Most mature heavy industries have a commodity layer and a premium layer. Cars for example. Traditionally, small hutchbacks had margins so tight that large auto conglomerates would make no money from them directly. But they would share parts with pricier models.
AI seems on its way to be a largely commoditized heavy industry. This may be good news for the economy as a whole. There would be smaller markets for specialized premium models. If you think you need THAT model, then you have to truly and fully pay up.
If you're fine with the commodity layer, then you have many options heavily competing on price. The sheer scale and efficiency of AWS, Google or Microsoft will enable them to sell compute cheaper than the average company could do in-house.
Interestingly, the only two companies with a defensible moat in all this are Samsung and SK Hynix. You cannot have AI without memory, be it an open-source or a frontier model, cloud or in-house. And from 2028 on those two will offer HBM5 (Micron a year or two later), which will give them the technological lead that produces those moats. And possibly around the same time they will offer the next innovation: cHBM.
These moves were inspired by a succession of stories: Samsung and SK Hynix have extremely low forward earnings compared to US semi stocks, even Sandisk - up. Memory is cyclical and the additional capacity coming online in 2028 is gigantic - down. The next story to come will be about next generation memory. The stock market is not governed by true fundamentals, never has been.
NVIDIA has been accused of neglecting the quality of their non-AI GPU division. Which *was* their core business before the AI boom took off.
The last six video cards (GPU's) I've purchased are NVIDIA, because of well-deserved superiority over their chief competitor AMD. But my latest, an RTX 5080, is not backwards compatible with a lot of older apps. And the company shows little intention of patching these incompatibilities (assuming it's even possible).
Bottom line here, is even for NVIDIA products, it's buyer beware.
One small thing re: your footnote: the way that Korean leveraged ETFs are constructed they end up actually having to sell shares, not just futures. For those who want a deep dive into the financial mechanics, I write about it here: https://liquidcontext.substack.com/p/twice-the-risk-twice-the-excitement
One difference between the AI boom and frequently cited analogs like the fiber and railroad booms, is that each new data center seems to be meeting current demand. The fiber laid down just before the dotcom crash was "dark", built in anticipation of demand that did eventually come, but came too late for a lot of the original builders of it.
But even with all the new data centers built so far, GPUs for rent are still scarce. Are there new data centers sitting dark somewhere? Kimi 3 might be bad news for closed-model frontier labs but you still have to run that inference somewhere and the Kimi model is not particularly token-efficient.
Just look at SpaceX who is now getting about $1.4B a month renting out GPUs to Anthropic and other companies, more than it takes in from StarLink. There is obviously high demand still for GPUs.
What is the likely sex ratio of the traders in the scenarios that have created what is characterized as a bubble? As a demographic, young men are the most likely to gamble. How are the young men in South Korea doing these days? Perhaps this is another factor to consider.
"A whole lot of Koreans used leveraged ETFs and other borrowing methods to borrow huge amounts of money"
I'm a pretty unsophisticated investor. If I came to believe that some stock was going to rocket upward in value, I'd just buy that stock. I wouldn't even think to look into leveraged ETFs or "other borrowing methods", and if they were brought to my attention, I'd consider them suspicious and exotic. So how are all these unsophisticated investors even learning about these options let alone being convinced to mass adopt them to such an extent that they can create a market-wide bubble?
My mindset is kind of like yours; If I believe in a product, whether it's a stock or an index or Bitcoin or whatever, I just buy it. Leveraged derivative plays aren't even on my radar.
These young guys are a different breed though. My favorite restaurant server in San Diego, my favorite bartender in Vegas, my favorite Uber driver, and two of my favorite valets -- they all trade stocks via leveraged ETF instruments. So when these young guys tell you that they're long stocks, they don't mean they bought an index fund or an actual stock, they mean they're long UPRO (for example), a 3x ETF, or some kind of leveraged semiconductor fund. It's a different world now. As a retired financial services guy (53), it's kind of baffling. No country for old men, as they say.
The rise and fall of Samsung and SK Hynix stocks are totally expected—if memory prices ballooned in a frenzy, a reality check was bound to follow. The AI boom is no different from any past bubble in internet, mobile, or cloud infrastructure. Nvidia losing 90% of its market cap wouldn't even be unprecedented.
What strikes me is actually how slow cheap alternatives like DeepSeek and Kimi and more others took to emerge; market forces should have accelerated this sooner. AI today is simply too expensive and energy-hungry. If we apply First Principles and compare current AI hardware to the human brain—a carbon-based neural network running on roughly 20 watts—the GPU-heavy, high-power-consumption route reveals a fundamental efficiency bug. Yet, driven by FOMO, tech giants are trapped in a data-center arms race regardless.
It reminds me of a famous prediction from the mobile boom when someone claimed, "In a few years, chips will be sold at the price of sand." Back then, many actually nodded along. History is full of dramatic swings fueled by hype and doom. But eventually, common sense prevails: memory shouldn't be astronomically expensive, chips shouldn't cost a fortune, and AI infrastructure shouldn't burn a city's worth of power. I wonder how many years it will take for the world to return to sanity, where AI matures into cheap, invisible, ubiquitous utility—like water or electricity.
Running open source models doesn’t reduce the amount of memory, GPUs or power needed for AI inference, even if run on premises instead of in a data center. It doesn’t reduce the demand for compute for training by much either, although it moves most of it to China , although those model makers cheat by getting a head-start by distilling OpenAI and Anthropic models which were trained in the USA.
Completely agree! That's a great observation. It reminds me of how electricity replaced steam but ended up burning even more coal overall.
What really matters now is driving the unit cost down. The insane energy and memory overhead we see today is proof that our current GPU-centric infrastructure is brute-forcing compute at terrible energy efficiency.
Transitioning to domain-specific architectures—like Google's TPUs, FPGAs, and custom NPUs—is essential to lowering power consumption. Diversifying chip architectures to tackle the power density wall seems like the only logical next step for the industry. The shift toward cheaper, more energy-efficient silicon is bound to happen, wouldn't you agree?
One cannot underestimate mob psychology operating among financial professionals. I once worked for a high leverage real estate investment firm that soared in the 80s and crashed in the 90s. It amazed me how brokers scrambled to get a piece of our deals
And for the hyperscalers remember the depreciation & amortisation associated with all this spend will start feeding through the P&L over the coming years. Something to bear in mind when thinking about valuations.
One thing I wonder is, in the US, who is on the other side of the debt? Who is lending Oracle and IMB the money? And those companies won't go bankrupt, but more worryingly, who is lending money to more marginal borrowers? In 1873, wasn't the biggest problem not that rail-road companies suffered, but that banks lost money lending to them? Is there evidence of big lenders being heavily exposed this time?
You might find this video VERY interesting and relevant to this and several other of your posts. Why Abundance Is an Illusion with Jeff Currie
It's about finance and fundamentals.
In this episode, Nate is joined by Jeff Currie for a wide-boundary look at what happens when the buffers that have suppressed energy price signals for fifty years finally run dry. Using his decades of experience as a former commodity strategist at Goldman Sachs and as a current senior advisor at The Carlyle Group, Jeff walks through why the "crack spread" between crude and refined products just hit its highest level in three decades. He also describes why draining strategic reserves is, in actuality, simply a bet that scarcity can be avoided rather than solved – in Currie’s eyes, the West's refusal to admit scarcity since the 70s has left it structurally unprepared, particularly compared to China's security-driven build-out of nuclear, solar, and battery capacity. He also lays out the "Grand Bargain" underlying the postwar dollar system, wherein the U.S. protects global sea lanes in exchange for global trade running through New York. Jeff explains why a failure to reopen the Strait of Hormuz could unravel this arrangement, bringing forward consequences that would land hardest on middle-class Americans' access to credit and consumption.
Most mature heavy industries have a commodity layer and a premium layer. Cars for example. Traditionally, small hutchbacks had margins so tight that large auto conglomerates would make no money from them directly. But they would share parts with pricier models.
AI seems on its way to be a largely commoditized heavy industry. This may be good news for the economy as a whole. There would be smaller markets for specialized premium models. If you think you need THAT model, then you have to truly and fully pay up.
If you're fine with the commodity layer, then you have many options heavily competing on price. The sheer scale and efficiency of AWS, Google or Microsoft will enable them to sell compute cheaper than the average company could do in-house.
Interestingly, the only two companies with a defensible moat in all this are Samsung and SK Hynix. You cannot have AI without memory, be it an open-source or a frontier model, cloud or in-house. And from 2028 on those two will offer HBM5 (Micron a year or two later), which will give them the technological lead that produces those moats. And possibly around the same time they will offer the next innovation: cHBM.
These moves were inspired by a succession of stories: Samsung and SK Hynix have extremely low forward earnings compared to US semi stocks, even Sandisk - up. Memory is cyclical and the additional capacity coming online in 2028 is gigantic - down. The next story to come will be about next generation memory. The stock market is not governed by true fundamentals, never has been.
NVIDIA has been accused of neglecting the quality of their non-AI GPU division. Which *was* their core business before the AI boom took off.
The last six video cards (GPU's) I've purchased are NVIDIA, because of well-deserved superiority over their chief competitor AMD. But my latest, an RTX 5080, is not backwards compatible with a lot of older apps. And the company shows little intention of patching these incompatibilities (assuming it's even possible).
Bottom line here, is even for NVIDIA products, it's buyer beware.
One small thing re: your footnote: the way that Korean leveraged ETFs are constructed they end up actually having to sell shares, not just futures. For those who want a deep dive into the financial mechanics, I write about it here: https://liquidcontext.substack.com/p/twice-the-risk-twice-the-excitement
One difference between the AI boom and frequently cited analogs like the fiber and railroad booms, is that each new data center seems to be meeting current demand. The fiber laid down just before the dotcom crash was "dark", built in anticipation of demand that did eventually come, but came too late for a lot of the original builders of it.
But even with all the new data centers built so far, GPUs for rent are still scarce. Are there new data centers sitting dark somewhere? Kimi 3 might be bad news for closed-model frontier labs but you still have to run that inference somewhere and the Kimi model is not particularly token-efficient.
Just look at SpaceX who is now getting about $1.4B a month renting out GPUs to Anthropic and other companies, more than it takes in from StarLink. There is obviously high demand still for GPUs.
This post (and the excellent accompanying slide deck) is quite on point here: https://www.exponentialview.co/p/the-state-of-the-ai-economy
What is the likely sex ratio of the traders in the scenarios that have created what is characterized as a bubble? As a demographic, young men are the most likely to gamble. How are the young men in South Korea doing these days? Perhaps this is another factor to consider.
I interpret this as a nice affirmation of the "VT and chill" approach.
"A whole lot of Koreans used leveraged ETFs and other borrowing methods to borrow huge amounts of money"
I'm a pretty unsophisticated investor. If I came to believe that some stock was going to rocket upward in value, I'd just buy that stock. I wouldn't even think to look into leveraged ETFs or "other borrowing methods", and if they were brought to my attention, I'd consider them suspicious and exotic. So how are all these unsophisticated investors even learning about these options let alone being convinced to mass adopt them to such an extent that they can create a market-wide bubble?
My mindset is kind of like yours; If I believe in a product, whether it's a stock or an index or Bitcoin or whatever, I just buy it. Leveraged derivative plays aren't even on my radar.
These young guys are a different breed though. My favorite restaurant server in San Diego, my favorite bartender in Vegas, my favorite Uber driver, and two of my favorite valets -- they all trade stocks via leveraged ETF instruments. So when these young guys tell you that they're long stocks, they don't mean they bought an index fund or an actual stock, they mean they're long UPRO (for example), a 3x ETF, or some kind of leveraged semiconductor fund. It's a different world now. As a retired financial services guy (53), it's kind of baffling. No country for old men, as they say.
Almost like sports gambling apps, and probably just as easy to throw your money away.
The rise and fall of Samsung and SK Hynix stocks are totally expected—if memory prices ballooned in a frenzy, a reality check was bound to follow. The AI boom is no different from any past bubble in internet, mobile, or cloud infrastructure. Nvidia losing 90% of its market cap wouldn't even be unprecedented.
What strikes me is actually how slow cheap alternatives like DeepSeek and Kimi and more others took to emerge; market forces should have accelerated this sooner. AI today is simply too expensive and energy-hungry. If we apply First Principles and compare current AI hardware to the human brain—a carbon-based neural network running on roughly 20 watts—the GPU-heavy, high-power-consumption route reveals a fundamental efficiency bug. Yet, driven by FOMO, tech giants are trapped in a data-center arms race regardless.
It reminds me of a famous prediction from the mobile boom when someone claimed, "In a few years, chips will be sold at the price of sand." Back then, many actually nodded along. History is full of dramatic swings fueled by hype and doom. But eventually, common sense prevails: memory shouldn't be astronomically expensive, chips shouldn't cost a fortune, and AI infrastructure shouldn't burn a city's worth of power. I wonder how many years it will take for the world to return to sanity, where AI matures into cheap, invisible, ubiquitous utility—like water or electricity.
Running open source models doesn’t reduce the amount of memory, GPUs or power needed for AI inference, even if run on premises instead of in a data center. It doesn’t reduce the demand for compute for training by much either, although it moves most of it to China , although those model makers cheat by getting a head-start by distilling OpenAI and Anthropic models which were trained in the USA.
Completely agree! That's a great observation. It reminds me of how electricity replaced steam but ended up burning even more coal overall.
What really matters now is driving the unit cost down. The insane energy and memory overhead we see today is proof that our current GPU-centric infrastructure is brute-forcing compute at terrible energy efficiency.
Transitioning to domain-specific architectures—like Google's TPUs, FPGAs, and custom NPUs—is essential to lowering power consumption. Diversifying chip architectures to tackle the power density wall seems like the only logical next step for the industry. The shift toward cheaper, more energy-efficient silicon is bound to happen, wouldn't you agree?
One cannot underestimate mob psychology operating among financial professionals. I once worked for a high leverage real estate investment firm that soared in the 80s and crashed in the 90s. It amazed me how brokers scrambled to get a piece of our deals
And for the hyperscalers remember the depreciation & amortisation associated with all this spend will start feeding through the P&L over the coming years. Something to bear in mind when thinking about valuations.
Good post!
One thing I wonder is, in the US, who is on the other side of the debt? Who is lending Oracle and IMB the money? And those companies won't go bankrupt, but more worryingly, who is lending money to more marginal borrowers? In 1873, wasn't the biggest problem not that rail-road companies suffered, but that banks lost money lending to them? Is there evidence of big lenders being heavily exposed this time?
Corporate debt, so lenders who bought their bonds on in the hook right?
Makes sense. I'm just wondering if these lenders are heavily concentrated
You might find this video VERY interesting and relevant to this and several other of your posts. Why Abundance Is an Illusion with Jeff Currie
It's about finance and fundamentals.
In this episode, Nate is joined by Jeff Currie for a wide-boundary look at what happens when the buffers that have suppressed energy price signals for fifty years finally run dry. Using his decades of experience as a former commodity strategist at Goldman Sachs and as a current senior advisor at The Carlyle Group, Jeff walks through why the "crack spread" between crude and refined products just hit its highest level in three decades. He also describes why draining strategic reserves is, in actuality, simply a bet that scarcity can be avoided rather than solved – in Currie’s eyes, the West's refusal to admit scarcity since the 70s has left it structurally unprepared, particularly compared to China's security-driven build-out of nuclear, solar, and battery capacity. He also lays out the "Grand Bargain" underlying the postwar dollar system, wherein the U.S. protects global sea lanes in exchange for global trade running through New York. Jeff explains why a failure to reopen the Strait of Hormuz could unravel this arrangement, bringing forward consequences that would land hardest on middle-class Americans' access to credit and consumption.
https://www.youtube.com/watch?v=ij1_uxiXmm8
For a much deeper dive on this run up and subsequent crash, see: https://youtu.be/nJtL9MBVj48?si=gEyITkTU-dbay9OR
does anybody know roughly how they arrived at that 2.5 trillion number?