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...
It varied a lot by region. At the mom and pop ISP I worked at, we went from paying around $1,500/month for a T1 to $500 to, eventually around $100/month for the T1 loop to the customer plus a few grand a month for an OC12 SONET ring that we used to backhaul the T1 (and other circuits) back to our datacenter.
But, all of it was driven by the Telecommunications Act requirement for ILECs to sell unbundled network facilities - all of the CLECs we purchased from were using the local ILEC for the physical part of the last mile for most (> 75%) of the circuits they sold us.
One interesting thing that happened was that for a while in the late 90’s, when dialup was still a thing, we could buy a voice T1 PRI for substantially less than a data T1 ($250 for the PRI vs $500 for the T1.) The CLEC’s theory was our dialup customers almost all had service from the local ILEC, and the CLEC would be paid “reciprocal compensation” fees by the ILEC for the CLEC accepting calls from them.
In my market, when the telecommunications act reform act was gutted, the ILEC just kept on selling wholesale/unbundled services to us. I think they had figured out at that point that it was a very profitable line of business if they approached it the right way.
Regarding the price of connection, it's also worth mentioning that while T1 and other T-channel and OCx connection remains in high use, 1996-1999 is also the period where DSL became readily available & was a very fine choice for many needs. This certainly created significant cost pressure on other connectivity options.
Monopolies gum up the system, reward the institutional capital rather than innovation capital, and prevent new entrants from de-ossifying and being the renewing forest fire.
We've been so lax on antitrust. Google, Apple, Meta, Amazon - they all need to be broken up. Our economy and our profession would be better for it.
Innovation should be a treadmill.
YC and a16z want this.
If it's true that this regulation was what helped jumpstart the internet it's an interesting counterpoint to the apocalyptic predictions of people when these regulations are undone. (net neutrality comes to mind as well)
I've never heard anyone claim before that just having these laws on the books for a small period of time is "enough".
Why would it be enough? This legislation prevents monopolies from abusing position, therefore we will repeal it the moment it turns out to be useful?
Yeah, it takes time to consolidate power again, that does not mean the legislation is not good.
But the price war was inevitable. And the telecoms bubble was highly likely in any case.
Telecoms investment was a response to crazy valuations of dot-com stocks.
Fiber networks were using less
than 0.002% of available capacity,
with potential for 60,000x speed
increases. It was just too early.
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.
I'd argue we very certainly will. Companies are gobbling up GPUs like there's no tomorrow, assuming demand will remain stable and continue growing indefinitely. Meanwhile LLM fatigue has started to set in, models are getting smaller and smaller and consumer hardware is getting better and better. There's no way this won't end up with a lot of idle GPUs.
Has it?
I think there is this compulsion to think that LLMs are made for senior devs, and if devs are getting wary of LLMs, the experiment is over.
I'm not a programmer, my day job isn't tech, and the only people I know who express discontent with LLMs are a few of programmer friends I have. Which I get, but everyone else is using them gleefully for all manner of stuff. And now I am seeing the very first inklings of completely non-technical people making bespoke applets for themselves.
From OpenAI, programming is ~4% of chatGPTs usage. That's 96% being used for other stuff.
I don't see any realistic or grounded forecast that includes a diminishing demand for compute. We're still at the tip of adoption...
Nvidia is betting the farm on reinventing GPU compute every 2 years. The GPUs wont end up idle, because they will end up in landfills.
Do I believe that's likely, no, but it is what I believe Nvidia is aiming for.
2% of it is dedicated to thinking.
My guess is that as a species, we will turn a similar percentage of our environment into thinking matter.
If there are a billion houses on planet earth, 2% of it are 20 million datacenters we still have to build.
That is nothing. Coding is done via text. Very soon people will use generative AI for high resolution movies. Maybe even HDR and high FPS (120 maybe?). Such videos will very likely cost in the range of $100-$1000 per minute. And will require lots and lots of GPUs. The US military (and I bet others as well) are already envisioning generative AI use for creating a picture of the battlespace. This type of generation will be even more intensive than high resolution videos.
The cost/benefit analysis doesn't add up for two reasons:
First, a refactored codebase works almost the same as non-refactored one, that is, the tangible benefit is small.
Second, how many times are you going to refactor the codebase? Once and... that's it. There's simply no need for that much compute for lack of sufficient beneficial work.
That is, the present investments are going to waste unless we automate and robotize everything, I'm OK with that but it's not where the industry is going.
I've seen lots of claims about AI coding skill, but that one might be able to improve (and not merely passably extend) a codebase is a new one. I'd want to see it before I believe it.
Other things might need to be done in two stages. You might ask the agent to first identify where code violates CQRS, then for each instance, explain the problem, and spawn a sub-agent to address that problem.
Other things the agent might identify this way: multiple implications, use of conflicted APIs, poor separation of concerns at a module or class level.
I don't typically let the agent do any of this end to end, but I would typically manually review findings before spawning subagents with those findings.
And you can't really hack / outsmart feedback loops.
Just because something is conceptually possible, interaction with the real rest of the world separates a possible from an optimal solution.
The low hanging fruits/ obvious incremental improvements might be quickly implemented by LLMs based on established patterns in their training data.
That doesn't get you from 0 to 1 dollar, though and that's what it's all about.
LLMs are a great tool. But, the real world is far too nuanced to be captured in text and tokens. So, LLMs will be a great productivity boosting tool like a calculator or a spreadsheet. Expecting it to do more is science fiction.
This is the fundamental error I see people making. LLMs can’t operate independently today, not on substantive problems. A lot of people are assuming that they will some day be able to, but the fact is that, today, they cannot.
The AI bubble has been driven by people seeing the beginning of an S-curve and combining it with their science-fiction fantasies about what AI is capable of. Maybe they’re right, but I’m skeptical, and I think the capabilities we see today are close to as good as LLMs are going to get. And today, it’s not good enough.
A year ago they need an extensive harness to get silver, and two years ago they could hardly multiply 1000x10000.
Terence Tao tweeted yesterday about using GPT5 to help quickly solve a problem he was working on.
Because the main reason for the price premiums in AI-class GPUs are the gobs of insanely fast memory, and that is very much not underutilized. AI companies grab GPUs with as much memory (at the fastest memory bandwidth) as possible and underclock the GPU to save on power. Linus Tech Tips had a great video about the H200 that touched on this this week: https://www.youtube.com/watch?v=lNumJwHpXIA
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.
Bezos is just saying shit to generate hype. All these executives are just saying shit. There is no plan. You must treat these people as morons who understand nothing.
Anyone who knows even the slightest details about datacenter design knows what moving heat is the biggest problem. This is the exact thing that being in space makes infinitely harder. "Datacenters in space" is an idea you come up with only if you are a moron who knows nothing about either datacenters or space.
If nothing else this is the singular reason you should treat AI as a bubble. All of the people at the helm of it have not a single fucking clue what they're talking about. They all keep running their mouth with utter nonsense like this.
Why does "Not needing labor at all" need to be in space?
We are a long way from that. At least 10 years, probably never gonna happen.
I agree, but would like to maybe build out that theory. When we start talking about the mechanisms of the past we end up over-constricting the possibility space. There were a ton of different ways the dotcom bubble COULD have played out, and only one way it did. If we view the way it did as the only way it possibly could have, we'll almost certainly miss the way the next bubble will play out.
How much is a 10 year old GPU worth? Where is the “dwdm but for GPUs?”.
There truly are interesting times and we have the benefit of being in them.
> How much is a 10 year old GPU worth? Where is the “dwdm but for GPUs?”.
From other sources cited in TFA it seems GPUs won't last 3 years, let alone 10! But I think we know what the "DWDM for GPUs" is -- it's the processing efficiency gains that we've seen over the last few years which keeps driving the per-token prices sharply down.
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)
You're looking for advancement in carriages unaware of the 'automobile' that made 5g and ftth deployment at scale possible.
The article cites anecdotal 1-2 years due to the significant stress.
This didn't last that much longer and many places were trying to diversify into managed services (data dog for companues on Orem network and server equipment,etc) which they call "unregulated" revenue.
Add written an things business, irrational exuberance can kill you.
As the AI spending bubble gives out, Nvidia's profit growth will slow dramatically (single digits), and slamming into a wall (as Cisco did during the telecom bubble; leading up to the telecom crash, Cisco was producing rather insane quarter over quarter growth rates).
AC, ups, generators not to mention the severs.
That's the thing with fiber it was still useful. The cards at either end are easy to add, waaaayyy cheaper and higher perf (they're were no cards on end of dark lines) 15 years later.
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.
There's also plenty of argument to be made that it's already here. AI can hold forth on pretty much any topic, and it's occasionally even correct. Of course to many (not saying you), the only acceptable bar is perfect factual accuracy, a deep understanding of meanings, and probably even a soul. Which keeps breathing life into the old joke "AI is whatever computers still can't do".
If you speak with AI researchers, they all seem reasonable in their expectations.
... but I work with non-technical business people across industries and their expectations are NOT reasonable. They expect ChatGPT to do their entire job for $20/month and hire, plan, budget accordingly.
12 months later, when things don't work out, their response to AI goes to the other end of the spectrum -- anger, avoidance, suspicion of new products, etc.
Enough failures and you have slowing revenue growth. I think if companies see lower revenue growth (not even drops!), investors will get very very nervous and we can see a drop in valuations, share prices, etc.
This is entirely on the AI companies and their boosters. Sam Altman literally says gpt 5 is "like having a team of PhD-level experts in your pocket." All the commercials sell this fantasy.
An extraordinary claim for which I would like to see the extraordinary evidence. Because every single interview still available on YT form 3 years ago ...had these researchers putting AGI 3 to 5 years out ...A complete fairy tale as the track to AGI is not even in sight.
If you want to colonize the Solar System the track is clear. If you to have Fusion, the track is clear. AGI track ?
Simple as this - as to why its just not possible for this to continue.
Funnily enough when you spend some months thinking into this intensively the result is that a monetary investment into the company that will bring about the singularity / AGI is the most irrational thing one can do.
If the enterprise is successful and the singularity/AGI is benign you won't need money anymore, if the experiment fail the possibility of things going rogue is very high, or even the panic from a possible series of rogue events.
So for the first time the rational thing would be to either spend those money to learn poker/chess/videogames or whatever game we will play with each other to feel cool while the AI takes care of everything else, or maybe outright spend money on coke and strippers given the chance of doomsday.
Their margin is ridiculous and they are still unable to meet demand.
Obviously its a bubble but thats meaningless for anyone but the richest to manage.
The rest of us are just ants.
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?
Of course, cern is still going to use their FPGA hyper-optimized for their specific trigger model for the LHC, and apple is gojng to use a specialized low power ASIC running a quantized model for hello Siri, but I meant the majority usecase.
I think that there are plenty of competitors in the "LLMs with open weights" space to essentially make the models a commodity, so all that is left is the compute cost and there is no way that someone will be running a datacenter in a way that is cheaper than "the computer that I already have running on my desk".
I expect models will get larger again once everyone is doing their inference on B200s, but the RL training budget is where the insatiable appetite sits right now.
If you tell me that people are pouring all that money into data centers because they believe that most applications will use some form of LLM or VLM as the main driver of machine-to-machine and machine-to-person interface, I'd be more inclined to buy it. But then I'd respond that it seems that LLMs are reaching a point of diminishing returns and the big next move is to make it easy and faster to distill/fine-tune the LLMs for specific business needs, which is something that should be possible to do with the existing infra already (I guess?)
> 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?
NVidia could protect itself against OpenAI bankrupcy by adding a clause to the lease saying that if OpenAI goes bankrupt, Nvidia gets its GPUs back. So the risk would only be that the lease would be aborted sooner than expected.
At the end Nvidia retains ownership of what are probably very low value assets.
Contrast that with car leases: there is a robust market for used cars.
Nvidia is in effect financing the GPUs by not requiring the full payment up front.
Do the lease payments add up to the total cost?
EDIT: Interesting note
> CPUs historically have 5-10 years of useful life , while GPUs in AI datacenters last 1-3 years in practice , despite 6-year accounting assumptions.30,31 Evidence from Google architects shows GPUs at 60-70% utilization survive 1-2 years , with 3 years maximum.31 Meta’s Llama 3 training experienced 9% annual GPU failure rates , suggesting 27% failure over 3 years.31
This is surely the most important line in the piece? In what world would this much demand not lead to alternatives emerging?
(Assuming the upside, yes if the demand is not there in two years then yes it’s all going to burn)
https://open.spotify.com/episode/2ieRvuJxrpTh2V626siZYQ?si=2...
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.
The PS3 is the last console to have actual specialized hardware. After the PS3, everything is just regular ol' CPU and regular ol' GPU running in a custom form factor (and a stripped-down OS on top of it); before then, with the exception of the Xbox, everything had customized coprocessors that are different from regular consumer GPUs.
I hope that's where we are, because that means my experience will still be valuable and vibe coding remains limited to "only" tickets that take a human about half a day, or a day if you're lucky.
Given the cost needed for improvements, it's certainly not implausible…
…but it's also not a sure thing.
I tried "Cursor" for the first time last week, and just like I've been experiencing every few months since InstructGPT was demonstrated, it blew my mind.
My game metaphor is 3D graphics in the 90s: every new release feels amazing*, such a huge improvement over the previous release, but behind the hype and awe there was enough missing for us to keep that cycle going for a dozen rounds.
* we used to call stuff like this "photorealistic": https://www.reddit.com/r/gaming/comments/ktyr1/unreal_yes_th...
But the way how it stayed niche shows how it's not just about new gameplay experiences.
Compare with the success of the Wii Sports and Wii Fit, which I would guess managed it better, though through a different kind of hardware that you are thinking about ?
And I kind of expect the next Nintendo console to have a popular AR glasses option, which also would only have been made possible thanks to improving hardware (of both kinds).
I could be very wrong, obviously.
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.
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.
Calling a VC a "domain expert" is like calling an alcoholic a "libation engineer." VC blogs are, in the best case, mildly informative, and in the worst, borderline fraudulent (the Sequoia SBF piece being a recent example, but there are hundreds).
The incentives are, even in a true "domain expert" case (think: doctors, engineers, economists), often opaque. But when it comes to VCs, this gets ratcheted up by an order of magnitude.
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.
Certainly it suggests that “this time is different” without saying it in a quotable fashion.
The metrics it provides seem useful. What are the metrics it is missing?
But the answer is, "kinda"? There are similarities, but the AI buildout is worse in some ways (more concentration, GPU backed debt) and better in others (capacity is being used, vendors actually have cash flow).
The conclusion:
> Unlike the telecom bubble, where demand was speculative & customers burned cash , this merry-go-round has paying riders.
Seems a little short sighted to me. IMO, there is a definite echo, but we are in the mid-late stage, not the end stage.
It's simply not fair to compare Lucent at the end of a bubble with Nvidia in the middle, and that is what the author did.
If you haven't listened to the referenced interview between Thompson and Kedrosky, I'd do so: https://www.theringer.com/podcasts/plain-english-with-derek-...
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.
No - it's very hard to successfully bet against anything in finance, and VCs and non-public investments are particularly hard. When you go long, you simply buy something and hold it until you decide to sell. If you short, you have to worry about borrowing shares, paying short fees, and having unlimited risk.
How would you even begin to bet against OpenAI specifically? The closest proxy I can think of is shorting NVDA.
There's also nobody whose job it is to make big one-time shorts. Like you said, it's a once in 20-30 years opportunity, so no one builds a hedge fund dedicated to sitting around for decades waiting for that opportunity. There will certainly be exceptions, and maybe they'll make a Big Short 2 about the scrappy underdogs who saw the opening and timed it perfectly. But the vast majority of Wall Street desperately wants the party to continue.
They are not in any corner. They rightly believe that they won't be allowed to fail. There's zero cost to inflating the bubble. If they tank a loss, it's not their money and they'll go on to somewhere else. If they get lucky (maybe skillful?) they get out of the bubble before anyone else, but get to ride it all the way to the top.
The only way they lose is if they sit by and do nothing. The upside is huge, and the downside is non-existent.
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...
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
Really?! I'm not used to chips having such a short lifespan.
Additionally - GPUs have multiple components. Which parts are at 60-70% load, the SM unit or the memory controller? If you're throttling the GPU but not the memory, it makes perfect sense why you're burning the damn card out...
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?
Nvidia has also begun trying to enter the custom silicon sector as well, but it's still largely dominated by Broadcom, Marvell, and Renesas.
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.
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.
We've technically been able to play board games by entering our moves into our telephones, sending them to a CPU to be combined, then printing out a new board on paper to conform to the new board state. We do not do this because it would be stupid. We can not depend on people starting to do this saving the paper, printer, and ink industries. Some things are not done because they are worthless.
If you’re a board game player then you are more than capable of imagining possibilities well beyond this.