There’s one key difference in my opinion: pre-.com deals were buying revenue with equity and nothing else. It was growth for growth’s sake. All that scale delivered mostly nothing.

OpenAI applies the same strategy, but they’re using their equity to buy compute that is critical to improving their core technology. It’s circular, but more like a flywheel and less like a merry-go-round. I have some faith it could go another way.

> they’re using their equity to buy compute that is critical to improving their core technology

But we know that growth in the models is not exponential, its much closer to logarithmic. So they spend =equity to get >results.

The ad spend was a merry go round, this is a flywheel where the turning grinds its gears until its a smooth burr. The math of the rising stock prices only begins to make sense if there is a possible breakthrough that changes the flywheel into a rocket, but as it stands its running a lemonade stand where you reinvest profits into lemons that give out less juice

OpenAI invests heavily into integration with other products. If model development stalls they just need to be not worse than other stalled models while taking advantage of brand recognition and momentum to stay ahead in other areas.

In that sense it makes sense to keep spending billions even f model development is nearing diminishing return - it forces competition to do the same and in that game victory belongs to the guy with deeper pockets.

Investors know that, too. A lot of startup business is a popularity contents - number one is more attractive for the sheer fact of being number one. If you’re a very rational investor and don’t believe in the product you still have to play this game because others are playing it, making it true. The vortex will not stop unless limited partners start pushing back.

The bigger threat is if their models "stall", while a new up-start discovers an even better model/training method.

What _could_ prevent this from happening is the lack of available data today - everybody and their dog is trying to keep crawlers off, or make sure their data is no longer "safe"/"easy" to be used to train with.

They can also buy out the startup or match the development by hiring more people. Their comp packages are very competitive.

But, if model development stalls, and everyone else is stalled as well, then what happens to turn the current wildly-unprofitable industry into something that "it makes sense to keep spending billions" on?

I suspect if model development stalls we may start to see more incremental releases to models, perhaps with specific fixes or improvements, updates to a certain cutoff date, etc. So less fanfare, but still some progress. Worth spending billions on? Probably not, but the next best avenue would be to continue developing deeper and deeper LLM integrations to stay relevant and in the news.

The new OpenAI browser integration would be an example. Mostly the same model, but with a whole new channel of potential customers and lock in.

If model development stalls, then the open weight free models will eventually totally catch up. The model itself will become a complete commodity.

It very well might. The ones with most smooth integrations and applications will win.

This can go either way. For databases open source integration tools prevailed, the commercial activity left hosting those tools.

But enterprise software integration that might end up mostly proprietary.

Because they’re not that wildly unprofitable. Yes, obviously the companies spend a ton of money on training, but several have said that each model is independently “profitable” - the income from selling access to the model has overcome the costs of training it. It’s just that revenues haven’t overcome the cost of training the next one, which gets bigger every time.

> the income from selling access to the model has overcome the costs of training it.

Citation needed. This is completely untrue AFAIK. They've claimed that inference is profitable, but not that they are making a profit when training costs are included.

I've also seen Open AI and Anthropic say it's pretty close at least. I'll try to follow up with a source.

There is something about an argument made almost entirely out of metaphors that amuses me to the point of not being able to take it seriously, even if I actually agree with it.

As much as I dislike metaphors, this sounded reasonable to me. Just don't go poking holes in the metaphor instead of the real argument.

Indeed, poking holes in the metaphor is like putting a pin in a balloon, rather than knocking it out of the park by addressing the real argument.

Yeah, except you can keep on squeezing these lemons for a long time before they run out of juice.

Even if the model training part becomes less worthwhile, you can still use the data centers for serving API calls from customers.

The models are already useful for many applications, and they are being integrated into more business and consumer products every day.

Adoption is what will turn the flywheel into a rocket.

Well, the thing is that that kind of hardware chips quickly decrease in value. It's not like the billions spend in past bubbles like the 2000s where internet infrastructure was build (copper, fibre) or even during 1950s where transport infrastructure (roads) were build.

Data centers are massive infrastructural investments similar to roads and rails. They are not just a bunch of chips duct taped together, but large buildings with huge power and networking requirements.

Power companies are even constructing or recommissioning power plants specifically to meet the needs of these data centers.

All of these investments have significant benefits over a long period of time. You can keep on upgrading GPUs as needed once you have the data center built.

They are clearly quite profitable as well, even if the chips inside are quickly depreciating assets. AWS and Azure make massive profits for Amazon and Microsoft.

There's at least one contributor here on HN that believes growth in models is strictly exponential: https://www.julian.ac/blog/2025/09/27/failing-to-understand-...

I think that, at best, that description boils down to Nvidia, Oracle, etc inventing fake wealth to build something and OpenAI building their own fake wealth by getting to use that new compute effectively for free.

There are physical products involved, but the situation otherwise feels very similar to ads prior to dotcom.

The same way the stock market invents a trillion dollars of fake wealth on a strong up day?

That's capital markets working as intended. It's not necessarily doomed to end in a fiery crash, although corrections along the way are a natural part of the process.

It seems very bubbly to me, but not dotcom level bubbly. Not yet anyway. Maybe we're in 1998 right now.

The stock market isn't inventing money. Those investing in the stock market might be, those buying on leverage for example.

Capital markets weren't intended for round trip schemes. If a company on paper hands 100B to another company who gives it back to the first company, that money never existed and that is capital markets being defrauded rather than working as expected.

I think it's worse. The US market feels like a casino to me right now and grift is at an all time high. We're not getting good economic data, it's super unpredictable, and private equity is a disaster waiting to happen IMO. For sure there are smart people able to make money on the gamble, but it's not my jam.

I don't tend to benefit from my predictions as things always take longer to unfold than I think they will, but I'm beyond bearish at present. I'd rather play blackjack.

More money is lost by bears fighting a bull market, than in actual bear market crashes.

I’ve made that mistake already.

I’m nervous about the economic data and the sky high valuations, but I’ll invest with the trend until the trend changes.

> It seems very bubbly to me, but not dotcom level bubbly.

Not? Money is thrown after people without really looking at the details, just trying to get in on the hype train? That's exactly how the dotcom bubble felt like.

Nvidia has a trailing PE of 50. Cisco was 200 At the height of the dotcom bubble.

Nowhere near that level. There’s real demand and real revenue this time.

It won’t grow as fast as investors expect, which makes it a bubble if I’m right about that. But not comparable to the dotcom bubble. Not yet anyway.

We shouldn't judge whether an indicator is stable or okay only by looking to see if its the highest historical value.

PE ratios of 50 make no sense, there is no justification for such a ratio. At best we can ignore the ratio and say PE ratios are only useful in certain situations and this isn't one of them.

Imagine if we applied similar logic to other potential concerns. Is a genocide of 500,000 people okay because others have done drastically more?

I’m not asking if it makes sense, I’m simply pointing out that by that measure this is much less extreme than 2000. As I stated, I think we’re in a bubble, so valuations won’t make much sense.

If you have a better measure, share it. I trust data more than your or my feelings on the matter.

Unless you have evidence that this measure of yours is a reliable predictor of how big a bubble is, it's on par with my gut feeling.

I sell you a cat for $1B and you sell me a dog for $1B and now we’re both billionaires! Whether the capital markets “want” that or not it’s still silly.

If we’re both willing to pay that in a free market economy, then we both leave the deal happy.

Things are worth what people are willing to pay for them. And that can change over time.

Sentiment matters more than fundamental value in the short term.

Long term, on a timescale of a decade or more, it’s different.

> If we’re both willing to pay that in a free market economy

The thing is: you've paid nothing - all you did was trade pets and played an accounting trick to make them seem more valuable than they are.

Is that not fraud?

Yes, it is fraud round tripping is fraud, whether the government is willing to prosecute it or not.

Both parties would need the $1B prior to the transaction for it to even potentially be meaningful, and still they just traded a cat for a dog and only paid each other on paper.

That ultimately wouldn't be a big deal if the paper valuation from the trade didn't matter. As it stands, though, both parties could log it as both revenue and expenses, and being public companies their valuation, and debt they can borrow against it, is based in part on revenue numbers. If the number was meaningless who cares, but the numbers aren't meaningless and at such a scale they can impact the entire economy.

No, only one party needed a billion dollars, they just needed to execute that trade first.

Actually, thinking more about it, if they both agreed the cat and dog were worth a billion dollars ahead of time, neither one would actually need that billion dollars as liquid cash, it could be an all-animal transaction.

Eventually when ChatGPT replaces Google Search, they will run ads, and so have that whole revenue stream. Still isn't enough money to buy the trillions worth of infrastructure they want, but it might be enough to keep the lights on.

That's an insightful point! Making insightful points like that one is taxing on the brain, you should consider an electolyte drink like Brawndo™ (it's got what plants crave) to keep yourself sharp!

Ugh I hate it so much, but you're right, it's coming.

One thing I've been contemplating lately is that from a business perspective, when your competitors expand their revenue avenues (generally through ads) you have three options: copy them to catch up, do nothing and perish, and lobby the government for increased consumer protections.

I've started to wonder why we see so few companies do this. It's always "evil company lobbying to harm the its customers and the nation." Companies are made up of people, and for myself, if I was at a company I would be pushing to lobby on behalf of consumers to be able to keep a moral center and sleep at night. I am strongly for making money, but there are certain things I am not willing to do for it.

Targeted advertising is one of these things that I believe deserves to fully die. I have nothing against general analytics, nor gathering data about trends etc, but stalking every single person on the internet 24/7 is something people are put in jail for if they do it in person.

Why would ChatGPT replace google search when search also has AI? At best they'd steal some of Google's market share, which I'd imagine would decline with embedded ads.

Two reasons:

1) Google Search is now 99% crap that nobody wants, and even the AI answers are largely crap,

2) I believe somebody is going to eventually realize that search engines are stupid and improve on them. The whole idea of a single text box where you type some words and the search engine reads your mind to figure out the one thing you wanted, and then gives you one generic answer, is crap. We've just been blind to this because we don't see any other answer to realize we've been getting crap.

If I type in "when did MMS come out", Google will tell me when the candy product M&M's came out. But I wanted to know when the Multimedia Messaging Service was released. At some point somebody is going to realize that you can't actually tell what the hell the person wants from these simple queries alone. The computer needs to ask you questions to narrow down the field. That's sometimes what happens in ChatGPT, but it can be greatly improved with simple buttons/drop-downs/filters/etc. I think it'll also be improved by more dynamic and continuous voice input for context. (I notice Google Search now has audio input; I wonder if that came in after ChatGPT? Wayback Machine shows it starting in mid-2024) When they eventually implement all this, and people realize it's a million times better than what Google has, then Google will be playing catch-up.

Dotcom scams included "vendor financing", where telecom equipment providers invested in their customers who built infrastructure:

https://time.com/archive/6931645/how-the-once-luminous-lucen...

The customers bought real equipment that was claimed to be required for the "exponential growth" of the Internet. It is very much like building data centers.

The assumption is that they have a large moat.

If they don't then they're spending a ton of money to level up models and tech now, but others will eventually catch up and their margins will vanish.

This will be true if (as I believe) AI will plateau as we run out of training data. As this happens, CPU process improvements and increased competition in the AI chip / GPU space will make it progressively cheaper to train and run large models. Eventually the cost of making models equivalent in power to OpenAI's models drops geometrically to the point that many organizations can do it... maybe even eventually groups of individuals with crowdfunding.

OpenAI's current big spending is helping bootstrap this by creating huge demand for silicon, and that is deflationary in terms of the cost of compute. The more money gets dumped into making faster cheaper AI chips the cheaper it gets for someone else to train GPT-5+ competitors.

The question is whether there is a network effect moat similar to the strong network effect moats around OSes, social media, and platforms. I'm not convinced this will be the case with AI because AI is good at dealing with imprecision. Switching out OpenAI for Anthropic or Mistral or Google or an open model hosted on commodity cloud is potentially quite easy because you can just prompt the other model to behave the same way... assuming it's similar in power.

> This will be true if (as I believe) AI will plateau as we run out of training data.

Why would they run out of training data? They needed external data to bootstrap, now it's going directly to them through chatgpt or codex.

As much ChatGPT says I’m basically a genius for asking it a good Vegan cake recipes, I don’t think that is providing it any data it doesn’t already have that makes it anyway better. Also at this point the massive increases in data and computing power seem to bring ever decreasing improvements (and sometimes just decline), so it seems we are simply hitting a limit this kind of architecture can achieve no matter what you throw at it.

ChatGPT chat logs contain massive amount of data teased out of people’s brains. But much of it is lore, biases, misconceptions, memes. There are nuggets of gold in there but it’s not at all clear if there’s a good way to extract them, and until then chat logs will make things worse, not better.

I’m thinking they eventually figure out who is the source of good data for a given domain, maybe.

Even if that is solved, models are terrible at long tail.

When I say models will plateau I don't mean there will be no progress. I mean progress will slow down since we'll be scraping the bottom of the barrel for training data. We might never quite run out but once we've sampled every novel, web site, scientific paper, chat log, broadcast transcript, and so on, we've exhausted the rich sources for easy gains.

Chat logs don’t run out. We may run out of novelty in those logs, at which point we may have ran out of human knowledge.

Or not - there still knowledge in people heads that is not bleeding into ai chat.

One implication here is that chats will morph to elicit more conversation to keep mining that mine. Which may lead to the need to enrage users to keep engagement.

The necessity of higher quality data from vetted experts is why Mercor just raised at 10B

I’mafraid I don’t share your optimism. I think we are more or less seeing the limitations of the transformer architecture.

Apple new M5 can run models over 10B parametres and if they give their new Studio next year enough juice, it can run maybe 30B local model. How long is it that you can run a full GPT-5 on your laptop or homeserver with few grands worth of hardware? What is going to happen to all these GPU farms, since as I understood they are fairly useless for anything else?

Very few people own top of the line Macs and most interactions are on phones these days. We are many generations of phones away from running GPT-5 on a phone without murdering your battery.

Even if that weren't true having your software be cheaper to run is not a bad thing. It makes the software more valuable in the long run.

Quantized, a top-end Mac can run models up to about 200B (with 128GiB of unified RAM). They'll run a little slow but they're usable.

This is a pricey machine though. But 5-10 years from now I can imagine a mid-range machine running 200-400B models at a usable speed.

They are pretty cheap compared to _actual_ costs of GPU farms, or buying A100 though. Of course not everybody will buy these machines, but everybody don’t really need high powered LLM’s either. Prob 13B Mistral can be trained to do your homework and pretend to he your girlfriend.

> critical to improving their core technology

It is at the very least highly debatable how much their core technology is improving from generation to generation despite the ballooning costs.

>they’re using their equity to buy compute that is critical to improving their core technology

That's only like 1/8th of a flywheel, though.

Wasn’t there also a bunch of telecom infrastructure created in the dot-com bubble, tangible products created, etc? Things like servers, telephone wires, underwater internet cables, tech-storefronts, internet satellites, etc.

so much fiber was run that in the US over 90% of it wasn't even used

> There’s one key difference in my opinion

The other difference (besides Sam's deal making ability) is, willing investors: Nvidia's stock rally leaves it with a LOT of room to fund big bets right now. While in Oracle's case, they probably see GenAI as a way to go big in the Enterprise Cloud business.

> Nvidia's stock rally leaves it with a LOT of room to fund big bets right now

And then what happens if the stock collapses?

Hence the emphasis on right now.

> I have some faith it could go another way.

I wonder how they felt during the .com era.

Yes, this time is different, trust big bro sama.

> OpenAI applies the same strategy, but they’re using their equity to buy compute that is critical to improving their core technology. It’s circular, but more like a flywheel and less like a merry-go-round. I have some faith it could go another way.

I'm commenting here in case a large crash occurs, to have a nice relic of the zeitgeist of the time.

Happy to have provided. I’m not an AI bull and not in any way invested in the U.S. economy besides a little money in funds, but I do try to think about the war of today vs the war of yesterday. Hopefully that’s always en vogue.

Gita Gopinath (imf’s economist) sounded the alarm on the scale of this - https://www.afr.com/wealth/investing/the-crash-that-could-to...

Stopped clock…

2020: https://www.youtube.com/watch?v=rpiZ0DkHeGE 2019: https://www.cadtm.org/spip.php?page=imprimer&id_article=1732...

Reminds me of Nassim Taleb who got one call right and then his career became consistently incorrect bubble predictions

Circular investments were also a compounding factor in the Japanese asset price bubble.

The practice was known as “zaitech”

> zaitech - financial engineering

> In 1984, Japan’s Ministry of Finance permitted companies to operate special accounts for their shareholdings, known as tokkin accounts. These accounts allowed companies to trade securities without paying capital gains tax on their profits.

> At the same time, Japanese companies were allowed to access the Eurobond market in London. Companies issued warrant bonds, a combination of traditional corporate bonds with an option (the “warrant") to purchase shares in the company at a specified price before expiry. Since Japanese shares were rising, the warrants became more valuable, allowing companies to issue bonds with low-interest payments.

> The companies, in turn, placed the money they raised into their tokkin accounts that invested in the stock market. Note the circularity: companies raised money by selling warrants that relied on increasing stock prices, which was used to buy more shares, thus increasing their gains from investing in the stock market.

https://www.capitalmind.in/insights/lost-decades-japan-1980s...

And I guess no one of the people who were doing that paid but the community paid the price for this scam ?

Here is a charitable perspective on what's happening:

- Nvidia has too much cash because of massive profits and has nowhere to reinvest them internally.

- Nvidia instead invests in other companies that use their gpus by providing them deals that must be spent on nvidia products.

- This accelerates the growth of these companies, drives further lock in to nvidia's platform, and gives nvidia an equity stake in these companies.

- Since growth for these companies is accelerated, future revenue will be brought forward for nvidia and since these investments must be spent on nvidia gpus it drives further lock in to their platform.

- Nvidia also benefits from growth due to the equity they own.

This is all dependent on token economics being or becoming profitable. Everything seems to indicate that once the models are trained, they are extremely profitable and that training is the big money drain. If these models become massively profitable (or at least break even) then I don't see how this doesn't benefit Nvidia massively.

Yup. Not just Nvidia. Just look at the quarterly results reported by Amazon, Google, Meta, Microsoft and Apple. Each one is reporting revenues never before seen in history. If you make 100 Billion a quarter you have to spend it on something.

These guys are running hyper optimized cash extraction mega machines. There is no comparison to previous bubbles, cause so no such companies ever existed in the past.

100 billion a quarter is Alphabet, right? Given how much click fraud there is, and that every org and business under the sun is held to ransom to feature on the SERP for their own name even — it’s tempting to say Google’s become a private tax on everything.

No, Apple also has 100 billion dollars in revenue despite floundering AI and running a very hardware dependent business.

> Given how much click fraud there is [etc]

It's easy for the techies to see the problems. But advertising results have been very measurable for a very long time by now. Larger advertisers can leave the details to their techies and still be very clear as to their advertising's productivity post-cost of doing business.

Odd how they are simultaneously having large layoffs even as reporting record revenues.

The question is where the profits are.

The layoffs at Amazon and Microsoft are not due to lack of profits. They’re massively profitable right now.

https://www.macrotrends.net/stocks/charts/MSFT/microsoft/ebi...

https://www.macrotrends.net/stocks/charts/AMZN/amazon/ebitda

They're "massively profitable" because they're laying off large portions of a major cost center - labor - and backloading uncoming data center construction costs. As those come due, and labor needs rise again, that profit disappears.

yes, these companies have a track record of disappearing profits :)

They have a track record of cornering a market and abusing their position, and also still somehow not being able to balance expenses and revenues to turn a profit that pleases shareholders. You get to decide if that's a problem with the company or the shareholders, I guess.

Amazon - 14,000 layoffs; significant

Microsoft - 14,000 (multiple rounds); significant

Meta - 600 layoffs; insignificant for company size

Google - "Several hundred layoffs"; insignificant for a company size

Apple - No layoffs

Source: https://techcrunch.com/2025/10/24/tech-layoffs-2025-list/

Google's Youtube unit is doing soft layoffs.

https://news.ycombinator.com/item?id=45766368

Also every one of them has hundreds of thousands of external contractors which are not reported anywhere.

And offshoring is also a huge cost-cutting effort everywhere.

So many such profitable companies are the best possible evidence for the need for drastic antitrust intervention. The lack of competition and regulation is leading to a massive drain on every other sector.

This bubble is caused by excess competition. There are 4 large companies who believe that a large new market is being created so each is investing large amounts without any evidence that there will be a single winner that dominates the future market. None of these companies has anything remotely resembling a monopoly except for Amazon in online retail.

Google: search, chrome, youtube

Microsoft: desktop software

Meta: social media

Maybe on some technical definitions of "monopoly" these aren't monopolies, but nothing remotely resembling a monopoly? come on maan

Monopolies being bad for free markets is a simplification. Substantial control over the market would be more accurate description of the issue.

How much more was worth USD at the beginning of the year?

What's shocking is the gulf between those companies and corporate 'normality'.

Eastern Airways, a UK airline, has just gone bust due to accumulated debts of £26 million. That's not even a rounding error for Google, yet was enough to put a 47-year-old company into bankruptcy and its staff out of work.

I think the only historical parallel to this disparity was the era of the East India Company.

> Everything seems to indicate that once the models are trained, they are extremely profitable

Some data would reinforce your case. Do you have it?

Here is my data point: "You Have No Idea How Screwed OpenAI Actually Is" - https://wlockett.medium.com/you-have-no-idea-how-screwed-ope...

Right. As far as I can tell, OpenAI, Grok, etc sell me tokens at a loss. But I am having a hard time figuring out how to turn tokens into money (i.e. increased productivity). I can justify $40-$200 per developer per month on tokens but not more than that.

There’s about 5M software devs in the US so even at $1000/year/person spend, that’s only $5B of revenue to go around. Theres plenty of other uses cases but focusing on pure tech usage, it’s hard to see how the net present value of that equates to multiple trillions of dollars across the ecosystem.

It's the first new way of interacting with computers since the iPhone. It's going to be massively valuable and OpenAI is essentially guaranteed to be one of the players.

Why is their product not palm? Or windows mobile?

It's not windows mobile because OpenAI was first and is the clear leader in the market. Windows mobile was late to the party and missed their window.

Palm is closer but it's a different world. It's established that Internet advertising companies are worth trillions. It's only in retrospect that what Palm could have been is obvious.

Barring something very unexpected OpenAI is coming out on top. They're prepaying for a good 5-10 years of compute. That means their inference and training for that time are "free" because they've been paid for. They're going to be able to bury their competition in money or buy them out.

Windows mobile by the time it looked like the iPhone was late to the party. But windows had been releasing a mobile os for a long time before that. Microsoft was first, they just didn’t make as good of a product as Apple despite their money.

OpenAI is also first, but it is absolutely not a given that they are the Apple in this situation. Microsoft too had money to bury the competition, they even staged a fake funeral when they shipped windows phone 7.

> Barring something very unexpected

Like the release of an iPhone?

Yep. It would have to be something that dramatic to render all the technology and infrastructure OpenAI has obsolete. But if it's anything like massive data training on a huge number of GPUs then OpenAI is one of the winners.

I'm waiting for my Google Glass smart glasses to be useful for anything other then annihilating the privacy of everyone around me

Blackberry was a big deal for a while, too

In 2006, I foresaw that the smartphone was going to exist and be significant. But at the time I carried a Sharp Zaurus which just identified me as a gigantic nerd, and the clearest company to invest in was... HP/Compaq for the iPAQ, in terms of the most forward-looking device.

Then Apple came out with the iPhone but it was not clear they were going to be the market leader. They were certainly not the first mover.

With OpenAI, I see even less of a moat. Someone else comes along, makes a better foundational model, and they've got a gigantic advantage. What does OpenAI have?

> Theres plenty of other uses cases

This is where the money is. Anthropic just released claude for excel. If it replaces half of the spreadsheet pushers in the country theyre looking at massive revenue. They just started with coding because theres so much training data and the employees know a lot about coding

I'm not trying to be annoying, but surely if you'd justify spending $200/developer/month, you could afford $250/month...

The reason I wonder about that is because that also seems to be the dynamic with all these deals and valuations. Surely if OpenAI would pay $30 billion on data centers, they could pay $40 billion, right? I'm not exactly sure where the price escalations actually top out.

No? That's a 25% expense increase. You just ate the margins on my product/service, and then some.

The equivalent of an additional $50 uber ride to the airport once a month can tank your business?

Per seat. And it's not $50 more thane baseline, it's $250. The $200 may have been justifiable where the $250 is not.

why would they sell you at a loss when they have been decreasing prices by 2x every year or so for the last 3 years? people wanted to purchase the product at price "X" in 2023 and now the same product costs X costs 10 times less over the years.. do you think they were always selling at a loss?

Inference cost has been going down for a while now. At what point do you think it will be profitable? When cost goes down by 2x? 5x?

This is behind a paywall. Is there a free link you can share ?

Would love to, and its normally what I do, but archive.is is currently down. At least here from the outer belt.

I can't read your hyperbolically titled paywalled medium post, so idk if it has data I'm not aware of or is just rehashing the same stats about OpenAI & co currently losing money (mostly due to training and free users) but here's a non paywalled blog post that I personally found convincing: https://www.snellman.net/blog/archive/2025-06-02-llms-are-ch...

The above article is not convincing at all.

Nothing on infra costs, hardware throughput + capacity (accounting for hidden tokens) & depreciation, just a blind faith that pricing by providers "covers all costs and more". Naive estimate of 1000 tokens per search using some simplistic queries, exactly the kind of usage you don't need or want an LLM for. LLMs excel in complex queries with complex and long output. Doesn't account at all for chain-of-thought (hidden tokens) that count as output tokens by the providers but are not present in the output (surprise).

Completely skips the fact the vast majority of paid LLM users use fixed subscription pricing precisely because the API pay-per-use version would be multiples more expensive and therefore not economical.

Moving on.

> Nothing on infra costs, hardware throughput + capacity (accounting for hidden tokens) & depreciation

That's because it's coming at things from the other end: since we can't be sure exactly what companies are doing, we're just going to look at the actual market incentives and pricing available and try to work backwards from there. And to be fair, it also cites, for instance, deepseek's paper where they talk about what their power foot margins are on inference.

> just a blind faith that pricing by providers "covers all costs and more".

It's not blind faith. I think they make a really good argument for why the pricing by providers almost certainly does cover all the costs and more. Again, including citing white papers by some of those providers.

> Naive estimate of 1000 tokens per search using some simplistic queries, exactly the kind of usage you don't need or want an LLM for.

Those token estimates were for comparing to search pricing to establish whether — relative to other things on the market — LLMs were expensive, so obviously they wanted to choose something where the domain is similar to search. That wasn't for determining whether inference was profitable or not in itself, and has absolutely no bearing on that.

> Doesn't account at all for chain-of-thought (hidden tokens) that count as output tokens by the providers but are not present in the output (surprise).

Most open-source providers provide thinking tokens in the output. Just separated by some tokens so that UI and agent software can separate it out if they want to. I believe the number of thinking tokens that Claude and GPT-5 use can be known as well: https://www.augmentcode.com/blog/developers-are-choosing-old... typically, chain of thought tokens are also factored into API pricing in terms of what tokens you're charged for. So I have no idea what this point is supposed to mean.

> Completely skips the fact the vast majority of paid LLM users use fixed subscription pricing precisely because the API pay-per-use version would be multiples more expensive and therefore not economical.

That doesn't mean that selling inference by subscription isn't profitable either! This is a common misunderstanding of how subscriptions work. With these AI inference subscriptions, your usage is capped to ensure that the company doesn't lose too much money on you. And then the goal is with the subscriptions that most people who have a subscription will end up on average using less inference than they paid for in order to pay for those who use more so that it will equal out. And that's assuming that the upper limit on the subscription usage is actually more costly than the subscription being paid itself, and that's a pretty big assumption.

If you want something that factors in subscriptions and also does the sort of first principles analysis you want, this is a good article:

https://martinalderson.com/posts/are-openai-and-anthropic-re...

And in my opinion it seems pretty clear that basically everyone who does any kind of analysis whether black box or first principles on this comes to the conclusion that you can very easily make money on inference. The only people coming to any other conclusion are those that just look at the finances of U.S. AI companies and draw conclusions on that without doing any kind of more detailed breakdown — exactly like the article you linked me, which now I have finally been able to read, thanks to someone posting the archive link, which isn't actually making any kind of case about the subscription or unit economics of token inference whatsoever, but is instead just basing its case on the massive overinvestment of specifically open AI into gigantic hyperscale data centers, which is unrelated to the specific economics of AI itself.

Your conclusion about training being the cost factor that will eventually align with profitability in the inference phases relies on training new models not being an endless arms race.

If the inference is profitable and training new models is actually an endless arms race that's actually the best outcome for nvidia specifically.

Only in the short term.

> Nvidia has too much cash because of massive profits and has nowhere to reinvest them internally.

Here's an idea: they could make actual GPUs used for games affordable again, and not have Jensen Huang lie on stage about their performance to justify their astronomical prices. Sure, companies might want to buy them for ML/AI and crash the market again but I'm sure a company of their caliber could solve that if they _really_ wanted to.

I also just don’t understand, as someone with no business experience, how they aren’t just pouring all of that money into enhancing their production capacity. That’s very clearly their bottleneck here.

Yes, I’m certain they are spending an astronomical amount on that already, but why not more? Surely paying more money for construction of more facilities still nets gain even if you run into diminishing returns?

Instead they set up this whacko tax laundering scheme? Just seems like more corporate pocket filling to me, an idiot with no business knowledge.

The bottleneck is TSMC, who also make chips for almost every other hardware vendor.

TSMC is indeed increasing their production capability as fast as possible, but it's not easy... chip foundries are extremely expensive, complex, and take serious expertise to operate.

It’s called seeding the market. If they can accelerate the growth of potential customers, it will be more profitable than just increasing production to serve existing customers.

Think of exponential growth — would you rather increase the base or the exponent?

Hedging their bets against a potential sudden downturn in consumption of their product, e.g., an AI bubble exploding? If they invest heavily in production capacity only to find that there is not commensurate consumption, then they'll have lost badly.

> as someone with no business experience, how they aren’t just pouring all of that money into enhancing their production capacity. That’s very clearly their bottleneck here.

They know this madness can't go on forever. The last thing they need is to be left with billions of dollars of unused capacity when the bottom falls out of this very stupid bubble.

> They know this madness can't go on forever.

If they thought that, would they invest massively in the companies that own all these GPUs? They would not. They would invest in anything ELSE they could think of.

But doing that would spook investors and possibly pop the bubble, so they look for ways to mitigate risk and fuel the bubble at the same time.

That might be the case if nvidia commonly invested in its customers - if they suddenly stopped doing that - but I don't think that was the case. On the contrary, these investments surprised because they were NOT the usual.

Why would they want to do that? The only sector that matter to nvidia is datacenter, its where 90%+ of their profits are. Making their consumer sector even less profitable just seems like a waste of time

How about positive mindshare? Regular people not growing up absolutely hating nvidia's guts and only begrudgingly buying their products. Also ensuring that a pretty big industry won't die from becoming too expensive.

Plus, diversification is good for when the bubble inevitably bursts.

But that's long-term thinking and we can't have that. People give Huang credit for having had a long-term vision on AI, but it feels like he definitely has blinders on right now.

The consumer gaming card market is minuscule in comparison to their primary market now, to the point where worrying about diversifying there probably doesn’t make sense. Nor does it really matter whether consumer gamers hate them. That is likely to have zero effect on their core customer now.

Underestimating compounding and secondary effects, especially while rationalizing the abandonment of their core market and capability is one of the most famous ways that big companies provide evidence of their downward spiral. I can feel the MBA energy from here.

Can you name any companies that suffered by switching focus away from one market where they dominate in order to also dominate a market that is 10x the size of the first market already, and growing faster?

If switching focus goes along with 'underestimating effects of ignoring core market and the treating competencies in it which lead to the market domination in other segments as legacy of the past which should be abandoned', then take your pick. A few off the top of my head: GE, Sony, HP, RCA, IBM.

Every specific situation is different, but the pattern I mentioned is easy to find. Here are three examples: RCA, GE, HP.

CUDA is where it is today because some years ago, people in grad school screwed around with CUDA on their personal gaming PCs

> How about positive mindshare?

Does anyone who can afford an nvidia card actually buy something else? Yes some people hate nvidia, but it's not like they have a hard time selling cards.

Consumers should already be glad that R&D for their products is being subsidized by OpenAI and Google

I'm just confused why people think token-based computing is going to be in such demand in the future. It's such a tiny slice of problems worth solving.

It's like how every big co these days is ML. It will transition to LLMs as well.

Just give it a few years.

Yep. Same vibes as “ha ha who needs internet connected appliances” (pretty much all appliances are internet connected now). And the apocryphal “there is a worldwide market for maybe 5 computers”.

No-one "needs" or even wants appliances to be connected to the internet. You claim that "pretty much all" appliances are internet connected, while almost none of the appliances in my house are.

I mean, who DOES need internet connected appliances? Are you implying these token-based techs will be shoved down our throats regardless of demand?

It is'nt just nvidia though.

If models are profitable once trained, isn't it weird that chatgpt and Claude have $200 tiers that still have usage limits?

>This is all dependent on token economics being or becoming profitable

What is the actual value of a token? If it were generated by a human expert in a given field? This should set an upper limit for now.

> - Nvidia also benefits from growth due to the equity they own.

aka this might be nvidia's next pivot. Contrary to gaming cards, the AI GPUs are productive assets. If nvidia feels that they will be very productive, then it makes sense that they invest broadly in the companies that are likely to make these profits. To share in the rewards while selling more GPUs.

I'd gander a guess that there's nothing tech specific here and that fraudulent schemes are well defined for the SEC and commercial courts to take action if something is not kosher

It's usually not actually fraud. It's the amazon reinvesting back into growth, except the unit economics don't work if everyone cashes out at the same time, and if anyone starts cashing out the growth stops and everyone cashes out before it's too late.

Exactly, everything old is new again. This was one of the drivers of the original dot-com bubble.

The original "Tech" boom was an infrastructure boom by the telecoms funded by leveraged debt. It was an overbuild mismatch with the market timing. If you brought forward the timeline to when that infrastructure was used (late 2000s) you probably would never have had the crash.

This boom is a data center boom with AI being the software layer/driver. This one potentially has a lot longer to run even though everyone is freaking out now. If you believe the AI is rebuilding compute then this changes our compute paradigm in the future. As well as long as we don't get an over leveraged build out without revenue coming in the door. I think we are seeing a lot of revenue come in for certain applications.

The companies that are all smoke and mirrors built on chatGPT with little defensibility are probably the same as the ones you are referring to in the current era. Or the AI tooling companies.

To be clear circular deal flow is not a good look.

I can see the both sides of bull and bear at this moment.

One interesting aspect of this is that, with the exception of OpenAI, all of the companies leading this boom generate massive amounts of income from other arms of their buinesses. I think this is one reason for the potentially longer run, since they can subsidize AI CapEx with these cash flows for quite a while.