[nopinsight]

> nobody is sure if even their metered pricing is profitable

This is most likely wrong. Lab executives insist that serving tokens is profitable. It's the cost of training next-gen models that requires them to keep raising ever larger rounds. More importantly, many independent providers price tokens of open-weight models at a fraction of Anthropic's prices.

[atwrk]

But are they actually profitable, or do they employ creative accounting where only parts of overhead expenses are counted against all of inference revenue, similar to what Uber did?

OpenAI's numbers show that they definitely are not profitable on inference, and even worse, revenue growth scaled linearly with inference cost from 2024 to 2025, which means they can't outgrow this problem. See https://www.wheresyoured.at/oai_docs/

[therealdrag0]

Does it matter if it’s creative accounting? Uber is a great example of a company that everyone was certain would fail because it was unprofitable and now it succeeded and is profitable.

[armonster]

Uber didn't have ever-increasing costs though.

[baq]

If they shut down all training today they’d be absolutely printing money for the next couple quarters and then die with a bang once the other lab releases the next frontier to the public.

[mattmanser]

Try doing some inference with local models.

I'd be surprised if they're making money on inference just from that. There's no way someone paying $20 p/m and using it all day is not spending way more on even just the electricity for tokens, let alone the capex.

[beepbooptheory]

I don't really get the last bit. It's hard to imagine what a new fangled "frontier model" could do that would blow anyone out of the water. Like what does this look like? Really good benchmarks? Who cares about that anymore?

[layer8]

Not hallucinating anymore would be a good start.

[shimman]

How? They're already burning $2 bills to make $1, court documents shown that Anthropic has already been lying around revenue (claimed to have made $19 billion when it's actually $5 billion to date [1]).

Not hard to believe they're lying about other things when they've been lying about the capability of their products since inception.

[1] https://www.reuters.com/commentary/breakingviews/anthropic-g...

[thereitgoes456]

That is not what the article says, it says $19B ARR.

I don’t necessarily see a contradiction. $19B run rate, achieved very recently, is actually consistent with $5B lifetime earnings, because their growth curve is so sharp. Zitron is not good at math.

[shimman]

Didn't link to Zitron site but if you can't see how dishonest it is to say you have $19b ARR when the reality is you have only a total of $5b IDK what to tell you. Says more about how you think and why you think it's okay for corporations to be misleading.

[s1artibartfast]

Seems natural to me too. ARR is understood as the current rate. It would be more misleading to say 5b ARR.

Its like asking how fast a car is moving.

[MattRix]

This is not lying, that is just what run rate revenue means! It makes sense to use as a metric when a company’s user base is growing as fast as Anthropic’s is.

[shimman]

It makes sense to be extremely misleading about actual accounting figures? In what world is it okay to say you have $19b in ARR when you have only ever generated $5b for the entire duration of your company's existence?

Did Enron start a business school I'm unaware of something?

[dragonwriter]

> In what world is it okay to say you have $19b in ARR when you have only ever generated $5b for the entire duration of your company's existence?

In the same world that it makes sense to say that your current speed is 57mph when you've only driven 15 miles since starting the trip.

[baq]

sir if you say a number is $19B and everyone who is invested knows what it means, is there a problem?

[B56b]

So just ignoring the link entirely, cool cool cool

[martinald]

Yes I wrote a detailed article about this Forbes claim. https://martinalderson.com/posts/no-it-doesnt-cost-anthropic...

Key points - if you compare it to openrouter costs for ~similar sized models it is ~90% gross margin.

And this claim came from Cursor - not Anthropic!

[shafyy]

Not counting training models as part of your gross margin is just creative accounting. It's an inherent part of being able to provde the service for OpenAI, Anthropic etc.

Even so, their subscriptions are significantly cheaper than the token pricing via API. So at some point they will need to get rid of subscriptions or increase the subscription prices dramatically... And that's assuming their current token pricing is actually profitable. Which it probably isn't.

Lastly, I would not trust one word that comes out of an executive of an AI company (or any other large company, for that matter).

[mrbungie]

> Lab executives insist that serving tokens is profitable.

Maybe marginally profitable, but right now they need to give out subsidies for people to use their products (Antigravity, Codex, Claude Code et al) in an actually useful manner that prevents churn and at the scale they need to justify usage growth forecasts, which they need to keep the wheel turning.

Probably if you look at the users who exclusively use the simple chat box interfaces (i.e. ChatGPT, Gemini in UI, Claude in UI) plans it is actually profitable, but I'd also say that's not where most of the usage comes from.

I'd love to actually look at both usage + profitability from each user segment to see if their PxQ growth expectations from non-enterprise usage make any sense.

> Many independent providers price tokens of open-weight models at a fraction of Anthropic's prices.

Are those open-weight models as good as Anthropic? Are they the same parameter class?

[est31]

It's a loss leader but this is normal. Same has happened with Uber, Airbnb, Amazon, etc. Using VC money to buy marketshare and once you have it, you can milk it.

The question is more around the moats that these companies have and it seems to me while their models are amazing technology, they don't really have a moat. The open/chinese models still continuously catch up to the american ones.

[hirako2000]

And what possible moat. It isn't hard to foresee that in just a couple of years, models outpacing the latest frontier tech we have today will run on consumer hardware. With open source workflows anyone can pull in to run, providers won't see a penny.

Another scenario is that dense models get replaced entirely, in which case the likelyhood of OpenAI and co pioneering the concept is pretty slim. They will be left with billions worth of infrastructure which cost them 10 times that 2 years earlier, faced with the reality touched by the article: liquidate.

[zozbot234]

> Are those open-weight models as good as Anthropic? Are they the same parameter class?

Are they as good as Anthropic was one year ago? That's more like it. They don't have to be just as good, they just need to be the most worthwhile for the price. If frontier models are only providing a negligible advantage for what they charge, that absolutely matters.

[techpression]

I wouldn't trust those claims from any private companies, even public ones play the most insane tricks in earnings calls to inflate numbers or heck, just make up new ones.

I'm not saying they're wrong, but I don't take much stock in their words.

[sunaurus]

The point is that you can’t just serve tokens without also training the next models. It’s an inseparable part of your costs, so naturally you can’t be profitable unless the price you are charging ALSO covers training.

[dash2]

Is that right? I think that you can serve tokens without training the next models. It would be bad strategy, but it would work. So it's an important question, are they covering their operating expenditure? If they are the business has legs (and it will be worth spending a lot to train the next models). If not, maybe not.

[camdenreslink]

If a major model provider were to just halt progress on developing new and improved models, the open weight alternatives would catch up in a couple years.

They would have a period of great margin, followed by possibly zero margin as enterprises move to free options.

They would have to come up with a lot of great products around the inferior models to justify charging at that point.

[leoc]

Also, an out-of-date model which doesn't know about last year's world events, hit songs and new JS libraries is a depreciating asset even before you consider low-cost competitors catching up. So you'd presumably have to do some training just to keep the model up to date at the current quality level (unless you completely give up and just sweat the assets). And on the other side of that coin: over the next few years, do the latest, biggest models continue to generate user-perceived real-world improvements sufficient to keep users wanting the latest and greatest?

[dash2]

> If a major model provider were to just halt progress on developing new and improved models, the open weight alternatives would catch up in a couple years.

That's why it would be bad strategy.

[yorwba]

There are companies that already do nothing but serve tokens using models trained by others. Just running infrastructure and collecting a reasonable fee for their troubles. It's only a bad strategy if you want to claim to investors that you'll gain monopoly market share if only they could give you a few more billion dollars.

[chasd00]

i don't think it will work, it's too easy to switch models. When google comes out with a new model people will just switch. I think Google wins in the long run, they have the money to just wait until everyone else goes bankrupt and they also have the Apple contract and therefore the mobile market.

[leoc]

And apparently the most efficient training and inference thanks to their TPUs, IIUC?

[gedy]

Buying and driving a new car off the lot costs the manufacturer nothing at that moment, but what happens before that is important to account for.

[phantom784]

Do tokens just cover ongoing operating costs, or are they also able to pay back the cost of training that model originally?

[pier25]

So these companies will be profitable if training stops? Is that even a real possibility?

[danaris]

Any given company could stop training tomorrow, and, as some others have said here, they'd be generating quite a bit of profit until their models visibly fell behind, however long that ended up taking, at which point they'd probably just fall over completely.

Over the whole industry? No; they can never, ever stop training, or they'll cease to be useful at all very soon.

Training is what keeps the models up-to-date on current events, which includes new programming languages, frameworks, and techniques. It's already been observed that using LLM assistance on some types of programming is much more effective than others, based on how well-represented they are in the training data: if everyone stopped training tomorrow, and next month a new programming language came out, none of them would ever be able to help you program in that new language.

This can be extended to other aspects of programming, too. If training stopped, coding assistants would gradually start giving you wrong answers on how to implement code for APIs, frameworks, and languages that continued to evolve, as they will always do, in much subtler (and likely harder-to-debug) ways than how they'd deal with a new language whose existence they don't even know about.

[naravara]

The impetus to continue training at the pace they are is driven by the competition. So if the money starts drying up, then they’ll naturally slow down because they’ll have to figure out how to do more with less.

I suspect that once the models hit a point of “good enough” for certain use cases companies will start putting R&D focus in other areas that may be less expensive. Like figuring out how to run more efficiently, UI/UX conventions that help users get what they’re trying to accomplish in fewer steps, various kinds of caching of requests, etc. So the cost to serve tokens over time should only come down, and will probably start coming down more rapidly as the returns to model training slow down.

That’ll probably be a while though, because each successive model tends to be a lot better than the last.

[WarmWash]

What's interesting to note is that the "intelligence" labs can squeeze out of an H100, an almost 4 year old GPU, is dramatically higher than what they got out of it in 2022.

It hints that once these labs get a good enough "everyday model", they can work on efficiency so they can serve these models on old hardware. Which is almost certainly already happening.

[pier25]

> So if the money starts drying up, then they’ll naturally slow down because they’ll have to figure out how to do more with less.

Meanwhile companies like Google will keep investing on training...

Anthropic's CEO has suggested all AI companies should slow down training but obviously this is only beneficial for companies that can't afford to keep training.

[hbn]

> UI/UX conventions that help users get what they’re trying to accomplish in fewer steps

If we can expect the past 15 years of software UI/UX history to continue, it's more likely they'll spend the money on making the UI/UX more confusing, removing features, and making basic tasks take more steps than they do today.

[naravara]

That’s because the past 15 years were dictated by Web 2.0 companies that make their money off keeping you glued to the screen.

A AI assistant would work more like Planet Fitness where the goal is to figure out how to convince you to keep paying them while using the facilities as little as possible.

A big part of that might just be steering you towards repos of existing solutions to the problem you’re trying to solve rather than helping you vibe code a solution yourself. Over time they’ll be able to accrue a whole pile of canned functions that’s all automatically documented and audited and it’ll be able to plug and play those rather than having to rewrite.

The security implications of this give me a headache to contemplate to be honest.