People tend to equate this to the railroad boom when saying that infrastructure spending will yield durable returns into the future no matter what.
When the railroad bubble popped we had railroads. Metal and sticks, and probably more importantly, rights-of-way.
If this is a bubble, and it pops, basically all the money will have been spent on Nvidia GPUs that depreciate to 0 over 4 years. All this GPU spending will need to be done again, every 4 years.
Hopefully we at least get some nuclear power plants out of this.
The vast majority of the dot-com comparison that I personally see are economic, not technological. People (or at least the ones I see) are claiming that the bubble mechanics of e.g. circular trading and over-investments are similar to the dot-com bubble, not that the AI technology is somehow similar the internet (it obviously isn’t). And to that extent we are in the year 1999 not 1995.
When this article are claiming both sides of the debate, I believe only one of them are real (the ones hyping up the technology). While there are people like me who are pessimistic about the technology, we are not in any position of power, and our opinion on the matter is basically a side noise. I think a much more common (among people with any say in the future of this technology) is the believe that this technology is not yet at a point which warrants all this investment. There were people who said that about the internet in 1999, and they were proven 100% correct in the months that followed.
While I mostly agree with the article's premise (that AI will cause more software development to happen, not less) I disagree with two parts:
1. the opening premise comparing AI to dial-up internet; basically everyone knew the internet would be revolutionary long before 1995. Being able to talk to people halfway across the world on a BBS? Sending a message to your family on the other side of the country and them receiving it instantly? Yeah, it was pretty obvious this was transformative. The Krugman quote is an extreme, notable outlier, and it gets thrown out around literally every new technology, from blockchain to VR headsets to 3DTVs, so just like, don't use it please.
2. the closing thesis of
> Consider the restaurant owner from earlier who uses AI to create custom inventory software that is useful only for them. They won’t call themselves a software engineer.
The idea that restaurant owners will be writing inventory software might make sense if the only challenge of creating custom inventory software, or any custom software, was writing the code... but it isn't. Software projects don't fail because people didn't write enough code.
>> Consider the restaurant owner from earlier who uses AI to create custom inventory software that is useful only for them. They won’t call themselves a software engineer.
I have a suspicion this is LLM text, sounds corny. There are dozens open source solutions, just look one up.
“But the fact that some geniuses were laughed at does not imply that all who are laughed at are geniuses. They laughed at Columbus, they laughed at Fulton, they laughed at the Wright brothers. But they also laughed at Bozo the Clown.”
Because some notable people dismissed things that wound up having profound effect on the world, it does not mean that everything dismissed will have a profound effect.
We could just as easily be "peak Laserdisc" as "dial-up internet".
> 1. Economic strain (investment as a share of GDP)
> 2. Industry strain (capex to revenue ratios)
> 3. Revenue growth trajectories (doubling time)
> 4. Valuation heat (price-to-earnings multiples)
> 5. Funding quality (the resilience of capital sources)
> His analysis shows that AI remains in a demand-led boom rather than a bubble, but if two of the five gauges head into red, we will be in bubble territory.
This seems like a more quantitative approach than most of "the sky is falling", "bubble time!", "circular money!" etc analyses commonly found on HN and in the news. Are there other worthwhile macro-economic indicators to look at?
It's fascinating how challenging it is meaningfully compare current recent events to prior economic cycles such as the y2k tech bubble. It seems like it should be easy but AFAICT it barely even rhymes.
Nice article, but somewhat overstates how bad 1995 was meant to be.
A single image generally took nothing like a minute. Most people had moved to 28.8K modems that would deliver an acceptable large image in 10-20 seconds. Mind you, the full-screen resolution was typically 800x600 and color was an 8-bit palette… so much less data to move.
Moreover, thanks to “progressive jpeg”, you got to see the full picture in blocky form within a second or two.
And of course, with pages was less busy and tracking cookies still a thing of the future, you could get enough of a news site up to start reading in less time that it can take today.
One final irk is that it’s little overdone to claim that “For the first time in history, you can exchange letters with someone across the world in seconds”. Telex had been around for decades, and faxes, taking 10-20 seconds per page were already commonplace.
Not sure the dial-up analogy fits, instead I tend to think we are in the mainframe period of AI, with large centralised computing models that are so big and expensive to host, only a few corporations can afford to do so. We rent a computing timeshare from them (tokens = punch cards).
I look forward to the "personal computing" period, with small models distributed everywhere...
I like to think of it more broadly, and that we are currently in the era of the first automobile. [0]
LLMs are the internal combustion engine, and chatbot UIs are at the "horseless carriage" phase.
My personal theory is that even if models stopped making major advancements, we would find cheaper and more useful ways to use them. In the end, our current implementations will look like the automobile pictured below.
I'm not a big google fan, but I really like the "Google AI Edge Gallery" android app [0]. In particular, I've been chatting with the "Gemma-3n-E2B-it" model when I don't have an internet connection, and it's really decent!
Mainframes still exist, and they actually make a lot of sense from physics perspective. It's good idea to run transactions in a big machine rather than distributed, the latter is less energy efficient.
I think the misconception is that things cannot be overpriced for reasons other than inefficiency.
Most of the big services seem to waste so much time clunking through updating and editing files.
I'm no expert but I can't help feeling there's lots of things they could be doing vastly better in this regard - presumably there is lots to do and they will get around to it.
Funny how this guy thinks he knows exactly what's up with AI, and how "others" are "partly right and wrong." Takes a bit of hubris to be so confident. I certainly don't have the hubris to think I know exactly how it's all going to go down.
I recall the unit economics making sense for all these other industries and bubbles (short of maybe tulips, which you could plant…) . Sure there were over-valuation bubbles because of speculatory demand, but right now the assumption seems to be “first to AGI wins” but that… may not happen.
The key variable for me in this house of cards is how long folks will wait before they need to see their money again, and whether these companies will go in the right direction long enough given these valuations to get to AGI. Not guaranteed and in the meantime society will need to play ball (also not a guarantee)
> If you told someone in 1995 that within 25 years [...] most people would find that hard to believe.
That's not how I remember it (but I was just a kid so I might be misremembering?)
As I remember (and what I gather from media from the era) late 80s/early 90s were hyper optimistic about tech. So much so that I distinctly remember a ¿german? TV show when I was a kid where they had what amounts to modern smartphones, and we all assumed that was right around the corner. If anything, it took too damn long.
Were adults outside my household not as optimistic about tech progress?
Recently, in my city, the garbage trucks started to come equipped with a device I call "The Claw" (think Toy Story). The truck drives to your curb where your bin is waiting, and then The Claw extends, grasps the bin, lifts it into the air and empties the contents into the truck before setting it down again.
The Claw allows a garbage truck to be crewed by one man where it would have needed two or three before, and to collect garbage much faster than when the bins were emptied by hand. We don't know what the economics of such automation of (physical) garbage collection portend in the long term, but what we do know is that sanitation workers are being put out of work. "Just upskill," you might say, but until Claw-equipped trucks started appearing on the streets there was no need to upskill, and now that they're here the displaced sanitation workers may be in jeopardy of being unable to afford to feed their families, let alone find and train in some new marketable skill.
So no, we're in the The Claw era of AI, when business finds a new way to funge labor with capital, devaluing certain kinds of labor to zero with no way out for those who traded in such labor. The long-term implications of this development are unclear, but the short-term ones are: more money for the owner class, and some people are out on their ass without a safety net because this is Goddamn America and we don't brook that sort of commie nonsense here.
I feel like this article is too cute. The internet, and the state of the art of computing in general has been driven by one thing and one thing alone: Moore’s Law. In that very real sense, it means that the semiconductor and perhaps more generally even just TSMC is responsible for the rise of the internet and the success of it.
We’re at the end of Moore’s Law, it’s pretty reasonable to assume. 3nm M5 chips means there are—what—a few hundred silicon atoms per transistor? We’re an order of magnitude away from .2 nm which is the diameter of a single silicon atom.
My point is, 30 years have passed since dial up. That’s a lot of time to have exponentially increasing returns.
There’s a lot of implicit assumption that “it’s just possible” to have a Moore’s Law for the very concept of intelligence. I think that’s kinda silly.
My head canon is that the thing that preemptively pops the bubble is Apple coming out and saying, very publicly, that AI is a dead end, and they are dropping it completely (no more half assed implicit promises).
And not just that, they come out with an iPhone that has _no_ camera as an attempt to really distance themselves from all the negative press tech (software and internet in particular) has at the moment.
I would go so far as to say we are still in the computing dial-up era. We're at the tail end, maybe - we don't write machine code any longe, mostly, and we've abstracted up a few levels but we're still writing code. Eventually computing is something that will be everywhere, like air, and natural language interfaces will be nearly exclusively how people interact with computing machines. I don't think the idea of 'writing software' is something that will stick around, I think we're in a very weird and very brief little epoch where that is a thing.
how much does the correction here hew to making an AI model just look like standardized API calls with predictable responses? If you took away all the costs (data centers, water consumption, money, etc) I still wouldn't use an LLM as a first choice because it's wrong enough of the time to make it useless -- I have to verify everything it says, which is how I would have approached a task in the first place. If we put that analogy into manufacturing, it's "I have to QA everything off of the line _without exception_ and I get frequent material waste"
If you make the context small enough, we're back at /api/create /api/read /api/update /api/delete; or, if you're old-school, a basic function
People keep comparing the AI boom to the Dotcom bubble. They’re wrong. Others point to the Railway Mania of the 1840s — closer, but still not quite right.
The real parallel is Canal Mania — Britain’s late-18th-century frenzy to dig waterways everywhere. Investors thought canals were the future of transport. They were, but only briefly.
Today’s AI runs on GPUs — chips built for rendering video games, not thinking machines. Adapting them for AI is about as sensible as adapting a boat to travel across land. Sure, it moves — but not quickly, not cheaply, and certainly not far.
It works for now, but the economics are brutal. Each new model devours exponentially more power, silicon, and capital. It just doesn't scale.
The real revolution will come with new, hardware built for the job (that hasn't been invented yet) — thousands of times faster and more efficient. When that happens, today’s GPU farms will look like quaint relics of an awkward, transitional age: grand, expensive, and obsolete almost overnight.
Great analysis but one thing overlooked is that current gen advanced AI could in five or ten years (or less) be run from the smartphone or desktop, which could negate all the capex from the hyperscalers and also Nvidia, which presents a massive target for competitors right now. The self same AI revolution we’re seeing created right now could take itself down if AI tooling becomes widespread.
69 comments
[ 3.3 ms ] story [ 66.7 ms ] threadWhen the railroad bubble popped we had railroads. Metal and sticks, and probably more importantly, rights-of-way.
If this is a bubble, and it pops, basically all the money will have been spent on Nvidia GPUs that depreciate to 0 over 4 years. All this GPU spending will need to be done again, every 4 years.
Hopefully we at least get some nuclear power plants out of this.
When this article are claiming both sides of the debate, I believe only one of them are real (the ones hyping up the technology). While there are people like me who are pessimistic about the technology, we are not in any position of power, and our opinion on the matter is basically a side noise. I think a much more common (among people with any say in the future of this technology) is the believe that this technology is not yet at a point which warrants all this investment. There were people who said that about the internet in 1999, and they were proven 100% correct in the months that followed.
1. the opening premise comparing AI to dial-up internet; basically everyone knew the internet would be revolutionary long before 1995. Being able to talk to people halfway across the world on a BBS? Sending a message to your family on the other side of the country and them receiving it instantly? Yeah, it was pretty obvious this was transformative. The Krugman quote is an extreme, notable outlier, and it gets thrown out around literally every new technology, from blockchain to VR headsets to 3DTVs, so just like, don't use it please.
2. the closing thesis of
> Consider the restaurant owner from earlier who uses AI to create custom inventory software that is useful only for them. They won’t call themselves a software engineer.
The idea that restaurant owners will be writing inventory software might make sense if the only challenge of creating custom inventory software, or any custom software, was writing the code... but it isn't. Software projects don't fail because people didn't write enough code.
I have a suspicion this is LLM text, sounds corny. There are dozens open source solutions, just look one up.
Because some notable people dismissed things that wound up having profound effect on the world, it does not mean that everything dismissed will have a profound effect.
We could just as easily be "peak Laserdisc" as "dial-up internet".
> 1. Economic strain (investment as a share of GDP)
> 2. Industry strain (capex to revenue ratios)
> 3. Revenue growth trajectories (doubling time)
> 4. Valuation heat (price-to-earnings multiples)
> 5. Funding quality (the resilience of capital sources)
> His analysis shows that AI remains in a demand-led boom rather than a bubble, but if two of the five gauges head into red, we will be in bubble territory.
This seems like a more quantitative approach than most of "the sky is falling", "bubble time!", "circular money!" etc analyses commonly found on HN and in the news. Are there other worthwhile macro-economic indicators to look at?
It's fascinating how challenging it is meaningfully compare current recent events to prior economic cycles such as the y2k tech bubble. It seems like it should be easy but AFAICT it barely even rhymes.
That is the real dial-up thinking.
Couldn't AI like be their custom inventory software?
Codex and Claud Code should not even exist.
A single image generally took nothing like a minute. Most people had moved to 28.8K modems that would deliver an acceptable large image in 10-20 seconds. Mind you, the full-screen resolution was typically 800x600 and color was an 8-bit palette… so much less data to move.
Moreover, thanks to “progressive jpeg”, you got to see the full picture in blocky form within a second or two.
And of course, with pages was less busy and tracking cookies still a thing of the future, you could get enough of a news site up to start reading in less time that it can take today.
One final irk is that it’s little overdone to claim that “For the first time in history, you can exchange letters with someone across the world in seconds”. Telex had been around for decades, and faxes, taking 10-20 seconds per page were already commonplace.
I look forward to the "personal computing" period, with small models distributed everywhere...
LLMs are the internal combustion engine, and chatbot UIs are at the "horseless carriage" phase.
My personal theory is that even if models stopped making major advancements, we would find cheaper and more useful ways to use them. In the end, our current implementations will look like the automobile pictured below.
[0] https://group.mercedes-benz.com/company/tradition/company-hi...
[0] https://play.google.com/store/apps/details?id=com.google.ai....
I think the misconception is that things cannot be overpriced for reasons other than inefficiency.
I'm no expert but I can't help feeling there's lots of things they could be doing vastly better in this regard - presumably there is lots to do and they will get around to it.
The key variable for me in this house of cards is how long folks will wait before they need to see their money again, and whether these companies will go in the right direction long enough given these valuations to get to AGI. Not guaranteed and in the meantime society will need to play ball (also not a guarantee)
That's not how I remember it (but I was just a kid so I might be misremembering?)
As I remember (and what I gather from media from the era) late 80s/early 90s were hyper optimistic about tech. So much so that I distinctly remember a ¿german? TV show when I was a kid where they had what amounts to modern smartphones, and we all assumed that was right around the corner. If anything, it took too damn long.
Were adults outside my household not as optimistic about tech progress?
The Claw allows a garbage truck to be crewed by one man where it would have needed two or three before, and to collect garbage much faster than when the bins were emptied by hand. We don't know what the economics of such automation of (physical) garbage collection portend in the long term, but what we do know is that sanitation workers are being put out of work. "Just upskill," you might say, but until Claw-equipped trucks started appearing on the streets there was no need to upskill, and now that they're here the displaced sanitation workers may be in jeopardy of being unable to afford to feed their families, let alone find and train in some new marketable skill.
So no, we're in the The Claw era of AI, when business finds a new way to funge labor with capital, devaluing certain kinds of labor to zero with no way out for those who traded in such labor. The long-term implications of this development are unclear, but the short-term ones are: more money for the owner class, and some people are out on their ass without a safety net because this is Goddamn America and we don't brook that sort of commie nonsense here.
Because we all know how essential the internet is nowadays.
We’re at the end of Moore’s Law, it’s pretty reasonable to assume. 3nm M5 chips means there are—what—a few hundred silicon atoms per transistor? We’re an order of magnitude away from .2 nm which is the diameter of a single silicon atom.
My point is, 30 years have passed since dial up. That’s a lot of time to have exponentially increasing returns.
There’s a lot of implicit assumption that “it’s just possible” to have a Moore’s Law for the very concept of intelligence. I think that’s kinda silly.
And not just that, they come out with an iPhone that has _no_ camera as an attempt to really distance themselves from all the negative press tech (software and internet in particular) has at the moment.
If you make the context small enough, we're back at /api/create /api/read /api/update /api/delete; or, if you're old-school, a basic function
The real parallel is Canal Mania — Britain’s late-18th-century frenzy to dig waterways everywhere. Investors thought canals were the future of transport. They were, but only briefly.
Today’s AI runs on GPUs — chips built for rendering video games, not thinking machines. Adapting them for AI is about as sensible as adapting a boat to travel across land. Sure, it moves — but not quickly, not cheaply, and certainly not far.
It works for now, but the economics are brutal. Each new model devours exponentially more power, silicon, and capital. It just doesn't scale.
The real revolution will come with new, hardware built for the job (that hasn't been invented yet) — thousands of times faster and more efficient. When that happens, today’s GPU farms will look like quaint relics of an awkward, transitional age: grand, expensive, and obsolete almost overnight.
>We’re in the 1950s equivalent of the internet boom — dial-up modems exist, but YouTube doesn’t.