> Another company, which then bought the largely finished canal, cheap, they had potential profit... until tech change (railroads).

Absolutely right. Somebody may buy the DCs. The tech changes very very fast, the DCs may be entirely worthless. US has high energy costs, it can never be cost competitive.

Why Data Centres Will Change the World and Bankrupt Their Builders — John Hempton https://www.youtube.com/watch?v=GY5YyPLfqp0

Useful (like solar, airlines, fiber), but it won't be profitable. High capital cost, low marginal cost, subject to competition and with technical improvement --> these four factors guarantee businesses will never be profitable, and they go bust.

Also, (setting aside Standard Oil’s business practices), this is a common meme in oil exploration historically, and even today. Early prospectors were honestly pretty stupid, and their principal differentiation was risk tolerance.

There a lot of parallels with AI…it doesn’t take a genius to see the long-term potential, but the risk is in implementation, so it’s a mad rush to the finish line by a wide variety of less scrupulous early players. Rockefeller got his big start by sweeping up failing producers who were all capital weak, and I’m sure something similar will happen here during the inevitable bust.

I can do the hard, risky work of drilling wells / deploying AI, or I can sell you a vision of that future, today. The most dangerous technologies in terms of grifting are the ones that everyone can understand.

I'm reminded of vertical hydroponic farming in old silicone fabs in Japan...

https://www.japantimes.co.jp/business/2026/07/09/companies/p...

Restaurants and bars!

This pattern repeats across many industries with large up-front costs.

Capital cycle theory: https://www.youtube.com/watch?v=2J2Fb1bBufA&t=577s

[dead]

A large number of companies failed in the dotcom crash, but the founders did pretty well if they'd made it to an exit event. Don't forget that it was a stock market bubble imploding. The founders and early employees had already sold shares in the IPOs that put the companies on exchanges. It was the investors who bought stock that lost money.

The after effects were that it was virtually impossible to raise capital for a while afterwards, so the next round of founders struggled. But then interest rates went to practically zero and money got cheap so it was all good.

Wait. Are you talking companies, or people? Companies come and go. They are replaceable virtual containers for people. I don’t think it’s appropriate to apply the term “poverty” to companies.

The Yahoo founders left with never-fly-commercial money before re-joining the billionaire club by investing the money they had. Others that survived post-IPO include eBay and Dell.

There were high profile failures but we have had those recently too eg. pet.com (dot com) and color.com (post-dot com, domain has since been repurposed)

eBay, Priceline (now Booking.com) did very well too

I know numerous multimillionaire dot-com techies. I guess they're not "top players".

Mark Cuban?

Jim Cramer lost big but has continued spouting nonsense ever since pretty successfully

Your logic is backwards.

NVIDIA will take a huge hit and we will probably end up with anti circular financing laws.

Frontier labs are losing money and are most likely to go bankrupt and be absorded by bigger companies.

Hyperscalers will take a huge beating but they are ultimately profitable and they already control the frontier labs. So after a hard crash they will be fine. A lot of their debt is actually offloaded onto pension funds, insurance companies, etc. Those risk bankruptcy.

Wait... with hyper-scalers, we are talking about Google, Meta, Amazon, Microsoft, are we not?

And do these not all have established business models that are ridiculously profitable, and huge cash reserves?

How the heck are THEY vulnerable, and the "bare" frontier labs like Anthropic and OpenAI with no preexisting business models or profits, but equally ridiculous borrowing are supposedly not?

Its profitability ain’t.

> AI one day must fly to space, be called Mother and discover aliens.

We're discussing economics, not science fiction.

Consider the wisdom of speculating on the incredibly fast adoption of the internet and mobile technologies, and on that basis investing heavily in whatever mobile carrier had the fastest network.

The speculation that the AI investment sector depends on is not just that adoption is incredible and that the value it generates is high, it's ALSO that the companies will be able to capture a significant amount of that value generation for themselves via fees or some other mechanism.

Are the LLM vendors more comparable to Google/Apple in this scenario, or more comparable to AT&T and Bell?

> The reality is that the adoption is incredibly fast and that revenues of the AI labs are like doubling by the quarter.

They’re still losing gigantic amounts of money. Whether that is because they’re “flooring it” as it were and growth will eventually catch up, or because the business idea is fundamentally flawed and just temporarily alive because of investor risk appetite, is the fundamental question underpinning whether it’s a bubble or not, and nobody knows the answer to that question for sure. That’s how bubbles work.

> The reality is that the adoption is incredibly fast and that revenues of the AI labs are like doubling by the quarter.

How is that compared to the speed at which money is thrown in them? And how much of the profit downstream they might indeed potentially capture with high retention rate?

Bubble can’t be assessed on mere quarter profit expansion over short period (relative to large company group life expectancy).

Their point is (and I tend to agree), that LLM AI is not different at the fundamental level. The idea is that e.g. coding capability will be a standard feature of future computing hardware, not a very expensive service you rent. Anthropic is the current-gen mainframe operator selling time-shared access. In the future, that capability, too, will just be a commodity, i.e. a part of the silicon in your notebook or smartphone or whatever will replace those.

>It needs human-in-the-loop.

Not AGI Pilled. NGMI

>I wonder if physical technological progress is also the same as information treatment progress. Everybody is happy having food and materials faster. But AI bringing you solutions faster is also taking away your job, passion..

and, apparently, our brains https://www.media.mit.edu/publications/your-brain-on-chatgpt...

> But AI bringing you solutions faster

Is it?

Every single vibe-coded software I've ever used looks pretty on the outside but has massive, glaring holes in functionality that make it unusable.

It's like having food faster but then you find out it's actually full of silicone and microplastics.

Who is using AI besides developers and scammers? Certainly the majority of people don't have codebases to write.

Also GPT-2 was 7 years ago, not 3.

> I just wish we lived in a world without the speculative bubbles

Speculative bubbles, I.E. risk is where much of the capital comes from to develop these ideas to start. I’m not sure you can have one without the other.

When you say “increasingly undeniable” do you mean that it’s already plainly undeniable and it’s getting more so? Or do you mean it is still a bit deniable but is threatening to become undeniable soon?

Is the utility increasingly undeniable? What does data actually show?

A tech workers union or a stronger regulatory state could help with this but neither party has been that interested in either of those things for the last 50 years.

> I've been saying for awhile now that the utility is increasingly undeniable.

If that was true it is strange that you would find yourself saying as much for a while now, given it sounds most appropriate in response to someone denying the utility of slop generators.

[dead]

Bubbles are caused by reckless stock purchases leading to insane, unjustifiable valuations, regardless of the company's prospects.

Hysteria + widespread financially illiteracy means there are lots of people prepared to buy AI stocks at any price. They're adamant to get in (FOMO).

Suppose OpenAI will be a successful company into the long term, and you bought its stock at a $5T valuation. After several reporting seasons and subsiding hype, its true worth turns out to be closer $1T. OpenAI could still be enormously successful, but anyone you bought at much over $1T valuation is going to get hurt.

Nah.

Every single railroad company failed. That repeated several times.

The ones that existed on the end were created by assembling the capital brought on the bankruptcy of the original ones. Most of those failed too.

That's how a bubble is. No original investor got out ahead.

How do you know we aren't already through the exponential part of the S-curve and about to enter the AGI winter (i.e. AGI will require a fundamentally different technology than LLMs that is still decades off)?

If no one is willing to pay increasingly high costs of training frontier models, we'll still have the models that we have now that can be fine tuned for incremental improvements. Still tons of utility in that, but little investor upside.

> I suspect this opinion may come across as naive for the myriad software engineers whose job has become pressing the enter key

FYI Claude Code’s default is Auto Mode these days

> appears from many angles to be an extremely steep exponential start of what we can probably agree is an S-curve.

In every bubble some say this. And it could be the exact same herd behavior (mentality) that creates a bubble.

They did go bankrupt, just a bit later :)

For lots of use cases you don't want the BEST model, you want the model that gets the job done for as little money as possible.

> Claude has a different personality and performance than other models for example

This kind of thing amounts to a LoRa applied to your model. There's no reason (apart from potential IP infringement) that someone else couldn't produce a "Claude LoRa" for their model (or, more likely, given the dislike for Claude's writing style, Anthropic could offer a "non-Claude LoRa" for their own models)

Claude’s personality is the worst aspect of the product!

"Cursor" Grok has the exact same (non-)personality as Claude.

GPT has the same performance. And performance is definitely a commodity. Why wouldn't it be?

Maybe

>AI models can be deployed on the same datacenter as the old one

This hasn't historically been true. The clusters have been getting bigger along with the models over time. And now we may even want to have these computers have rapid access to extra compute or caches of the internet or other data stores.

That's not the difference. The difference is the distinction between core infra, B2B services, and B2C services.

The companies that sold core infra during dot com mostly died, because they invested in fiber etc far too quickly. Likewise with server factories like Sun.

The B2B and B2C survivors were patchy. A few did extremely well but there were a lot of casualties.

The biggest winners offered B2G - business to government - disguised as search and social media. They were actually selling individual user profiling and targeted ads, and that turned out to be the biggest money maker.

The reason Anthropic and OpenAI appear wildly overvalued is because their offering isn't AI in some abstract sense, it's the next generation of surveillance and persuasion technology. But now with corporate surveillance in addition to the traditional profiling and targeting.

You can tell a lot about what someone is doing - and the industry is doing in general - by looking at everyone's code bases. You don't need to train on them or steal keys to extract immense value.

Potentially there's also the option to eventually make claims on corporate IP developed with AI tech.

Anyone who thinks the companies who pulped and torrented whole libraries of books for their training data would never, ever, pinky promise, try to extract value from the code they're generating is being very naive.

> AI models can be deployed on the same datacenter as the old one

Not if the datacenters are in space, you can upgrade the models for a little while but once you need to upgrade the silicon you need to launch new hardware.

And data centers are practically free to build, are welcomed everywhere, cost little to run, and never need to have their infrastructure upgraded!

The market can remain 3.14159...

Why?

The major problem with VR is that it simply doesn't have a good use case beyond the initial wow effect. AI certainly does.

> I wonder why so few people make the comparison to the recent VR hype.

AI is already a gigantically transformative technology, with a staggering amount of paid consumption, and virtually everyone I know uses it every day, where I don't know anybody who regularly uses VR in any form.

The ship has absolutely sailed on "is AI a world-changing technology", where "are the leading AI companies worth what's been paid for them" is a completely open question.

Because VR failed to deliver at all. The experience is narrow, isolating, and frankly pretty horrible. People get motion sickness. The set of games is limited never mind the set of non-gaming activities. AR similarly has a lot of issues like narrow FOVs, low resolution, and brightness issues.

VR is a toy with a few very narrow areas of application in industry. That does not describe the current status of AI.

It didndump a lot of money into it, improving some things and headset availability. But arguably a most of it went to useless stuff no one asked for, like paying developers to build shovelware for Quest or their useless corporate Metaverse.

Dot com crash is a mush better comparison. Over promise and under delivery in a immature market.

Yes. That's the situation in 2026 in most of the world. But in the 19th (after the railway bubble - https://en.wikipedia.org/wiki/Railway_Mania) and early 20th century, railways were actually very profitable.

> Almost everywhere I the world railways are state owned and heavily subsidised and still aren’t that attractive to customers.

That's a remnant of the past, where only a nation state could afford massive projects such as creating a railway line crisscrossing throughout a massive country.

Since the last couple of decades we see private companies investing in railway lines, including high-speed railway corridors. We also see a major liberalization push in Europe to decouple railway operators from infrastructure maintainers, and now there are plenty of private companies operating railway services across the world. In Spain alone there are 2 or 4 private companies providing high-speed railway connections between major urban centers.

I think this is untrue for Japan.

There isn't much money to be made in agriculture but food is very attractive to customers.

Based on what data?

Yep, trains had a good run, but highways and automobiles took over. On-prem inference is the automobile.

I know where you're coming from, but we probably should not be talking about AI in this way. I don't think it's helping or serving its purpose.

I think civilization will have time to react and adapt, panic leads to haphazard action, we need deliberate consistent action.

AI wouldn’t work without any of that above stuff though.

Their monopolies got broken and the profits disappear in a commodity market.

If nvidia or tsmc got split up i doubt its stock is going to stay at as a high a p/e.

This is exactly the kind of talk that led to all the other bubbles. Amazing how infected people get.

You know what was actually a bigger invention? Language. Farming.

Imagine a world without those. Oh, wait: tomorrow people won’t know how to use language, will they? And they won’t have water to farm with, so there you go.

Timing is all that matters here.

Humans living forever is inevitable. Galactic colonization, inevitable. Over a long-enough time scale.

AI is inevitable on the same time scale, sure. But can it actually survive as a sustainable business given what is needed to keep it relevant—will people pay what it costs to support said business—not clear. Or, at least in the next five years.

The second generation AI companies will be built on different assumptions and will likely be slower-growing, and that’s when you’ll see it truly become ubiquitous. That could be a decade though.

> because AI is inevitable.

is anyone else weirded out by folks chanting this like a mantra? imagine saying this for any other emergent tech. "NFTs are inevitable", "crypto is inevitable", "excel is inevitable", etc.

inevitability != usefulness/profitability

> everyone is assuming it's a winner take all market.

I guess this is the big gamble, the idea that if you just spend enough money as fast as you can you can become the gods of AI forever and can just charge whatever you want for your magic beans.

But what is the moat? If in reality these things become commodities and just compete with each other to spit out cheap tokens then there's no way the margins are there.

A "failure to misunderstand" is a fitting turn of phrase.

People not only ignore second order effects, they also don't think about what would actually happen when a "winner" really "took all".

While AI is a tremendously useful technology, business types wildly misunderstand it and its consequences. Greed eats brain.

It's probably closer to 45-60%.

https://epoch.ai/data-insights/ai-datacenter-cost-breakdown

https://www.csis.org/analysis/impact-tariffs-ai-data-center-...

> since everyone is assuming it's a winner take all market. That's probably close to correct.

why? I keep seeing this over and over again but I've not seen an explanation. I would really appreciate some insight into this line of thinking, since it's not obvious to me. What is the benefit of holding hundreds of billions of computer hardware when it's not proven that it's even needed.

Power generation is also extremely expensive ongoing - either through gas turbines that eat fuel or on-grid power that datacenters will need to pay to capitalize (depending on ratepayer agreements of course - but clearly people aren’t stoked about subsidizing these)

I read somewhere that the expected 3-5 year lifespan of inference GPUs was not holding up to scaring: in practice, the GPUs have the same lifespan as any datacenter CPU or RAM. I think that estimate was based on Bitcoin mining ASICs, which wasn't an apples-to-apples comparison.

The problem is likely the GPU price, NVidia has a very large margin - but it doesn't seem like the the use case supports it. If NVidia cut the GPU price by 2-4x then these problems likely go away.

I wouldn't be surprised if most hyperscalers are betting on cheaper GPUs on their next upgrade through either pricing pressure on NVidia, competition from AMD/Broadcom or in-house chips.

> And they probably only have a 3-5 year primary lifecycle.

I think in 3-5 years close to 100% of current GPUs/RAM will still be in service. New hardware matters only if gains in efficiency are large enough to put you out of business because the competition gets them. Extreme case of that are bitcoin miners. With AI inference there are way too many factors that influence the profitability. It's not like you can't run new model without the new hardware. It can just run on more of the old one. 3-5 years for GPU are estimates when the main application was gaming with new games with new requirements for running on a single GPU being constantly released.

The gamble: gunning for market supremacy before investor cash dries up.

The real gamble: betting there is a large enough pool of c-suite that doesn't fundamentally grasp the concept of vertical integration that they're willing to pay to train their competition. Because if you think AI companies aren't training models how to do everything your company does with your data you are deeply delusional.

The real real gamble: praying sovereign governments stay anesthitized to the national security implications of a handful of extra-national organizations operating the conputer security equivalent of WMD with no oversight.

Fun part ... all those GPUs are going to shredder after use.

> The GPUs are probably 80% of the total cost or more. And they probably only have a 3-5 year primary lifecycle.

There's plenty of data indicating that inference at typical API prices is profitable after accounting for the cost of GPUs. However that doesn't account for the arms race of training new models while performing R&D.

but is it a winner takes all market? currently we have models that are close-but-not-quite SOTA which are still heavily used and very profitable because they are _good enough_ (i'm talking about deepseek, of course).

or is the "winner takes all" mentality based on the assumption of being the first to achieve the technological singularity?

One thing that has to be considered with takes like this not assuming that things wont' get more efficient cost, power, or performance wise, when that's the only thing that's happened with hardware and software, ever.

Normally you'd replace the device because there is more power efficient one on the market so running one at current energy prices costs more than buying new, more power efficient one and saving power.

IF prices of GPU/RAM come back to pre bubble costs, we might see the 5 year cycle of replacement that's relatively common in normal datacenters pre-ai bubble.

At current prices ? those GPUs that are there now will be running there in 10 years still.

Everything happening right now hinges on the idea that there's a singularity approaching.

If you think that, then all this greed and stupidity makes sense.

If not, well, it's just greed and stupidity init.

Jensen has said the primary lifecycle is now 10 years

This is my heuristic for the capex split and effective lifetime of the asset it creates (I'm no expert of course, this is based on various reading):

- 20% Capex has ~20-year lifetime: The shell (land, building, cooling gear, rack mounts, power lines etc.).

- 80% Capex has ~3-year lifetime: The computers, including bulk commitments to power and maintain them.

And interest rates are why this split is a ticking time bomb. The sheer quantum of investment to build one is almost entirely debt funded (~90% of total outlay). I would guess the endgame would be bad even if the split for these entirely different asset classes is 50/50, which would be very generous, based on what I've understood.

Datacenter funding is extremely leveraged. For comparison, only some banks are allowed that much leverage for that size of capital, that too only if they have federal reserve backing. Even in such a case, the retail account holder (pseudo-investor, because our deposits are loans to someone else) has practically no downside protection. If the bank collapses you will lose everything.

1 Gigawatt * 1 year = 8760 GWh * 0.8 load factor * $0.2 per kWh = $1.35B of only electricity.

I definitely get "Community-Adjusted EBIDTA" vibes from that number.

(for those who don't remember WeWork, this was their made-up variant of EBIDTA that they tried to use in their S1 filing before crashing out, here's a random article on it https://privateequitymodels.com/blog/community-adjusted-ebit...)

This appears to be from this weird graph on page 137 of their S-1, "Significant increase in revenue per dollar of compute". It's not really explained well and has very not only WeWork but also Enron and even DotCom ("number of eyeballs on the page") vibes.

It _could_ be the cost of running inference to respond to requests which bring in revenue, however the cost is described as "compute and infrastructure expenses", while revenue is described as "in-period revenue". I don't see these numbers matching anything else in the filing, however I do not claim to go over the entire document.

So to be perfectly honest I'd just ignore this metric completely :)

Please, Giancarlos. Let's all behave with the times... I think it's culture-appropriate to say "You're absolutely right. Datacenters aren't more clients. We build. What if they don't come? It's not just a one-off. It's a massive gamble."

/s

(edit: fix name)

I reckon it depends on how hot they run GPUs for how long. Running any chip at clock or overclocked (likely) will burn it out faster than "business as usual" workloads.

Defraying of capex also depends on how the hyperscalers scrap used hardware.

AFAIK, all hyperscaler hardware is effectively a consumable. Disks, memory, boards, chips; all the high-value pieces do not get refurbished into new computers for sale into secondary markets. The commercial and legal liability is too high---leaking their own proprietary IP as well as leaking customer IP, not to mention secrets and hardening configs and so forth.

I can't recall where I read it, but if a datacenter is modeled as a bunch of black-box hot-swappable shipping containers, then for pre-LLM workloads, per-container unit economics works only as long as a container delivers over 30% capacity. After that you got to swap it out, and send the black box into a furnace to recover them rare earth metals for the next generation of chips.

To a first approximation, high-throughput, gpu-maxxing math workloads (LLM inference) are surely harder on chips and memory, than are bread-and-butter SaaS / data lake / ETL / RDBMS workloads. Speculating further, I'd guess that the ~40% operating margin of datacenters is due to the relative compute-friendliness of non-math workloads for lack of better words.

> The $2 for every $1 inference cost quote from Anthropic I presume includes all opex

No, it does not, its just revenue minus compute. Per their S1, Anthropic has never turned an operational profit.

That's because there's a temporary supply side crunch.

> And have good resale value for other use cases.

Do they? Is H200 $-competitive with an ASIC for crypto? Bearing in mind they're 4 figures.

No horsing around here, just money math that says interest rates on debt sunk into datacenters will far outstrip the most optimistic revenue projections as of 2028-29-ish, when the bulk of the debt will come a-calling.

Also, inverse Pascal's Wager, because horse v/s motorcar is a false dichotomy... they are bound by the same laws of physics (faster horse, less shit, more noise and more smoke), merely different economic realities.

Hyperscaler bet is "Superintelligence" and/or "AGI", or we will make a God. That is a technological singularity. There will be a "before God" and "after God" world, and we will very likely never be able to understand what the AGI/Superintelligence's world looks like. (Because to understand what it is like to be a bat, you have to become the bat, but at that point your meaning of "understanding" itself changes entirely).

So I believe ten trillion dollars and a bunch of pirated books and source repos, and well pirated everything, mostly of very dubious quality and provenance, will not result in a God. I've been acting accordingly.

I mean Ford didn't get a trillion dollars because he is to be only manufacturer of any and all transportation. There are still horses, and Ford and Dodge(Anthropic) are not even the biggest manufacturers.

On the contrary. Anthropic states in their S-1 that they are moving to building their own compute (ie see page 139, Compute Investments):

>As of and subsequent to September 30, 2026, we have entered into purchase commitments and lease agreements, including short-term commitments, as well as leases not yet commenced, with an aggregate committed spend of approximately $895.4 billion. These commitments include approximately $447.5 billion expected to be incurred before December 31, 2030, and substantially all of the remainder is expected to be incurred between 2031 and 2045. Approximately $69.8 billion of these commitments are due in the 12 months following September 30, 2026. A significant portion of this total committed spend relates to Anthropic Direct, and we expect it to represent a growing share of our compute investments over time. The flexibility from Anthropic Direct allows us to backfill the short-term compute we procure during periods of high demand with more cost-effective, long-term compute as our dedicated infrastructure comes online.

So they decide to change the way they procure compute (rent vs buy) for which they will have to raise a little shy of $1T in the next 20 years, with $70B due in the next 12 months (and $450B in the next 4 years). This adds significant risk because they have not only to execute the build-up (which they haven't done before) but also raise the funds to pay for it, all while continue to operate and grow the core business.

I mean that's not entirely true

https://openai.com/index/building-ai-infrastructure-with-the...

Yeah, Open AI and Anthropic may be buying the bulk of their compute from other providers but they are now moving into building their own infrastructure directly. There is a scale beyond which it makes more sense to build vs. rent

It's "personal bet-ium" against LLM-AI as business [0] (NOT LLM-AI as technology).

Personally, still haven't found cause to alter my opinions, stated here:

- After "AI": Anticipating a post-LLM science & technology revolution

- Riff: LLMs are Software Diamonds

- Halting AI (as in, the Halting Problem, not stopping it dead in its tracks --- a foolish idea).

- Riff: Classifying Tools for Thought (what, relatively speaking, is the intelligence level of LLM tech, in my book... It is intelligence. But what kind?)

Cf: https://www.evalapply.org/tags/ai/

---

[0] Kind of stupid to bet oneself against ten trillion dollars backed by absolutely massive corporations as well as world superpowers, but it takes all kinds to make the world :shrug:

Sorry, did you have a real rebuttal here to OP or did you just want to say the tired phrase "copium" and drive off?

Massive burn: the people I disagree with have motivated reasoning.[1] Would it be cope to deny that there is a bubble because one might have invested a lot in AI?

[1] https://news.ycombinator.com/item?id=50422043

> The hackernews crowd just can't emotionally handle the coming economic disruption.

The “hackernews crowd” has a lot of highly-paid SC wage workers, even venture capitalists (whatever that is about).[2] Personally though I just have a normal job and don’t want to be disrupted out of how I earn my living. Now you can go on and gloat. Gloat about how you think that I will become economically destitute. This is an open invitation.

[2] But I don’t even buy that the hackernews crowd is coping about the success of AI. Most seem to be euphoric about outsourcing their coding. Any and all LLM-technical posts get hundreds of replies by eager agent managers. While the LLM-political posts get a mix of enthusiasts and haters. It’s like the LLM-technical submissions live in an alternate dimension.

Even Anthropic is having their most recent debt priced at 10% interest. Wall Street is seeing them as pretty high risk.

> stranded assets will have a shelf life

This is what keeps me up at night. That fiber that got overbuilt in 1999 is still there spitting photons. That Vera Rubin gathering dust on a shelf will be a paperweight in 5 years.

> margins shrinking at Anthropic

Why is this? Competing on price, or compute somehow getting more expensive?

To me these are arguments _against_ the bubble popping. The market reacted to all these signals. The semis took the brunt of it but even there it wasn't really dramatic. These are signs of a relatively healthy environment.

> What's weird to me is that you don't see a lot of people making deals that seem way too good to be true

You gotta be kidding. HuggingFace, a website to download files, was sold for 13B.

> Plateauing of model capability improvement

On what are you possibly basing this? The fact that there hasn't been a new model release in 6 weeks?

Yeah, seems like there is a GPU to engineer sweet spot with models like GLM. This could become a more popular pattern for some organizations. Likely also the reason gov. attempts to intervene

> anthropic make ~$2 for every $1 of compute

That was true a quarter ago, when Elon still wasn't charging them. Now that he is they're making about $0.96 for ever $1 of compute.

100%, great additions. Someone should make a tracker site for this.

With big technological shifts like this, it's important to keep in mind that the ability to do more with less resources can pan out in two different ways: (1) Keep doing what you're already doing, and lay off the people you no longer need, or (2) Do more than you did with the same amount of resources.

The answer depends on the company, but I think many companies are still figuring out whether and how they can exercise this new capacity. We don't know where it will land yet at a societal level. It's a parameter shift, and it will take some time. Pray that you live in a place with sensible labor laws and safety nets.

No. The biggest layoffs are happening at companies that need to free up additional capex for infrastructure buildout, because they're not getting ROI from their existing capex, and they're having trouble finding more money.

I think that's actually going to moderate the bust, by giving the big companies the moat they need.

It performs the same function as regulation, replacing "the scale to navigate and manipulate the regulatory environment successfully" with "the scale to navigate and manipulate the political environment".

It's essentially grassroots regulation instead of federal regulation.

(I don't think it will stop the bust, but moderate it, allowing more companies to survive.)

Agreed. This applies even to simple words like “velocity”. When you keep seeing the same key words/expressions used in support of the same argument, you know the person is parroting an idea as opposed to sharing an original thought.

> We need a temporary moratorium on the words

There is no such thing as a moratorium on words. It's another phrase that people like to repeat (a meme) until the excitement or prestige of using it has waned.

That said, i agree with you that word-salad investobabble is a sign that the bubble has put many people in a raving fever. This current bubble manifests like the cryptocurrency mania.

Agreed.

Also the sadder slice is all the X people token maxxing with no hope of distribution. Cool things are cool, but getting buyers is getting harder.

Still useful. Especially if Chinese GPUs operating on Nvidia software become performant and accessible, thereby undercutting the price gouging that we're seeing with western chips.

The big issue will be access (Re: BYD)

> those servers have about a 5 year lifetime and then are likely to be replaced

This made sense in the past when they were not prohibitively expensive, but other than them breaking down why replace them on the 5 year schedule now? Yes, the newer stuff is better but the jumps from this point forward seem more incremental than revolutionary hardware wise. And even then, a company can simply swap out racks as parts fail.

> those servers have about a 5 year lifetime

The real ramp up started in 2025, most servers will live at least until 2030. If the bubble pops it will be much earlier than that.

They actually are compatible, because the idea of replacing workers with AI in large numbers isn't compatible with the kind of economic growth these valuations are predicated on. We are not living in pre-industrial revolution times, where most people are wasting huge human effort on food growth and physical tasks, having room to move into more white collar jobs.

So, if you have massive unemployment because AI really is cheaper than labor, you'll get a massive reduction in revenue for virtually every industry, collapsing all of these valuations. Even for B2B companies, Google isn't going to be making much money selling ads to unemployed people who can't afford to buy anything.

They're compatible because they are both driven by hype, not reality.

Liabilities in stocks are still liabilities. It's money they will not get if they go public. You can't just ignore these.

Big money (Wall Street) will be looking at GAAP equivalent accounts for their real valuations. They'll happily go along with public figures looking healthier than reality, providing they can exit investments or lock themselves in to be paid back first, when offloading to the public or in the case of bankruptcy.

GAAP rules exist for a reason. One of the main reasons being the investors should have a clear and consistent view of the business before they decide to invest. The "non-GAAP adjusted" might as well read "But if we exclude the numbers we don't like our future looks so much brighter".

Most of the "laws" of economy in this form say often as a footnote "all else being equal", which remains to be seen. Otherwise you can justify any attempts at price decrease in any industry the same way.

> it’s quite rational that demand increases as price decreases

This is not necessarily wrong, but neither is it right; it is underspecified. Even if we were ignoring other factors and taking into account only the assumptions of a classical economic model, the shape of the demand curve still depends on the size of the market and the price elasticity of demand.

Unless the number of users is already pretty much saturated. Which it probably is if we are honest by this point.

> I don’t think this is naive at all; it’s quite rational that demand increases as price decreases, which does seem inevitable.

But that's not what OC is saying, isn't it? Frontier models becoming better and cheaper at the same time are not a given.

Near-frontier models that have 99% of the capabilities of frontier models, at a tenth of the cost of frontier models, might ironically be what pops the bubble.

But which companies will capture that value? It could be like semiconductors where TSMC makes most of them but captures very little of the value.

People keep repeating this but is it true? In natural sciences, for example, there are endless applications. Anthropic models have only recently reached the level what I would call useful. Still, they are far from perfect and there's a lot of room for growth. Is science a very small niche? I don't think so.

> But the AI industry doesn't consist of just Anthropic.

That is not the claim. Author uses Anthropic's filing as an example indicating growth of the whole industry. While questionable on its own, it is also not given in full, which I pointed out.

The competition is also a problem tho. Both of them are betting the house pretty much on winner takes all. If competition reduces growth, that's a sign that they both can't fulfill all their promises which ipso facto means one or both is overvalued.

Isn't that worse because you then get into neocloud companies like Coreweave and Nebius who have taken insane debt to build out data centers and if the demand sinks for those they'll go under

Assuming you have done that, what's your conclusion?

As a kid maybe I don’t get it, a giant ecosystem? Of frontier model providers?

A "gigantic ecosystem" of two frontier model providers

All of the data up to this thread show that there is not enough revenue on the current uptake trajectory to service this debt, nor will there be (through some combination of generative AI not being that revolutionary from a productivity perspective and open weight options).

Could this change? Maybe. That's the bet.

Agree. CDS on the debt of several players here is getting expensive. Debt markets are getting less promiscuous quite quickly, just when even more money is needed.

It could be that a slowing of data centre build (and hence a slowing in the finance needed) takes the heat out of the situation avoiding a dramatic burst. Even then there will be casualties. If someone knows who and when ... they could do very well. More likely, Microsoft and Google will buy up anything distressed that's worth having.

The debt thing is amazing since it ignores the interest on the debt and cherry picks Google's revenue from 2019 for some reason (which is then adjusted for inflation for the comparison).

If you have 20 years to pay off 5 trillion dollars, it doesn't come out to a neat 250 billion per year as that suggests there's 0% interest and no servicing cost. Servicing costs would probably be at least 3 times the stated figure, corresponding to 10% interest paid. Some of the debt may obviously carry less interest, but the financing is opaque (as the author admits to), we have seen interests rates creep across the economy and big tech has had hits to their credit ratings.

I hope the Liberal Party can find a better candidate than the author of this piece ...

Without a requirement for DC's to add more power to the grid I don't think the price is going down, they're essentially adding to demand and the supply isn't particularly elastic. That being said this is a good opportunity for renewables, they're the cheapest form of power.

For the time being this thesis hasn’t worked out though. H100’s are only going up in price and becoming more scarce [1]. At what point would you update if now going on 3 years into H100 lifecycle this doesn’t seem close to happen.

Jevons' Paradox. If it becomes more efficient, cost goes down, demand goes up, use rises, revenue rises.

Look at DeepSeek! They have done full system optimization on their model, architecting it so that it would be cheap to serve. I am proud to call DeepSeek a world class tech engineering firm in a way that 99% of the tech world is not. They are a 1% company.

People are buying v100s on ebay to run LLMs and Chinese manufacturers make custom PCIE boards for them. Older cards might not be useful for training but they're perfectly capable of doing inference work.

> The current-gen hardware is expected to have only 2-3 years of useful life.

Expected by whom and why?

Google hasn't retired any of its now ten-year-old TPUs, or any of the subsequent generations. Maybe TPUs are special or Google doesn't understand compute economics but either argument requires data.

If your argument is depreciation, you're confusing a tax concept (depreciation) with useful life. Yes, they're supposedly related, but in practice, the relationship is rather loose. (Residential housing depreciates in 28 years yet typically lasts 2-3x as long and 4x is not rare.) Google's V1 TPUs

no, it's 10y. you just repurpose.

also, power and cooling agreements are usually 10y too.

Having dealt with many family members just like this, I can tell you they certainly can use Google for search, they just don't want to. When forced to they use it fine.

If I were to guess, a lot of people, particularly older people, see computer use outside the web browser as an amateur might a chop-saw; a dangerous tool not to be used by inexperienced hands, so an AI that can also directly do the tasks for them might solve that problem (assuming it came preinstalled or someone else did that for them). Though then comes the question of whether they'd be willing to pay for it

I'd say the exact opposite. People who have no idea how technology works shouldn't touch AI at all. "AI can understand human language" is fundamentally false, and it leads people to actually believe AI is "AI" and not just a token predictor.

These people WILL fall victim of their own ignorance and upload all their private information to an AI chatbot.

In the US, roughly 95% of AI users do not pay for it. So, they are sold, but only for the free tier. Technology is there, money are not.

Based on my experience, you’re right. I know a woman in her late 60s that is basically tech illiterate (can barely send an email), and she has been making effective use of the ChatGPT free plan. AI has been helping her cut through the noise of navigating the web in 2026 (which in her defense is often a bewildering experience). It can’t help her with everything yet, but it’s a clear boost to her independence online.

> Give your neighbors something like Alexa powered with better AI

Didn't they literally try this and it failed miserably?

"Alexa but better" is not a profitable business. Amazon profits from Alexa being in people's homes, not from people buying Alexa.

i dont know about that, they could've adopted using google which you could already search things in pretty much natural language and even got youtube snippets how fix your problem... so i'm not convinced they will adopt

There is no "honest" AI though. It is always biased towards its prompt and ultimately its training data.

Compare flights? You have to assume it is objective and not tainted by corporate bias or backroom deals(aka ads and sponsored content), unless you selfhost and find an objective model.

Search Wikipedia? You have to constantly check that it is giving you straight facts instead of being constrained by its community and safety guidelines.

The problem with so many of these usecases is treating AI as a super intelligent Vulcan who is always logical and objective aka Spock, when we are better off treating it as if it is a salesman.

Also, their neighbors are also the exact exemplars of what two generations later will look like. There isn't a lot of hard data on what AI usage might look like, but there is data on tech literacy, and the numbers are clear: We trained precisely -ONE- generation that actually understands computers. The incoming one will happily gobble up all the AI, same way they gobbled up willful ignorance of filesystems. Much like there is a button that corresponds to "gimme shot videos", there will be a button that corresponds to "Listen to what I'm about to say, and do it".

Demand will not stop in the next 20 years.

> AI can understand human language.

It is a bit comical how the HN crowd spent the last number of decades trying to convince the audience that a programmer's value was actually in their system thinking, not their arcane knowledge of a particular language, but now that the language barrier is removed it turns out it was just the arcane language knowledge after all...

Although I'd posit that the old school HNers were the ones actually right. From what I see in the average Joes is that, while they can communicate with AI, they don't really see how to leverage it. Granted, the younger generations grew up in a school system that tried to promote system thinking more so than in the past, so that gap will likely shrink over time.

Now tell them it costs $100/month and, eh, not so much.

By "understand" you must mean uncannily appear to understand while creating a plausible human readable response. Except when the knowledge remixer glitches and produces nonsense, or worse still, gaslight the user with confident misinformation. At which point they've got problems much more baffling than how to print out their tax return with the printer they haven't used in three months.

[dead]

Building the middle system that can take any model as a back end, based on cost/performance/client choices/etc, and API style present the same interface and functionality to end users is huge right now.

Local LLM usage is bouncing across major runtimes, custom runtimes and forks plus custom quants and variations of open models. llama-swap and some other tools integrate those and provide metrics but as the field expands can't keep up even with AI support.

I need to look at tailscales dashboard/etc again. I started on that when it launched but couldn't get my runtimes to go through it. I think the easiest tie in for metrics/etc is grafana + Prometheus but you need that stood up, and then a custom parsing job for each runtime based on metrics it reports and processing that data into useful dash.

I would love to see more standards here like OpenAI endpoints with metrics and information to link into. ./metrics is great but not standardized.

This is the point. hmontazeri is imagining where AI still has left to grow as a consumer product, but AI is currently more of a business product. Even without a new consumer use case, it's easy to see demand for tokens keep going up.

> But this is the major thing were ai will be introduced into every single layer of software we have.

How is that even possible? I thought they already put AI chats into everything. Maybe AI chat inside an AI chat.

Is using AI to open a zip file and not interact with the neighbor supposed to be an improvement? Maybe they just like interacting with their neighbor.

Exactly!

I said it two years ago already: An LLM is the perfect user interface.

The problem though: It will make it a lot easier to do harder things which will devaluate jobs.

having to write "hey LLM-du-jour please unpack these files!" os strictly less efficient/more time consuming than clicking "extract" on a menu. I have hard time what usecase you have in your mind. (Unless the user is blind that is)

I'm sorry but if somebody born in the 1980s cannot or will not double click on a zip file to open it they also cannot or will not ask Claude to open it for them.

[dead]

price?

The main criticism is, imo, that these models cannot think and have no logic. This is proven time and time again. Every day something new comes out, like counting Rs or the ”drive or walk“ test etc. These are the ”fixed“ or rather band-aided on a case by case basis, but nobody seems to be able to fox the underlying problem.

Having done my master‘s thesis on Transformer LMs, I think this is a fundamental limitation of the architecture.

These models can solve math problems and then fail to count the number of problems they solved. Not being able to count to 68 is not intelligence.

Saying it is a 'next token predictor' should not diminish what LLMs are capable of. It is incredible what can be achieved by a simple rule like this.

“Predicting the next token” is a bad argument, as it completely ignores all the emergent properties of transformer models. It’s like saying that discreet electrical impulses from sodium/ potassium transporters is the beginning and end of what neuroscience describes.

Next token predictor is not a very good argument. A submarine does not need to be a fish to swim.

LLMs cannot, by their core design, predict next word correctly every time. That is the main thing you are missing. Another thing is that I know that I am not predicting next word based on othr words in my "trainin data", there is much deeper understanding of most subjects that I am using to "predict next word". This may not be true for everyone, I cannot know for sure, but it might be possible that there are people out there who are just biological LLMs, and in that case I am certain those people will not be able to predict next word Einstein would say either.

“It” is not “thinking” for christ’s sake, it’s a massive statistical regression curve fit to a massive, frozen library of text. Unless you’re thinking of “thinking mode” which still is just a template that tells the model to ask questions, which are asked one token at a time, turtles all the way down.

It's a statistical machine, so you only predict a statistical Einstein, ie. something that might appear quite like Einstein and even can be fairly accurate in predicting the next word he types, but which is not him, and will be unable to predict the words that are produced not just statistically but with certain intelligence, direction and intuition.

> If you can predict the exact next word Einstein is about to write, every time, you’re as intelligent as Einstein!

Einstein is not hold in such esteem because he was really could at spewing out text or formal proofs or theorems, but because he was able to invent something no-one else could. Some people say he advanced physics 100 years wuth his theories in relativity, whether it is true or not.

So even if you can predict what Albert is going to write next, you are still merely copying old Albert. They are amazing at processing large amounts of data, naturally, but creating something new and surprising we have yet to see evidence of (the solutions of the millennium problems seem to be brute-forcing a solution based on fuckton of data to me, but I am not a mathematician so I guess it is up in the air still).

> I never understand the “only a next token predictor” argument

Yeah, that's typically the case for people who have no clue about LLMs or AI.

We have a world model.

> Is there any evidence that you yourself are more than a 'next action predictor with a multi-modal context window'?

I love the age of LLMs because people will post stuff like “the fact that I absolutely, unqualifiedly, fundamentally do not understand two separate things indicates that I understand at least one and maybe even both of them more than you do“ and then do that cool move where they pretend they were gonna high five you but don’t

it is a golden age of Posting

A theoretically perfect next token prediction machine is just an LLM with a temperature of 0, which performs worse than one with a non-zero temperature (i.e. imperfection)

The "prediction" has nothing to do with predicting the future, it just means choosing the most likely next token for existing human text

A theoretically perfect next token prediction machine would be uncomputable.

This is not how LLMs work.

In 2007, It was regular people who lost their homes and their leveraged investment properties. Regular people are not leveraged into AI startups today. If the AI bubble pops, regular feel people will feel the direct financial consequences significantly less than the '07 housing crisis.

And the valuations made no sense in 10 yea....oh.

You make it sound reasonable then.

That $400B for video and phones though. That's like a $1k average for every adult and child in the country. It's twice what I spend, and I'm not on a family plan.

I more than agree on the value, but not on the price. There's currently a competitive market, including free models that are pretty impressive if not "frontier".

You mean they hope to be the next Facebook/Google that knows everything about you and call sell that data to advertisers? Yay :-|

Yeah, we need a lot more kinds compilers like LEAN. Give the accountants, physicists, chemists, biologists, and so on the tools where they can verify the results.

> it would have sounded ludicrous to suggest that soon, practically every single software application would have some networking component

Still does.

Open a restaurant. Is it a McDonalds? Yes -> you instantly have lots of customers No -> you don't have lots of customers

People know McDonald's. They don't know your random new chain.

This is called a moat.

An economy of scale, which McDonalds surely has as GP suggested in various areas, supports durable profits. Seems like a moat to me.

Land is the classic cornered resource. Again a source of durable profits. You can’t open up a franchise restaurant on land owned by McDonalds; the profits from traffic earned in that location are not accessible to you.

CUDA is not a moat because it happens to be unique. CUDA provides a moat for NVIDIA due to the high switching costs and network effects it provides.

McDonalds to my knowledge does not have any network effects dynamics going on however they have entrenched switching cost mechanisms in the form of loyalty rewards programs.

If you think McDonalds has no moat you’re dreaming.

MCD has a huge real estate moat.

Is there evidence they are unambiguously making profit off inference when all the supply costs of that inference are considered? It does not seem likely.

> I don't think you read the Anthropic filing. It is substantially profitable if you exclude the one type of debt which was actually equity.

Anthropic filling shows negative EBITDA, how is that profitable? Revenue < (computing + personal + selling expenses). That is not a business, that is a money burning furnace.

That before even taking into account the huge amount of money they promised to datacenters (i.e. "build that for us, and we would rent it for 5 years", non cancellable).

Did you read the amended filings with Q3? Because they are not profitable in Q3 either. And their debt due is about to increase substantially. This is all before they inevitably have to issue billions of dollars in bonds with insane yields. 450 billion in commits due over the next 5 years or so.

I use Claude a lot at work and Deepseek for personal stuff, and they feel very comparable to me. I haven't found a task using Deepseek that makes me wish I had Claude - especially given how expensive Claude is.

A lot of people use Opus more than Fable, so yeah, a "good enough model" definitely exist.

This is important. I think people would pay a large premium for a marginal increase in intelligence. Primarily because intelligence is correlated with exponential utility which is often ignored.

For example, I'd much rather pay for the highest intelligence model when designing something or developing something because the problems it can preempt will save me more money in the long run.

The whole issue is that intelligence (as measured in AA benchmarks for example) is not linear utility

It also assumes Anthropic can maintain the lead. If anything the competition both in the USA and in China seems to be getting closer and closer to parity with Claude.

> This remains to be seen, though, and people's behaviour paying for Claude Code and Codex suggests at least for now people are willing to pay.

I feel like this will depend on the user.

McDonald has a giant moat: Consistency

I can pop into any McDonald across the world and be 99.9% sure that the meal will be clean and taste like what I except McDonald to taste.

I don’t love fast food but I’ve been to a lot of them, across a wide enough range of country , to say that every other fast food chain struggles hard with that.

Corals has no moat. The foo business is notoriously competitive with thin margins.

If that’s what Ai looks like they are in trouble.

It means that the companies have to be highly cost-competitive, they can't win market share and then raise margins. If (as is the case today) astra is bad value for money compared to opus-5-5, then I'm switching, and if (as is the case today) for lots of stuff gemini can do the low hanging fruit, I'm switching. And all of this is happening before my company has really put any boundaries on our spend, in the future I'm going to be on a budget and that's really going to impact my decisions.

Yes, this reasoning has been used to justify the absolutely gigantic amount of spending going into AI right now.

I'm not convinced it's true. It just seems to be taken for granted that demand for intelligence is unlimited and that its cost limited, which just doesn't seem true to me.

The trouble that that idea is that there comes a point where the generated AI needs machines to test the code that is generated... and it actually puts a limit on the growth of AI.

You mention that training can be suspended, which is true, but don't acknowledge that puts them behind in the arms race against both other frontier companies as well as second line cheaper providers.

Without a moat, training is an ongoing cost that they can't afford to not pay

Training is not infrastructure. It’s not capex, not “investment”. It’s really an ordinary running cost of what they do, for the foreseeable future, and it’s absolutely enormous. They couldn't even dial back training in a winner-takes-it-all scenario, because training of smaller, domain specific models would remain a way to compete.

Not if they've bought the land, started/finished building them and bought the GPUs already, but just can't get power. See the case of Oracle.

Some people have a weird bias against anime profile pictures. It's mostly ageism.

Sure, but we are a lot smarter now aren't we? So everyone can prepare accordingly.

But whenever people talk about AI Bubble, it sounds like they expect AI to stop working and impact getting reversed.

Yeh - same. But I read “other” as “leisure things”. It’s not like we’re all having a nice time elsewhere, we’ve just compounded the amount of stuff we can deliver so we deliver even more. It’s the polar opposite of the dream where the tech does the nasty shit so we can all go chill.

I could only find these numbers in articles on something called crypto briefing dot com, not sure how the author got this number from the S-1.

When using traditional business metrics, such as operational net income, we'll see that Anthropic has been in the red for the entirety of its history, at times to the tune of $3B per quarter, except in Q2 2026.

When looking at amended S-1 (with Q3 numbers), page 23, "Consolidated Statements of Operational Data", we can see that for 2026 YTD (Q1 - Q3) net loss was $103B; operating loss was $1,260B.

Anthropic also disclosed up to $895B (multiyear) commitments for infrastructure build-up.

Basically, the company's financial results do not allow for any reasonable valuation (which would be more than 0), leave alone $2T or even $1T. If this is not an indication of a bubble I don't know what is.

That’s 52% gross margin. Not much higher than the overall S&P 500. If you break it down by sector, Real Estate, Health are, Financials (and of course Tech overall) have higher gross margins at the sector level.

That $1 of compute is not taking into account R&D spending, infrastructure, etc.

And that's their only revenue stream

You assume these companies telegraph their motives, they could just as easily be feeding this narrative so they can buy more of the company at a lower price.

This should be a top level comment

Or it's like no one wanted to believe Ed Zitron, who has been telling everyone everything they need to know about this real-life, actual AI bubble....

> If the hyperscalers stopped building new infra tomorrow, they would just roll in the profits available from what exists today!

They owe hundreds of billions in loans on datacenters that aren't built yet and don't currently have capacity for the new users and additional tokens they promise are coming. If they don't build capacity the increase of demand would cause more outages and higher prices which would crater both their use and valuation. They have to grow or die.

> They intend to offer about $100B in stock, their valuation is at 1T

That's one company, the aggregate debt of the US AI industry approaches 1T.

> Most AI companies are profitable aside from expenditures to build new infrastructure

Which they have to complete to make the profit they promised which they need to pay off the loans they already have.

> A good GPU to run a reasonable model is $10k+

Today. Due to supply and demand. Funny thing, supply and demand. It's very reliable. High demand makes supply short for a while, but then causes supply to go up. When supply goes up, eventually demand goes down, which brings prices down. The chip factories and economies of scale that will be complete in a few years will send prices crashing, as they have before. Which might bankrupt the remaining chip companies.

"Reasonable model" isn't a thing. Many different models of different size have capacity to do many tasks. I have a laptop from 3 years ago that runs local AI models. My phone from 5 years ago runs local AI models. And any strong CPU can run models (no GPU needed). Speech to text, image/video generation, image classification, research, analysis, computer use, troubleshooting, etc, all works on tiny models. I think you don't understand the tech here. Remember that many people spend a grand every year just for a newer iPhone/Samsung. They'll spend the same for one with better AI and more RAM.

Here [1] are some various benchmarks for a RTX 4090 on models like Qwen 3.8, where it usually hits around 43 tokens/s. You can find similar vids on YouTube showing it live for comparison, 'proof' if you doubt, etc. And yes, it is frontier tier.

[1] - https://getrigbase.com/gpus/rtx-4090/