https://www.sec.gov/Archives/edgar/data/2130357/000162828026...
Income at page F-7 (about 350 overall):
(units in Kilo-millions)
Revenue in first 2026 9-months 33.2
Compute and infra expenses 18.8
Personel 4.8
Sales 5.5
Administrative expenses 5.3
Operating earnings (loss) (1.2)
Q-3 only, this is calculated substracting first half from 9 first months:
Revenue in Q3 16.9
Compute and infra expenses 9.9
Personel 1.7
Sales 2.3
Administrative expenses 3.6
Operating earnings (loss (0.6)
The problem with AI is that it's still dependent on the skill of the prompter.
Ah so when people don’t have the same outlook as you we’ve gone from being lazy to now it’s a skill issue.
It's not exactly laziness. It's that most people don't understand basic financial concepts like fixed versus operating costs.
Companies like Anthropic and OpenAI have invested huge sums in building/training models. Those are the headline numbers everyone talks about. What they don't understand is that the gross profit margins on AI model inference are very high.
The question is whether these companies ever get to the point where the profit from inference justifies (recoups) the investment made in training. If they have to keep investing vast sums in training new models, and/or competition causes the margins on inference to collapse, they're in trouble. But if training becomes less costly and inference margins don't compress too much, it's a totally different story.
Can we be sure of this? The second they tried to hike the prices to be profitable, they lost a lot of clients, or a lot of clients lowered they token expenditure. In any market margins tend to go to zero. I am using frontier models, but only because they are dirty cheap if not free. If they raise the price, I just switch to a cheaper model, start exploring the option of renting a GPU-VPS or buying my own card. Or reducing my AI usage, that currently includes stupid useless things like "translate this cooking recipe from french... for free".
> But if training becomes less costly and inference margins don't compress too much, it's a totally different story.
They are being priced as the "ifs" are certainties, and not only a possibility. If anything goes a millimetre below perfect for them, their finances collapse. High margins attract competitors, and unless you have a huge moat (and their only moat is expensive training), your high margins vanish.
This just isn't true. You're talking about the textbook model of perfect competition, which is not the real world.
Commodity-like markets do tend to see outsized margins get competed away, but many markets aren't commodity-like.
Markets with network effects, economies of scale, high switching costs, significant brand differentiation, high legal and regulatory barriers to entry and heavy collusion/concentration don't function the way you claim, and there are a large number of major companies, many of them tech-related, that have delivered above-average returns year after year for many years because it's impossible for new entrants to flood the market and simply undercut them.
It's way too early to predict how this will play out for the leading AI labs like Anthropic.
The irony of your comment is that you're still using frontier models even though there are already much cheaper options available.
1. Model Build/Training costs are escalating if anything
2. They are in an arms race with competitors and can't afford to stop building new models - if they lose the perception of leading, they lose lots of customers overnight.
3. Inference margin is under attack from more specialised models and open weights models.*
4. They have no moat given how simple the interface to the products is - a bunch of text files and text prompts, and little prospect of building a moat.
So those all run counter the possibility of coasting on previous research. To me the most promising avenue for making money for these companies is to replace Google, not AGI or taking over the world's knowledge work. That would mean they are overvalued.
* https://strandsagents.com/blog/amazon-saves-millions-replaci...
> Companies like Anthropic and OpenAI have invested huge sums in building/training models. Those are the headline numbers everyone talks about. What they don't understand is that the gross profit margins on AI model inference are very high.
Only when you ignore that the models they provide inference for have an extremely high cost and pretend they don't depreciate. They keep having to train new models. I don't think they recovered the cost for a model when a new one releases.
It is kind of cute to say other people "don't understand basic financial concepts like fixed versus operating costs" when you seem to not understand basic financial concepts such as depreciation and amortization.
Also, and this is important, we don't really know their true compute cost. We don't know if their compute with the likes of Google and Amazon were initially subsidized, and for how long it might be. In Q2 they also got to not pay SpaceX for compute for a while.
Gross margin is revenue minus cost of revenue (COGS). COGS is basically the compute that serves inference (rented capacity, or depreciation on owned GPUs used for inference, plus expenses like support, etc.).
Model training is normally treated as R&D. Under US GAAP, R&D is expensed as it occurs and is below gross profit as an operating expense. It is not COGS.
You mention "depreciation" but depreciation applies to physical assets like GPUs. The GPU depreciation that belongs in gross margin is already in COGS. A trained model would be an intangible asset, so that is amortization, not depreciation. Even that only applies if the training cost was capitalized, and under GAAP it generally wouldn't be. The full cost is expensed when it's incurred. Nobody is "pretending models don't depreciate." Their cost is counted all at once but not in gross margin because that's simply not what gross margin measures.
Obviously, models lose value fast in an economic sense. They get replaced within a year or two and competition limits what they can earn. But that's an argument about whether the R&D expense is sustainable. It doesn't mean that gross margins are being misstated.
"Inference has high gross margins" and "training is extremely expensive" can both be true, which oddly seems to be something many people have trouble comprehending. But this is the crux of the questions that will determine whether these businesses are viable long-term:
1. Will training costs fall over time, or at least grow more slowly than revenue?
2. Does the gross profit a model earns from inference before it's replaced exceed what it cost to train?
3. Will inference margins hold up as competition and open-weight models push prices down?
You have a legitimate point about compute costs. If the deals with Google, Amazon or anyone else are priced below market, that directly inflates gross margin, because it lowers COGS. But we don't know those terms, so that's just speculation, and it's still a different argument from depreciation.
To reiterate what I wrote:
> The question is whether these companies ever get to the point where the profit from inference justifies (recoups) the investment made in training. If they have to keep investing vast sums in training new models, and/or competition causes the margins on inference to collapse, they're in trouble. But if training becomes less costly and inference margins don't compress too much, it's a totally different story.
They are when they say that inference is incredibly profitable. It is incredibly profits exist only when you pretend that models have 0 cost to be created (since they are a digital asset that can't depreciate), and R&D is a imaginary thing that exists only as a fixed cost of doing business that doesn't need to be recovered.
Anthropic released Opus 4.6 in February 2026 and deprecated it in June 2026. Was it anywhere close to recover the "R&D" costs during those few months?
More importantly, how long a model needs to be current until it recovers "R&D"? Their economic viability hangs on this, and the compute costs both to train and servr models.
A lot of people like to vaguely gesture at the possibility of compute costs coming down, but it is in fact coming up with the rising cost of hardware (and possibly with sweetheart compute deals being replaced by contacts where providers actually have to make money).
Expensing the cost of training a model isn't treating it as free. The whole cost is charged to the income statement the year it's incurred, which is actually faster than depreciation would ever recognize it. Calling something R&D does not make the cost "imaginary". It means the cost is counted. It's just not in gross margin.
We are actually asking the same basic question (can the investment made ever be recouped?). What you're struggling with is the accounting concepts and terminology. The costs of creating the models are a different part of the income statement than the costs of serving inference.
Again, "inference has high gross margins" and "training is extremely expensive" can both be true at the same time.
In absolutely none of my comments did I state that R&D never needs to be recovered. In fact, I stated repeatedly that the entire crux of whether these businesses are viable long-term comes down to whether the costs of creating the models is eventually recouped.
> Anthropic released Opus 4.6 in February 2026 and deprecated it in June 2026.
First off, Opus 4.6 is still available, even though Anthropic recommends migrating to 5.5. Google Cloud says it will be on offer until February 2027. So revenue for this model didn't drop to $0 overnight.
https://platform.claude.com/docs/en/models/opus-4-6/overview
More importantly, your comment implies that models cease to have value when the next version is released. This completely ignores the reality of model training, which is that the subsequent models almost certainly built on the research, data and techniques of 4.6. And 4.6 was built on the research, data and techniques of the models that came before it.
So the whole notion that point versions of these models are standalone investments that cost billions to train and then become worthless the minute a new point version is released is incredibly naive. It's just not how model training works.
> A lot of people like to vaguely gesture at the possibility of compute costs coming down, but it is in fact coming up with the rising cost of hardware (and possibly with sweetheart compute deals being replaced by contacts where providers actually have to make money).
You have absolutely no evidence of "sweetheart compute deals" so this is pure speculation.
Furthermore, there's a flaw in your logic. It's entirely possible for chips and memory to get more expensive while the cost per token served falls due to more efficient hardware and better inference software. What matters here is the cost per unit of output, not the price of a GPU.
Bottom line: inference has high gross margins and these businesses have had to invest massive sums. If inference profits grow sufficiently faster than costs, or costs decrease sufficiently over time, these businesses could be viable and perhaps incredibly successful. If that doesn't happen, their viability is questionable.
Lots of people have an opinion about this but as this discussion proves, lots of people also don't understand the basic accounting required to make a cogent argument one way or the other.
The problem I have with this is that the claim that "inference has high gross margins" is paraded around as something that automatically makes all this make sense.
My concern here is something that you described with more nuance, the cost of training cannot just set aside as if it doesn't need to be recovered. R&D makes sense if we consider that research and development eventually gets to a model that doesn't need to be trained anymore, and can be served at relatively low cost indefinitely. Otherwise the constant need to train new models needs to be taken into account; you can only keep revenue by training new ones.
> You have absolutely no evidence of "sweetheart compute deals" so this is pure speculation.
I agree that this is speculation. But it is based on actual sweetheart deals taking place (such as SpaceX not charging for compute during a period in Q2) and on Anthropic being very unclear if their compute cost is based on market rates or subsidized by providers who are also investors in the company.
> Furthermore, there's a flaw in your logic. It's entirely possible for chips and memory to get more expensive while the cost per token served falls due to more efficient hardware and better inference software. What matters here is the cost per unit of output, not the price of a GPU.
I agree that my statement does not take into consideration hypothetical improvements in efficiency, but I dispute that the increased cost of hardware is not a menace to their viability.
You may have a major breakthrough that halves the compute use for a task. It is meaningless if the hardware is now 3 times more expensive.
Except that this isn't the argument I'm making, and I don't think most of the people who are pointing it out are claiming that inference being sold at high gross margins means that you don't have to consider the massive investments that have been made.
The more common narrative is that these companies are losing massive amounts of money and will never be able to make a profit. And that's where looking at the investment in training versus the economics of inference is useful because, once again, inference actually does have high gross margins.
> R&D makes sense if we consider that research and development eventually gets to a model that doesn't need to be trained anymore, and can be served at relatively low cost indefinitely.
This just isn't true. If revenue grows at a sufficient enough rate versus training costs, and/or training costs decrease substantially enough, the numbers could work out over time.
People who are claiming that it's mathematically impossible are going on vibes. And just to be clear: I'm not claiming that it working out over time is guaranteed, or the likely outcome. But it is possible and the mechanics by which it would work out (if that comes to pass) is pretty simple.
> I agree that this is speculation. But it is based on actual sweetheart deals taking place (such as SpaceX not charging for compute during a period in Q2) and on Anthropic being very unclear if their compute cost is based on market rates or subsidized by providers who are also investors in the company.
Regarding SpaceX: my understanding is that as it was ramping up its capacity, which was not optimized and capable of being delivered in full at the time, it provided concessions, including discounts and deferred billing, to large customers. What SpaceX got from this in return was a massive backlog (over $14 billion worth) of non-cancellable contracts.
The concerns about Anthropic investors are legitimate, but again, you can only speculate. If your entire thesis ignores plausible scenarios that aren't so negative (like the possibility that compute is being sold at or near market rates), you have to be honest: the thesis is biased to support the argument you want to make.
If I was going to invest in any of these companies, I would want to know more about the terms of these deals but I don't think anyone who hasn't actually seen the agreements should make confident arguments based on what they think must be in them.
As a side note: if you really want to be exact, we'd need to know how compute purchased at below cost affects the income statement. If the below-market pricing was for compute related to training, it would not be the same on the income statement as compute related to inference. Whereas below-market pricing for inference compute would hit COGS and improve gross margins.
> You may have a major breakthrough that halves the compute use for a task. It is meaningless if the hardware is now 3 times more expensive.
What matters is the price per unit of compute. "The hardware is more expensive" only threatens viability if the price per unit of compute is rising, not just the price per chip.
You're throwing around random numbers to make a point but there is real data on this subject. For example, Epoch AI published a report a couple of months ago that found the cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023 even as hardware and memory prices rose by tens of percent. On one of its science benchmarks the cost per question at a fixed score fell 725x in less than 18 months.
It did also find that large training runs have grown by 2-3x per year, which is the crux of everything, once again: training costs versus inference costs.
> People who are claiming that it's mathematically impossible are going on vibes.
You are doing the same thing you claim I am, just on the opposite side of the road.
In truth this is sentiment based on the current narratives.
To be clear, I don't think it is theoretically impossible to be able to recoup the cost of training by selling inference, but I do think it is impossible with the current rate new models have to be trained and with current cost they charge customers. I believe that they will have to increase prices pretty dramatically for this to make some sense.
> As a side note: if you really want to be exact, we'd need to know how compute purchased at below cost affects the income statement. If the below-market pricing was for compute related to training, it would not be the same on the income statement as compute related to inference. Whereas below-market pricing for inference compute would hit COGS and improve gross margins.
My whole point is that this is unclear. Anthropic's S-1 mentioned how much that they paid for compute, but the rate they pay is totally opaque.
If they pay near the market price this is a nothingburger. If they pay, say, half the market price, now we can see their already eye-watering training costs increase significantly and their high profit margins to suddenly not look as seductive anymore.
Again, my annoyance is how prople pretend that Anthropic has a clear path to become profitable because of ther incredible revenue growth and profit margins (even if you are not doing so, this is the pervasive narrative in forums like HN). I sustain that this is far from settled.
Except that I'm not on the opposite side of the road. I've simply pointed out that the mechanics by which Anthropic could be viable long-term are pretty simple; not that this is what is going to happen.
Frankly, I don't think anyone can predict what will happen. This is a fast-moving, complicated stew of technology, finance and politics.
> My whole point is that this is unclear. Anthropic's S-1 mentioned how much that they paid for compute, but the rate they pay is totally opaque.
> If they pay near the market price this is a nothingburger. If they pay, say, half the market price, now we can see their already eye-watering training costs increase significantly and their high profit margins to suddenly not look as seductive anymore.
Pricing in commercial agreements is typically considered competitively sensitive and can be redacted from contracts filed with the SEC but only if it isn't material. Anthropic still has to disclose known trends that are likely to affect its results so if below-market compute were inflating its margins you'd expect to see that in the S-1.
The more detailed related-party disclosures probably aren't required yet because Amazon and Google mostly hold non-voting preferred stock and convertible notes. The S-1 says they'll each hold more than 5% after the IPO conversions. At that point, their transactions with Anthropic fall under the related-party rules so we'll learn a lot more after the IPO if there are any "issues" there.
> The more detailed related-party disclosures probably aren't required yet because Amazon and Google mostly hold non-voting preferred stock and convertible notes. The S-1 says they'll each hold more than 5% after the IPO conversions. At that point, their transactions with Anthropic fall under the related-party rules so we'll learn a lot more after the IPO if there are any "issues" there.
Whoa, that I really didn't know. Something for everyone to keep their eyes on then, because this is actually very opaque, and may dramatically change what we know about their numbers. Or not.
Isn’t it plausible that this is conceptually wrong? What if a model’s revenue-generating potential only lasts for something like a year due to, loosely speaking, competition? Doesn’t this make model development more akin to, say, the costs of annual agricultural planting/harvesting/delivery cycles, or a tv network’s annual (back in the day at least…) production cycle for its main content?
Under US GAAP, production costs are capitalized when they're incurred and then amortized into cost of revenue over the period when the content actually earns revenue. Because content usually earns more revenue early and then tapers off, the amortization is usually front-loaded.
The one problem with this is that not all training actually goes into what's sold to customers. There are experiments, failed runs, various studies, etc. So you would need to come up with a way to cost only the final production run (and perhaps post-training) to come up with a number akin to content cost.
ASC 730 basically favors expensing anything uncertain, so unless you have a really good, defensible way of breaking out these costs and separating them, doing it this way would go against the current accounting standards.
Would this be analogous to, say, paying lower level employees to read endless scripts and write coverage, purchasing some of those scripts, producing but then canning failed pilots, etc. obviously some of this is outdated, but I’m at least thinking more of the pre-streaming era of television production pipelines, in which there certainly was a non trivial amount of content produced that never reached consumers.
Thanks for your more detailed response; I obviously know nothing about the real intricacies of GAAP so it’s interesting to understand the analogues better.
If you bought a script, that would be capitalized as an asset, and written off if abandoned. And so on.
The difference is that film and TV have an industry-specific standard (ASC 926). AI companies fall under the general R&D and software rules and those don't allow capitalization when there's significant uncertainty about whether or not the project will work. There's a new software standard (ASU 2025-06) that makes this explicit with a "significant development uncertainty" test. Because you don't know the outcome of a model training run until it's done the training costs are generally expensed as they are incurred.
Obviously there's room to debate whether the rules should be updated and maybe they will be in the future. But they are what they are now. Expensing model training as R&D isn't an accounting choice companies like Anthropic have made on their own. It's what the rules require. So obviously, understanding what the rules are and where different costs show up on the financial statement, it's clear that inference gross margin alone doesn't speak to whether the business is ultimately viable. But the high inference gross margins shouldn't be dismissed as meaningless or the product of some sort of accounting trickery either.
No problem if this is so far off topic as to no longer be interesting to you… I just don’t often engage with somehow who can comment on Hollywood accounting with actual knowledge of relevant standards!
There's a formal process for these rules to be made and it's designed to be responsive to the needs/requests of businesses, investors, auditors, etc. While I wouldn't say there's any room for differences of opinion, given the rule-making process, I don't think there are any fist fights taking place over these. :)
the gross margins on growing potatoes is very high, if you exclude all the other costs needed to get the potatoes to market.
Think about it, you take some potatoes you had left over from last yeah, dump them in the ground and wait 9 months. Boom pure profit.
This of course excludes the cost of the land, machinery needed to plant, till and harvest, storage, pesticide, fungicide and herbicide.
Then the transport, cleaning and packaging.
If you arbitrarily cut out all the bits of the business needed to make revenue, everything is a profit.
A company selling potatoes would not include land in COGS. Land is a a fixed asset. Even if a farm rented/leased the land, the payments would fall under OpEx, not COGS.
Seeds, pesticide/fungicide/herbicide, direct labor, water and electricity for irrigation, etc. would all be COGS.
Would it make sense to spend $1 billion on land to grow $1 million worth of potatoes every year? Probably not. But that doesn't mean you couldn't sell potatoes on your $1 billion land at a positive gross margin if revenue exceeded COGS.
This is the AI debate that is impossible to have when people refuse to educate themselves about basic financial accounting concepts.
ARR based on October revenue 400% higher than GAAP revenue.
Net Revenue Retention: 1,000% After all one seed potato creates 10 more potatoes - infinite replication! We estimate soon we will 10x the no of potatoes in the world.
EBITDA - the severe frost last year is a one-off non-recurring event, fertiliser is an R&D cost, seed is a growth investment. Land rent is not considered a core cost and therefore removed. Depreciation, interest, taxes all taken off the figures because somehow they don't count.
No man. It’s that, like you’ve explained further in your comment, they haven’t become net profitable yet. Being profitable in 1-2 quarters is a good trend but it doesn’t matter if you are still behind on mountains of debt or investors wanting their money back and your market is one where people talk about shit that happened 4 months ago as obsolete.
There is no proof that the current profit they are making on inference is durable.
If VC money gets tight, they would have to issue debt, that pay interests. With a current negative EBITDA, their bond will be junk at IPO, paying higher interests and having a lot less market than AAA debt (e.g. Google has access to that market). Both things combined translate to "running out of cash".