Glad to see Tom's blog on HN - as usual a great write up. A number of us have been chatting about this for several months now, and the take is fairly sober.
Meta commentary but I've grown weary of how commentary by actual domain experts in our industry are underrepresented and underdiscussed on HN in favor of emotionally charged takes.
TLDR: Lucent was committing various forms of accounting fraud and had an unhealthy cash flow position, and had their primary customers on economically dangerous ground. Nvidia meanwhile appears to be above board, has strong cash flow, and has extremely strong dominant customers (eg customers that could reduce spending but can survive a downturn). Therefore there's no clear takeaway: similarities but also differences. Risk and a lot of debt as well as hyperscalers insulating themselves from some of that risk... but at the same time as lot more cash to burn.
By that I mean, those were the last consoles where performance improvements delivered truly new experiences, where the hardware mattered.
Today, any game you make for a modern system is a game you could have made for the PS3/Xbox 360 or perhaps something slightly more powerful.
Certainly there have been experiences that use new capabilities that you can’t literally put on those consoles, but they aren’t really “more” in the same way that a PS2 offered “more” than the PlayStation.
I think in that sense, there will be some kind of bubble. All the companies that thought that AI would eventually get good enough to suit their use case will eventually be disappointed and quit their investment. The use cases where AI makes sense will stick around.
It’s kind of like how we used to have pipe dreams of certain kinds of gameplay experiences that never materialized. With our new hardware power we thought that maybe we could someday play games with endless universes of rich content. But now that we are there, we see games like Starfield prove that dream to be something of a farce.
Some great insights with some less interesting in there. I didn’t know about the SPVs, that’s sketchy and now I wanna know how much of that is going on. The MIT study that gets pulled out for every critical discussion of AI was an eye roll for me. But very solid analysis of the quants.
How much of a threat is custom silicon to Nvidia remains an open question to me. I kinda think, by now, we can say they’re similar but different enough to coexist in the competitive compute landscape?
Looking at the last chapter of the essay, there was a lot of illegal activity by lucent in the runup to the collapse. Today, We won’t know the list of shady practices until the bubble bursts. I doubt Tom could legally speculate, he’d likely be sued into oblivion if he even hinted at malfeasance by these trillion dollar companies.
I worked at a mom and pop ISP in the 90s. Lucent did seem at the forefront of internet equipment at the time. We used Portmaster 3s to handle dial up connections. We also looked into very early wireless technology from Lucent.
Something I wanted to mention, only somewhat tanget. The Telecommunications Act of 1996 forced telecommunication companies to lease out their infrastructure. It massively reduced the prices an ISP had to pay to get T1, because, suddenly, there was competition. I think a T1 went from 1800 a month in 1996, to around 600 a month in 1999. It was a long time ago, so my memory is hazy.
But, wouldn't you know it, the Telecommunication companies sued the FCC and the Telecommunications Act was gutted in 2003
I think the fundamental issue is the uncertainty of achieving AGI with baked in fundamentals of reasoning.
Almost 90% of topline investments appear to be geared around achieving that in the next 2-5 years.
If that doesn’t come to pass soon enough, investors will loose interest.
Interest has been maintained by continuous growth in benchmark results.
Perhaps this pattern can continue for another 6-12 months before fatigue sets in, there are no new math olympiads to claim a gold medal on…
Whats next is to show real results, in true software development, cancer research, robotics.
I am highly doubtful the current model architecture will get there.
With all the major players like Amzn, Msft and Alphabet going for their own custom chips and restrictions on selling to China it will be interesting to see how Nvidia does.
I personally would prefer China to get to parity on node size and get competitive with nvidia. As that is the only way I see the world not being taken over by the tech oligarchy.
Are these companies developing InfiniBand-class interconnects to pair with their custom chips? Without equivalent fabric, they can’t replace NVIDIA GPUs for large-scale training.
At a "telecom of telecom" we (they) were still lighting up dark fiber 15 years later (2015) when mobile data for cell carriers finally created enough demand. Hard to fathom the amount of overbuild.
The only difference is fiber optic lines remained useful the whole time. Will these cards have the same longevity?
Fiber networks were using less
than 0.002% of available capacity,
with potential for 60,000x speed
increases. It was just too early.
I doubt we will see unused GPU capacity. As soon as we can prompt "Think about the codebase over night. Try different ways to refactor it. Tomorrow, show me your best solution." we will want as much GPU time at the current rate as possible.
If a minute of GPU usage is currently $0.10, a night of GPU usage is 8 * 60 * 0.1 = $48. Which might very well be worth it for an improved codebase. Or a better design of a car. Or a better book cover. Or a better business plan.
One of the things before AI in the market was that capital had limited growth opportunities. Tech, which was basically a universe of scaled out crud apps, was where capital would keep going back into.
AI is a lot more useful than hyper scaled up crud apps. Comparing this to the past is really overfitting imho.
The only argument against accumulating GPUs is that they get old and stop working. Not that it sucks, not that it’s not worth it. As in, the argument against it is actually in the spirit of “I wish we could keep the thing longer”. Does that sound like there’s no demand for this thing?
The AI thesis requires getting on board with what Jenson has been saying:
1) We have a new way to do things
2) The old ways have been utterly outclassed
3) If a device has any semblance of compute power, it will need to be enhanced, updated, or wholesale replaced with an AI variant.
There is no middle ground to this thesis. There is no “and we’ll use AI here and here, but not here, therefore we predictably know what is to come”.
Get used to the unreal. Your web apps could truly one day be generated frame by frame by a video model. Really. The amount of compute we’ll need will be staggering.
Knowing history of past bubbles is only mildly informative. The dotcom bubble was different than the railroads bubble etc.
The only thing to keep in mind is that all of this is about business and ROI.
Given the colossal investments, even if the companies finances are healthy and not fraudulent, the economic returns have to be unprecedented or there will be a crash.
Great points. I am bullish on AI but also wary of accounting practices. Tom says Nvidia's financials are different from Lucent's but that doesn't mean we shouldn't be wary.
The Economist has a great discussion on depreciation assumptions having a huge impact on how the finances of the cloud vendors are perceived[1].
Revenue recognition and expectations around Oracle could also be what bursts the bubble. Coreweave or Oracle could be the weak point, even if Nvidia is not.
This reminds me of SGI at the peak of the dot-com bubble.
SGI (Silicon Graphics) made the 3D hardware that many companies relied on for their own businesses, in the days before Windows NT and Nvidia came of age.
Alias|Wavefront and Discreet were two companies where their product cycles were very tied in the SGI product cycles, with SGI having some ownership, whether it be wholly owned or spun out (as SGI collapsed). I can't find the reporting from the time, but it seemed to me that the SGI share price was propped up by product launches from the likes of Alias|Wavefront or Discreet. Equally, the 3D software houses seemed to have share prices propped up by SGI product launches.
There was also the small matter of insider trading. If you knew the latest SGI boxes were lemons then you could place your bets of the 3D software houses accordingly.
Eventually Autodesk, Computer Associates and others eventually owned all the software, or, at least, the user bases. Once upon a time these companies were on the stock market and worth billions, but then they became just another bullet point in the Autodesk footer.
My prediction is that a lot of AI is like that, a classic bubble, and, when the show moves on, all of these AI products will get shoehorned into the three companies that will survive, with competition law meaning that it will be three rather than two eventual winners.
Equally, much like what happened with SGI, Nvidia will eventually come a cropper due to the evaluations due to today's hype and hubris not delivering.
The biggest issue with Nvidia is their revenue is not recurring but the market is treating their stock as it were, which is correlated with all semi stocks, with a one-time massive CAPEX investment lasting 1-2 years.
Simple as this - as to why its just not possible for this to continue.
The smartest finance folks I know say that this “irrational exuberance” works until it doesn’t. Meaning nobody really thinks it’s sustainable, but companies and VCs chasing the AI hype bubble have backed themselves into a corner where the only way to stop the bubble from bursting is to keep inflating the bubble.
The fate of the bubble will be decided by Wall Street not tech folks in the valley. Wall Street is already positioning itself for the burst and there’s lots of finance types ready to call party over and trigger the chaos that lets them make bank on the bubble’s implosion.
These finance types (family offices, small secret investment funds) eat clueless VCs throwing cash on the fire for lunch… and they’re salivating at what’s ahead. It’s a “Big Short” once in 20-30 years type opportunity.
> However, what’s become clear is that OpenAI plans to pay for Nvidia’s graphics processing units (GPUs) through lease arrangements, rather than upfront purchases
I wish someone here could explain it to a dummy like me. Nvidia tells OpenAI: heres some GPUs, can you pay for them over 5 years. How is this an "investment" by Nvidia? That reference keeps calling this an investment, but what they describe is a lease agreement. Why do they call it an investment? What am I missing?
I wonder if the buying customers of Nvidia are going to find the self’s left with the overcapacity. Certainly people are waking up to LLM challenges and as budgets focus more on useful applications, smaller language models, how much of that demand will remain.
Also, depreciation schedules beyond useful life of an asset may not be fraud but I’d call it a bit too creative for my liking.
The one thing I don't understand is this assumption that demand for GPUs for training is going to keep growing at the rate they grew so far.
I get the demand for new applications, which require inference, but nowadays with so many good (if not close to SOTA) models available for free and the ability to run them on consumer hardware (apple M4 or AMD Max APUs), is there any demand for applications that justify a crazy amount of investment in GPUs?
The telecom bubble built infrastructure for something that didn’t exist, they built anticipating the need for high didn’t come in time.
The gpu bubble is different. Nvidia is actually selling gpus in spades. So it’s not comparable to the telecom bubble. Now the question remains how many more gpus can they sell? That depends on the kind of services that are built and how their adoption takes off. So now is it a bubble or just frothy at the top? There is definitely going to be a pull back and some adjustment, but I cannot say how bad it is
34 comments
[ 3.1 ms ] story [ 42.3 ms ] threadMeta commentary but I've grown weary of how commentary by actual domain experts in our industry are underrepresented and underdiscussed on HN in favor of emotionally charged takes.
By that I mean, those were the last consoles where performance improvements delivered truly new experiences, where the hardware mattered.
Today, any game you make for a modern system is a game you could have made for the PS3/Xbox 360 or perhaps something slightly more powerful.
Certainly there have been experiences that use new capabilities that you can’t literally put on those consoles, but they aren’t really “more” in the same way that a PS2 offered “more” than the PlayStation.
I think in that sense, there will be some kind of bubble. All the companies that thought that AI would eventually get good enough to suit their use case will eventually be disappointed and quit their investment. The use cases where AI makes sense will stick around.
It’s kind of like how we used to have pipe dreams of certain kinds of gameplay experiences that never materialized. With our new hardware power we thought that maybe we could someday play games with endless universes of rich content. But now that we are there, we see games like Starfield prove that dream to be something of a farce.
How much of a threat is custom silicon to Nvidia remains an open question to me. I kinda think, by now, we can say they’re similar but different enough to coexist in the competitive compute landscape?
Something I wanted to mention, only somewhat tanget. The Telecommunications Act of 1996 forced telecommunication companies to lease out their infrastructure. It massively reduced the prices an ISP had to pay to get T1, because, suddenly, there was competition. I think a T1 went from 1800 a month in 1996, to around 600 a month in 1999. It was a long time ago, so my memory is hazy.
But, wouldn't you know it, the Telecommunication companies sued the FCC and the Telecommunications Act was gutted in 2003
https://en.wikipedia.org/wiki/Competitive_local_exchange_car...
Almost 90% of topline investments appear to be geared around achieving that in the next 2-5 years.
If that doesn’t come to pass soon enough, investors will loose interest.
Interest has been maintained by continuous growth in benchmark results. Perhaps this pattern can continue for another 6-12 months before fatigue sets in, there are no new math olympiads to claim a gold medal on…
Whats next is to show real results, in true software development, cancer research, robotics.
I am highly doubtful the current model architecture will get there.
I personally would prefer China to get to parity on node size and get competitive with nvidia. As that is the only way I see the world not being taken over by the tech oligarchy.
The only difference is fiber optic lines remained useful the whole time. Will these cards have the same longevity?
(I have no idea just sharing anecdata)
If a minute of GPU usage is currently $0.10, a night of GPU usage is 8 * 60 * 0.1 = $48. Which might very well be worth it for an improved codebase. Or a better design of a car. Or a better book cover. Or a better business plan.
AI is a lot more useful than hyper scaled up crud apps. Comparing this to the past is really overfitting imho.
The only argument against accumulating GPUs is that they get old and stop working. Not that it sucks, not that it’s not worth it. As in, the argument against it is actually in the spirit of “I wish we could keep the thing longer”. Does that sound like there’s no demand for this thing?
The AI thesis requires getting on board with what Jenson has been saying:
1) We have a new way to do things
2) The old ways have been utterly outclassed
3) If a device has any semblance of compute power, it will need to be enhanced, updated, or wholesale replaced with an AI variant.
There is no middle ground to this thesis. There is no “and we’ll use AI here and here, but not here, therefore we predictably know what is to come”.
Get used to the unreal. Your web apps could truly one day be generated frame by frame by a video model. Really. The amount of compute we’ll need will be staggering.
The only thing to keep in mind is that all of this is about business and ROI.
Given the colossal investments, even if the companies finances are healthy and not fraudulent, the economic returns have to be unprecedented or there will be a crash.
They are all chasing a golden goose.
The Economist has a great discussion on depreciation assumptions having a huge impact on how the finances of the cloud vendors are perceived[1].
Revenue recognition and expectations around Oracle could also be what bursts the bubble. Coreweave or Oracle could be the weak point, even if Nvidia is not.
[1] https://www.economist.com/business/2025/09/18/the-4trn-accou...
SGI (Silicon Graphics) made the 3D hardware that many companies relied on for their own businesses, in the days before Windows NT and Nvidia came of age.
Alias|Wavefront and Discreet were two companies where their product cycles were very tied in the SGI product cycles, with SGI having some ownership, whether it be wholly owned or spun out (as SGI collapsed). I can't find the reporting from the time, but it seemed to me that the SGI share price was propped up by product launches from the likes of Alias|Wavefront or Discreet. Equally, the 3D software houses seemed to have share prices propped up by SGI product launches.
There was also the small matter of insider trading. If you knew the latest SGI boxes were lemons then you could place your bets of the 3D software houses accordingly.
Eventually Autodesk, Computer Associates and others eventually owned all the software, or, at least, the user bases. Once upon a time these companies were on the stock market and worth billions, but then they became just another bullet point in the Autodesk footer.
My prediction is that a lot of AI is like that, a classic bubble, and, when the show moves on, all of these AI products will get shoehorned into the three companies that will survive, with competition law meaning that it will be three rather than two eventual winners.
Equally, much like what happened with SGI, Nvidia will eventually come a cropper due to the evaluations due to today's hype and hubris not delivering.
Simple as this - as to why its just not possible for this to continue.
The fate of the bubble will be decided by Wall Street not tech folks in the valley. Wall Street is already positioning itself for the burst and there’s lots of finance types ready to call party over and trigger the chaos that lets them make bank on the bubble’s implosion.
These finance types (family offices, small secret investment funds) eat clueless VCs throwing cash on the fire for lunch… and they’re salivating at what’s ahead. It’s a “Big Short” once in 20-30 years type opportunity.
> However, what’s become clear is that OpenAI plans to pay for Nvidia’s graphics processing units (GPUs) through lease arrangements, rather than upfront purchases
I wish someone here could explain it to a dummy like me. Nvidia tells OpenAI: heres some GPUs, can you pay for them over 5 years. How is this an "investment" by Nvidia? That reference keeps calling this an investment, but what they describe is a lease agreement. Why do they call it an investment? What am I missing?
Also, depreciation schedules beyond useful life of an asset may not be fraud but I’d call it a bit too creative for my liking.
Time will tell.
I get the demand for new applications, which require inference, but nowadays with so many good (if not close to SOTA) models available for free and the ability to run them on consumer hardware (apple M4 or AMD Max APUs), is there any demand for applications that justify a crazy amount of investment in GPUs?
https://capitalgains.thediff.co/p/vendor-financing
The gpu bubble is different. Nvidia is actually selling gpus in spades. So it’s not comparable to the telecom bubble. Now the question remains how many more gpus can they sell? That depends on the kind of services that are built and how their adoption takes off. So now is it a bubble or just frothy at the top? There is definitely going to be a pull back and some adjustment, but I cannot say how bad it is