analyses of the whole mess and why it is a mess are getting better. Hopefully the inevitable collapse and return to sanity (with a few major players rightfully dead and soon to be forgotten) is not too far ahead.
Dude discovers railway mania (Score:1)
Feed the hype (Score:1)
That's not how anything works (Score:2)
A bubble is "a good or fortunate situation that is isolated from reality or unlikely to last". What's good about it, profit for those who are profiting. Why's it isolated from reality, all three of those reasons. Why's it unlikely to last, reality is inexorable, no amount of ignoring it will cause it to change.
One bubble, at least three reasons why it's bubbling. Probably we could identify a bunch more, like nerd fantasy. One of the consequences of techbros being in a position to decide what society does wi
Nonsense (Score:2)
Re: (Score:1)
Not bubbles (Score:1)
There is one bubble with three facets. (Score:1)
Itâ(TM)s all speculation in different facets of the same bubble. Investor speculation in equities, business speculation in infrastructure and building markets and power, and broad social speculation about the disruption into the future.
When the internet bubble burst around 2001, similar forces were at play. And also: do you think the socio-economic impact of the internet today is less than what was being foretold in 1999? No. Some people had stupid ideas, there was a gold rush, a lot of people lost, ot
Bullshit (Score:2)
Those aren't the only bubbles... (Score:2)
There's also the subject-verb disagreement bubble.