Not because there is no path to profitability (they make a ton of money on inference), they just spend a lot on R&D.

Afaik Anthropic still loses money for their main product in this space: Claude Code and their Max plans.

This became immediately clear to me over the weekend when I used Opus via API key. I had it review the code for my (relatively small) personal blog to create an AGENTS.MD - it cost me $3.26.

same here... The API costs are absolutely insane for any real usage. This is either high prices to make sure no profitable competitor to claude workspace or other agent system emerges, or heavily sponsoring of their own soluions.

Api cost need not correlate with running cost.

Not really. They are burning money on hardware, resources and payroll without meaningful return prospects.

> they make a ton of money on inference

So it is stated, but is it actually true? I am not convinced.

Besides, it's not as if they can suddenly stop training models, the moment you do that you've spelled a death sentence for profitablity because Google and open source will very quickly undercut a 15 year break even timeline.

It's widely reported and acknowledged as true.

Well, the only people with any ability to acknowledge it have a massive incentive to do so, and I've been around the block enough times to know that startups will use every trick in the book to paint a rosy financial picture, even when it's extremely misleading or occasionally just straight up lies. In the current climate of AI hype my skepticism is even greater.

I'll believe it when I see it.

Where and by who? Critical context missing here.

Dario Amodei @ 18:05

https://www.youtube.com/watch?v=GcqQ1ebBqkc&t=1088s

The CEO hyping his product and the viability of his business during an interview with Stripe does not, at least to me, qualify as “widely reported and acknowledged”

Agreed, the revenues are big.. but very small next to the datacenter bills.. even if a fraction of which are being used for inference, it's hard to argue they even break even. That's before all the other costs (Super Bowl ads, billions in compensation).

from what i understand, the issue with inference is it doesn't scale as user count grows the way traditional saas scales. In typical saas adding users requires very little additional capacity. However with inference, supporting more users requires much more capacity to be added. I don't know if it's quite linear but it certainly requires more infrastructure to support additional LLM users than say a web application.

And the existing infrastructure routinely struggles for several of the well known players. You can literally tell when it's getting bogged down by workload. And that's after all the absurdly large datacenters we've already established at significant expense (to both the corporations and the average person).

With coding (it's not really coding per se that matters imo it's more like dynamic logic writ large) it's a land grab strategy. They want to get established as the de facto standard and get a whole bunch of people on their platform so by the time they need to "get profitable" they have a captive audience, a leg-up on other labs. It's a tale as old as time, that's why ubers used to be cheaper than cost.

It's a strategy as old as time, but it's a strategy that usually fails. Spending a lot of money on customer capture only works when customers are actually solidly captured. Most markets have fairly heavy competition and customers will only stay captured as long as there is no substantial cost to staying captive.

Take Uber as an example: yes they've raised prices to become profitable, but not to the insanely profitable levels they could if they had a true monopoly. People will stay on Uber when the competition is still at a roughly equivalent price, but will switch if Uber raises its prices enough.

Uber Eats is different, since its a 3 sided market where the cost is paid by the restaurant rather than the user.

AI appears it's going to be more like Uber the car service. Claude can charge $200/month, but charging $2000/month seems unlikely to work. I'm sure many would be willing to pay $2000/month if they had no alternative, but there are alternatives.

> it's a strategy as old as time, but it's a strategy that usually fails

I like to call this the "Yahoo Effect"

> They want to get established as the de facto standard and get a whole bunch of people on their platform so by the time they need to "get profitable" they have a captive audience, a leg-up on other labs. It's a tale as old as time, that's why ubers used to be cheaper than cost.

Some of that is seeking to kill competitors before they can get established. That's normal and has been around for generations, if not since trading was invented.

But most of what we've seen during the "enshitification age" has been to burn money until you achieve a critical mass of users. However, this only really applies to social platforms where the point of it is communicating with people you know. That's the lock-in. You convinced Grandma to join Bookface and now you feel bad leaving if she doesn't leave at the same time, and more importantly, who wants to join Google Square if nobody else uses it?

That's not going to work for AI platforms.

What I do see potentially working is one method that email platforms use to lock in users: having tons of data you can't export/migrate. If you spent lots of time training your AI by feeding it your data, that's going to make it harder to leave.

So far none of them have capitalized on this (probably due to various technical reasons) but I expect it to start eventually.

The lock-in of email platforms is the address. With IMAP you can extract the messages right away and migrate. Yet, you would still have to check the old mailbox for stray emails that you must tell to reach you on the new address. And continue doing so for years or risk missing some critical email.

Coincidentally, bringing your own address that can be migrates away is somewhere between impossible and expensive.

No, you can do it on all the major providers for either no or low cost.

Disregarding the grandfathered free accounts, own domain is $7.20/user/month on gmail, €5/month on Proton. On microsoft that's business tier feature and AFAIK not supported at all on Yahoo.

Zoho Mail Lite is $1/user/mo when billed annually.

https://www.zoho.com/mail/zohomail-pricing.html

A few DNS hosting companies still bundle in a few free email mailboxes with registration costs but that is becoming more rare.

Yes They are just pivoting to stuff that loses money more slowly but maybe has a path to profits eventually…

Some of these AI companies that promised AGI are going to find out that they're actually IDE plugin subscription companies

I think it remains to be seen if LLMs are even 25% as good at everything else as they are at coding.. which is fine, if they focus and stop promising the world.

That alone is huge, if they let go of their egos about putting the entire white collar class out of work..

Coding is a small minority of total generated tokens. It's easy swimming in tech waters all day to think Claude is the pack leader because it writes excellent code, but the reality is that tokens are overwhelmingly coming from OpenAI and Google doing mostly stuff like "Make this e-mail sound nicer" and "What's a cheap vacation spot with warm turquoise waters"

> "Make this e-mail sound nicer" and "What's a cheap vacation spot with warm turquoise waters"

Right but I think a lot of these use cases aren't replacing any jobs because it wasn't anyones job. It's just a little polish on existing work (did spell correction in Word kill jobs?) or the stuff that voice assistants have been promising for 10 years.

Both of those things both were and are jobs. They're called secretaries and travel agents.

Jobs that have already been killed is my point

Together that's about four million American jobs so I'd disagree those jobs have "already been killed".

Nvidia CEO said we already have agi:)

Ad generated income

Frontier model developers don't make money, but inference providers do. For open weight models there is a healthy market of inference providers that operate profitably without VC backing.

Such as? Where do we find these open weight model providers? Why is hardly anyone talking about them or sharing links (here or elsewhere) if they are so wonderful and profitable?

Go to https://models.dev/ and you're going to see plenty of providers.

OpenRouter makes it easy to use them, just add credits to your account.

I thought this was common knowledge to anyone looking to use an inference API, but it seems it isn't. Well, even AWS is in this business with Bedrock.

Why is hardly anyone talking about basic web hosting provides or sharung links (here or elsewhere) if they are so wonderful and profitable?

Because few people really care much about the commodity hosting world. They're not making waves, they're just packaging things made by others for a low-ish cost. They're also not very consumer-focused, as they're a bit lower level than what most people prefer to think about. It doesn't mean they don't exist or that they're not profitable though, just not headline-reaching numbers in the end.

CoreWeave's cash flow do not look too healthy.