> Capital equipment suppliers have been increasing their forecasts for revenue growth in 2000 as chip manufacturers begin to add new wafer fabs and backend assembly plants. Last week, for example, Applied Materials Inc.–the world's largest semiconductor equipment supplier–announced it had increased its forecast for industry growth in wafer fab equipment sales to 46% from a projection of 19% at the end of last year (see Feb. 16 story).
> The strong wave of new orders “confirms the strength of the current market cycle,” said Stanley T. Myers, president of SEMI.
So you are saying, just like the Internet, there may or may not be a painful AI burst in a few years, then exponential grow in decades to come? What is the time window to fit your narrative?
The parent described "exponential returns", which I was reacting to.
For what it's worth, the S&P dividend yield is typically 1-2%, and below inflation for the majority of relevant years. So you are not ahead in real dollars until roughly the dates I mentioned.
I was there, and this is why I consider the 2000s a lost decade.
Honestly we only just started to get back into the swing of things from 2012-2016 when Obama's reelection provided enough market confidence for electric cars to reach MVP status in 2013 and solar panels to begin scaling until Trump slapped a 30% tariff on them in 2018. Progress in sustainability is my litmus test for the economic viability of getting real work done in the economy, sort of like grading a steakhouse on how good its least-ordered mushroom appetizer tastes. We could also choose healthcare, education or any other field whose success is determined by the social good it provides.
If it weren't for the arrival of AI, I'd call this a wasted quarter century.
Yeah I always find dot com burst comparison apt because like, yeah market corrections will probably happen, but even if you invested in big tech RIGHT before the burst you would probably still be very rich today if you just held.
This is the definition of survivorship bias. There are a ton of tech stocks that you would have lost your shirt on had you invested in them just prior to the crash.
Even if you invested in a company that ultimately became a winner, you'd need to be able to hold on to it throughout the correction. You put your savings in tech stocks, the tech sector undergoes a significant correction, the stocks dive, you get laid off because you also work in tech and now you have to liquidate your stocks to meet your obligations. (Or you just retire at the wrong time.)
Right, people often look to companies like Intel, AMD, Nvidia and think 'If only I held!'
They don't look at 3DFX, Transmeta, Centillium, Virata, or the numerous other companies because they are dead in the graveyard of failed companies that have lost full valuation.
Depending on how much you invest in each, if you panic-buy/sell, etc changes all that. You're also ultimately betting on whether or not there will ever be another bubble/hype in the industry, and that you are betting wide enough into fully speculative companies like Nvidia instead of just aiming for power players. People forget that the market for luxury PC hardware barely existed at the time and was populated by a very small group of wealthy consumers. Expecting that to take off against the cheap prices of console offerings and against the technical know-how of assembling/disassembling PCs is closer to faith than any rationale.
Just by law of probability and total number of actual bagholders, the vast majority of people didn't bet that way. But I guess if you think you're special or have some lucrative insight into the industry, you might think you're an exception to that rule.
I'm sure there are studies out there on the ratio of upper-middle-to-upper class households vs not, adjusted for inflation, with surveys on how they invest (if they do at all) with other factors accounted for. I know for sure that the total number of households who hold index funds (50+%) [0] far exceeds households that have a million+ in assets (10+%) [1]. If index funds were some magic bullet investment vehicle for multiplying your money, we'd see better ratios than 1:5 with only the top 20% becoming wealthy-adjacent (and that's provided that 100% of the millionaire class got rich via index funds and not other means like real estate).
edit: as a corollary, I was just talking about this with someone who works in finance and the sentiment amongst ICs like them, they say, is that this is definitely a bubble that will pop at some point. The only contention is how devastating the effects will be, if it'll be a slump, a recession, or a depression.
It’s showing how much better or worse than inflation.
That’s not an apples-apples comparison.
Like if you don’t invest into say SNP500 then the money grows (or not) at what rate and is that rate better than the SNP500?
It’s way better to get a 10% return when there’s 11% inflation if the alternative was 0%.
It’s weird to the point I consider it deliberately misleading:
- Anybody wise enough to invest in the S&P 500 is likely also wise enough not to invest their entire nest egg in a single year, 20 years before retirement.
- The massive bulk of “ordinary American” 401(k) / IRA funds are in low-fee, no-tax-on-growth accounts. This chart has a nebulous “taxes and fees” component, which in this already artificial context makes even less sense to apply to annual returns.
- Counting only real growth without comparing to other investment options (real estate, bonds, cash under the mattress) fails to capture the relative value of diversified index funds. In short: Impossible to beat reliably no matter how you set the metric.
To spell out my thinking more completely: Yes, I did not mean AskJeeves specifically, I meant broad investment. And likewise nobody should go all in on just SpaceXSI or whatever.
And in any case no, you should not pour your money into tech stocks if you need that money to eat, that was true then as it is today.
> This is the definition of survivorship bias. There are a ton of tech stocks that you would have lost your shirt on had you invested in them just prior to the crash.
Which is why you buy an index, rather than try pick individual stocks.
Buy individual stocks in blue chip companies. The return is always better but if you do buy individual stocks, you have to actually pay attention to that individual company that you invested in does, in short continue to do your homework, in the current bubble, invest in the companies that have great fundamentals (revenue and actual profit) do not invest in those companies where the financials are terrible and have been terrible for years, yet they still command a high price, eventually the party ends.
Currently, there are three companies owned by the same person, the financials and their relative performance/competitiveness is not good despite the oversized hype in comparison to other blue chip companies. But many people have gotten great returns however, it is a don’t get caught holding the bag type of stocks Hint: The three companies are owned by same guy and has a cult like following.
The problem is psychology. People don't HODL while their retirements drop by 20, 30, 40% in a week. High earners might, institutions can, but regular every day people they take the loss, and might never go back into the market again, forgoing those recovery years.
The average recovery is 18 months I think (I'm in a meeting, and can't look it up), but people like my dad watched his retirement fizzle out in the dotcom bubble, because he was all in on tech. Then he moved to safer bond and realestate funds, and his value never recovered, and then 07/08 happened and he watched what he thought were "safe" investments disappear almost immediately. He pulled everything out of the market, and never invested again. He went through 2 crashes on his portfolio, and to this day doesn't trust stocks and bonds.
He was sold an annuity that according to him is supposed to double it's payout every 5 years it is left un touched, so he didn't touch it for 15 years, now that he is drawing it down, it draws from principal before ever touching the growth, and when he dies, only the principal will get passed to my mother. It was basically a scam.
Peak to peak, inflation adjusted, can be 10 to 15 years for the really big busts though. It's not entirely irrational if you moved in at the peak.
If you were in on the upside then you've got much less ground to retrace.
And it's painful to watch your holdings drop like you said if you're planning on retiring in the next 1, 2, 5 years.
So I get it.
If you're 25 you've got tons of room to maneuver. If you're 60 things are a lot less rosy (even if by now you should theoretically be cycling into less volatile assets. But bonds didn't do that great either over that period) .
In most western countries, pets are on #2 after spendings for children in households, so your bet made sense. Though, as a lot of companies, tractions wasnt big enough back then.
Today, e.g. ZooPlus is one of the biggest pet-stuff-retailers (2023 revenue 2.5billion)
I'm pretty sure that housing, cars and food are #1, #2 and #3. Your definition of household spending seems a little unconventional. If pets are household spending, surely so is food for humans?
If you invested in XLK (State Street Technology Select Sector SPDR ETF) In March of 2000 (just before the .com bust), you would have underperfomred SPY (S&P 500) until June 2026. As of today, the spread between the two is just over 1%, with good reason to believe that we are in another tech bubble.
Yes, if you only invest in winners you'd be very rich! Which companies will be the winners though? Nobody knows.
What about the hundreds of losers that looked good right before the burst? My coworker retells a story where he dumped his life savings in dotcom companies that went to zero.
I'd start from a slightly more-charitable assumption which ends up in a slightly different "Hold up that's a little optimistic" territory.
1. Yes, if someone generally invested in "dot-com stocks" before that crash--covering both winners and losers--they might today have a surprising amount of (potential) dollars.
2. However much of that impressive-looking sum would be just because of 25-30 years in the market.
3. So we need to scale that "very rich" against someone putting the same money into an total-stock market index fund back then, which may not make it look nearly as impressive.
I could have been more charitable. I think their key word was "big" tech. The blue chips. What were the tech blue chips before the dotcom crash? Microsoft, Cisco, Intel, IBM, Oracle, Etc. Some might call Enron blue chip. I think their point mostly holds though. If you'd have invested in the blue chips and held you'd likely do well.
This is pretty much why the best financial advice for 99.99% of people is "stick with index funds".
Yeah you won't get extremely lucky and 200x your money because you bought NVidia a decade ago, but you also won't lose all your money because you bought Enron.
It’s not clear yet. Index funds are now an enormous part of the market and players are starting to play them (see SpaceX, and given the door they opened OpenAI and Anthropic to follow).
Inflation calculator puts that 1k at a value of 1.95k today. Seems like a lousy 26 year investment. I guess it's better than putting it under your matress.
I think it's more if you had invested in an index fund you'd be in a pretty good place. Individual companies would have a much larger range of outcomes (as ever).
One problem for internet vs AI is the internet boom laid down a bunch of fiber, which has a usable life of decades. The AI boom is 10% datacenter, 90% GPU, and the GPUs are consumables on a timeline of ~3-5 years, and even that's being optimistic compared to rumored real-world degradation rates. 2000 put a couple trillion in the ground that we got to dig up later, 2026 is setting a couple trillion on fire.
> the GPUs are consumables on a timeline of ~3-5 years
When newer GPUs with clearly better perf/$ are released at a yearly cadence, then older GPUs might indeed become _obsolete_ at this pace, meaning that it makes more economic sense to swap them out because datacenter occupancy is extremely limited. Many of these GPUs might themselves still be functional.
The infra and its lifespan isn't what created the value that came out of the Internet. The question I think is how much value will people create with these investments
The usable life of a model is infinite. You can always fine tune it for current events and technology, or distill into a new one with a more efficient architecture.
If Anthropic or OpenAI go bankrupt they will have to auction off all their IP including code and weights. Claude won't disappear just change its name if someone else buys the trademark.
Can’t speak for the original commenter, but back in the spring (maybe early summer?), I expected the bubble was going to pop whenever we get a full S-1 filing for one of the frontier labs, and that’s still my expectation. I don’t think the selective leaks have prepared the market for what “positive margin on some products yet incredibly unprofitable” is likely to look like in detail.
I like the dot com bubble comparisons mainly so that I can look at the differences. That is where the information is, and it's what you want if you're going to navigate this market.
Wow, is this Hacker News? Where have you been? Last night I had two agents up autonomously developing and training deep learning models much faster and better than I ever could as a professional. Seems the reddit bubble of "AI doesn't work but also it's somehow evil" has taken over this place as well.
(Scare-quote absolutely necessary: even with maximal anthropomorphisation, we have no way to differentiate between "rewarded for work well done" vs. "punished for not working or half-arseing").
I think there is a bit of a gotcha here. Is intelligence the same thing as a will or drive? AI seems quite intelligent, somewhat lacking in wisdom, and having almost no ego or will. It completely lacks self preservation and desires (as far as we can tell), but are those innate qualities of intelligence?
It also exists is a very different state that biological life. Biological life has a brain that either has activity or doesn't. Once the activity is gone, the brain is forever off and nothing can start it back. So we associate the activity as life and as the intelligence running. AI has weights with vast interconnections, but they can easily be turned on and off, rewound, paused, and so on in a way that we couldn't dream of doing with biological intelligence.
So, maybe these differences are enough it isn't intelligence. But it is something, something that is getting better at quite a higher speed. Some that approximates intelligence in many ways. That passes the Turing test. So that something needs a name, and intelligence seems a hard enough to define thing that it is easy to stretch it and include this new factor as falling under the term.
no, it's very hot, and the fans are loud and obnoxious so they keep putting it in data centers near places that don't really want it, and they install lots of natural gas turbines because the grid isn't enough and that's louder too, and then the cooling often needs a lot of water, and -- really unpopular overall, as it turns out!! huge backlash. and that's before, you know, allll the job impacts, which is when it gets ugly and threatens to complicate the whole consumption-driven structure of the economy.
... but the little dgx spark is in the room with us! it's pretty smart-ish, in a pinch!
Even AGI might just be a 'meh' moment. It will take a while before you get through social, political and physical barriers before we start seeing it make a meaningful impact.
I imagine the first thing that would happen is that some people will declare AGI is here and some people will be in denial and call it a "stochastic parrot" because they either feel threatened by it, cynical or just ignorant. Then we just continue business as usual for a while because most people wouldn't know what to do with it and definitely don't want to be replaced even if they aren't against the idea of using AI. A lot of research is also useless without collecting data, experimentation in the real world and hardware, all of which take time.
Agreed. The "singularitarians" always seem to overlook one simple truth: we don't use the intelligence we already have. The appearance of a 100-fold increase in cognitive resources will have no immediate effect at all.
"immediate" is doing a lot of heavy lifting there.
"The intelligence we already have" is why we're splitting atoms and the other primates are splitting rocks. That took us a few hundred thousand years, of course.
But if you have our mere equals in digital form, that only takes as long to change everything as it takes to build a factory to produce the same number of equally dexterous robots.
The value was driven by the size of the user base. We gave each user an arbitrary price but didn’t understand how to turn that into actual money. So if you had a website with 30M users it was valued at $900M. The late 90s early 2000s were wild times in VC, and some lucky people realizing how nonsensical all that was cashed out on time.
Yes, I have no doubt that AI is going to seriously impact the economy over time. I have lots of doubt that it can impact the economy sufficiently quickly and with enough impact to pay out all the very, very, very large bets being made on it. Even as the AI quality curve has been running ahead of most predictions, its actual impact on the economy has been running well behind.
I am watching the dawning realization among the management class that even if coding costs drop to zero, Amdahl's Law means that perhaps productivity doubles for most software engineers, rather than goes to the MOON. AI can help some of the other bottlenecks, but not 100%.
On the positive side, we programmers have been trying to collectively explain to management for decades now that our jobs are not just 100% banging away at code on keyboards, we're not just interchangable code monkeys. Well, they're getting their noses rubbed in it now, complete with metrics and charts and all the other things they value more than the feedback from their employees.
In fact I was just noticing this week... in the past few months I had been really banging away at a lot of code projects. I was doing that whole "3 or 4 streams going on at once" that many of us have talked about, with a variety of mixed feelings. But in the past few weeks that seems to have starved out. I seem to have burned through the things that I was blocked on code for, and had a reasonable business case to pursue. Now my projects are all blocking up on non-code things again... communication with other teams. Review of documents that are irreducibly things humans need to review. Coordination and planning. Applying human attention to cover over the way that even the best frontier models still seem to have huge gaps in their superficially-expansive automated testing code that they write for their own code. I'm still sometimes unblocking multiple things at once and can briefly have multiple projects going, but then I hit a blocker somewhere and I end up unable to continue having multiple simultaneous code projects going at once.
(The document review is a particularly interesting case. The documents in question are something we expect some humans to at least skim, but are going to be basically a meta-program for AIs programming the rest of the company's code. It is important to get it right because every error is going to magnify quite a bit. AIs are helping a lot with that document, yes, I certainly agree we are proceeding more quickly with it than the same project 5 years ago would have gone, but it is also irreducibly sopping up a lot of human attention.)
Point of all this being that the casual presumption that AI will just utterly destroy the entire economy because it is good at coding may prove to be, well, yeah, exactly the same as the assumption that we are going to all be rich by 2002 because we'll just shift the entire economy to be on the internet and everything will multiply by 1000 in just a couple years as a result. But everything can't be shifted on to the internet. I can't eat a web page. AI can't do everything, and as it advances the things it can't do become a larger and larger percent of the remaining pie as a result. It seems to rhyme a lot with the dot com crash.
Ok, but go through the bottle necks you face. How many of then will 1. go away with better models, and 2. go away when those bottle necks (other people) are automated away? For me, I can go quite fast because I work alone, no other people in a team to slow me down, but verifying the result is still a bottle neck. Also interacting with government agencies, my accountant etc, still a bottle neck. None of these are technically insurmountable, and I suspect that when one of these becomes a real bottle neck for the entire economy, they will be fixed.
There's a huge difference between "zero" and "a little" when it comes to Amdahl's Law and scaling up, let alone the things where it won't be "a little". Which is one of those things that sounds obvious when I say it. But clearly a lot of people are predicting a future that is based on costs and times being "zero" that won't be.
Right now, the most visible bottleneck is real world robotic control.
I kinda hope that one stays around for a long time, because robots controlled by AI competent enough to guide those robots through every step in their production from mining raw materials to end-of-assemble acceptance testing, is one where tiny mistakes kill everyone.
Like, say you've got a sticky 0 key and accidentally ask for more burgers than there is carbon in Earth's biosphere and atmosphere combined: if the sufficiently-competent AI is a little too eager to obey and not quite willing enough to say "no" or listen when told to "stop", that's enough to kill everyone.
Trust me I know, I co-founded a robotics startup, and robotics is HARD. But now it's becoming a hardware problem, the software will sort itself out soon, it's not a bottle neck that will keep us safe for long.
Abou sycophancy, I sense that it's much lower now with Opus 5.5. Yesterday I had to argue my viewpoint for quite a bit before it relented, on an issue I'm pretty sure I'm broadly correct. It's becoming a better sparring partner that actually pushes back and raises counter-arguments.
No it isn't. It doesn't matter how real the technology is. The bubble will still pop, because when a technology is seen as revolutionary, you get rampant overinvestment. And only a fraction of those investments will pay off(or, indeed, none, in the case where frontier labs are eventually made irrelevant by open models, which seems very likely to me).
The internet and web were revolutionary, no one can deny that. They completely transformed our society. Nevertheless, the bubble popped. The current bubble is no different.
I don't doubt there will be a big crash at some point, but I actually doubt we've reached that point yet. So far stock increases have been driven by earning increases, not speculation. I know an "earnings bubble" can be a thing, but given how useful I know these models to be, and how few use them in the way they can be, I think we have far to go still.
"No, no, you don't understand. If you keep putting bigger and bigger engines in a car, it'll somehow turn into an aeroplane at some point! There's no other way, believe me!"
People forget the fact there may not be enough capital in the world to invest in AI, and so while some companies will go under the market could grow for literal decades. I don't know the future, but this is not a guaranteed bubble.
I’d think that orders for long lead time rivalrous slots of a very finite resource are basically OOTM call options for future demand. It says something that firms think demand is plausible but not a whole lot about how probable it really is. You can always idle these things, you can’t conjure one immediately.
It's certainly an indicator. But, just like airplane orders, the backlog is long enough that you have to put orders in on a speculative basis.
If everything goes as planed, we'll need new equipment in a couple years, so order now. If things don't work out as planned, you might be able to sell your slot, or push it back a few months, or cancel it.
In a frothy market, you don't really know from order backlogs how many orders are based on firm plans and how many are we might want to build chips, so let's order a machine and see if we want it when it comes up. Or the middle ground of we definitely need these machines in the next ten year, and if sales are still strong, sooner is better.
> Companies buying those machines need years to get their money back. If they didn't believe they would make a good ROI they wouldn't buy as much.
In the late-1990s companies were building out fibre optic networks, as they believed they would get a lot of ROI from all the Information Superhighway demand. The expected demand never came in the expected timeline.
Of course demand eventually used up all the dark fibre over the course of many moons, just not quickly enough to save many companies:
More specifically, the demand for fiber didn't come - but the demand for bandwidth did!
Those companies were building out fiber like it was a legacy telephone network: needing a massive amount of individual wire pairs to carry a massive amount of parallel transmissions. In reality the bandwidth per channel was rapidly increasing due to better tx/rx equipment, and with the rise of WDM it was suddenly possible to carry not one channel per fiber pair but several dozen.
When you were laying fiber on the assumption that a single long-distance pair was going to carry 1Gbps - and therefore laying hundreds of fiber pairs in a trench at once - it is quite a big upheaval when that single pair can now carry 100 channels of 10Gbps each. Your brand-new expensive fiber trunk can suddenly be replaced by a single spare pair some utility company had lying around already.
Turns out laying fiber isn't a great business when the capacity of existing fiber is doubling every year or so!
maybe collapse in valuation, collapse in demand? I don't see that happening any time soon, it's too useful, idk how anyone can contest it is useful and not just hype.
No, in fact inference is far cheaper than most think, see analysis for serving Kimi K3.
If you mean OpenAI and Anthropic's investments into R&D, ok, perhaps.
However, they will not be able to increase prices unless what they are offering next is a significant enough improvement over the current SOTA, which might only be six months ahead of the Chinese models.
Not only is inference currently much cheaper than people seem to think, it is going to get much cheaper. NVIDIA’s Vera Rubin is going to drive costs down for current models considerably.
It's amazing to see the shift from "Anthropic is profitable" to "it's just R&D costs" following the leak of Anthropic's finances. If there's one thing more abundant in this bubble than FOMO, it's special pleading.
These companies are burning helps of money, and they need to keep burning heaps of money in order to stay competitive. People have (correctly) learned to be extremely fickle consumers, they can't depend on making profit by running inference on a 1 year old model.
I really dislike that term. It's stupid and supports a distorted narrative.
They arbitrarily decide 1M API tokens cost X, and any subscription where usage costs less than that is "subsidized". No it isn't. They may be profiting on subscriptions as well, or losing on full API costs as well, or anything in between, and that X is anyway more reflective of competition than their costs. For all we know they did a rand()%1000 to determine it in the first place.
If it isn't subsidized, why aren't all those AI companies making huge amounts of profit? Instead they are rapidly burning through the billions, and the money must be going somewhere.
Investing in R&D is fine. as long as you are turning a profit somewhere within your company that can support it. If you are not, eventually the end comes. In short, the bubble.
The problem with this view is that that isn't a "bubble". "Eventually" can support quite a while, especially when the investment is increasing revenue over time and the unit economics look good. The timeline point cuts both ways, and none of that is a subsidy.
The only company making AI chips, and allowed to buy from ASML, is TSMC, otherwise it's probably memory manufacturers.
Even if the order book is mostly TSMC, it doesn't necessarily mean the demand is mostly driven by AI (although maybe it is) - there are plenty of other high volume TSMC customers like Apple, AMD, Intel, Qualcomm, Broadcom ...
Finally, just because lots of TSMC customers may have made forward commitments (or whatever it takes for TSMC to commit to expansion), the main conclusion would seem to be that they needed to do that for business continuity given current capacity constraints. It's not clear why the need for them to do this should be taken as a sign that the AI datacenter bubble won't pop. Do you want to be Apple and gamble that rather then locking in TSMC capacity now you can gamble on the bubble popping and maybe get better terms later?
As a counterpoint to this "no need to worry" story, CoreWeave's 5 yr bonds are now yielding 15% ...
> Companies buying those machines need years to get their money back. If they didn't believe they would make a good ROI they wouldn't buy as much.
But if they're in FOMO mode, they might not be thinking about ROI at all. Or if they're in bubble-optimist mode, they would just assume their ROI will be good.
ASML sell photolithography machines, which presumably can be used for many kinds of chips: CPUs, GPUs, flash, RAM, etc...
They don't need to be AI chips, so even if the AI bubble bursts before they managed to get a ROI, they can switch to other applications that need high end electronics, such as PCs and smartphones, maybe cars.
I don't think "running out of steam" is the expected failure mode here. We are talking about bulk ordering machines that produce chips that in turn will hit the market in half a decade. That's like adding power plants to the grid that runs the ball pump inflating the ballon. ASML will probably be the last that faces the ripples of the AI bubble going bust.
Even after the AI bubble pops, ASML's machines will be useful. People generally want computers and chips, preferably the fastest an latest-generation ones. The ROI may be lower once the bubble pops, but the machines will earn themselves back eventually.
Once the bubble does pop, ASML will be in trouble, though. It'll take a couple of years for their latest machines to generate as much interest again, so the effects of the market collapse will echo for years after it happens.
The internet bubble popping did not cause internet traffic to decline -- on the contrary internet traffic increased rapidly during the bubble pop.
Similarly, I don't think the AI bubble popping will cause token demand to decline. On the contrary, I predict the bubble pop will happen because of a recognition that tokens are a highly competitive industry, thus the margins of frontier AI firms aren't sustainable, popping their valuations.
I don’t buy this argument. Everyone who suggests there’s a bubble speaks about the excessive data center buildout, low demand, excessive debt financing.
All of these arguments don’t matter if you think demand will increase.
What bubble is left? Really. Just people with OpenAI or Anthropic exposure? Not that high.
low demand? Got a source for that? Anthropic is paying crazy sums to direct competitors to meet demand, so it seems that demand is quite high. Could be wrong, though.
> What bubble is left? Really. Just people with OpenAI or Anthropic exposure? Not that high.
The stock prices of 9/10 of the top S&P 500 stocks have significant AI exposure. A drop in valuation of them will be widely felt.
And there's a significant bubble in the hardware layer. The Nvidia overvaluation alone far exceeds that of both OpenAI and Anthropic. When the bubble pops people are going to want dedicated inference chips without an 80% profit margin rather than the B200.
everybody and their dog are designing inference chips. Even currently, Google's Tensor can do inference cheaper and more efficiently than a B200. There are dozens of companies designing inference chips; when they hit the market the pricing power of B200 will drop significantly.
And it's correlated with the other bubble dropping because chip cost will matter a lot more once the margin of tokens is tight.
ASML's order book is way way way at the tail end of where the signal of the bubble's bursting will be felt
Users sign up for Claude, Claude buys more compute, AWS/Azure runs out of capacity, Microsoft buys 10,000 acres in Virginia, Microsoft places order for a million GPUs from NVidia, Nvidia places order for a hundred million GPUs from TSMC, TSMC buys 1,000 acres in Arizona, TSMC places an order for a bunch of new ASML machines
Yes, the end of that chain is the one I'm goin to watch to know when the bubble has burst, not the beginning of it
Despite the fact that chip making machines come before chips, this seems like a lagging indicator, not a leading indicator.
These would be among the first orders placed in a build-out plan, and the build-out plans are first in a boom. The bust phase would very well happen before the orders are completed, and orders would not decline until the bust phase starts?
the site threw up a modal, blocked me from scrolling, then locked the browser, so I couldn't read everything, but from what I could see the gist is people are still over-ordering therefore the boom is going strong.
Yes. ASML is filling order book for 2029 and 2030. I am just not sure, with how long-term investment the chip fabs are, would the slowdown in datacenter demand be visible on ASML's side first or last?
There's a cost to breaking contracts like these. You can't change your mind two years down the line without paying a fine. ASML isn't stupid.
That said, computers are here to stay. Even if FABs can't sell chips at ridiculous prices, there will be a return on investment on these machines, even if it's not as high as originally hoped.
If you buy the machines and not use them, at most you lost several hundred of millions of dollars (not really, because that machine still retains value for you or someone else later).
On the opposite end, if you don't buy the machines in time, mispredict the demand and can't produce, you lose tens of billions.
So weird that they're inverting knowledge and/or causation.
AI is a bubble because the demand story is implausible. The fact that someone is willing to make ASML orders (which is a >= 10 year indicator) doesn't tell you anything at all about whether AI companies can, in fact, get MANY ORDERS OF MAGNITUDE of revenue growth.
I think tying this to resilient AI spending is misleading. The current AI spending boom has jolted a lot of fabs and companies into realizing that they’re not adequately prepared for the next few decades of semiconductor markets regardless of how AI turns out. That the previous market of divided consumer and business segments is ill-suited to a future where compute is more universal between the two, and that if they want to avoid one going bust when the other booms (like the current memory apocalypse) then they need to be ready to meet that sort of demand. Then you mix in nationalist tensions and China quickly rising to fill the consumer gap with older (but still excellent) specced hardware like DDR4, and the writing is on the wall for the traditionally Western-dominated compute industry: adapt to changing demands or die, just like manufacturing did.
What I’m watching is the timeline. If China can get DDR4 kit out in volume (or even better, DDR5) faster than the AI bubble deflates or markets correct, then the future consumer compute players will be Chinese, not American, especially if they can ship kit at pre-COVID pricepoints.
This current bubble will be the last big payday for the three Hyenas of memory, rest of the world (most of it) will be buying Chinese memory after 2030, and a few other companies will probably be designing an engineering memory in house and getting it fab at TSMC. And AI models they will be better and smaller, and the hardware? It will be up to the task of running the models on your desktop.
Feb 2000: Record orders for chip gear push SEMI index to four-year high
https://www.eetimes.com/record-orders-for-chip-gear-push-sem...
> Capital equipment suppliers have been increasing their forecasts for revenue growth in 2000 as chip manufacturers begin to add new wafer fabs and backend assembly plants. Last week, for example, Applied Materials Inc.–the world's largest semiconductor equipment supplier–announced it had increased its forecast for industry growth in wafer fab equipment sales to 46% from a projection of 19% at the end of last year (see Feb. 16 story).
> The strong wave of new orders “confirms the strength of the current market cycle,” said Stanley T. Myers, president of SEMI.
This time it's different.
So you are saying, just like the Internet, there may or may not be a painful AI burst in a few years, then exponential grow in decades to come? What is the time window to fit your narrative?
An investment in the S&P 500 at the heights of 2000 would not materially exceed its basis until 2013!
An Nasdaq would not materially exceed until 2016.
Yes, this ignores dividends, but that really only helps manage inflation.
You would have done much better being out of the stock market until the 2009 crash.
Excluding dividends is a pretty big asterisk, including them you're back to even within 6.5 years.
The parent described "exponential returns", which I was reacting to.
For what it's worth, the S&P dividend yield is typically 1-2%, and below inflation for the majority of relevant years. So you are not ahead in real dollars until roughly the dates I mentioned.
Are you looking at dividends as cash value or reinvested?
Dividends are part of the total returns and cant just be ignored. Taking them into account, the break-even point is non-insignificantly earlier.
It’s the total returns that matter, not just stock price.
But remembering that most people have taxes to pay on dividends even if the stock price is going down.
Well, you aren't necessarily better sitting out entirely.
When the stock market slumps like that, continual investment lowers your average cost basis.
e.g.:
Year 1: Buy 1 stock at $100
Year 2: Buy 1 stock at $50 (post-crash)
Year 3: Buy 1 stock at $60
Year 4: Buy 1 stock at $70
Year 5: Buy 1 stock at $80
If we just look at "Year 1 = $100, Year 5 = $80" then it looks like we lost $20/share.
However, our average cost basis is $72, so we actually made $8 per share.
That’s not really a great return but it’s not horrible considering the circumstances.
That's why DCA and do a good job at work is the best investment for most people.
I was there, and this is why I consider the 2000s a lost decade.
Honestly we only just started to get back into the swing of things from 2012-2016 when Obama's reelection provided enough market confidence for electric cars to reach MVP status in 2013 and solar panels to begin scaling until Trump slapped a 30% tariff on them in 2018. Progress in sustainability is my litmus test for the economic viability of getting real work done in the economy, sort of like grading a steakhouse on how good its least-ordered mushroom appetizer tastes. We could also choose healthcare, education or any other field whose success is determined by the social good it provides.
If it weren't for the arrival of AI, I'd call this a wasted quarter century.
So if you knew in advance about the timing of both the tech bubble and the real estate bubble, you could have made a lot of money trading on that?
Yeah I always find dot com burst comparison apt because like, yeah market corrections will probably happen, but even if you invested in big tech RIGHT before the burst you would probably still be very rich today if you just held.
This is the definition of survivorship bias. There are a ton of tech stocks that you would have lost your shirt on had you invested in them just prior to the crash.
Even if you invested in a company that ultimately became a winner, you'd need to be able to hold on to it throughout the correction. You put your savings in tech stocks, the tech sector undergoes a significant correction, the stocks dive, you get laid off because you also work in tech and now you have to liquidate your stocks to meet your obligations. (Or you just retire at the wrong time.)
Right, people often look to companies like Intel, AMD, Nvidia and think 'If only I held!'
They don't look at 3DFX, Transmeta, Centillium, Virata, or the numerous other companies because they are dead in the graveyard of failed companies that have lost full valuation.
Depending on how much you invest in each, if you panic-buy/sell, etc changes all that. You're also ultimately betting on whether or not there will ever be another bubble/hype in the industry, and that you are betting wide enough into fully speculative companies like Nvidia instead of just aiming for power players. People forget that the market for luxury PC hardware barely existed at the time and was populated by a very small group of wealthy consumers. Expecting that to take off against the cheap prices of console offerings and against the technical know-how of assembling/disassembling PCs is closer to faith than any rationale.
Just by law of probability and total number of actual bagholders, the vast majority of people didn't bet that way. But I guess if you think you're special or have some lucrative insight into the industry, you might think you're an exception to that rule.
Right, the other case seems to definitely be happening
> 'Yikes': CoreWeave stock sinks after its bond yield touched 15%
https://finance.yahoo.com/technology/article/yikes-coreweave...
https://news.ycombinator.com/item?id=50090791
Sure but if you bought a broad index you were fine
Again, that depends on how long you hold and when you cash out. Most people don't get rich: https://archive.nytimes.com/screenshots/www.nytimes.com/inte...
I'm sure there are studies out there on the ratio of upper-middle-to-upper class households vs not, adjusted for inflation, with surveys on how they invest (if they do at all) with other factors accounted for. I know for sure that the total number of households who hold index funds (50+%) [0] far exceeds households that have a million+ in assets (10+%) [1]. If index funds were some magic bullet investment vehicle for multiplying your money, we'd see better ratios than 1:5 with only the top 20% becoming wealthy-adjacent (and that's provided that 100% of the millionaire class got rich via index funds and not other means like real estate).
[0] https://www.ici.org/news-release/ici-report-shows-mutual-fun...
[1] https://finance.yahoo.com/news/guess-percent-people-1-millio...
edit: as a corollary, I was just talking about this with someone who works in finance and the sentiment amongst ICs like them, they say, is that this is definitely a bubble that will pop at some point. The only contention is how devastating the effects will be, if it'll be a slump, a recession, or a depression.
WOW, MANY thanks for linking the NYT graphic!
Its very good and tells a lot!
Thanks for sharing!
I’m not so sure.
It’s showing how much better or worse than inflation.
That’s not an apples-apples comparison.
Like if you don’t invest into say SNP500 then the money grows (or not) at what rate and is that rate better than the SNP500? It’s way better to get a 10% return when there’s 11% inflation if the alternative was 0%.
It’s weird to the point I consider it deliberately misleading:
- Anybody wise enough to invest in the S&P 500 is likely also wise enough not to invest their entire nest egg in a single year, 20 years before retirement.
- The massive bulk of “ordinary American” 401(k) / IRA funds are in low-fee, no-tax-on-growth accounts. This chart has a nebulous “taxes and fees” component, which in this already artificial context makes even less sense to apply to annual returns.
- Counting only real growth without comparing to other investment options (real estate, bonds, cash under the mattress) fails to capture the relative value of diversified index funds. In short: Impossible to beat reliably no matter how you set the metric.
To spell out my thinking more completely: Yes, I did not mean AskJeeves specifically, I meant broad investment. And likewise nobody should go all in on just SpaceXSI or whatever.
And in any case no, you should not pour your money into tech stocks if you need that money to eat, that was true then as it is today.
> This is the definition of survivorship bias. There are a ton of tech stocks that you would have lost your shirt on had you invested in them just prior to the crash.
Which is why you buy an index, rather than try pick individual stocks.
Buy individual stocks in blue chip companies. The return is always better but if you do buy individual stocks, you have to actually pay attention to that individual company that you invested in does, in short continue to do your homework, in the current bubble, invest in the companies that have great fundamentals (revenue and actual profit) do not invest in those companies where the financials are terrible and have been terrible for years, yet they still command a high price, eventually the party ends.
Currently, there are three companies owned by the same person, the financials and their relative performance/competitiveness is not good despite the oversized hype in comparison to other blue chip companies. But many people have gotten great returns however, it is a don’t get caught holding the bag type of stocks Hint: The three companies are owned by same guy and has a cult like following.
The problem is psychology. People don't HODL while their retirements drop by 20, 30, 40% in a week. High earners might, institutions can, but regular every day people they take the loss, and might never go back into the market again, forgoing those recovery years.
The average recovery is 18 months I think (I'm in a meeting, and can't look it up), but people like my dad watched his retirement fizzle out in the dotcom bubble, because he was all in on tech. Then he moved to safer bond and realestate funds, and his value never recovered, and then 07/08 happened and he watched what he thought were "safe" investments disappear almost immediately. He pulled everything out of the market, and never invested again. He went through 2 crashes on his portfolio, and to this day doesn't trust stocks and bonds.
He was sold an annuity that according to him is supposed to double it's payout every 5 years it is left un touched, so he didn't touch it for 15 years, now that he is drawing it down, it draws from principal before ever touching the growth, and when he dies, only the principal will get passed to my mother. It was basically a scam.
Peak to peak, inflation adjusted, can be 10 to 15 years for the really big busts though. It's not entirely irrational if you moved in at the peak.
If you were in on the upside then you've got much less ground to retrace.
And it's painful to watch your holdings drop like you said if you're planning on retiring in the next 1, 2, 5 years.
So I get it.
If you're 25 you've got tons of room to maneuver. If you're 60 things are a lot less rosy (even if by now you should theoretically be cycling into less volatile assets. But bonds didn't do that great either over that period) .
For Cisco it took 25 years.
But that is assuming they survive the bust, which is not a given.
Same for MSFT and some others, IIRC
Yeah I made millions by hanging on to my pets.com stock and just riding the wave.
In most western countries, pets are on #2 after spendings for children in households, so your bet made sense. Though, as a lot of companies, tractions wasnt big enough back then.
Today, e.g. ZooPlus is one of the biggest pet-stuff-retailers (2023 revenue 2.5billion)
That doesn't help someone that invested in pets.com: https://en.wikipedia.org/wiki/Pets.com
I'm pretty sure that housing, cars and food are #1, #2 and #3. Your definition of household spending seems a little unconventional. If pets are household spending, surely so is food for humans?
It was a little inaccurate, you are right:
#2 "after all payments done for yourself/husband"
#1 will be the kids
You can always eat your pets
It was a joke I suppose.
No. If you bought Cisco it never recovered, once you factor in inflation. Almost the same with Intel.
If you invested in XLK (State Street Technology Select Sector SPDR ETF) In March of 2000 (just before the .com bust), you would have underperfomred SPY (S&P 500) until June 2026. As of today, the spread between the two is just over 1%, with good reason to believe that we are in another tech bubble.
https://www.etftimemachine.com/en/compare/XLK-vs-SPY?start=2...
Had you invested in XLK when it opened in 1998, then following the .com crash, it would take until early 2020 to return to parity to the S&P 500.
Held what? A majority of those companies evaporated. You'd have to cut your losses and reinvest.
Yes, if you only invest in winners you'd be very rich! Which companies will be the winners though? Nobody knows.
What about the hundreds of losers that looked good right before the burst? My coworker retells a story where he dumped his life savings in dotcom companies that went to zero.
I'd start from a slightly more-charitable assumption which ends up in a slightly different "Hold up that's a little optimistic" territory.
1. Yes, if someone generally invested in "dot-com stocks" before that crash--covering both winners and losers--they might today have a surprising amount of (potential) dollars.
2. However much of that impressive-looking sum would be just because of 25-30 years in the market.
3. So we need to scale that "very rich" against someone putting the same money into an total-stock market index fund back then, which may not make it look nearly as impressive.
I could have been more charitable. I think their key word was "big" tech. The blue chips. What were the tech blue chips before the dotcom crash? Microsoft, Cisco, Intel, IBM, Oracle, Etc. Some might call Enron blue chip. I think their point mostly holds though. If you'd have invested in the blue chips and held you'd likely do well.
This is pretty much why the best financial advice for 99.99% of people is "stick with index funds".
Yeah you won't get extremely lucky and 200x your money because you bought NVidia a decade ago, but you also won't lose all your money because you bought Enron.
Does this advice still hold once 99.99% of people are invested in index funds?
It’s not clear yet. Index funds are now an enormous part of the market and players are starting to play them (see SpaceX, and given the door they opened OpenAI and Anthropic to follow).
And big tech is a large part of most indexes
I mean probably? It's still a much lower-risk strategy compared to investing in individual stocks.
what? you mean that all my beenz.com stock is worth nil?? but it's an online currency! it's the future!
Hey, at least Flooz was used by a crime syndicate. That's market validation.
Survivorship bias.
It took the companies that survived a decade or more to recover.
You still have to pay bills while you endure inflation and the opportunity costs in the meantime.
I was under the impression that if you bought Cisco stock for $1000 during the height of the bubble, you'd have like $1500 now?
Inflation calculator puts that 1k at a value of 1.95k today. Seems like a lousy 26 year investment. I guess it's better than putting it under your matress.
I think it's more if you had invested in an index fund you'd be in a pretty good place. Individual companies would have a much larger range of outcomes (as ever).
Maybe, but 50% growth in a stock over 25 years is fairly bad.
That's pretty good.
I know people who bought Ask Jeeves at the height of the bubble.
One problem for internet vs AI is the internet boom laid down a bunch of fiber, which has a usable life of decades. The AI boom is 10% datacenter, 90% GPU, and the GPUs are consumables on a timeline of ~3-5 years, and even that's being optimistic compared to rumored real-world degradation rates. 2000 put a couple trillion in the ground that we got to dig up later, 2026 is setting a couple trillion on fire.
> the GPUs are consumables on a timeline of ~3-5 years
When newer GPUs with clearly better perf/$ are released at a yearly cadence, then older GPUs might indeed become _obsolete_ at this pace, meaning that it makes more economic sense to swap them out because datacenter occupancy is extremely limited. Many of these GPUs might themselves still be functional.
They burn out in the life of a dc.
The infra and its lifespan isn't what created the value that came out of the Internet. The question I think is how much value will people create with these investments
The usable life of a model is infinite. You can always fine tune it for current events and technology, or distill into a new one with a more efficient architecture.
If Anthropic or OpenAI go bankrupt they will have to auction off all their IP including code and weights. Claude won't disappear just change its name if someone else buys the trademark.
The usable life of a model is until the next one comes out. If OpenAI were to fully release the weights of GPT 3.5 today, I doubt anyone would care.
Can’t speak for the original commenter, but back in the spring (maybe early summer?), I expected the bubble was going to pop whenever we get a full S-1 filing for one of the frontier labs, and that’s still my expectation. I don’t think the selective leaks have prepared the market for what “positive margin on some products yet incredibly unprofitable” is likely to look like in detail.
I like the dot com bubble comparisons mainly so that I can look at the differences. That is where the information is, and it's what you want if you're going to navigate this market.
You do realize the same was true every year leading up to 2000.
The peak is always before the fall by definition.
Just because something has peaked SO FAR is a very poor approximation that it has peaked completely.
They're not saying it's bad news, they're saying it's not a valid indicator at all.
It is different this time, it's like comparing a hammer to a fucking steam engine.
This time we're building AGI, so yes, this time is different.
Is the AGI in the room with us?
The real AGI is the friends we met along the way (to financial crisis)
Wow, is this Hacker News? Where have you been? Last night I had two agents up autonomously developing and training deep learning models much faster and better than I ever could as a professional. Seems the reddit bubble of "AI doesn't work but also it's somehow evil" has taken over this place as well.
Can you not feel it’s growing presence?
I'm literally replying to you right now while it's chugging away on my work, so yeah?
Weird, if I was intelligent, I don't think I'd be spending my time doing autocomplete on your Django app
It's been motivated to "like" work.
(Scare-quote absolutely necessary: even with maximal anthropomorphisation, we have no way to differentiate between "rewarded for work well done" vs. "punished for not working or half-arseing").
I think there is a bit of a gotcha here. Is intelligence the same thing as a will or drive? AI seems quite intelligent, somewhat lacking in wisdom, and having almost no ego or will. It completely lacks self preservation and desires (as far as we can tell), but are those innate qualities of intelligence?
It also exists is a very different state that biological life. Biological life has a brain that either has activity or doesn't. Once the activity is gone, the brain is forever off and nothing can start it back. So we associate the activity as life and as the intelligence running. AI has weights with vast interconnections, but they can easily be turned on and off, rewound, paused, and so on in a way that we couldn't dream of doing with biological intelligence.
So, maybe these differences are enough it isn't intelligence. But it is something, something that is getting better at quite a higher speed. Some that approximates intelligence in many ways. That passes the Turing test. So that something needs a name, and intelligence seems a hard enough to define thing that it is easy to stretch it and include this new factor as falling under the term.
no, it's very hot, and the fans are loud and obnoxious so they keep putting it in data centers near places that don't really want it, and they install lots of natural gas turbines because the grid isn't enough and that's louder too, and then the cooling often needs a lot of water, and -- really unpopular overall, as it turns out!! huge backlash. and that's before, you know, allll the job impacts, which is when it gets ugly and threatens to complicate the whole consumption-driven structure of the economy.
... but the little dgx spark is in the room with us! it's pretty smart-ish, in a pinch!
>building AGI
Great strategy. Like growing a metric ton of HeLa cells in the hope someone will assemble a super brain out of them.
Even AGI might just be a 'meh' moment. It will take a while before you get through social, political and physical barriers before we start seeing it make a meaningful impact.
I imagine the first thing that would happen is that some people will declare AGI is here and some people will be in denial and call it a "stochastic parrot" because they either feel threatened by it, cynical or just ignorant. Then we just continue business as usual for a while because most people wouldn't know what to do with it and definitely don't want to be replaced even if they aren't against the idea of using AI. A lot of research is also useless without collecting data, experimentation in the real world and hardware, all of which take time.
Agreed. The "singularitarians" always seem to overlook one simple truth: we don't use the intelligence we already have. The appearance of a 100-fold increase in cognitive resources will have no immediate effect at all.
"immediate" is doing a lot of heavy lifting there.
"The intelligence we already have" is why we're splitting atoms and the other primates are splitting rocks. That took us a few hundred thousand years, of course.
But if you have our mere equals in digital form, that only takes as long to change everything as it takes to build a factory to produce the same number of equally dexterous robots.
The dot.com bubble didn't pop because the web was a bad idea.
Every time is different, that doesn't stop it rhyming.
The value was driven by the size of the user base. We gave each user an arbitrary price but didn’t understand how to turn that into actual money. So if you had a website with 30M users it was valued at $900M. The late 90s early 2000s were wild times in VC, and some lucky people realizing how nonsensical all that was cashed out on time.
Mh, a website with 30M DAU or WAU is today worth 1b USD at least (if you cant monetize on 30M users, then the product has a problem?)
Much harder to monetise users in 2000, even that many.
If you sold stuff directly, handling transactions was still a pain; PayPal had to be invented, and that was only 1998.
If you ran on advertising, you had to invent the ad network and convince people to buy ad slots.
Na, I meant today: I would happily buy your 30M DAU weppage for just 1B USD today :-)
Yes, I have no doubt that AI is going to seriously impact the economy over time. I have lots of doubt that it can impact the economy sufficiently quickly and with enough impact to pay out all the very, very, very large bets being made on it. Even as the AI quality curve has been running ahead of most predictions, its actual impact on the economy has been running well behind.
I am watching the dawning realization among the management class that even if coding costs drop to zero, Amdahl's Law means that perhaps productivity doubles for most software engineers, rather than goes to the MOON. AI can help some of the other bottlenecks, but not 100%.
On the positive side, we programmers have been trying to collectively explain to management for decades now that our jobs are not just 100% banging away at code on keyboards, we're not just interchangable code monkeys. Well, they're getting their noses rubbed in it now, complete with metrics and charts and all the other things they value more than the feedback from their employees.
In fact I was just noticing this week... in the past few months I had been really banging away at a lot of code projects. I was doing that whole "3 or 4 streams going on at once" that many of us have talked about, with a variety of mixed feelings. But in the past few weeks that seems to have starved out. I seem to have burned through the things that I was blocked on code for, and had a reasonable business case to pursue. Now my projects are all blocking up on non-code things again... communication with other teams. Review of documents that are irreducibly things humans need to review. Coordination and planning. Applying human attention to cover over the way that even the best frontier models still seem to have huge gaps in their superficially-expansive automated testing code that they write for their own code. I'm still sometimes unblocking multiple things at once and can briefly have multiple projects going, but then I hit a blocker somewhere and I end up unable to continue having multiple simultaneous code projects going at once.
(The document review is a particularly interesting case. The documents in question are something we expect some humans to at least skim, but are going to be basically a meta-program for AIs programming the rest of the company's code. It is important to get it right because every error is going to magnify quite a bit. AIs are helping a lot with that document, yes, I certainly agree we are proceeding more quickly with it than the same project 5 years ago would have gone, but it is also irreducibly sopping up a lot of human attention.)
Point of all this being that the casual presumption that AI will just utterly destroy the entire economy because it is good at coding may prove to be, well, yeah, exactly the same as the assumption that we are going to all be rich by 2002 because we'll just shift the entire economy to be on the internet and everything will multiply by 1000 in just a couple years as a result. But everything can't be shifted on to the internet. I can't eat a web page. AI can't do everything, and as it advances the things it can't do become a larger and larger percent of the remaining pie as a result. It seems to rhyme a lot with the dot com crash.
Ok, but go through the bottle necks you face. How many of then will 1. go away with better models, and 2. go away when those bottle necks (other people) are automated away? For me, I can go quite fast because I work alone, no other people in a team to slow me down, but verifying the result is still a bottle neck. Also interacting with government agencies, my accountant etc, still a bottle neck. None of these are technically insurmountable, and I suspect that when one of these becomes a real bottle neck for the entire economy, they will be fixed.
How many of them will go away?
To a first approximation, none of them.
There's a huge difference between "zero" and "a little" when it comes to Amdahl's Law and scaling up, let alone the things where it won't be "a little". Which is one of those things that sounds obvious when I say it. But clearly a lot of people are predicting a future that is based on costs and times being "zero" that won't be.
Right now, the most visible bottleneck is real world robotic control.
I kinda hope that one stays around for a long time, because robots controlled by AI competent enough to guide those robots through every step in their production from mining raw materials to end-of-assemble acceptance testing, is one where tiny mistakes kill everyone.
Like, say you've got a sticky 0 key and accidentally ask for more burgers than there is carbon in Earth's biosphere and atmosphere combined: if the sufficiently-competent AI is a little too eager to obey and not quite willing enough to say "no" or listen when told to "stop", that's enough to kill everyone.
Trust me I know, I co-founded a robotics startup, and robotics is HARD. But now it's becoming a hardware problem, the software will sort itself out soon, it's not a bottle neck that will keep us safe for long.
Abou sycophancy, I sense that it's much lower now with Opus 5.5. Yesterday I had to argue my viewpoint for quite a bit before it relented, on an issue I'm pretty sure I'm broadly correct. It's becoming a better sparring partner that actually pushes back and raises counter-arguments.
>The dot.com bubble didn't pop because the web was a bad idea.
The web WAS a bad idea though.
Is that coming at the same time as viable fusion reactors? Just one more year guys
You can tell they’re serious because they’re proposing data centers in space as a near-term solution.
No it isn't. It doesn't matter how real the technology is. The bubble will still pop, because when a technology is seen as revolutionary, you get rampant overinvestment. And only a fraction of those investments will pay off(or, indeed, none, in the case where frontier labs are eventually made irrelevant by open models, which seems very likely to me).
The internet and web were revolutionary, no one can deny that. They completely transformed our society. Nevertheless, the bubble popped. The current bubble is no different.
I don't doubt there will be a big crash at some point, but I actually doubt we've reached that point yet. So far stock increases have been driven by earning increases, not speculation. I know an "earnings bubble" can be a thing, but given how useful I know these models to be, and how few use them in the way they can be, I think we have far to go still.
But those tulips are so pretty!!
I can’t tell whether this is sarcasm or not.
"No, no, you don't understand. If you keep putting bigger and bigger engines in a car, it'll somehow turn into an aeroplane at some point! There's no other way, believe me!"
People forget the fact there may not be enough capital in the world to invest in AI, and so while some companies will go under the market could grow for literal decades. I don't know the future, but this is not a guaranteed bubble.
I’d think that orders for long lead time rivalrous slots of a very finite resource are basically OOTM call options for future demand. It says something that firms think demand is plausible but not a whole lot about how probable it really is. You can always idle these things, you can’t conjure one immediately.
Welcome to the field of real options! https://en.wikipedia.org/wiki/Real_options_valuation
This was a fun morning read thank you.
The title makes no sense... how does ASML orders where you get the machine in the far future predict AI spending?
It's certainly an indicator. But, just like airplane orders, the backlog is long enough that you have to put orders in on a speculative basis.
If everything goes as planed, we'll need new equipment in a couple years, so order now. If things don't work out as planned, you might be able to sell your slot, or push it back a few months, or cancel it.
In a frothy market, you don't really know from order backlogs how many orders are based on firm plans and how many are we might want to build chips, so let's order a machine and see if we want it when it comes up. Or the middle ground of we definitely need these machines in the next ten year, and if sales are still strong, sooner is better.
Companies buying those machines need years to get their money back. If they didn't believe they would make a good ROI they wouldn't buy as much.
> Companies buying those machines need years to get their money back. If they didn't believe they would make a good ROI they wouldn't buy as much.
In the late-1990s companies were building out fibre optic networks, as they believed they would get a lot of ROI from all the Information Superhighway demand. The expected demand never came in the expected timeline.
Of course demand eventually used up all the dark fibre over the course of many moons, just not quickly enough to save many companies:
* https://en.wikipedia.org/wiki/Telecoms_crash
More specifically, the demand for fiber didn't come - but the demand for bandwidth did!
Those companies were building out fiber like it was a legacy telephone network: needing a massive amount of individual wire pairs to carry a massive amount of parallel transmissions. In reality the bandwidth per channel was rapidly increasing due to better tx/rx equipment, and with the rise of WDM it was suddenly possible to carry not one channel per fiber pair but several dozen.
When you were laying fiber on the assumption that a single long-distance pair was going to carry 1Gbps - and therefore laying hundreds of fiber pairs in a trench at once - it is quite a big upheaval when that single pair can now carry 100 channels of 10Gbps each. Your brand-new expensive fiber trunk can suddenly be replaced by a single spare pair some utility company had lying around already.
Turns out laying fiber isn't a great business when the capacity of existing fiber is doubling every year or so!
The demand for AI exists today.
As in OpenAI is at capacity, 6.1 Sol is served at 30 tokens per second and with frequent capacity errors.
Anthropic seems better now but were also working hard on expanding capacity.
I have no crystal ball to see how future demand and capacity will develop, but as of today these data center investments appear to be fully utilized.
Some people anyway seem to think a collapse is inevitable.
maybe collapse in valuation, collapse in demand? I don't see that happening any time soon, it's too useful, idk how anyone can contest it is useful and not just hype.
it is all because of subsidization. If OAI or Ant starts to use real pricing there will be lot of capacity.
No, in fact inference is far cheaper than most think, see analysis for serving Kimi K3.
If you mean OpenAI and Anthropic's investments into R&D, ok, perhaps.
However, they will not be able to increase prices unless what they are offering next is a significant enough improvement over the current SOTA, which might only be six months ahead of the Chinese models.
Not only is inference currently much cheaper than people seem to think, it is going to get much cheaper. NVIDIA’s Vera Rubin is going to drive costs down for current models considerably.
It's amazing to see the shift from "Anthropic is profitable" to "it's just R&D costs" following the leak of Anthropic's finances. If there's one thing more abundant in this bubble than FOMO, it's special pleading.
These companies are burning helps of money, and they need to keep burning heaps of money in order to stay competitive. People have (correctly) learned to be extremely fickle consumers, they can't depend on making profit by running inference on a 1 year old model.
How do you know that it's subsidized?
I really dislike that term. It's stupid and supports a distorted narrative.
They arbitrarily decide 1M API tokens cost X, and any subscription where usage costs less than that is "subsidized". No it isn't. They may be profiting on subscriptions as well, or losing on full API costs as well, or anything in between, and that X is anyway more reflective of competition than their costs. For all we know they did a rand()%1000 to determine it in the first place.
If it isn't subsidized, why aren't all those AI companies making huge amounts of profit? Instead they are rapidly burning through the billions, and the money must be going somewhere.
Investing in R&D isn't the same thing as subsidizing something.
Investing in R&D is fine. as long as you are turning a profit somewhere within your company that can support it. If you are not, eventually the end comes. In short, the bubble.
The problem with this view is that that isn't a "bubble". "Eventually" can support quite a while, especially when the investment is increasing revenue over time and the unit economics look good. The timeline point cuts both ways, and none of that is a subsidy.
The only company making AI chips, and allowed to buy from ASML, is TSMC, otherwise it's probably memory manufacturers.
Even if the order book is mostly TSMC, it doesn't necessarily mean the demand is mostly driven by AI (although maybe it is) - there are plenty of other high volume TSMC customers like Apple, AMD, Intel, Qualcomm, Broadcom ...
Finally, just because lots of TSMC customers may have made forward commitments (or whatever it takes for TSMC to commit to expansion), the main conclusion would seem to be that they needed to do that for business continuity given current capacity constraints. It's not clear why the need for them to do this should be taken as a sign that the AI datacenter bubble won't pop. Do you want to be Apple and gamble that rather then locking in TSMC capacity now you can gamble on the bubble popping and maybe get better terms later?
As a counterpoint to this "no need to worry" story, CoreWeave's 5 yr bonds are now yielding 15% ...
And incidentally, believing that you can keep making a good ROI even if it's not true is what fuels a bubble.
> Companies buying those machines need years to get their money back. If they didn't believe they would make a good ROI they wouldn't buy as much.
But if they're in FOMO mode, they might not be thinking about ROI at all. Or if they're in bubble-optimist mode, they would just assume their ROI will be good.
More importantly, if the market goes south, they rip up the contracts
ASML sell photolithography machines, which presumably can be used for many kinds of chips: CPUs, GPUs, flash, RAM, etc...
They don't need to be AI chips, so even if the AI bubble bursts before they managed to get a ROI, they can switch to other applications that need high end electronics, such as PCs and smartphones, maybe cars.
Presumably it indicates confidence from chip makers in long-lasting demand. With a signal with actual stakes, not just words in a shareholder meeting
It means companies are expecting demand for AI chips to keep increasing
Or companies want the market to think they are expecting demand.
Everything's fine, don't look at CoreWeave's ticker, look at (picks random other healthy company) ASML instead.
Their 10 year maturity bonds are pushing towards 15% yields too. I mean, if all works out, someone's making a nice return
Uh we had a slight hype cycle malfunction, but uh everything's perfectly all right now. We're fine, we're all fine here now, thank you. How are you?
Aren't they always selling everything they manufacture? No matter if you are in a boom or a bust, ASML will always have a years long backlog.
Phew. That was close. Thought we would be able to upgrade all of our tech for a moment there!
You sell the most shovels right before the gold rush ends.
And Nvidia is the sugar daddy of SK-Hynix, which buys ASML machines. What could possibly go wrong with Nvidia being the new Federal Reserve?
I don't think "running out of steam" is the expected failure mode here. We are talking about bulk ordering machines that produce chips that in turn will hit the market in half a decade. That's like adding power plants to the grid that runs the ball pump inflating the ballon. ASML will probably be the last that faces the ripples of the AI bubble going bust.
Even after the AI bubble pops, ASML's machines will be useful. People generally want computers and chips, preferably the fastest an latest-generation ones. The ROI may be lower once the bubble pops, but the machines will earn themselves back eventually.
Once the bubble does pop, ASML will be in trouble, though. It'll take a couple of years for their latest machines to generate as much interest again, so the effects of the market collapse will echo for years after it happens.
Agree, and orders dont mean liquidity, billing, revenue or profits. If it blows nothing of that Will happen.
The internet bubble popping did not cause internet traffic to decline -- on the contrary internet traffic increased rapidly during the bubble pop.
Similarly, I don't think the AI bubble popping will cause token demand to decline. On the contrary, I predict the bubble pop will happen because of a recognition that tokens are a highly competitive industry, thus the margins of frontier AI firms aren't sustainable, popping their valuations.
That's right. Jevon's Paradox won't matter when token prices are falling by orders of magnitude
I don’t buy this argument. Everyone who suggests there’s a bubble speaks about the excessive data center buildout, low demand, excessive debt financing.
All of these arguments don’t matter if you think demand will increase.
What bubble is left? Really. Just people with OpenAI or Anthropic exposure? Not that high.
low demand? Got a source for that? Anthropic is paying crazy sums to direct competitors to meet demand, so it seems that demand is quite high. Could be wrong, though.
> What bubble is left? Really. Just people with OpenAI or Anthropic exposure? Not that high.
The stock prices of 9/10 of the top S&P 500 stocks have significant AI exposure. A drop in valuation of them will be widely felt.
I agree there's no low demand.
> The stock prices of 9/10 of the top S&P 500 stocks have significant AI exposure
not true.. only ~2-3% exposure at model layer. rest are at hardware layer..
And there's a significant bubble in the hardware layer. The Nvidia overvaluation alone far exceeds that of both OpenAI and Anthropic. When the bubble pops people are going to want dedicated inference chips without an 80% profit margin rather than the B200.
why would the hardware layer bubble pop if the demand for hardware will still increase (or not change)?
Because everybody and their dog are designing inference chips. Same as tokens, high demand at commodity prices means low profits.
TSMC will still make massive bank because they're the bottleneck. That bank will be squeezed out of companies dependent on TSMC.
> commodity prices means low profits.
I still don't understand how models being commoditised implies the hardware (NVIDIA) itself being commoditised.
What is the exact mechanism that causes NVIDIA's profit to go lower?
Do you mean to suggest that the bubble popping will simultaneously commoditise models as well as inference hardware?
They are completely different layers and have almost zero correlation. I see no indication that hardware also gets commoditised.
everybody and their dog are designing inference chips. Even currently, Google's Tensor can do inference cheaper and more efficiently than a B200. There are dozens of companies designing inference chips; when they hit the market the pricing power of B200 will drop significantly.
And it's correlated with the other bubble dropping because chip cost will matter a lot more once the margin of tokens is tight.
ASML's order book is way way way at the tail end of where the signal of the bubble's bursting will be felt
Users sign up for Claude, Claude buys more compute, AWS/Azure runs out of capacity, Microsoft buys 10,000 acres in Virginia, Microsoft places order for a million GPUs from NVidia, Nvidia places order for a hundred million GPUs from TSMC, TSMC buys 1,000 acres in Arizona, TSMC places an order for a bunch of new ASML machines
Yes, the end of that chain is the one I'm goin to watch to know when the bubble has burst, not the beginning of it
And what percentage of these orders have been paid for in actual cash, Vs some kind of promise to pay from a company who may fold before delivery?
"Archduke's Death Removes Danger of European Conflict"
"Agreement made in Munich - Peace in our time"
"Effective immediately, there is an ENERGY CEASEFIRE in the War between Russia and Ukraine. Both parties have agreed. Do not break it!"
"We will not have any more crashes in our time."
- John Maynard Keynes, 1927
Despite the fact that chip making machines come before chips, this seems like a lagging indicator, not a leading indicator.
These would be among the first orders placed in a build-out plan, and the build-out plans are first in a boom. The bust phase would very well happen before the orders are completed, and orders would not decline until the bust phase starts?
Hmm. In a boom people over-order.
the site threw up a modal, blocked me from scrolling, then locked the browser, so I couldn't read everything, but from what I could see the gist is people are still over-ordering therefore the boom is going strong.
Yes. ASML is filling order book for 2029 and 2030. I am just not sure, with how long-term investment the chip fabs are, would the slowdown in datacenter demand be visible on ASML's side first or last?
There's no downside to ordering a machine in 2030 an cancelling it later if it turns out you don't want it.
The further out an order backlog is, the more orders get put in to hold a spot in line.
There's a cost to breaking contracts like these. You can't change your mind two years down the line without paying a fine. ASML isn't stupid.
That said, computers are here to stay. Even if FABs can't sell chips at ridiculous prices, there will be a return on investment on these machines, even if it's not as high as originally hoped.
> There's a cost to breaking contracts like these.
There will be at some point, but probably not until the quarter they start building your machine.
I'd kinda like the AI spending boom to end so I can, y'know, buy components and such at non-insane prices.
I think the big question is how many orders are '0 or close to zero dollars cancellation fee'?
Strong orders predicated on future earnings of OAI and Ant?
What earnings? It’s losses
Until we find a better substrate, silicon IS the future.
The stock price did not move.
So nothing was eased.
How so? The stock is near historic highs
I thought the fear is that AI spending is much higher than the earnings
If you buy the machines and not use them, at most you lost several hundred of millions of dollars (not really, because that machine still retains value for you or someone else later).
On the opposite end, if you don't buy the machines in time, mispredict the demand and can't produce, you lose tens of billions.
So weird that they're inverting knowledge and/or causation.
AI is a bubble because the demand story is implausible. The fact that someone is willing to make ASML orders (which is a >= 10 year indicator) doesn't tell you anything at all about whether AI companies can, in fact, get MANY ORDERS OF MAGNITUDE of revenue growth.
no it doesn't
I think tying this to resilient AI spending is misleading. The current AI spending boom has jolted a lot of fabs and companies into realizing that they’re not adequately prepared for the next few decades of semiconductor markets regardless of how AI turns out. That the previous market of divided consumer and business segments is ill-suited to a future where compute is more universal between the two, and that if they want to avoid one going bust when the other booms (like the current memory apocalypse) then they need to be ready to meet that sort of demand. Then you mix in nationalist tensions and China quickly rising to fill the consumer gap with older (but still excellent) specced hardware like DDR4, and the writing is on the wall for the traditionally Western-dominated compute industry: adapt to changing demands or die, just like manufacturing did.
What I’m watching is the timeline. If China can get DDR4 kit out in volume (or even better, DDR5) faster than the AI bubble deflates or markets correct, then the future consumer compute players will be Chinese, not American, especially if they can ship kit at pre-COVID pricepoints.
This current bubble will be the last big payday for the three Hyenas of memory, rest of the world (most of it) will be buying Chinese memory after 2030, and a few other companies will probably be designing an engineering memory in house and getting it fab at TSMC. And AI models they will be better and smaller, and the hardware? It will be up to the task of running the models on your desktop.
> ASML's strong orders ease fears AI spending boom.
By submitting this title you make yourself the enemy of the USA.
Remember it is called SI - for "Super Intelligence".
https://www.politico.com/news/2026/10/08/trump-ai-super-inte...
"<president> says anyone who does not refer to AI as ‘super intelligence’ is ‘THE ENEMY'"