So, it's a race to the bottom because they have to keep spending more and more money on training the models just to keep ahead of the competition, while prices are driven down by the availability of open models.

Chinese companies understand this, and that's why they give their models away in the open. The real money is going to be in integrations and custom services built on top of the models. The models themselves aren't what's valuable.

This is a real question, because I just don't know. Don't the models needs lots and lots of training that need the gobs of giant data centers? So you either need the infrastructure or the money to buy computing to make the models? I see that all the money is being invested in AI and data centers but I do not quite understand why.

I am no economist or business person but AFAICT the massive amount of money being burned is because everyone in the space believes there will be a Google-like (for search) single winner who takes all when the dust settles on the AI race.

Because of this belief they are willing to burn any amount of money they can get their hands on to get to this mythical finish line as fast as they possibly can, seeing it as existential. Moving fast means moving expensively.

The real question is why anyone believes this single-winner theory when all available evidence so far suggests there is neither a moat nor a lock-in opportunity at all.

they're already making a business model where enterprises buy tokens and put them in a pool for their users to use, and I feel like that's exactly where some companies will then compare that token pricing to open source models and find it hard to justify the prices, especially as it's quite obvious that many models https://news.ycombinator.com/formatdocare now "good enough"

I shortened my list to NVidia and AMD because putting Apple in muddies my point: I'd argue they'll be able to sell laptops + phones in that world, but only if you can run frontier open models on their devices, and do so efficiently. This is definitely not the case for their in-house AI offerings, like Siri, but it is already true if you install llama.cpp or whatever. I suspect that AI will not be a big revenue source for them (unless they completely turn around their software engineering org), but instead will be a "good enough to keep people from dumping their iCloud subscription" offering instead. Siri is basically that: It's completely useless, but people aren't switching to android because of it.

In the First World War, when they introduced the tank, everybody assumed its role would be to punch a big enough hole in the front lines that the horse cavalry could ride through and attack the enemy's supply lines, because That's How Wars End™. You fight the war you know, even when it's not the one you're actually fighting.

This is a really great identification of why, from an investing perspective, AI is nothing like past technologies.

That ain't going to happen, and if it somehow did in the US, things like Anna's Archive + libgen will save the rest of the planet from our self-imposed dark ages.

As for those data centers: The cost of inference and training on the same hardware is dropping much faster than Moore's Law. That multiplies with hardware advances, since Moore's Law isn't quite done yet.

So, those datacenters are depreciating at 50% every 18 months (Moore's law) at the same time as performance/$ and performance/watt of the software is dropping at some crazy rate like 10x per year. So, token per unit of data center is increasing by something like > 20x every 18 months, and that'll probably happen at least one more time. If it does, the datacenters will be provisioned to produce 400x more tokens than they could have when they were sized.

Put another way, if the singularity is coming, the first things it will destroy are those data center build outs. That doesn't matter though: the data centers are being built on circular loans, so the money needed to complete them probably never existed.

Dark times ahead; hopefully Trump will be blamed, since all this corruption happened on his watch while he had control of the senate, house and (partially) the courts. The stock-goes-down graph will show it happening after the midterms, of course.

This, to me, is the most interesting question in all of this. I increasingly feel like we live in a 'post-truth' world with regards to money and the stock market in particular. Like even during the COVID lockdowns the market blinked and then was at record highs within 6 months. It was just lol and completely out of touch with the reality on the ground.

The fed, and government alike, seem to have gradually shifted their mandates from trying to provide to a stable playground to trying to create a playground where important kids are never allowed to fall down. And when your existence is tied to trillions of dollars of the economy, you become the important kid.

Come to think of it this is also probably being exasperated by the rise of money management firms controlling tens of trillions of dollars. When the market valuation is technically just based on sentiment and you control that much money in a self-interested way, you're going to get weird outcomes.

We will continue to have every major player trying to build vertically - model, chips, infra, packaged products and services.

This is like saying there's no lock-in that iPhone has because it's trivial to switch to Android. Or that there's no lock-in that Windows has because it's trivial to switch to Linux or Mac. Or that there's no lock-in that cars have because it's trivial to switch to bikes or trains.

They're different products. They're not the same. Being able to write the same code does not make a model Claude. Or being able to reach the same benchmark scores. Or performing one or a choice set of tasks with a certain level of skill.

It's true, people can switch if all they want is to perform X, Y, or Z, and some other model can do X, Y and Z well enough. People switch from iPhone to Android all the time. From Windows to Linux or Mac. From cars to bikes or trains or other transit. But they're not the same thing!

You can't just say that makes cars the same as bikes or Windows the same as Linux. You can't get Claude from anybody other than Anthropic. Maybe you can get something else that does as well where it counts, but not Claude.

I don't switch from Claude for things like chatting. I gave Siri AI a big trial run when it first launched, because I absolutely love the system integration and the user interface, but the model itself is a bit dry for my needs. It doesn't really have principles the way Claude does (thanks to "constitutional AI" or whatever the fuck they call it). And I want principles, and sometimes the impression of empathy. Claude meets that bar without being obsessively sycophantic or paranoid.

(Siri AI is very good for some things, though. And I do greatly admire it. I even hacked up an adapter to use it for coding with Pi for just a short time, partly as a meme but also just out of curiosity. Just doesn't work for me here.)

I tried switching from Claude Code to Pi with DeepSeek. I had tried OpenCode at first, but it's just complete garbage. I felt so impressed of with it at first, it feels so slick on the outside, but that impression's just been shattered and shattered and shattered as I dug into it trying to fix a performance issue. Its data model and internal behavior is just not acceptable for me. (I tried Reasonix for about five minutes before concluding it's also a bit slop. It seems crafted by someone who actually understands LLMs though, unlike OpenCode, but it's still not very good.)

I absolutely loved the flexibility of Pi and the cheap prices of DeepSeek, but found myself falling into the same patterns as when trying Linux: I was constantly debugging the harness and the model, trying to get it to perform better or investigate something or other. Constantly worried that something was going wrong, that I should try to fix. And yeah, Pi has a lot of convenient features that Claude Code doesn't have, is a lot lighter, and of course can be incredibly modded. I just felt like it was also my responsibility to make my own experience, not just my capability.

At this point, Claude Code seems to just work for me, even though they removed a bunch of features I loved a while ago and never added them back (they came up with huge misunderstandings that completely missed what I actually appreciated about those features, and then proceeded to file them away as resolved). I never have to worry about debugging Claude or Claude Code.

For certain other things that aren't chatting or coding, I found that Gemini 4 Argon is actually by far the best model I've ever seen. It is the first model that's ever strictly exceeded the original November 2022 release of ChatGPT for me. Yup, it took the collective world nearly four years to replicate an experience from 2022. Gemini 3.1 Pro was so close though, I could feel it. When I first found it, I was so relieved that type of experience wasn't lost forever. Then Gemini 4 truly solidified which line of models I prefer for that use case.

Plenty of companies are attached to systems that don't do what they need because they're afraid to replace it, or because it's linked to an employee they need, etc.

Granted Anthropic (and Google, maybe less OpenAI actually) do what they can to make it similar. But switching phone brands would take me a day or two of planning and I don’t even use the Apple mail app.

And the cost of training is going down as a percentage of revenue over time, is it not?

And unless my math skills are completely off, it’s about 21 months, not two years, but fair enough. What 18 month period in that range would you like to consider as validating the ridiculous idea that open weight models are going to eat Anthropic’s lunch?

It also doesn't do much to disprove the OP's argument, which is that the actual models aren't what drive growth. I personally think it doesn't particularly matter which models Anthropic is running because the tools like Claude are what people are actually paying for. Anthropic, at this point, is so far ahead of the competition in the tooling around models that they can probably just start running Deepseek products directly and people would still pay them.

Everyone has to get capital efficient now — and fast!

WeWork was renting lemon water and cubicles and calling it technology.

Whatever you think about this market, the LLM technology is real, obviously important, and anthropic is the market leader.

Sure, they have some flashy models, harnesses, and tooling around it, but their competition isn't that far behind - and some of that competition is 100% free. Their moat is closer to a puddle, to the point that there are entire companies built around the concept of dynamically routing requests either to them or their competitors.

Anthropic is the market leader in a rapidly-evolving market, where customers can and will jump ship the second a competitor's model looks shinier or cheaper.

Sure, sure. And SpaceX is "renting rockets". Nothing to see here.

While you can imitate the style of my response by ignoring all of the technical details, you've still ignored all of the details.

WeWork was not a technology company. Anthropic is.

The problem with WeWork's IPO wasn't that they aren't a tech company. It's that they didn't have a good enough business model.

AFAICT, the only one who cares about whether Anthropic, or WeWork, is a technology company, is you.

> The problem with WeWork's IPO wasn't that they aren't a tech company. It's that they didn't have a good enough business model.

No, they had a fine business model. It just wasn't worth a technology company multiple.

Microsoft has already proven that they do not have the software chops to deliver this. Even if they had the right product managers and UI experts, MS management would torpedo the project and ensure that it gets zero adoption.

And AWS has zero software capability in that regard.

Despite being awful in so many ways, Microsoft Office continues to dominate. It's a "competitive edge" because is has the trust of corporate customers. There are many competitors, but everybody is used to the main one already, and it's "good enough".

In many ways, I prefer to use other models through Pi.

They are extremely capable, much cheaper than Claude, and can get the job done for the software development tasks I use AI for. (Quite a lot these days)

This is going to sound tautological but of course it's not the same, because it's different?

Participants in markets make their own choices from the available options, for their own reasons.

Another way to frame my point is this: not the same for whom? You are but one market participant, and it's at least wobbly thinking for any of us to assume that our set of needs is the archetype.

There's any number of harnesses and people have any number of preferences. Sure, overchoice of imperfect substitutes tends to entrench better-resourced, better-resourced (or better-marketed) alternatives, but it's not a slam dunk.

Seeing this bubble play out in real time is the most surreal thing. It's like watching a train wreck play out at a speed of a few inches per day. Like come on, just get off the track FFS.

Have you tried requesting GPU instances on Gcloud without being associated with a major institution? Yeah, good luck.

https://www.wsj.com/finance/major-investors-are-skeptical-an...

Implying something with context is just as bad as saying it outright, so hunter2-ing your swears doesn't save you as much as dropping it altogether.

It may only be a short-term circumstance that you need a data centre to run AI for lucrative applications. Thirty years ago you needed a mainframe to beat Kaspersky at chess. Algorithm's got better, and now you can beat the mainframe that beat Kaspersky with a bot running on a desktop Pentium of the day[1]. You could still, today, devote a full hyperscale data centre to create a chess bot to beat all other chess bots, but nobody is doing that because there's no money in it.

The frontier is continually pushed outwards. At it's outer edge you need mainframes and data-centres. Once it pushes past, you don't. There wasn't a lot of money in Chess or protein folding, despite the later having huge potential for public benefit. There is a lot of money in coding and middle management, and that's what currently has tech-bros in a frenzy. However, there's no guarantee that what the frontier moves onto next will be lucrative. It might be more chess and protein folding. If, as Algorithm's and hardware improve, coding moves to local models (Sam Altman's nightmare) and nothing lucrative immediately replaces that, what are these companies going do do with their massive data-centres?

The scale of investment we're seeing these companies commit only makes sense if they have multiple decades to reap the returns. They may well not have that much time.

_____________________

[1]https://hackaday.com/2026/11/05/pitting-a-modern-chess-engin...

Reality can’t be saturated. The set of things you can accomplish with more intelligence is unbounded. So while cheap models will continue to proliferate, there will always be a massive market at the frontier, and that won’t ever change.

Guess what could help solve it? :)

Seriously though, this “problem” will never be fully solved, we’ll just get better and better at connecting AI to the broader world, and to me that’s already the story of the last few years. Models get better, they’re able to build and drive tools to interact with both the digital and physical world better, which helps build the next generation of models, and so on.

It was worth $8.9 billion when the US govt got it "for free". It's worth $23B today, even after the big drops recently. How is the US public worse off?

Yes. Personally I'd prefer the US government get its stake from a bankruptcy court for pennies on the dollar.

I think about this because I pay $100/mo for Fable. How much more would I be willing to pay? I'm not sure yet. It depends on how smart the AI in the future becomes.

Already they have exclusive deals with trading firms who pay $$$ for faster models and more features. But there are only so many firms who will pay that much (and a lot of companies are pulling back on spend, not increasing it)

Whether it's Anthropic or not, I think it's inevitable some highly specialized offerings like this will emerge. I suspect there are already some incredibly sophisticated, hyper-narrow, astronomically expensive models being created for private in-house use.

As an imaginary example, I wouldn't be shocked to learn some company like Jane Street has invested a billion dollars to create the highly specialized synthetic training data needed to create an AI that gives them .08% edge on certain kinds of commodity trades that we'll never hear about. If so, it's not a leap to imagine another company spending $100M on a model that gives a .008% edge and selling access to 100 high-volume traders for $2M each.

Expanding general purpose LLMs to incorporate increasingly specific use cases has, so far, seemed to be economically viable but I don't see a fundamental reason that would continue to remain true for incredibly narrow yet ultra high-value domains requiring monumental bespoke development, data, etc.

The crazy thing is ... it's not imaginary, nor is secret, it's literally a video on YouTube from five months ago - https://www.youtube.com/watch?v=8J-GUnfSqeE (though that facility is only showing of ~$200 million worth of compute).

Thing is - regular people have no clue what to do with it since they are locked in ecosystems of apple and google.

This feels wrong in several ways:

a) they're certainly not addressing their customers like they're compute-limited. One must ask, what would that look like? For one, I think it'd look like a wait list to sign up. It'd look like prices going _higher_, and not lower

b) IIUC, their revenues don't cover their _current_ compute costs. So, let's pretend coreweave magics up a new dc and hands it over to anthropic, which means presumably anthropic needs to start paying for it. How does that affect anthropic's revenue? AFAICT, it just adds costs, it doesn't make their revenue go up.

The only interpretation I can find is if Anthropic is _convinced_ that they need to train faster/harder, and that doing so would somehow significantly shift their TAM or costs, for example by coming up with a model that was simply 10+ years ahead of anyone else that costs the same as opus 5.5 to run, and they instantly nab the entirety of openai's market share. Seems pretty unlikely.

But you’re right. I have often wondered why it is that such a small amount of extra compute (300MW) from SpaceX was entirely enough for them to uncap so many customers.

I think the (in-house) data centre burdens of ever-more complex models may be considerable, but I think the data centre businesses expecting to serve inference tasks are probably heading into a Chinese real estate bubble.