ASML's strong orders ease fears AI spending boom is running out of steam
reuters.com
reuters.com
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.
Every time is different, that doesn't stop it rhyming.
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.
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.
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.
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.
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.
The web WAS a bad idea though.
Great strategy. Like growing a metric ton of HeLa cells in the hope someone will assemble a super brain out of them.
... but the little dgx spark is in the room with us! it's pretty smart-ish, in a pinch!
(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").
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.
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.
"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 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 know people who bought Ask Jeeves at the height of the bubble.
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.
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.
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.
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.
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) .
But that is assuming they survive the bust, which is not a given.
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.)
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.
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).
[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.
Its very good and tells a lot!
Thanks for sharing!
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%.
- 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.
> 'Yikes': CoreWeave stock sinks after its bond yield touched 15%
https://finance.yahoo.com/technology/article/yikes-coreweave...
Which is why you buy an index, rather than try pick individual stocks.
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.
Today, e.g. ZooPlus is one of the biggest pet-stuff-retailers (2023 revenue 2.5billion)
#2 "after all payments done for yourself/husband"
#1 will be the kids
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.
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.
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.
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.
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.
It’s the total returns that matter, not just stock price.
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.
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.
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.
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.
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.
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:
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.
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.
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.
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.
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!
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.
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% ...
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.
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?
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.
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.
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.
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.
> 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.
> 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..
TSMC will still make massive bank because they're the bottleneck. That bank will be squeezed out of companies dependent on TSMC.
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.
And it's correlated with the other bubble dropping because chip cost will matter a lot more once the margin of tokens is tight.
The further out an order backlog is, the more orders get put in to hold a spot in line.
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 will be at some point, but probably not until the quarter they start building your machine.
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
- John Maynard Keynes, 1927
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.
So nothing was eased.
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.
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'"