Are Anthropic’s AI Economics Finally Working?

[00:00:00]

Mikhail: are AI economics finally starting to make sense? What we've learned recently with Anthropic public filing, or S1 filing, is that we're starting to see numbers that actually add up instead of like what we've assumed, because let me be clear, we haven't seen any real numbers. We've had assumed numbers and leaked numbers, but we've never seen a public-- one of these massive companies go public with their revenue numbers that we can trust.

This is a public filing that has to be vetted, that has to be audited because it's allowing them to be listed on the stock exchange as a public company, and we can buy shares in the public company, right? So this is a legit offering, so they have to go through these numbers pretty intensely. These must be legit, otherwise they could be, you know, sued for fraud,

And what we're seeing with the latest release is that there is revenue, and [00:01:00] revenue is strictly outpacing compute spending. Now, preface that with there is a ton of debt. I'm not gonna get into specific numbers 'cause I'm gonna bore you with the numbers, and I don't understand all the numbers either.

I just know the correlations between the numbers because I read the summaries. All I know is that the 2026 revenue for Anthropic is 14x what it was a year earlier, the first half of 2026, I should be clear. And compute and in- infrastructure expenses obviously also rose, but they only rose by 3.8x, which actually fell, like the amount of, like the amount of increase fell from 201% the year before to 55% the, the 2026 first year.

So that means that expenses are still rising, but they're rising a lot slower, and revenue is rising exponentially faster. So we're seeing this correlation of like, hey, they're profiting from inference with [00:02:00] the revenue, and they're able to, get more efficiency out of their current hardware and the hardware that they're, they're, they're making, right?

They're, they're, they're starting to optimize. We talked about this many times on the show where I'm like, "For the past six years, all the AI companies cared about was intelligence." That's all they cared about. All of the money, all the research, all the R&D went into how to make the models more intelligent.

Well, I think over the last maybe year, year and a half, they're starting to shift that where they're spending a significant amount of money on, compression, optimization, hardware acceleration, et cetera, et cetera. There were so many knobs to turn to make it cheaper that I was almost positive that they would be able to figure it out because they hadn't, they had d- they had done nothing.

They were ju- they literally were just like, "5090s? Oh, yeah. Let's get to 5090s and just throw compute at them with no consideration or very minimal consideration for how to actually, like, optimize them." [00:03:00] They were like... I listened to a podcast on, uh, like the, the new gigafactories that Elon was creating or the new giga data centers, these massive data centers that he was creating, and they were just finding buildings, like, just like old construction, old, uh, warehouses, and they were just literally rolling racks of GPUs into them, just r- like, they weren't, like, retrofitting them or anything.

It was just racks with wires fricking everywhere just trying to get them up. Like, there's no efficiency there. Like they're burning, they're just trying to get compute up. There was no time for efficiency. There... It's a, it was an intelligence war. But now we're seeing the first steps and the first shots fired in the efficiency wars, and with this IPO filing, there is hope And clear evidence that they are able to profit from inference.

So they, there is a healthy margin. Like, we're, we're paying a lot more, like API costs specifically, and potentially even subscription costs. I haven't actually done the, the, like, that in-depth research, 'cause this came out really recently, uh, to see if the subscription is also maybe covered to a [00:04:00] cer- to a cer- certain degree by these costs, by, by the revenues.

'Cause they're making margin. They're not, they're not operating at 0% margins anymore. They are making a, a healthy margin. And that poses a lot of questions that I wanna throw at you, Matt, and we can discuss. the first one is, is the bubble not popping really? Like, does this change the argument that AI companies cannot make money because inference is too expensive?

Matt: I don't necessarily think that that, that this is gonna stop the bubble. I think it's positive news for the economics of AI companies, and I think it proves that they could potentially be profitable. But we also have to remember the amount, like you just said, of, of inefficiencies that were put in place in order to build things to this amount.

And I also recently listened to a, a Diary of a CEO episode with Mark Cuban, and I, I don't wanna put words in his mouth, but His idea behind AI and some of his ideas are really interesting to me, and one of them is that the AIs [00:05:00] themselves are not getting smarter. The way we make them smarter, the way that we make them do more, is we throw more compute at them, and that doesn't really reflect intelligence.

Again, I'm paraphrasing what he said or how I understood it. I, I'm gonna link to that episode inside of this, the show notes so you can listen to the episode. I, I, I enjoyed it myself, but I think that that's really interesting is I ... Exactly what you were describing. They're going into these warehouses, these construction zones, whatever, and they're throwing, they're throwing compute at these things. Meanwhile, a human can go from having the exact same organs and parts on board, if you will, and can learn get better at multiple things. Can get better at doing exercise, can get better at reading, can get better at biking around. Can get better at all these things. Meanwhile, "Oh, I want my, need my AI to be faster.

Give me another $10 million in GPUs. Give me another, you know, $2 billion in RAM. Give me another $5 million in, in all these other computing things. Give me another [00:06:00] $2 million in networking infrastructure." More, more, more, and the actual AIs are not getting more and more intelligent. So I think that it Maybe there, maybe it strengthens the, the wall of the bubble if there is one. I'm not an economic expert. But I think that there's a lot more to this story, obviously, than what we're gonna be tackling in just a Web News episode here. Because there's a lot of debt that's been taken on. There's a lot of public questions, a lot of public protests against these AI data centers. There's a lot of public protests against the use of AI in general. There's a lot of questions as to what the regulations will do, and my whole point there is is that it doesn't just take good economics for a bubble to not pop, but go away. It, it can be popped by bad regulation that comes in, bad in the c- to the actual AI companies where it blocks their training or it, it hurts how they do things or it hurts how they [00:07:00] buy hardware and now all of a sudden maybe the debt collectors come out and they say, "Well, you were training at this pace.

You're not training at that pace anymore. Where's my money?" And pop goes the bubble. Who knows? Who knows?

Mikhail: I, I definitely agree with you. I, I don't think this is a direct line to this being a bubble pop. I think if anything, this proves that a company with infinite resources is able to figure shit out. but the reality is, is that there's a lot of smaller companies that won't be able to figure this out, and they're re- they're solely relying on the invested money that they have.

That might not be enough to figure it out. the amount of money these companies like OpenAI and Anthropic are burning is astronomical. Like, H1, we, we still had massive losses in the first half of 2026. 80, 80.74 billion dollars in losses in the first half of 2026. The- [00:08:00] ridiculous numbers.

Absolutely ridiculous numbers, right? So the only thing that I see this staving off is this is a push to be like, to show investors and the public, because this is a public offering, so the public is going to be investing money. They are going to net a big chunk of change from the public offering. Let me be clear.

they're gonna make tons of money, or not make, but like be- get up tons of money for their shares It's a, it's a way for them to show that, hey, we've got... We, we have a path, right? Like, this is a path forward. There is money to be made here. We don't have... I don't think they have a direct que- answer to, like, their astronomic evaluation yet.

I think all of their answers is, like, super intelligence, and that's not really, like... Oh, what does that mean? Like, what... How does that, how does that actually make money? No one knows that answer, right? Like, no one, no one has a very direct line for that, and I think that is a very high risk factor for the bubble popping is, is we don't have a direct line to, the valuation making sense.

I think the valuation they're [00:09:00] trying to get to is 2 trillion, which would make it the biggest IPO offering ever. Public IPO, like, a, a first public IP- IPO offering ever. I, I don't know if they're gonna get to 2 trillion, but that's what they're trying to. It's gonna be probably somewhere between 1.5 and 2.

Ridiculous. Like, absolutely crazy amounts of money. But this is like a, "Hey, we're on a, a track now." Like, there, there's something to be said about the fact that they have figured out a way to make money from a product that didn't exist even, like, two or three years ago. It, you know, like, from an investment point of view, you're like, "Okay, maybe I trust these guys a little bit more.

Maybe I'm willing to invest a little bit more money because I see the outcomes that they're, that they're getting to, even though I don't see the poten- like, I don't see the direct line to the potential outcome, like, of the valuation. Like, they... I see that they are able to figure it out." But these are these big companies.

Like, OpenAI's probably gonna have something similar, if not worse potentially. I ha- I have a feeling that they're delaying their [00:10:00] public offering because they're trying to make their numbers make sense, like Anthropic has. Um, but regardless, it's gonna be probably similar But the other smaller companies, it's gonna be tough.

And the bubble popping doesn't mean that all of the companies go down. It, it probably means that majority of companies go down. So I still think there's a high risk of a bubble pop. It's not gonna be a complete bubble pop like we've talked about. It's gonna be a contraction. Whatever you wanna call it.

It's probably gonna happen, in a more contained sense. Probably not on, probably not Anthropic and OpenAI, or at least both of them, for sure. They're gonna be around. We can see their economics now. Now, another question for you, Matt. do you think they would ever pass the savings that they're getting onto us?

Do you think the costs will, like, for us, will always be rising? Like, we've seen that the-

Matt: of, in terms of like the plans to

Mikhail: The plans, the, the, the tokens, like the, the straight API costs, like do you think we'll ever, they'll ever go down now that we've seen that they can make money on it? Init- initially our [00:11:00] assumption was that they were, they were, even, even with us paying them, they were burning the money on the, for that compute.

That was what everyone was saying, you know, three, four months ago. Everyone was saying that. But now we see that that's not the case. So could they pass the cost back down to us? OpenAI just released a $500 subscription. That seems like a lot. Um, would we ever be, would they ever be like, "Oh, we're removing it, going back down to just the $200 and we're giving you more usage"?

Matt: I could see it due to other people innovating in the space.

Mikhail: Mm-hmm

Matt: I, I don't know what's gonna happen with this whole super intelligence idea or how intelligent they're gonna be able to make these things realistically, because at the end of the day, I mean, we could, we could order more graphics cards today than have ever been made in the history of mankind. We could, we could do that, and it would be very detrimental to the economy, and there'd be a whole bunch of other different things, but we could, we could technically do it and slowly but surely produce those and then throw it all at a model, and it could still not be a super [00:12:00] intelligence. It could still not be all that impressive. I think, personally, I think one of the things that is going to absolutely disrupt them is innovation, meaning random, random agents that just come out at the consumer level, something like Muse. Uh, there's even the Muse pendant and things. I'm not saying that those are the end-all be-alls. My point is, is those are innovations that are coming in in different ways, using the technology in different ways.

Having Muse just have a Linux box in the cloud for anyone who just signs up currently, like, whether that's profitable or not profitable or whatever, you know, remains to be seen. But my, my point is, is that I think we're gonna see a lot of different innovations come in. There's been some AI accessories, some AI little PDA handheld things.

There's been AI pins earlier, and some of these places have not done well. Some of them have gone bankrupt, some of them are whatever. But there's in-- the point is that there's a ton of different ways that these companies are, quote unquote, being attacked from a competitive standpoint, [00:13:00] where it's possible that we might start seeing, man, everyone has this, this, like, not a smart spe- uh, actually, there's a perfect example. Back in the day, smart speakers weren't anywhere. Now they're kind of in a lot of people's houses, like, quite a bit. Maybe you're gonna have, like, a little smart dock that has a little local model that runs just over your Wi-Fi and does a lot of your local AI computing and then only calls upon the cloud, calls upon the big amount, the big, powerful data centers when it needs that, if, and, and it intelligently decides when to do that.

We saw something similar at a very rudimentary level. something similar when, might've been iOS fifteen or something, when Siri, a lot of Siri's computing went on device. Now, obviously, Siri is not an LLM or, I know that they have Apple

Mikhail: They have that, yeah.

Matt: in the day, Siri was not in the same idea.

The-- it was not-- Siri was not an LLM. It is

Mikhail: No

Matt: definition of what we're talking about today. My point is, [00:14:00] though, is they brought that on device. entirely possible that there, it'll become a thing where if you want this AI, here's a little smart speaker looking thing, you put it near your router, and this thing will m- will do a lot of your compute for you because even Sam Altman said there might be almost a utility charge for AI where it's pay as you go, and it's kinda like your electrical bill or your water bill where it's, like, a thing that we actually use a utility. Well, this would be a way that people would say, "Okay, I'm just gonna try to save some money here." And

Mikhail: Угу

Matt: this already with m- standard utilities, low flow- Toilets, and more energy efficient light bulbs More energy efficient appliances. The list goes on and on. And so I can see them being hit in varying ways, which will then, to go back to your question, which will then naturally cause them to have to rethink how they price things

Mikhail: Yeah, competition. You're, you're 100% right. Any sort of competition is going to make it so that the prices have to go down. And n- now that we know that they're able to get the compute down, [00:15:00] I think it's, I think it's inevitable that they're gonna have to compete with the DeepSeeks of the world when they figure stuff out, if they figure stuff out, the GLMs, the open source models that come out that are able to compete with them on a direct level, right?

'Cause right now they are able to compete, but there's certain things that, like, just it's not the same, it's not the same. I can't replicate an Opus, especially the latest Opus, with any of these older open source models. At some point they might get to the point where they-- we can, and that will change things quite a bit.

And that, that in my opinion is also another huge risk for the bubble popping, is if they aren't able to innovate, for example, like they aren't able to create these new devices that they could sell or the new workflows or the utility, capabilities of AI, and the open source cheaper models catch up, their margins are gonna go down significantly and quickly, right?

on device models I think are still ... [00:16:00] I mean, there are, there are some decent ones, let me be clear. Like, like Qwen has a couple decent ones out there right now that are able to maybe compete with like a lower level Luna model or something like that. So there's certain things that you could definitely already pass off to those models.

But on the grand scale, they're not there and, uh, due to hardware-- due to the current hardware situation where like the worldwide shortages of RAM and stuff like that, I think that almost moats the capabilities of local AI models for a little while, because we're just not able to get significant enough hardware into consumers' hands to create a local, um, a good local experience.

Matt: Yeah, but

Mikhail: Like,

Matt: are also burning the candle at both ends though, because you

Mikhail: we are

Matt: how many, how many users are giving prompts that actually require a lot of computing power? Probably not a lot. Many of them are asking for help in daily things. [00:17:00] They're... or they're asking for things that are, that could have been asked to Google Search even five or 10 years ago. 'Cause people are not all in the tech space. They're not all engineering the next b- greatest app. They're,

Mikhail: Yeah

Matt: s- questions like g- find me the best burger joint and, organize it by enter in a metric here. Find me the best burger joint and compare all the Yelp and the Google Map reviews and give me a PDF.

That's not a super difficult thing for it to do. My point is, is that if that's using, let's say, 10 compute units right now, in a year that might use five compute units, and then in 10 years it might use one compute unit. So then we're gonna be using less computing to begin with, and that will... How mu- how big of a percentage?

I, I think it's a lot. How big of a percentage of the everyday user's prompts, much are th- are gonna fall under that, what I'll call a mundane level? I've mentioned on this podcast before that computing has become mundane, meaning back in the day we [00:18:00] had to have a video conference, we gotta go to the conference room, we gotta get WebEx fired up, we gotta have all this stuff.

They gotta have a WebEx set up. It's gonna be a whole thing. Now, I could be driving, down the highway and the person sitting beside me in the passenger seat could be doing a video conference from their phone. They're conferenced into four different countries. o- they're doing stuff OTA.

They're doing... They're using the 5G. They don't have a big boardroom. They don't need all that stuff. Look how mundane, in terms of computing, look how mundane it's become. Look how quick we can communicate now. Look how quick we can send photos and video and things like that now. My point there is even when you and I were going to college, Mike, which realistically wasn't that long ago, we used a lot of USB sticks. No, but it, it wasn't that long ago in the end of the day here, we were using a lot of USB sticks. Cloud computing, cloud storage just kind of took over suddenly and now it's just a part of everything. Oh yeah, there's Steam Cloud storage. Oh, there's Xbox Cloud storage.

There's Nintendo game saves cloud storage. Everything's in the cloud. Don't worry about it. And then [00:19:00] that's just mundane and we just move on.

Mikhail: I think it's been over 10 years since we've been-- since we graduated college

Matt: this AI thing, and it's only been a couple of years since it's really hit the mainstream,

Mikhail: Absolutely. Yeah

Matt: 20% of the timeframe that you're talking about. Look

Mikhail: Yeah

Matt: look how quickly the conversation changed. If there was no AI stuff right now, we'd be talking about how to frigging code in PHP or something right now

Mikhail: Yeah, I think if they could survive 10 years, they'll figure, they'll probably figure it out. I, I, like, my worry is that the bubble pops or the bubble could start popping within, like, a year and a half or two years, which is ... I don't know. I feel like the timelines have now changed completely h- of how I evaluate anything lately.

Like, it's

Matt: because of this initial offering?

Mikhail: No, no, no, not because of the-- just in general because of AI. Like how AI's come onto the scene, how it's disrupted the economies and different perspectives. Like, you know, the, the, the shortages of RAM and all that. Like all of that kind of contributes to me just [00:20:00] like, I don't know, I can't predict anything in the next year, let alone two, let alone five.

Like 10 years seems like a lifetime away at this point in terms of progress from all these different sec-sec-sectors. And I th- I think I almo-- I, I 100% agree with you that in 10 years' time they will have probably reduced the 10:1 compute units. I,

Matt: It's

Mikhail: guarantee you

Matt: mundane

Mikhail: gonna happen. Yeah.

Matt: up Google Maps stuff. It's gonna be-

Mikhail: Yeah.

Matt: Look, look at even how we ingest media. When we were in

Mikhail: Yeah

Matt: there were video stores on every corner. From the beginning of college, it was approximately a four-year period. From the beginning of college to the end of college, the entire VHS video industry c- basically collapsed. the large lar- the

Mikhail: Yeah

Matt: chains largely collapsed. People stopped renting and started streaming, and we had streaming years and years and years prior

Mikhail: mhm

Matt: even, right? But this is on the forefront. Money's being poured into this. Like [00:21:00] back then, streaming was held back by people not having high-speed internet and yada, yada, yada.

Infrastructure gets better. It's not just the infrastructure that, that gets better 'cause people always want the infrastructure in ca- in the case of internet to become more reliable and faster obvi- and cheaper, of course, right? Of course. So that happened to some extent. At the very least, it's, I don't know if it's cheaper, but it's definitely more reliable, and it's definitely faster.

So therefore, it's more accessible 'cause it, uh, there are decent plans out there, even at slower speeds, and those slower speeds are way faster than what they were 10 years ago. The point of that being is no one thinks, "Oh, b- I have faster internet, so now I'm gonna stream Netflix," which was new at the time.

They think, "I'm gonna go do this, use the same websites that I did before, but they're gonna be faster than ever." That's what the consumer thinks. But then the entrepreneurs think, "Oh my God, I have this streaming platform. I can now get this into more homes than ever," and that completely tanks the industry.

There's, it's too... There's too many parts to an economy, there's too much innovation. There's too many mistakes in order to, [00:22:00] reliably predict this stuff, but there are patterns. Like, things

Mikhail: Yep

Matt: are easier to compute today. That's something that, that almost always, if not, has always happened

Mikhail: Yeah, so if we're, if we're looking at the patterns, then realistically what's gonna happen is costs have gone up now. As the competitors come into spa- come into the space more and stuff gets more commoditized, the cost will have to go down. As the com- competitors start buying each other up, similar to what we've seen with the streaming space and the TV space, they're going to start price fixing and setting costs higher, and so the cost will start going back up.

So we're gonna see this, like, wave of, costs will go down, it'll get commoditized, and then they'll try to figure out ways, whether it be through, you know, conglomerates or whatever they wanna do, to bring the cost back up to get their margins higher, et cetera, et cetera, et cetera. That's ... I mean, if we're looking historically, probably the case, the, of, of how this is gonna go.

maybe we could take advantage of the time when the costs are down, great, now you [00:23:00] know. but we don't know. you're right, like, it is impossible to predict. Like, I think TLDR, probably the cost will go down at some point soon

Matt: that if there is a bubble that pops, we have to look at the, the value to life that these things have. But you also have to look at the goal that these-- Like, we're talking largely, uh, is this a normal IPO? Maybe legally speaking, but

Mikhail: It's diff- it, it's not, it's not completely normal yet

Matt: trying to just make, you know, the greatest app ever and is trying to make the greatest app ever and get as many users as possible and yada, yada? That's part of it. That's part of it for sure. They

Mikhail: Yeah

Matt: this to the, to the consumer. But there's also almost a research angle to these companies where they're trying to get to this super intelligence.

They're trying to push the, the, the margin. They're trying to get to these different goals, and there's been tons of talk in those areas about what are they trying to do. They're trying to automate themselves. They're trying to push into this area. They're trying to push into that area. They're trying to...

It's not all about profit. Maybe at the end of the day when the dust settles, and it, it may be it will be all about profit. But right [00:24:00] now it's about I need to get the smartest model out there because my competitors are chasing me and, and, you know, smarter, more capable, faster, et cetera, et cetera, et cetera, is what's selling it currently

Mikhail: I think that's it. Yeah, and I, and I think there's so much unknown in that space that no matter how much you predict, they just don't know, right? Like, they don't know what the superintelligence would even bring. Like, they're betting superintelligence is gonna make them money. Will it?

Matt: Is it even gonna happen?

Mikhail: Yeah, is it even gonna happen?

Like, they don't know. Like, it...

Matt: the code in, of, of the singularity?

Mikhail: Yeah

Matt: If they crack the code of the singularity, now the m- now the business changes. Do

Mikhail: Yes Nope. That, yeah, it's a completely, it's a completely new thing, and they don't have the answers to it, neither do we. And, like, it's a very different ... I, I agree that this is a very different offering than a regular IPO,

Matt: It's not an

Mikhail: for ... Yeah, exactly.

Matt: well

Mikhail: And you're, you're ... The, the [00:25:00] difference here is that, like, there could be legislation that comes out that sig- significantly destroys the value of the IPO.

Let me be very clear. So I'm not ... This is not financial advice. Don't go investing your money into Anthropic just from listening to this episode. do your own research. But, like, there, there's a lot of ex- external factors that could happen, or even internal. They're retaining over 50% ownership even through the IPO, so there, there's no way for a, a board, an, like a, uh, to be formed from the IPO or, like a c- uh, you know, an Elon/Peter Thiel conglomerate try to take ownership of the company.

They're making it

Matt: not possible due to them owning the majority of the

Mikhail: Correct. Yeah, they're, they've, they've established it so that the seven cofounders own 50.5%.

Matt: Unless they

Mikhail: So

Matt: turn one of the people that are within that 50 some odd percent or what have you. There's

Mikhail: far as I understand, they have, they have, um, they have a fight... Like, they, they have a way to stop that. Anyway, so regardless, they, they're, they, they've made sure that there's no way [00:26:00] for a hostile takeover to happen, at least in the legal terms.

But that doesn't mean that the government can't take over. That doesn't mean that there, like, legislation can't happen. That doesn't mean that internally they can be like, "We've made super intelligence and it's evil. We're gonna shut it down." Like, that's part of why they're doing this, so that they can't be forced by customers, by shareholders to push out a model that has the potential to destroy everyone.

At least that's what they're saying, but... And I kind of believe them to a certain degree. So, like, there's a lot of external factors i- other than just profits that are happening here. You're right, Matt, and it's, it's a very complicated process. At the very least, it's very interesting and, I will have a link, we will have a link in the show notes to the S1 filing that you can go look through yourself or prompt an AI to explain it to you in whatever way you want.

And, uh, yeah, it's, yeah, it's a, it's, it's something that's out there now, and we know that there is profit to be made. It's not all just, rainbows and unicorns

Matt: All right. Well, I think that that is this edition of the web [00:27:00] news, unless you had anything else to add, Mike. I think you faded out there. I think it'll c- be corrected in the, the, the, the edited version, but you, like, disappeared a little bit. I think your internet cut out. So I'm not sure if you were finished or if...

I hope I didn't cut you off there, but

Mikhail: was absolutely finished. That was a natural fade-out. But yeah, uh, that's it for the web news then. Yeah,

Matt: that's a good cover-up. I like

Mikhail: yeah.

Matt: fade out.

Mikhail: Natural fade-out. Thank you for listening to the web news. Uh, we'll see you next week. Goodbye

Matt: Goodbye