I admit to only being in this industry for three decades now, and only designing and implementing the thermal/power control algo of an AI chip family for three years in that time, but it's the first time I hear of chips "wearing under high intensity use".
He writes as if only datacenters and network equipment remain after the AI bubble bursts. Like there won't be any AI models anymore, nothing left after the big training runs and trillion-dollar R&D, and no inference served.
> Most of the money is being spent on incredibly expensive GPUs that have a 1-3 year lifespan due to becoming obsolete quickly and wearing out under constant, high-intensity use.
How about chips during the dotcom period? What was their lifespan?
Without moralizing or assuming the worst intentions of oligarchs, globalists, evil capitalists, and so on, I still don’t understand how a consumption based economy continues to fund the build out (oil->Saudi Arabia->LPs->OpenAI) when the technology likely removes the income of its consumers. Help me understand.
I hope we'll get back to building things that actually matter -- solutions that help real people; products that are enjoyable to use, and are satisfying to build.
As the noise fades, and with luck, the obsession with slapping "AI" on everything will fade with it. Too many hype-driven CEOs are chasing anything but substance.
Some AI tools may survive because they're genuinely useful, but I worry that most won't be cost-effective without heavy subsidies.
Once the easy money dries up, the real engineers and builders will still be here, quietly making things that work.
Altman's plea -- "Come on guys, we just need a few trillion more!" -- and that error-riddled AI slide deck will be the meme that marks the top of the market.
Also, the eco-system plays the biggest controlling role in the bubble and its aftermath. Ecosystem of social, political and business developments. Dotcom aftermath still had the wind from all the ecosytem trends that brought the dotcom back with bigger force. If the post AI hype world still has high priority for these talking bots, then maybe it's comparable to dotcom. If the world has other bigger basic issues that need attention, then yes, it could become a pile of silent silicon.
AI chips and bespoke data centers are closer to tulips than rail or fiber in terms of deprecated assets. They're not fungible stranded assets with long shelf life or room for improvment. Bubble bursting also points to current model towards AI is inherently not economically viable, i.e. we'll still be inferencing off existing models but won't be pouring 100billions into improving them. TLDR much more all or nothing gambit than past infra booms.
Compares China's current 2025 energy capability with US at the end of 2023. Then starts throwing 2025 uptake numbers for solar and wind into the 2023 numbers as if having a TWO YEAR GAP in the numbers dont matter at all.
One key difference in all of this is that people were not predicting the dotcom bubble, so there was a surplus left after it popped. It was a surprise.
This AI bubble already has lots of people with their forks and knifes waiting to capitalize on a myriad of possible surpluses after the burst. There's speculation on top of _the next bubble_ and how it will form, even before this one pops.
Local/open-weight models are already incredibly competent. Right now a Mac Studio with 256GB can be found for less than $5000, and an equivalent workstation will likely be 50% cheaper in a year. If anything that price is higher because of the boom, rather than subsidized by a potential bubble. It can run a 8bit quant of GPT-OSS 120B, or 4bit quant of GLM-4.6 using only an extra 100-200W. That energy use comes out to about 100 Joules or 1/4 Wh per query and response, and is already competitive with the power efficiency of even Google's offerings.
I think that people doing work in many professions with these offline tools alone could more than double their productivity compared to their productivity two years ago. Furthermore if the usage was shared in order to lower idle time, such as 20 machines for 100 workers, the initial capital outlay is even lower.
Perhaps investors will not see the returns they expect, but it is difficult to image how even the current state of AI doesn't vastly change the economy. There could be significant business failures among cloud providers and attempts to rapidly increase the cost of admission to closed models, but there's essentially no possibility of productivity regressing to a pre-AI levels.
The AI "bubble" won't burst just like the "internet bubble" didn't burst.
the dotcom bubble was a result of investors jumping on the hype train all at once and then getting off of it all at once.
Yes, investors will eventually find another hype train to jump on, but unlike 2000, we have tons of more retail investors and AI is also not a brand new tech sector, it's built upon the existing well established and "too big to fail" internet/ecommerce infrastructure. Random companies slapping AI on things will fail but all the real AI use cases will only expand and require more and more resources.
OpenAI alone just hit 800M MAU. That will easily double in a few years. There will be adjustments,corrections and adaptations of course but the value and wealth it generates is very real.
I'm no seer, I can't predict the future but I don't see a massive popping of some unified AI bubble anytime soon.
Cell phones have been through how many generations between the 80s and now? All the past generations are obsolete, but the investment in improving the technology (which is really a continuation of WWII era RF engineering) means we have readily available low cost miniature comms equipment. It doesn’t matter that the capex on individual phones was wasted.
Same for GPUs/LLMs? At some point things will mature and we’ll be left with plentiful, cheap, high end LLM access, on the back of the investment that has been made. Whether or not it’s running on legacy GPUs, like some 90s fiber still carries traffic, is meaningless. It’s what the investment unlocks.
I think we will enter a neo-Luddite era at some point post-AI boom where it suddenly becomes fashionable to live one’s life with simple retro style technology, and social networks and much of the internet will just become places for bitter old people to complain amongst themselves and share stupid memes. Social media was cool when it was more genuine, but it got increasingly fake, and now with AI it could reach peak-fake. If people want genuine, what is more genuine than the real world?
It will become cool for you to become inaccessible, unreachable, no one knowing your location or what you’re doing. People might carry around little beeper type devices that bounce small pre-defined messages around on encrypted radio mesh networks to say stuff like “I’m okay” or “I love you”, and that’s it. Maybe they are used for contactless payments as well.
People won’t really bother searching the web anymore they’ll just ask AI to pull up whatever information they need.
The question is, with social media on the decline, with the internet no longer used for recreational purposes, what else are people going to do? Feels like the consumer tech sector will shrink dramatically, meaning that most tech written will be made to create “hard value” instead of soft. Think anything having to do with movement of data and matter, or money.
Much of the tech world and government plans are built on the assumption that people will just continue using tech to its maximum utility, even when it is clearly bad for them, but what if that simply weren’t the case? Then a lot of things fall apart.
I wrote a similar article (not published yet), but my conclusion was "Free GPUs for everyone" or at least cheap ones. Right now H100 are very specialized for the AI pipeline, but so were GPUs before the AI boom. I expect we will find good use for them.
If there is a downturn in AI use due to a bubble then the countries that have built up their energy infrastructure using renewal energy and nuclear (both have decade long returns after the initial investment) will have cheaper electricity which will lead to a future competitive advantage. Gas powered power plants on the other hand require constant gas to convert to electricity. The price of gas would become the price of electricity regardless and thus very little advantage.
I think this article was written with AI, it has a that contrastive sentence signature.
And if you can't think of what to do with massive amounts of matrix multiplying compute, that's pretty sad IMO. Not to mention the huge amount of energy demand probably creating a peace dividend in energy generation for decades to come.
We have also gotten a lot of open models we wouldn't of had without the AI boom competition, not mention all the other interesting stuff coming out in the open model world.
China has already won the game. We developed media and wealth management as invididualistic, they approached the problems institutionally and built infrastructure at the pivotal point renewables became affordable. They made an army of technical engineers, we scattered innovation in comp sci/VC acceleration/automation and sat on our asses as asset managers. AI works in secluded modules like robotics and revision. As a general tool it's navel gazing energy sink. AI was a lure we fell for and wasted money and energy while smart players like Apple and China knew their reality.
GPUs will last much longer after the crash because there won't be any money available to replace them. You can either keep running the existing GPUs or throw them in the trash. The GPUs will keep running as long as they can generate enough revenue from inference to cover electricity. Tokens will become very cheap but free tokens might go away.
AI datacenters aren't that specialized. You can buy a 1 GW datacenter and put 300 MW of equipment in it and sell 2/3 of the gas generators. You'll have to buy some InRow units.
The AI stack isn't as proprietary as it sounds. GPU rental exists today and it's pretty interoperable. Ironically, Nvidia's moat has made their GPUs a de facto standard that is very well understood and supported by whatever software you want.
Centuries ago, the building materials for castles was taken by the locals and reused to build their houses with. Those data centres have a lot of sheet metal in them. Air conditioning units as well. The upgraded power networks around those data centres will alter where people live too. Oh and there's a lot of electric motors with copper wire in all the fans and hard drives those computers have.
32 comments
[ 3.0 ms ] story [ 45.5 ms ] threadMaybe we can finally have a Rosie from the Jetsons.
How about chips during the dotcom period? What was their lifespan?
As the noise fades, and with luck, the obsession with slapping "AI" on everything will fade with it. Too many hype-driven CEOs are chasing anything but substance.
Some AI tools may survive because they're genuinely useful, but I worry that most won't be cost-effective without heavy subsidies.
Once the easy money dries up, the real engineers and builders will still be here, quietly making things that work.
Altman's plea -- "Come on guys, we just need a few trillion more!" -- and that error-riddled AI slide deck will be the meme that marks the top of the market.
In 10 years GPUs will have a lifespan for 5-7 years. The rate of improvement on this front has been slowing down faster then CPU.
Compares China's current 2025 energy capability with US at the end of 2023. Then starts throwing 2025 uptake numbers for solar and wind into the 2023 numbers as if having a TWO YEAR GAP in the numbers dont matter at all.
This AI bubble already has lots of people with their forks and knifes waiting to capitalize on a myriad of possible surpluses after the burst. There's speculation on top of _the next bubble_ and how it will form, even before this one pops.
That is absolutely disgusting, by the way.
I think that people doing work in many professions with these offline tools alone could more than double their productivity compared to their productivity two years ago. Furthermore if the usage was shared in order to lower idle time, such as 20 machines for 100 workers, the initial capital outlay is even lower.
Perhaps investors will not see the returns they expect, but it is difficult to image how even the current state of AI doesn't vastly change the economy. There could be significant business failures among cloud providers and attempts to rapidly increase the cost of admission to closed models, but there's essentially no possibility of productivity regressing to a pre-AI levels.
the dotcom bubble was a result of investors jumping on the hype train all at once and then getting off of it all at once.
Yes, investors will eventually find another hype train to jump on, but unlike 2000, we have tons of more retail investors and AI is also not a brand new tech sector, it's built upon the existing well established and "too big to fail" internet/ecommerce infrastructure. Random companies slapping AI on things will fail but all the real AI use cases will only expand and require more and more resources.
OpenAI alone just hit 800M MAU. That will easily double in a few years. There will be adjustments,corrections and adaptations of course but the value and wealth it generates is very real.
I'm no seer, I can't predict the future but I don't see a massive popping of some unified AI bubble anytime soon.
Same for GPUs/LLMs? At some point things will mature and we’ll be left with plentiful, cheap, high end LLM access, on the back of the investment that has been made. Whether or not it’s running on legacy GPUs, like some 90s fiber still carries traffic, is meaningless. It’s what the investment unlocks.
...whether it is profitable is another matter
It will become cool for you to become inaccessible, unreachable, no one knowing your location or what you’re doing. People might carry around little beeper type devices that bounce small pre-defined messages around on encrypted radio mesh networks to say stuff like “I’m okay” or “I love you”, and that’s it. Maybe they are used for contactless payments as well.
People won’t really bother searching the web anymore they’ll just ask AI to pull up whatever information they need.
The question is, with social media on the decline, with the internet no longer used for recreational purposes, what else are people going to do? Feels like the consumer tech sector will shrink dramatically, meaning that most tech written will be made to create “hard value” instead of soft. Think anything having to do with movement of data and matter, or money.
Much of the tech world and government plans are built on the assumption that people will just continue using tech to its maximum utility, even when it is clearly bad for them, but what if that simply weren’t the case? Then a lot of things fall apart.
And if you can't think of what to do with massive amounts of matrix multiplying compute, that's pretty sad IMO. Not to mention the huge amount of energy demand probably creating a peace dividend in energy generation for decades to come.
We have also gotten a lot of open models we wouldn't of had without the AI boom competition, not mention all the other interesting stuff coming out in the open model world.
Typical pessimist drivel.
GPUs will last much longer after the crash because there won't be any money available to replace them. You can either keep running the existing GPUs or throw them in the trash. The GPUs will keep running as long as they can generate enough revenue from inference to cover electricity. Tokens will become very cheap but free tokens might go away.
AI datacenters aren't that specialized. You can buy a 1 GW datacenter and put 300 MW of equipment in it and sell 2/3 of the gas generators. You'll have to buy some InRow units.
The AI stack isn't as proprietary as it sounds. GPU rental exists today and it's pretty interoperable. Ironically, Nvidia's moat has made their GPUs a de facto standard that is very well understood and supported by whatever software you want.