If you are growing revenue at a high rate then taking profit is a misallocation of resources. That is short-term thinking. It is much better to reinvest in revenue growth.
You can take small profit now or much larger profit later. Insisting that companies need to be profitable even when growing revenue rapidly is failing the marshmallow test.
But we don't have visibility on the COGS until they IPO and file, right? So where is all this premature judgement coming from about their unit economics?
(not pointing the finger only at you, at least you identified that gross margins is the correct thing to look at rather than net profit!)
I think a lot of us are calling it out because it's a contrast to SaaS/ads businesses where incremental goods sold are practically free compared to R&D, so spending can be looked at as one-time investments. There's very little additional COGS per additional customer in those businesses, so the default assumption to treat AI like other tech businesses has a blind spot.
> But we don't have visibility on the COGS until they IPO and file, right? So where is all this premature judgement coming from about their unit economics?
It's the lack of visibility that causes the judgement; Were the numbers good, it's quite unlikely that Anthropic would be so reluctant to share them.
Were it just Anthropic doing this, it's not much evidence. But it's EVERYONE that obfuscates their numbers, even the publicly traded companies.
Why would Amazon and Microsoft obfuscate the revenues and costs of their AI products? Even their cloud numbers are less clear than desired. And beyond those two, why would the datacenter companies obfuscate their numbers, when everyone desperately needs them to raise debt and investment to build more DCs?
Pretty much the only company showing clear numbers is Nvidia & GPU orders. But immediately beyond that, it's all obfuscated. How many GPUs are sitting in datacenters? They ain't telling.
Isn't it illegal for publicly traded companies to hide such information from their shareholders? What do you mean by "obfuscated"? The numbers are disclosed, but you think they're unclear somehow?
But what they can (and do) do is structure (in the not financial jargon sense) the reports such that the given datapoint does not exist individually.
E.g. If you want to hide AI revenues or costs, the SEC won't let you just _not report them_, but you can just group the AI revenues with the SaaS/Cloud numbers under a new division, and report only the combined figure for that division.
This works especially well if the grouped components already fluctuate a bit, so one cannot simply substract known SaaS/Cloud numbers from the new total.
I buy 3. But 1, 2, and 4 rely on models continuing to improve at the same rate, such that you need the latest version to stay competitive. At the cut below frontier models, there's already robust competition between open source models, cheaper providers like Deepseek, more local AI alternatives, etc.
I think the case for the unit economics being fine starts to fall apart if you can't charge a large premium for your best in class model.
"Best in class" is a broad definition. Anthropic can easily charge a large premium for their "best in small class" and "best in medium class" models, for any given definition of "small" and "medium"
Inference has dropped by like 75% from a year ago. While anthropic does offer more tokens now for the same money, the value of the business is based on an expectation of future profits. There are dozens, if not hundreds of examples of companies being valued this way.
You could be right but the unit economics matter to this business in a way they don't for a SaaS or ads business, they're not free and you can't just point at revenue.
If a system can perform or positively augment the work done by a human, especially knowledge workers, then it’s got value, it’s just quite hard to put a finger on what the extent of that value is even now, let alone next month.
If you give me $100, I will give you back $101, funded by equity raises.
Very high growth, very high revenue, huge customer satisfaction.
We hope to be profitable one day, already foresee a mechanism to double profitability per transaction and also double the number of transactions our customers perform.
Now let’s see you do this with 40B. The f you can do that I will be intrigued, since it sounds like you solved a bunch of complicated economic problems.
Of course you can sell $1 for 90¢, but your unit economics look terrible. If you want to seriously critique Anthropic you need to explain why their unit economics are bad.
Given that they just filed a (confidential) S-1, we will get an answer soon enough.
I expect it's going to look a bit more like selling 38 billion dollars for 40 billion in revenue^ than your example.
^ Some other caveats about how they're marking their p&l, but I think if the growth continues and they have a durable moat^^ then this will look like Amazon and be able to pivot into higher margin stuff
^^ haha this is the biggest condition, but I'm optimistic
That's not how classical valuations worked though. I was taught the rough shorthand for valuing a company was profits / real_interest_rate, treating the company like a perpetuity. Revenue ain't in it. Now we have a bunch of "We'll make it up on volume companies!" like https://www.youtube.com/watch?v=CXDxNCzUspM
Net profit is the wrong metric for high growth companies. You want the unit economics.
If the unit economics look great (my understanding is they're at least fine-to-good), then they should not be taking profit because it is very much in their interest to do capex to grow their capacity so that they can continue to grow.
It's a knife's edge because if they invest too aggressively it could lead to bankruptcy, but so far they've actually been quite conservative and given their unit economics they can afford to pay $$$$ for expensive inference compute (and indeed that's why they've inked a bunch of deals in the past month or two)
What do you think they are, what do you want them to be, and how much revenue and growth do they need before low margins make for a monster of a company?
From all their desperation in making sure api keys are not used in contexts where they are not supposed to, I would say that they actually appear to have services where their profit is negative, if a customer is actually using their api to the limits they set, they lose money. They wouldn't have been this desperate in trying to shut off OpenClaw if it wasn't this way. Most companies that provide api infrastructure love when a killer app using their api is made by outsiders.
And while you can beat low margin with scale, there is the famous joke "we lose money on every sale, but make it up in volume".
If you scale a low margin operation, you can become giant. If you scale a loss making operation, you go bankrupt.
Yup - the subscriptions are a VC subsidy. They've been phasing their enterprise customers directly onto API-pay-per-usage pricing (hence the recent reports from Uber and Microsoft about phasing out Claude code). Rest assured, many of their customers are happy paying for the value they get from Claude code.
The subscription is the loss leader to show you how good it is. And people think it's good and worth paying for.
There is some reason to think their margins will improve, also: they couldn't really plan for the capacity they've needed so far this year, so they're paying through the nose for it. That's fine because they can pass the cost onto customers and give a more reliable service at cost. But in a few years, they should be able to get those costs under control (presuming some ops excellence. Something Google has in spades)
I don't know if it is a chess move but Elon certainly will take every possible opportunity to fk over Sam. Have to admit watching these two sociopaths duking it out is somewhat entertaining.
> Doesn't inference have very good profit margins* but all the losses come from training?
There is no source for this. Amodei just pulled a hypothetical explicitly distanced from Anthropic out of his ass and kickstarted some citogenesis when people half-remembered that number and started quoting it as truth.
The only material claim of Anthropic is that they would "turn an operating profit of $559 million in the June quarter ... The company might not remain profitable for the full year as it plans spending increases due to its vast computing needs." with an explicit disclaimer that: "It is unclear what accounting methods Anthropic has used to book revenue and costs, as the company isn’t yet required to follow the financial-reporting requirements of a public company."
This is the exact same quarter where xAI is giving them deeply discounted compute, as such the numbers cannot be projected out to the later quarters once Anthropic has to actually pay xAI for the compute they use.
Finally, there's the reality that were the revenue numbers any good, Anthropic would just publish them and leapfrog OpenAI. That they do not provide clear GAAP numbers suggests the numbers are bad.
This is a very interesting point. Is there a similar list in India, Japan, Korea or any other country? Will be interesting to compare them and see if some of those books on the other lists are translated.
You don't have all the relevant invoices etc at on time. Some of that takes quite awhile. Especially inter country purchases and sale transaction information.
This doesn’t get better when you have a quarterly or semiannual deadline. It’s just the scale might be smaller. However you would handle them in the same way and either disclose on an accrual basis or on a cash basis, but either way, you do it as you know it.
I think the most important is number 2. People are now looking for things that are made by humans. Most detest AI slop. And if they find out that you are peddling slop, you lost trust.
It seems to me that we will go through the same phases that chess went through when chess on computers became a thing. First, people thought that this will kill chess, then people start using it as a tool to play better chess. Now, chess is thriving, despite AI being used in chess. I can see a similar path with art. Using AI to generate ideas, still create art by humans.
Crucially, chest is also thriving as a spectator sport - and what's drawing the views is not the high-level matches: people are far more interested in more fast-paced and casual content, where the personality of the streamer can shine through.
On the other hand, absolutely nobody is watching livestreams of two chess bots playing each other. They might technically be better at chess, but that doesn't mean it makes for entertaining content.
reply