Those numbers intentionally exclude the single largest operating expense that Anthropic has: model training. [1]
So yeah, if they stop training models forever, Anthropic will probably start making a profit... until someone else with better models comes along to eat their lunch.
I am not even sure that is true. We don’t know everything that is included or excluded from their calculation. I suspect there is a lot of funny math going on to get to profitability. Remember there was a lot of similar talks and reports about SpaceX and how profitable they were. The reality was much worse. I just don’t believe it until I see it in the S1. Even then they can hide quite a bit.
Or if model training is more of a rollercoaster, where spending gets you to the top of the hill where you create a massive internal model which can then build the next version of itself for cheaper and cheaper amounts relative to human R&D costs. If Anthropic is first over that hill, they can race far ahead.
They have found a path to “profitability” iif you define “profitable” in a way that makes every early stage start-up that has at least one paying customer as “profitable”. Literally any start-up has a COGS lower than their income, but that doesn't mean anything at actual profitability given that the rest of their expenses dwarfs it.
I believe the entire basis of their profitable quarter was getting a discount on compute from Musk.
All these figures are so utterly weaselly. AAR is a made up measure to make them look good. If they cannot show GAAP numbers, they are hiding something. Full stop. While as private companies they are under no legal obligation to show us their books, their PR and intent to go public requires it.
If an AI company can exclude the cost of training the new models they release every three months from the business of whether they are profitable, it would be shocking if they weren't profitable. And the figure is tiny compared to the valuation they appear to be seeking, and may only be positive because of a short term boost.
Steve Eisman said the other day that he suspects part of Anthropic's rush to get to IPO is that their third quarter figures are terrible.
I'm insanely profitable each month if you exclude my mortgage and bills and shopping too.
My low level conspiracy theory is AI is encouraging habits of people not to read so noone can read statements like "we excluded our costs from our profit calculation"
I think we're going to suddenly see them greatly scale back training and try to sell inference-only, but they all know when they do that someone can jump up and outstrip them.
But only as long as training actually improves models significantly. As soon as those improvements stay below a certain threshold, the better move is to invest your R&D money into other things like harnesses or new tricks one can play with existing models and the immense cost of training is just not worth it to be 0.5% ahead.
I'm absolutely certain that we will reach that point, just not when. Could come sooner than we think though.
Don't get me wrong I still want the models intelligence to improve, but for all practical purposes we are already there this is why many people are already moving to cheaper/open source. There is still a case though for the 1% of queries that demand SOTA
In this version of conspiracy theory, all the labs secretly understood that training wasn't economically feasible anymore so they all jointly made it look like they were stopping for safety reasons.
Is there no end to this kind of lazy conspiracy theory
You don't need to communicate to coordinate. All these companies have the same business model, if it was financially advantageous for one of them to push that narrative, then it's financially advantageous for all of them.
> Well there certainly is at least some kind of viable business running large AI models for a fee.
ok, where are:
- The economies of scale?
- The network effects?
- The switching costs?
- The intangible assets (e.g. brand?)
Running AI models for a fee has none of these. At best, there are some economies of scale for running a datacenter, but OpenAI and Anthropic have none.
There are not as many network effects, intangible assets, or switching costs as other businesses. I believe there are economies of scale in terms of power and cooling and network bandwidth and the people who plug in cables and other such things.
The business logic is similar to the general transition to cloud. Corporations and individuals are better off paying someone else to manage physical hardware that they just access over the network. That is even more true of large, expensive, fancy AI GPUs than regular web servers.
OpenAI and Anthropic may both fail, or may not, I don't know. But I'm sure there is some kind of viable business running some kind of AI in the cloud.
> OpenAI and Anthropic may both fail, or may not, I don't know. But I'm sure there is some kind of viable business running some kind of AI in the cloud.
The problem is that the investment does not expect "some kind of viable business" ROI needs to be in the order of several trillion for this to make any sense.
Well sometimes when people invest money that does not work out.
That money may not be made back in the way people hope.
But there is a whole ecosystem of companies on OpenRouter etc. who have a viable commodity business serving Chinese open source models on GPUs. I'm sure at least some kind of business like that will survive even if OpenAI and Anthropic completely fail. And I'm sure AWS, GCP and Azure will end up having something like that too.
Amazon will be able to run large models for a fee, and make money on it. It's not a trillion dollar business, it may not even be a good business, but they'll be able to do it.
Sure, a supermarket will sell avocados if they make a profit, and won't sell them if they don't. That's a very different business model; the product is the infrastructure, not the AI.
Anthropic likely would not be saying, in October, that they planned to go public next month, if this were also true of their business.
In the last ~month, OpenAI announced a delay to its IPO and Anthropic put a relatively near-term range on its IPO date. These are very different signals.
The reason is that companies can choose the best timing to go public - when their financial look the best - and they do.
Anthropic trying to go public very soon is a good tell their financial look pretty decent. OpenAI postponing the IPO is a very good tell theirs look bad.
Anthropic appears to have found a path to profitability: https://www.forbes.com/sites/jonmarkman/2026/08/17/anthropic...
Those numbers intentionally exclude the single largest operating expense that Anthropic has: model training. [1]
So yeah, if they stop training models forever, Anthropic will probably start making a profit... until someone else with better models comes along to eat their lunch.
[1] https://www.morningstar.com/news/marketwatch/2026091414/the-...
In the discussion of a similar article, it was called PBBT - profit before bad things.
I am not even sure that is true. We don’t know everything that is included or excluded from their calculation. I suspect there is a lot of funny math going on to get to profitability. Remember there was a lot of similar talks and reports about SpaceX and how profitable they were. The reality was much worse. I just don’t believe it until I see it in the S1. Even then they can hide quite a bit.
Or if model training is more of a rollercoaster, where spending gets you to the top of the hill where you create a massive internal model which can then build the next version of itself for cheaper and cheaper amounts relative to human R&D costs. If Anthropic is first over that hill, they can race far ahead.
kind of puts it in 4K the reason behind "we must pace the frontier"
Maybe... just maybe companies find an equilibrium? Maybe companies reinvest in training because there's performance increase?
They have found a path to “profitability” iif you define “profitable” in a way that makes every early stage start-up that has at least one paying customer as “profitable”. Literally any start-up has a COGS lower than their income, but that doesn't mean anything at actual profitability given that the rest of their expenses dwarfs it.
I believe the entire basis of their profitable quarter was getting a discount on compute from Musk.
All these figures are so utterly weaselly. AAR is a made up measure to make them look good. If they cannot show GAAP numbers, they are hiding something. Full stop. While as private companies they are under no legal obligation to show us their books, their PR and intent to go public requires it.
These figures are EBBT.
Earnings Before Bad Things.
If an AI company can exclude the cost of training the new models they release every three months from the business of whether they are profitable, it would be shocking if they weren't profitable. And the figure is tiny compared to the valuation they appear to be seeking, and may only be positive because of a short term boost.
Steve Eisman said the other day that he suspects part of Anthropic's rush to get to IPO is that their third quarter figures are terrible.
I'm insanely profitable each month if you exclude my mortgage and bills and shopping too.
My low level conspiracy theory is AI is encouraging habits of people not to read so noone can read statements like "we excluded our costs from our profit calculation"
Yep, I agree. The only 'frontier' any of the big labs are racing towards is the frontier of financial ruin.
I think we're going to suddenly see them greatly scale back training and try to sell inference-only, but they all know when they do that someone can jump up and outstrip them.
But only as long as training actually improves models significantly. As soon as those improvements stay below a certain threshold, the better move is to invest your R&D money into other things like harnesses or new tricks one can play with existing models and the immense cost of training is just not worth it to be 0.5% ahead.
I'm absolutely certain that we will reach that point, just not when. Could come sooner than we think though.
Don't get me wrong I still want the models intelligence to improve, but for all practical purposes we are already there this is why many people are already moving to cheaper/open source. There is still a case though for the 1% of queries that demand SOTA
At that point they would lose all advantage stemming from their ability to boil the ocean though.
> I think we're going to suddenly see them greatly scale back training and try to sell inference-only
Remember a few weeks ago when all the AI labs said "we need to slow down, to uh, prevent destroying the world"?
Ding!
In this version of conspiracy theory, all the labs secretly understood that training wasn't economically feasible anymore so they all jointly made it look like they were stopping for safety reasons.
Is there no end to this kind of lazy conspiracy theory
You don't need to communicate to coordinate. All these companies have the same business model, if it was financially advantageous for one of them to push that narrative, then it's financially advantageous for all of them.
Or all the big ones are content with the current status quo and willing to compete amongst each other and want to shut the door to upstarts.
You see it time and time again.
So the more likely answer is they are on the verge of making super intelligence that could kill us all?
Well there certainly is at least some kind of viable business running large AI models for a fee.
These are useful and too big to run locally.
The ultimate size of that business in terms of revenues and profits may not match current expectations, but it's also not 0
> Well there certainly is at least some kind of viable business running large AI models for a fee.
ok, where are:
- The economies of scale?
- The network effects?
- The switching costs?
- The intangible assets (e.g. brand?)
Running AI models for a fee has none of these. At best, there are some economies of scale for running a datacenter, but OpenAI and Anthropic have none.
There are not as many network effects, intangible assets, or switching costs as other businesses. I believe there are economies of scale in terms of power and cooling and network bandwidth and the people who plug in cables and other such things.
The business logic is similar to the general transition to cloud. Corporations and individuals are better off paying someone else to manage physical hardware that they just access over the network. That is even more true of large, expensive, fancy AI GPUs than regular web servers.
OpenAI and Anthropic may both fail, or may not, I don't know. But I'm sure there is some kind of viable business running some kind of AI in the cloud.
> OpenAI and Anthropic may both fail, or may not, I don't know. But I'm sure there is some kind of viable business running some kind of AI in the cloud.
The problem is that the investment does not expect "some kind of viable business" ROI needs to be in the order of several trillion for this to make any sense.
Well sometimes when people invest money that does not work out.
That money may not be made back in the way people hope.
But there is a whole ecosystem of companies on OpenRouter etc. who have a viable commodity business serving Chinese open source models on GPUs. I'm sure at least some kind of business like that will survive even if OpenAI and Anthropic completely fail. And I'm sure AWS, GCP and Azure will end up having something like that too.
Amazon will be able to run large models for a fee, and make money on it. It's not a trillion dollar business, it may not even be a good business, but they'll be able to do it.
Sure, a supermarket will sell avocados if they make a profit, and won't sell them if they don't. That's a very different business model; the product is the infrastructure, not the AI.
Anthropic likely would not be saying, in October, that they planned to go public next month, if this were also true of their business.
In the last ~month, OpenAI announced a delay to its IPO and Anthropic put a relatively near-term range on its IPO date. These are very different signals.
> Anthropic likely would not be saying, in October, that they planned to go public next month, if this were also true of their business.
they are 50/50 at best.
IPO = It's Probably Overpriced.
The reason is that companies can choose the best timing to go public - when their financial look the best - and they do. Anthropic trying to go public very soon is a good tell their financial look pretty decent. OpenAI postponing the IPO is a very good tell theirs look bad.