0:00
All right, guys, welcome.
Search this video
0:01
Let's talk about the Anthropic S-1.
0:14
Once in your lifetime event, biggest day of human history.
0:19
Certainly in the history of capitalism, right, Joey? No hyperbole there. There's zero hyperbole.
0:25
What an amazing IPO it'll be.
0:28
How capitally efficient this growth curve was, how fast it happened. You can't ask for more.
0:36
Gonna come in at the, what, sixth most valuable company from the start of trading in the IPO in the whole world.
0:42
It'll be up there, I think yeah, Tokenomics I think covers seven of the top ten.
0:48
So this will be this and OpenAI will be trillion plus and Okay, what do you guys want to talk about first? Profitability?
0:55
I think that's a good place to start.
0:57
You know, compute, you know, they kind of give the compute margin numbers.
1:01
But I think even on the GAAP basis, you know, non-GAAP basis, you know, 10% plus margins already. Two quarters profitable. Really solid.
1:12
I think what we're seeing on the gross margin line just the past three years, like the scaling of compute and inference efficiencies is amazing.
1:20
And what they're able then to spend and increase, the training splits, really impressive.
1:27
So eighty-five percent inference gross margins, forty-one percent compute margins, and then extremely capitally efficient like sales and marketing and partnership thing too.
1:37
I think a lot of like software and tech companies spent a ton to go acquire a customer and you know, their net dollar retention number.
1:45
So a customer last year at this time that spent a dollar is now spending nine.
1:49
And in Q one that was five hundred percent.
1:53
So customer spending one was spending five.
1:54
And that all falls just down at the bottom line.
1:59
There's not much you have to do to upsell the usage base side of that.
2:03
So really solid headline numbers all around the business model.
2:10
Yeah, I would say really really interesting to me because from if you talk to folks in the datacenter market, right?
2:18
I'm the datacenter bro here.
2:19
Folks have been looking forward to this IPO for, you know, one being reason which is that Anthropic has the biggest compute demand of anyone, alongside OpenAI.
2:30
They're on par with you with hyperscalers and they can't deal directly with the folks actually building the datacenters, one because maybe they have a team too small, which they're building, but two is mostly financial.
2:42
They can't sign a like you can't fund a datacenter on Anthropic credit.
2:46
And so we saw the market kind of hating on Anthropic and OpenAI like naked exposure, whether it be on the datacenter side or on the GPU side over the last few weeks.
2:57
And so it, I think, is gonna be like very reassuring for folks who see the actual financials.
3:03
So the first thing I wanted to say is like talk to anyone on the datacenter market, everyone was just looking forward to this IPO.
3:09
They're like, dude, whenever these guys are gonna IPO, I'm gonna sell so much, so many megawatts to these guys.
3:15
It's all gonna be so easy.
3:17
Obviously that's contingent on Anthropic being investment grade.
3:21
And so when you look at some of these numbers, how much cash they have on the books, you know, like over a hundred billion dollars, how much they're gonna have in addition with what they're raising and the promises they have from like Google and others.
3:32
These guys are gonna have over 200 billion and as Joey said that's a company that in the last three years despite being very unprofitable barely burnt any cash so I think it's really reassuring.
3:43
Also the compute commitments frankly aren't that large which clearly shows you that they've shown some good like fairly flexible deals with folks like Amazon, Google, where instead of having to commit for like gigantic deals like Stargate, which is this huge three hundred billion dollar agreement for four and a half gigawatts.
4:02
They've managed to do something like a bit more efficient on the commitment side, which I think is overall great, good news and you know, good for AI infrastructure as a whole market.
4:14
Jeremie, can you walk us through?
4:15
I think a lot of people see the headline, five hundred, six hundred billion of compute commitments.
4:19
Can you walk us through a way that's not that large? Yeah.
4:23
Yeah, well good like people ask that and it's when you break it down it's not that much. Yeah, for sure.
4:30
If you look at recent deals that have been done, such as Anthropic Nscale, let's take this one.
4:37
The compute per gigawatt commitment is a hundred billion dollars.
4:41
And so there's a world where you would have to for like these half a trillion dollars would be just five gigawatts, which is really not enough for Anthropic to serve their roadmap, right?
4:53
Like what they have in terms of gigawatts contract that is actually materially higher than that.
4:58
And so their effective ratio of gigawatts contracted relative to commitments is really low.
5:03
And I think this gives us the right now the compute they have is like under five gigawatts.
5:11
That's gonna go up materially as we get into the end of next year.
5:15
Plenty of compute to you know accelerate revenue growth, cut cost of models to keep finding efficiencies, and so it's pretty clear all of that is gonna trigger an acceleration of revenue.
5:26
So these guys are gonna be in the multiple hundreds of billions of annual revenue next year.
5:30
And so yeah, it doesn't feel like that complicated for them to pay half a trillion dollars given the revenue they're making, free cash flow profile, strong balance sheet.
5:42
So I think overall like very reassuring, especially when you compare it to like, you know, what was the first freak out about compute commitments that was last year, that was OpenAI?
5:50
And people were saying, dude, you're making twenty billion annualized and you have one point four trillion. Right.
5:57
So I think we're in a much more reasonable ratios here, like you know, eighty billion ARR, half a trillion commitment, it's a lot, but there's two hundred billion dollars on the books post IPO.
6:07
So overall I think you know it's gonna be fairly easy for them.
6:10
And as I was saying, like they've done a mix of like long term deals like Nscale, but also more flexible shorter term deals with hyperscalers and folks like SpaceX, which they can cancel effectively anytime.
6:22
So overall pretty reassuring I would say. Can I wanna hop in?
6:30
Day in the history of capitalism.
6:33
I actually I just made that shit up in terms of history of capitalism, but it felt right.
6:39
I mean like look it's a big business, dude.
6:41
It is a big business, you know?
6:43
Like I remember once upon a time, you know, one billion run rate was like the grown up number and they obviously did a lot bigger than that.
6:52
I think the thing I thought was best on that is like incremental margins are 90% gross margin and 40% EBIT.
6:59
And that's like that's a best in class business.
7:02
Like forty percent EBIT is as good as it gets.
7:05
That's like as good, if not better, than the hyperscalers in terms of like if they get there, right?
7:09
Gross margins at 90% incremental, like it's very clear, it's on a very large base.
7:14
I think it's as clean as you possibly could want it to be.
7:17
And I think people we've been saying this shit for a long time that there's obviously really great incremental, like there's real revenue, right?
7:25
Like when you sell it, you're not selling a dollar for a dollar and ten cents.
7:29
You're selling a dollar for ten cents.
7:30
Okay, that's margin, right?
7:32
I think the biggest bears have always believed that there's a heavy subsidized subsidization.
7:38
Sorry, I can't say that word today.
7:40
And it's very clear there isn't.
7:41
So like I don't know what else you want to say.
7:43
I also think the fact that Ed Zitron hasn't tweeted is like probably the most interesting fact of all.
7:50
Like, you know, he like him calling Jensen a crook last week was like pretty based.
7:55
In terms of he's like you're watching the speed run of a short seller in real time. So yeah, we'll see.
8:02
I hope he's mentally okay, you know?
8:04
Might be a hard day to be Ed. Yeah.
8:08
So inference is a good business model, I guess. That's the net show. Yeah. Controversial.
8:16
Controversial, people actually want these tokens.
8:19
How about the customer concentration?
8:21
I mean, there was a comment in there saying that top five customers represent twenty five percent of revenue now, and then the CSP commitments, the marketplaces are representing another fifty percent.
8:30
So what's the leftover twenty five percent?
8:32
I mean, it's just a really big base, so the leftover twenty five percent is just still many billions, but about what you mean.
8:39
I think you think that it's either or mutually exclusive.
8:42
58% is the total, and they can it's like the total distribution, and then 25% can look through that, if that makes sense. They don't add together.
8:51
The entire top 10 customer could be 100% on the CSP, you know.
9:00
Think that some of the top five or top ten customers are actually going through the CSP marketplaces?
9:05
It implies that the numbers show that, right?
9:07
Because, to be clear, I really want to make sure, Jordan, we get this straight on the podcast.
9:11
That 58% of distribution is like their sources, right?
9:16
But I don't think when they're talking about their net customer, they are not talking about the CSP as the net customer.
9:21
They're talking about the look, you know.
9:23
I'm assuming there is some kind of pass through, meaning that when Anthropic gets billed, they get the customer who purchases through the distribution platform.
9:33
Aka Bedrock, right, so let's just cut this part out where I'm being stupid then and not talk about customer concentration.
9:40
Well no, we can talk about the five percent customer concentration.
9:43
No, it dude, the fact that you didn't you like it's important to know, people don't know that.
9:48
It's just the top five doing twenty-five, which is pretty good, honestly.
9:52
That's most of the companies are actually more concentrated than that.
9:54
I know the top five are doing the twenty five, and then there's an additional forty eight percent of revenue that goes through the CSP marketplaces, which does not include the top five customers because elsewhere in the S-1 they also talk about how they have a direct relationship with people who are making purchase commitments to get a discount long term, like any other big customer would do directly.
10:15
The 48% of revenue, none of the top five customers are going through Bedrock or Vertex today.
10:22
So then there's which makes sense, like those customers wouldn't buy through that channel.
10:28
So then you have that, you have the CSP marketplace, which we know is like going towards your traditional buyers, you know, just your large AWS accounts, your large Google GCP accounts that are buying through that.
10:42
A little more Fortune 500.
10:43
And then you have another, so that's 48.
10:46
Then you have another 25.
10:47
And then there's 27 left over.
10:49
And I think even the point there, and there are some really interesting stats on some of these customers, is you know, there's actually it's I think that's still rather concentrated.
11:00
I think if we look, we still look at any, you know, alternative data set, the top 1% of customers make up, you know, 70, 80% of the panel.
11:08
And they said this, the top five percent of Claude enterprise customers, users, I should say at these end companies, are spending eighteen thousand per user per year.
11:18
Your non-Claude Code user spends less than seventy five a month on enterprise.
11:24
You're talking to three, maybe four of the top five percent Claude from bio from that map. Topic was.
11:31
Two hundred and fifty ten million dollar customers on a trailing twelve month basis.
11:34
We I know we're in the million plus group, the nineteen hundred one of nineteen hundred.
11:39
We've gotta be getting close at some point.
11:41
We're one of nineteen hundred, yeah. Yeah.
11:43
At some point we'll be in that two hundred fifty, I would think. Yeah, some point. Yeah.
11:49
That's not a if Dylan's listening to this, that this is not a KPI, but I think it's a result.
11:54
It's a result of some other KPIs being enforced. Not that we have KPIs.
12:00
Okay, but talk about the actual customers.
12:04
They name customer A B C D.
12:06
Somebody who started as forty three percent of their revenue is went below ten percent, is now back up over ten percent.
12:12
We've got our estimates for who those people are, but let's just say you can't look at twenty three numbers, I don't think.
12:20
There were customers at OpenAI that were over ten percent in twenty twenty three that are nothing today.
12:25
So I don't know if the twenty twenty three numbers.
12:28
I mean, we know some of those large customers.
12:30
You know, today that are you know, there was a one, I think ten percent customer in you know Who do you think?
12:37
Let's do some speculation my meme was that it's Kimi, MiMo, GLM not just kidding no yeah yeah yeah five sorry that's five that's my meme Well, okay to be clear they're they I mean even I don't know if their KYC is good enough to completely cut them off. Right?
13:01
I'm sure someone at Moonshot is using Claude Code.
13:05
If you saw on the benchmark I forget which benchmark, I think on DeepSeek, they're using Claude Code to be the final verifier of like, you know, to verify all the results.
13:15
So they're using it obviously, but maybe not just distilling.
13:21
So who do you think top five is? What I'm saying. Like let's speculate.
13:24
We know Meta was the one that rationalized spend in Q three.
13:28
Or not rationalized spend.
13:30
They moved from an external provider to internal.
13:32
The tokens are still there from an industry perspective.
13:36
I think Microsoft is one, Cursor, we talked about, being another one.
13:41
NVIDIA, hedge funds like Jane Street, HRT. I think of a spender? Yeah, definitely.
13:47
I think they're probably the forty three percent that went lower.
13:50
Who else is using this much stuff in twenty twenty three?
13:54
My understanding is they were really late to getting access to Claude.
14:00
They were a Claude adopter in 24.
14:03
I don't think they were using it.
14:05
I don't think NVIDIA was using it in 23. That's my impression.
14:07
I might be long wrong, but I feel like Jane Street probably feels like the most rational, or maybe you can argue Google or something, I don't know. You know Google uses. That would be crazy.
14:18
I know Jane Street is producing a ton of cash flow right now, but they have such big commitments with OpenAI already.
14:25
It'd be crazy if they were also a top five and so big on Anthropic as well.
14:30
But I mean it's totally possible.
14:32
There's only so many people that can be spending tens of billions of dollars on these tokens.
14:37
Well, actually no, I was gonna say I feel like yeah, I was gonna say I don't know.
14:41
It feels we're gonna figure out who it is over time and I bet you it's gonna be like someone really obvious like that we just didn't think about. Accenture.
14:50
The point of clarification for me is that these top customers are not the CSP marketplaces because they have this direct relationship with people through the CSP marketplace if they're using Bedrock or Foundry or Vertex.
15:01
And I think that's also pretty interesting because if people end up having this direct relationship with Anthropic and they make these purchase commitments ahead of time the way they do with like an AWS MSA agreement, that's a big lock in. For using those models. I made another chart.
15:23
They said that in Q3, 20% of the revenue was from models released before April. So that's Opus 4. 6, Sonnet 4. 6, 4.
15:33
5, which they kind of said, okay, 80% is the latest model.
15:39
So people move to the latest models.
15:40
But I'm actually shocked that 20% of revenue is still stuck on models that are three months old.
15:46
How long does it take somebody to get off of the old models and onto the new ones?
15:49
And I would think that somebody buying through Bedrock, a Fortune five hundred who has to certify something, they're gonna have that lag time, whereas a startup's gonna be a little more nimble to move around.
15:59
And so both of these things just point to me as being really sticky.
16:04
If we're saying Microsoft is a top five customer then they're probably the one spending on all the Opus 4. 5 open.
16:12
Via via GitHub Copilot also, but Copilot 365 or something like that, the Cursor, Perplexity equivalent where they don't pass through the relationship to the end user.
16:22
It's their API key that's doing all the spending.
16:24
Microsoft has both the Foundry business where they do pass through the relationship, and then also the GitHub Copilot stuff where they consume so many tokens.
16:33
And you can probably see if they had more granular data, you could probably see Microsoft spend be high at the beginning, dip down when they have all these OpenAI options and then come back up now that they're actually more motivated to serve Anthropic and not tied so closely to OpenAI? We'll see. What else is hot?
16:54
What else is hot about this S-1?
16:57
Compute commitments, Jeremie.
16:58
Talk about the seventy billion dollars MPA. Sure.
17:05
They're just essentially gonna get even more money than they already have, which is mostly good news.
17:11
A bunch of that was all already known if you look through the filings of like Google and Amazon, but they've signed a bunch of like large fundraises.
17:19
For example there was a forty billion with Google, which have like multiple milestones, right?
17:26
And so like Google has delivered twenty billion, there was a remaining twenty billion, which is just gonna be more cash for Anthropic based on milestones. Such as going public.
17:36
So effectively like all of that is just additional funding.
17:40
Which again like as I was saying earlier just means these guys are gonna have more money on the books, gonna be easier to pay for stuff.
17:48
They got the Broadcom one as well which is forty two billion dollars which gonna enable them to buy a whole lot more TPUs.
17:55
And so yeah overall like the twenty twenty seven compute roadmap for Anthropic is fairly de-risked which kind of goes back to what we said in an article on a podcast a few weeks ago about like the DSR built out, yeah, this is most of what they're gonna get next year is de-risked, and it's gonna be huge.
18:14
So the implication is okay, they're gonna have all of that compute next year.
18:20
Then can they actually accelerate the revenue enough, right?
18:24
Anything that's still the question.
18:26
But maybe one thing that folks weren't aware of is like how much do they have on the books right now?
18:31
And now we know, and it tells. Hell yeah.
18:41
Rates are five per let's say rates are five percent a year, they're making five billion ARR, you know?
18:49
Two gets us to one trillion almost, Doug.
18:51
Yeah, dude, it's just a little bit of the way.
18:54
I honestly the thing that I really thought was most interesting in the whole S-1 is the crap that they put in or and like honestly I think they've been pretty consistent about this.
19:03
Dario's been saying this for a long time.
19:04
It's like we never want to overextend ourselves, right?
19:07
And then essentially just being like, We'll never miss your expectations is an insane thing to put in the S-1. Like that is insane.
19:15
I think that this the most nutty thing they could possibly do.
19:18
And then everyone is really worried and bared up about Q four because what they've been telling the bankers for ARR.
19:24
And you're like, they just told you they're never gonna miss your expectations.
19:28
They're like, We've talked to all these I Joey's frozen.
19:31
We've talked to all these people and all these stuff you should never be worth one more dollar than what you're actually worth.
19:37
And so we'll underpromise and overdeliver.
19:39
I was like, my god, dude. Yeah.
19:42
And don't overdeliver, Doug.
19:43
Yeah, I guess that was that maybe that's in their fit interview, dude. I mean it's crazy.
19:49
I think this talk about okay, wait, wait a second.
19:51
So this is a core principle of the business.
19:53
So all of this talk about cybersecurity ending the world and drugs and all of this stuff, this is them underpromising for the AGI that they're forecasting?
20:01
Ten percent chance of death of humanity. This is underpromising?
20:04
Yeah, apparently, according to Dario. It can get worse. It can get way worse.
20:09
You saw, did you guys see the head of cybersecurity?
20:12
He's like, if you think it's bad, it's gonna get way worse. Laughing is scary, man. Scary.
20:21
This is gonna be our last secure podcast.
20:25
I was, yeah, I was freaking out last week with all of the cybersecurity stuff and then interesting stuff over the weekend as Elon and Chamath are calling for the death of all neoclouds.
20:37
And then we have this counterweight to all of that, which is like we'll save it with capitalism.
20:44
I guess that's the thesis of today.
20:47
Yeah, the money looks good to me.
20:49
I think Elon is just Elon's smart and he knows when he has an advantage and he's trying really hard to lower his cost of capital at SpaceX.
21:02
And I think he's I mean, he is a neocloud dude, so he's saying that while being the largest neocloud in the world.
21:08
So, yeah, it's kinda talking book.
21:10
I would also say our view and I think it's starting to like folks are starting to be really worried about that is hey there might be an inversion of like right now it's datacenters that are more constrained than chips, but maybe by the end of next year it's the other way around.
21:25
And there you have this guy that says, Hey, I'm single-handedly gonna build up to twenty gigawatts, maybe just ten you know.
21:32
So he's gonna He's gotta make sure he has the chips, right?
21:36
And so one of the ways to make sure no one else has the chips what bro, one of the ways to make sure others don't buy the chips is just make them go broke. Yeah. Just break markets.
21:49
They can't fund them, all of them for you. That's happening.
21:52
Right now there's essentially a liquidity run on all the neocloud debt.
21:56
And so anyone who wants to place new neocloud debt isn't able to.
22:00
And so if he presses it, if he presses his advantage, he presses his foot at their throat and they essentially never get funding, that's all his opportunity, right?
22:09
Because then everyone's like, dude, I wanna invest, I wanna invest, I wanna do this SPV and SpaceX is the only game in town and he's like, Great, that's a better cost of capital for me, right? Elon is Yeah, exactly.
22:23
And then as he said, he wants to do half a trillion dollars, two trillion dollars of ARR by in the end of next year.
22:30
The less these guys can fund their stuff, the more he can just build it himself and score deals at fifty million bucks a megawatt a year and you know, just make a whole lot of money. Yeah.
22:43
Jeremie, how real do you think that is that fourteen dollar an hour GB300 contracts that are on demand and can be cancelled any month just continue to go because Anthropic's ARR just continues to grow and they can't bring enough compute online and so they just keep it.
22:56
I mean look, like there's many things that go into like understanding how sustainable this is.
23:02
But one of them is if the markets first of all, it's the basic supply versus demand.
23:09
If demand continues to like outperform supply growth, we've said very publicly we think supply growth is gonna be huge.
23:18
But hey, what if ARR and overall demand also outperforms that?
23:22
Then that's perhaps even gonna augment the premium, right?
23:26
That's a realistic scenario because who else has as much capacity available in twenty twenty seven? Right.
23:34
Now if the equilibrium stays the same, then I think there's still probably like availability to sell and obviously like if demand starts to fade then of course like things are more complicated.
23:46
Now there's a few other things to look at.
23:48
One, I mean we talked about compute commitments Kind the beauty of SpaceX contracts is it makes your it's like much more balance sheet efficient or like compute commitment efficient than signing a six year deal, right?
24:03
So you saw it scale at like $100 billion per gigawatt.
24:07
That's likely going up, so perhaps going to 120 or more.
24:11
And this guy actually 50 a year but 90 days.
24:16
So the commitment actually is extremely small, right?
24:19
It's like fifteen billion dollars a gigawatt.
24:22
So yeah, it's more expensive, margins go down, but on the other hand, for your books, for what people are looking at, it's actually quite efficient.
24:32
And the last thing is obviously like as the drivers of this market remain like these small companies that don't have yet full trust of creditors, of lenders, it's gonna be hard for them to fund an insane amount of supply.
24:47
I mean, even for hyperscalers is not that it that easy, right?
24:50
Like we saw tiny, tiny cracks on Meta, for example, when spreads went up.
24:54
So obviously like this environment is very positive to companies that can take a lot of risk.
25:02
And these guys are basically taking more risk than anyone else by just equity funding all of that stuff.
25:09
Like some that as well, but and you know, building stuff without a customer. Yeah.
25:14
Let me throw this chart on screen, Joey.
25:16
What's your take when you saw this chart that they put up about compute efficiency? I think it's great.
25:22
I think even looking at just inference compute efficiency, the revenue per megawatt, just on inference going from, you know, like sixteen million a megawatt last year to sixty plus you currently.
25:34
That tells even more the story of how great a business model just inference is.
25:39
And this is the entire fleet.
25:40
So what you need to do to continue to invest in new models. This is great.
25:46
Obviously in Q three it went down a little bit.
25:49
With some of the new SpaceX compute coming online and then a little bit less monetization via meta distillation and then some additional safety compute but I think this really tells the story of hey AI model businesses are really solid and they scale extremely well to Doug's point incremental margins are amazing.
26:11
Yeah, let me read the line from the S-1 that jumped out most to me about compute efficiency.
26:16
They said, Our compute efficiency is demonstrated by revenue that grew one thousand three hundred and sixty eight percent year over year for the three months ended June thirtieth, compared to compute and infrastructure expenses that grew three hundred percent during the same period.
26:32
Although to be fair, how much of that is like Anthropic being geniuses versus the actual market wide efficiency being higher, right?
26:42
I mean we run InferenceX, what do you think, Jordan?
26:45
Like we're seeing you know GB being 50x better than Hopper, for example, isn't that just reflection of the silicon and systems market overall?
26:57
Yeah, the new stuff is better.
26:58
I think maybe one of the other quotes that I can read from the S-1 is related to their fungibility across the fleet, which is super fascinating.
27:07
I think they do have a unique advantage to pick the right accelerator for the right job because they have TPU, Trainium, NVIDIA GPU, and AMD GPU all online, whereas other people are kind of catching up there.
27:20
And if you're gonna want to realize the true benefits of new hardware coming online, which is always more performant than the previous generation.
27:29
You need the software to be able to use it.
27:31
Who's got the best software for writing code for this stuff?
27:34
Probably the company with the top engineers and the best model.
27:38
So yeah, I think you can blame the underlying hardware for sure, but I don't think there's any reason to think that's going to slow down as they ramp up compute commitments that all bring on the latest and greatest, and they have a track record of being able to adopt whatever the best silicon is for the job.
27:55
I think also in that same vein, I think we also like maybe underestimated the models, to be honest with you, 'cause I think we were also in the camp where we're like, yeah, maybe not like the CUDA moat will never be breached, but maybe the thought processes with the was the kernel optimization would be a harder problem.
28:15
But we're watching like models, you know, reverse to engineer binary.
28:20
Which was the gold standard of the binary for gaming, which reverse engineering gaming was allegedly impossible because of all the anti-cheat stuff like that. Clearly it isn't.
28:30
So I think that, maybe we also underestimated the fact that Claude is good enough that Fable 5.
28:30
5 can probably port and it makes like the AMD versus NVIDIA versus whatever.
28:44
Actually not matter as much as it used to because the same model that can do Navier-Stokes is able to also program a GPU to definitely make their economics work out.
28:56
And I think by pushing to be they were the first to really push to be multi-chip.
29:02
And I think that strategy totally paid out for them versus like OpenAI was very much an NVIDIA first shop at almost anyone else's expense.
29:11
And I think that that's been part of the Anthropic catch-up story is supply availability.
29:16
And I think they continue to be really smart and intelligent with supply availability.
29:20
They want to have vendors that aren't just TSMC.
29:23
They want to make sure that every aspect is diversified so that no party has power over them and they have power over everyone else.
29:31
That's just good business and strategy.
29:35
And it definitely played out.
29:36
Like the model is the differentiator.
29:40
Yeah, I think I view it as a race where you have to start running at some point.
29:46
And some people took off earlier than Anthropic, like OpenAI or even Google DeepMind, made a bunch of commitments, spent a bunch of money, brought on compute, and maybe now that compute is less efficient than the stuff Anthropic has because they don't have the bloat of as many like H100s from 2022 when they weren't like doing anything or building any compute.
30:05
But then vice versa, if you Or a startup that's getting off the ground with a billion dollars of seed round in the last three months.
30:14
I mean, you're not profitable, you don't have the data flywheel, you're missing all of this compute.
30:19
Anthropic has a bunch of advantages over somebody who's trying to get started right now.
30:23
And so I think the dynamic is they kind of struck it perfectly where they both have this like financial flywheel as well as all the research and data that puts them so far ahead of somebody else that not only are they, you know, ahead.
30:36
They're also running faster.
30:39
So I don't know how anybody catches up from here.
30:42
Like, there's kind of no moats left in terms of, these guys have a bunch of researchers or compute or data that Anthropic doesn't have.
30:50
Like I think they're clearly in first or second place on any of those dimensions.
30:55
And then you have to make a case for any of these neolabs to catch them on any, maybe in specific industries where they have great researchers and great data.
31:03
But I don't know how anybody possibly builds more compute than them at this rate.
31:08
So as far as compute being an advantage, they're ahead of anybody else.
31:12
Except maybe Meta or Google have a flywheel that's better than the Claude API as a way to fund this build out with their ads business.
31:21
But even that I'm questioning as Claude goes into next year and starts to grow so quickly.
31:26
Yeah, I think the only places that they'll win or someone has the potential to win is in multimodal where I think Anthropic has always been nearsighted.
31:36
Like there's no other way to put it.
31:37
Like it's just not as good as let's say Astra for like vision.
31:42
But I think maybe you could argue that we love Aqua Voice, for example.
31:46
Really niche applications that are very specialized can maybe take tiny slivers, but I think the big pie in the sky is definitely Anthropic's five fly.
31:55
And I'm sure they'll let little things lose, but like in the same way that they can just, you know, vibe code a new model, they can catch up if they ever had the time or attention.
32:05
I think it's just doing the biggest opportunity first.
32:08
Look at the domain because let's say that inference efficiency or like system engineering on hardware was a domain that's important.
32:17
They point Claude at it, they spend some time on it, they come out on top and first, even though others were ahead of them.
32:23
Now they're gonna build a chip.
32:25
Is their CPU derivative chip gonna be better than jalapeno?
32:29
Is it gonna be better than what all the startups can come up with?
32:32
They've got Claude working for them.
32:34
That's a bit of an advantage.
32:36
They're Classifying everybody else where they can't work on designing chips using their models.
32:43
And so what's next, right?
32:45
Can they point that at the voice-to-voice interaction models?
32:48
Can they point that at the decision models where we saw Jev?
32:50
Can they point that at bio and start building cancer drugs?
32:54
I think the thesis that AI research and software engineering is fundamental and that.
33:07
Comes first before you win any of the other domains seems to be playing out exactly as their thesis was.
33:14
I don't see any reason to doubt that at this point.
33:18
Like they're gonna have more money, more compute, more researchers, more data than everybody else going into next year, it seems. Hundred percent agree.
33:26
Got two questions for you, Joey.
33:29
First one, you had this great piece maybe six months ago. I loved the chart. That was so good. But yeah, great job.
33:37
Like you had you beat this model of Anthropic financials was super accurate.
33:41
So I'm just curious, like how much of a surprise were the numbers that were there and relative to what we had.
33:47
And the second one I'm just curious is it any disappointments for you overall on Anthropic as a company, as a business?
33:54
Yeah, I think we had the margin story, right?
33:57
Obviously Q three was little tougher number internally for both OAI and Ant, but you know, I think we had a pretty good understanding of the business model.
34:07
So there wasn't anything massively you know, I think most stuff got leaked too that there wasn't a massive surprise here today.
34:14
And then sorry, you're gonna have to repeat the second. It was so clean.
34:18
Everything has been leaked, but the big it's like seeing all the different parts of the picture.
34:22
We see the whole picture and you're like You could nitpick a number in there, any of any numerical disclosure, I don't think you could really go in and nitpick or say, that's not really that good.
34:32
Analy off balance sheet or the debt commitment to the facilities, but yeah. We all knew that too.
34:39
Like it's not it's nothing crazy, dude.
34:41
Jeremie, what was the second question?
34:43
Yeah, any disappointments with regards to like the business overall in this S-1?
34:48
No, I think I don't know, if you're of the view that, hey, every the average employee at an enterprise is going to pick up their spend or even the median looks like what the seventy fifth percentile is today.
35:03
They don't have to get up to the ninety ninth percentile user.
35:06
There is a ton of runway here just in enterprise work.
35:11
I think the really, you know, interesting number I shared the one earlier on the top five percent of users are at eighteen thousand.
35:18
Per year and then the non Claude Code users are at seventy five sub seventy five a month.
35:22
The we're at zero point one percent of their Fortune five hundred OpEx budgets today.
35:28
So like under you know eight bucks a month per employee, like under eight bucks to under fifty per employee. And I yeah.
35:37
Time to apply the eight the 0.
35:37
85% ARR but per cohort SaaS number so 900 is just gonna be 0.
35:37
85 for the rest of the cohort's life.
35:48
Sorry, I'm just that's some real brain dead SaaS bubble stuff.
35:52
I don't know if you remember that.
35:53
Think I if you look at the cohorts, I think the other big thing is towards twenty seven, we can obviously talk about cyber.
35:58
We can talk about this traditional white collar work, 'cause Anthropic's obviously extremely, extremely heavily coding based.
36:05
There's a lot of other verticals.
36:06
These cohorts are going to rip.
36:08
We're going through the first Q four traditional enterprise IT budgeting cycle where AI and that's the point.
36:15
It's such a low percentage today of Fortune five hundred or global two thousand.
36:20
Most people that's not a top one percent customer.
36:22
What inning are we in, Joey?
36:25
I like Doug's comment today, we are still paving the foundation, the parking lot.
36:30
We're painting the lines, the asphalt is you know, we are that early, you know, and yeah the We're they're still grading, yeah.
36:41
They're still like yeah, they've got the dirt.
36:44
I was dying when you guys said this about being bored and parking.
36:48
The best part of this Jordan is we're already at like a hundred and twenty billion of ARR combined between the two frontier labs.
36:54
So if that's where we're at today, you know.
36:57
Maybe one day they can get to the enterprise SaaS revenue penetration of Microsoft Teams or Salesforce and start monetizing more than eight bucks a month per user per seat. Maybe when they grow up.
37:13
I that's a big thing and like, you know, some people might say, you know, that goes to 365 Cowork.
37:19
But I'm you know, there's a your suspicions?
37:24
No, there's as there'll always be demand for frontier intelligence.
37:30
And so as we look towards next year, I think like the pieces are in place to have a really strong twenty seven.
37:37
You know, we'll look at twenty eight when we get there, but very early innings. Beautiful.
37:47
Doug, what's your take on the market reaction when this came out?
37:49
How does this affect other things?
37:50
We talked a little bit about neoclouds and debt generally.
37:54
What do you think happens to everybody else when they come public?
37:57
I would think on the cleanest quarter or not the clean, the a pretty clean S-1, the reaction in semis would be more or semis and let's just say all the AI beneficiaries.
38:08
The land build out companies actually had were the number one winners today, I think.
38:13
So essentially like the datacenter side is really starting to pick up the bid.
38:17
But I it wasn't as strong as like the shit we've been seeing all year.
38:21
I think the AI trade is exhausted.
38:24
There's no other way to put it. The AI stocks are tired.
38:28
But we just gotta refresh a Q4 budget, dude. Who knows? They're so tired, dude.
38:34
They're you know, they've traveled so far.
38:36
They're so, you know, big Bro, this is gonna go viral on Twitter.
38:42
You're gonna see like Taiwan pick up your face in a video and say the AI trade is tired.
38:48
SemiAnalysis They're not gonna see it.
38:51
They're not gonna see it, man.
38:52
He's just like Batman with that light coming behind him.
38:55
We just talked to a shadow of Doug now. Yeah.
39:02
How about the okay, maybe the one letdown I can come up with is that they talked about the low usage over the holidays.
39:15
So for Q four, what sort of revenue impact are we expecting Christmas vacation to have on Claude Code usage?
39:23
Well they said it was a twenty percent decline in versus the two weeks prior.
39:27
So on December just on the month of December it's a ten percent cut, let's call it.
39:33
I imagine it'll be less pronounced this year. Why less?
39:37
So I think it gets worse, man.
39:40
I don't well, maybe we can see a Polymarket for this one.
39:43
But the line says consistent with this pattern for the two weeks ended January second, twenty twenty-six, total customer spend declined by approximately twenty percent compared to the prior two weeks ended December nineteenth.
39:54
I feel like the mix has gone towards consumption over the all-in products.
40:00
And so I think a lot of people log off over the Christmas holidays, right?
40:05
And they just don't use API token.
40:07
That's what's gonna happen to us.
40:08
Average Claude Code user though, Claude Code is more part of their daily routine now?
40:12
Even if they're working two hours or four hours a day than it was at this point last year?
40:17
Jeremie, are you gonna be on your phone Claude coding while you're having Christmas holidays? You have Muse now, bro. Now Muse is for this.
40:25
So Muse is actually taking vacation dollars out of Claude Code, man. So yeah, I don't know. I'm not sure though.
40:33
I'm, well, I use Perplexity Computer now. Maybe I'm a weirdo.
40:38
Just This is a call for frontier labs.
40:43
Please make the experience of using AI models on the phone and then continuing on the you know, on the computer and the phone again better.
40:51
'Cause you know, shout out Perplexity, they do really well.
40:55
I didn't see that as a material risk in their S-1 Jeremie moving to Perplexity.
40:59
I'm surprised they missed that one.
41:01
Yeah, you're gonna see it in the Q4 ARR. It might go down a bit.
41:05
They're gonna say it as another one off here after Meta cuts it. SemiAnalysis. We had a one off.
41:17
Meta left to use their own models and somebody named Jeremie Eliahou Ontiveros said SemiAnalysis is gone.
41:25
What's Okay, what models do you choose though?
41:28
Are you still using Claude or have you gone for some Codex models through Perplexity?
41:34
You just tell it use whatever state of the art model and sub agent max effort to figure it out. Just use the best stuff.
41:46
So generally I think the orchestrator is Fable though. Fight Fable point One. Yeah, yeah. Yeah, so Fable 5.
41:52
x hasn't even been considered here. We obviously saw Opus 5.
41:56
5 come in and that had a big impact. I love that model. Fable's still in 5. 1.
42:03
What happens when they just keep ripping that?
42:06
Do we see that before the next version of Astra?
42:08
I gotta expect so, right? Yeah, we'll see 5. 5 into the IPO.
42:16
You have to juice the first quarter numbers.
42:20
What happens then when they release Fable 5.
42:20
5 fast mode at $500 output tokens?
42:26
That's gonna be a good one.
42:28
They don't like to disappoint expectations is what I'm thinking, They are gonna release a faster tier in Q four two with Fi Five.
42:38
They see the under promise over deliver.
42:42
Cable six people are saying for Q four are so stupidly low. They're not believable.
42:47
It's dude, it's they're actually gonna they're gonna make a new class of model, Giga Fable.
42:54
The unreleased models coming out.
42:56
Honestly, okay, actually let's talk about look, you know what we haven't talked about is OpenAI. Is OpenAI kinda cooked?
43:03
Because Astra is their best.
43:05
Maybe their Bell model like improves or something like that, but like where Is OpenAI.
43:12
Because it's actually, Opus 5.
43:12
5 accelerating is pretty it comes at someone's direct expense.
43:19
And at least within the SemiAnalysis spend, it's very clear who it is. It's Astra.
43:23
What do we think about OpenAI?
43:25
Because they're kind of in they're caught in the middle of everything.
43:28
They're in the middle between consumer on Gemini and Meta, right?
43:33
They are like between Anthropic and open source models for enterprise, and they're just like kind of literally number two in everything. What's it number one?
43:49
I said it earlier when I was talking through Anthropic's justification for why they have a lead and why they have a moat.
43:54
I think OpenAI got running a little quick and they might be a little bit behind now.
43:58
They're certainly behind on ARR and I think they land some counterpunches in model capability and then Anthropic comes beyond them.
44:04
But if people make significant prepayment commitments to Anthropic, then that's tough for OpenAI.
44:11
But I mean, I can't see all of this and not just be bullish on both of them.
44:17
Like Isn't there a need for multiple models from multiple providers for different things?
44:22
I still use OpenAI models for lots of stuff that I don't think Claude is good enough at.
44:27
I find them more reliable for data analysis, for example. So I don't know.
44:33
I'm not an OpenAI bear by any means on this result of I feel like every day someone gets a vibe that hey today Claude is retarded, hey, today Astra is retarded and so it's becoming so easier to ship from one to another, man.
44:50
Did you see Thariq left Claude and it man.
44:52
That his first tweet is A is Opus 5. 5 dumber? So genius dude. Yeah.
44:58
So like I don't know, I feel like every week which we have like some big user shifting from one to another because the model feels dumber for whatever reason.
45:06
And so yeah, if both are good, like I don't know, why would OpenAI be done?
45:13
I just think in okay, so if the low it's like there are the lowest cost producer with the highest margin is Anthropic.
45:20
And this you see this in semiconductors, right?
45:23
Number two has less share and worse margins, right?
45:27
And so OpenAI is just gonna be mid like they're not done, okay?
45:29
Obviously ninety percent incremental gross margins tell you they can figure out a business, okay?
45:34
They could definitely they put their heads down.
45:36
I think they can figure it out, there's the up a few hundred billion dollars of scraps.
45:41
Yeah, there's some money to be made for them, okay?
45:44
But I just think in many ways, because of their place, it makes their positioning, their competitive positioning worse.
45:50
Because if you're in the lead, you just charge a premium margin and you get a premium product, you get premium, it signals that you're a higher quality luxury product.
46:00
OpenAI will now always be competing with Anthropic on price.
46:03
And that's just such a shitty place to be.
46:05
You don't wanna be, we've seen one of the big takeaways is that the number one thing is AI is about best in breed and if OpenAI isn't best in breed, then they're competing in best in price and that's just a worse place to be. So Yep. Overall agree. Okay.
46:26
I guess one of the kind of pushback I'm curious what you guys think, but what afterward is generally like OpenAI is much better right now at enterprise sales.
46:36
And they've been more successful at signing large contracts with these I call them like you know Boomer Fortune 500 enterprises.
46:43
And so like that's perhaps one bull case for OpenAI is hey, as enterprise spend starts to pick up as Joey was saying, is such a small portion today.
46:52
If you exclude this super power users and maybe that's good for OpenAI as the better enterprise sales platform.
47:01
Yeah, they've been OpenAI's been very bullish on their Q4 enterprise ramp.
47:05
You know, they're like Bedrock will we ramp a ton for them.
47:09
You know, just going on you know, Vertex as well.
47:13
Yeah, it's a different business model.
47:16
They don't have the third party share.
47:17
They haven't had, you know, it was like majority consumer up until a few months ago.
47:22
And they're back on the right track.
47:25
I think there's this argument that we're just ping ponging between model share going forward of who's best.
47:32
And, you know, there's certain things like if you're signing these one, three year enterprise deals, you're gonna see less of that.
47:38
And, you know, switching costs will be a little bit higher.
47:42
But yeah, it's like and then there's been a lot of news in the last week just on them being aggressive, discounting against Anthropic accounts as well to win some of that business.
47:53
So they're willing to fight for it.
47:56
And they've definitely taken enterprise a lot more seriously since July with the release of five six. And they executed well.
48:02
They had a really good Q three.
48:04
They're yeah, the business their margins don't look anything like Anthropic's, but it'll get there.
48:14
They have just too much consumer.
48:17
Yeah, they have a lot of consumer subscription, I think.
48:20
Too, you have to support all the free users.
48:22
If six percent of users, the billion, you know, weekly active users, if only six percent convert to paying, that's tough.
48:31
And then yeah, they had very little of this usage based like API revenue that we talk about monetizing, you know, at a hundred million, you know, having the potential to monetize at a hundred million per megawatt plus.
48:44
About government as a proxy for what's going on there.
48:48
They said something about government. Let me read the passage. Our custom right? Yeah, sub one percent.
48:53
Our customers include US and non-US government agencies.
48:53
Revenue from government agency contracts represented in aggregate approximately one percent of our revenue in each of the periods presented in this prospectus.
49:04
That's a huge market for everybody.
49:08
Of seats you can sell in yeah, in government.
49:11
Let's size this relative to what, all of GovCloud or something?
49:15
Yeah, yeah, look at GovCloud or yeah, FedRAMP or any of those programs, but then just look at I think the big thing you can look at for a lot of this work is just like 365 seats out there.
49:25
How many Microsoft 365 commercial seats are there out there?
49:29
And there's hundreds of millions.
49:31
And if you can just collect thirty bucks a month from a lot of people, regardless, like you can go sell to you know, Microsoft has these like they've been selling the last few years, like these lower tiers, like selling them to Walmart cashiers.
49:43
For enterprise usage, like people that don't think they've proven I think this has proven that it's worth more than thirty bucks a month though.
49:49
Even on those small scale. Yeah.
49:53
Start somewhere in like the general like I don't know, middle back office like role isn't you know, even front office role.
50:01
Those people aren't gonna start out spending even as they ramp anywhere close to like what coding would be or even like hundreds of dollars per month for a lot of these.
50:10
I mean when we talk to enterprises, a lot of like enterprises don't let people use frontier like you have to use Sonnet if you're like of a certain you know job type.
50:22
So models or the proletariat models, sorry.
50:26
You don't have to have the bourgeoisie models like Jeremie does.
50:31
Hundred percent bourgeoisie man. No other way.
50:34
Max effort, max of agents, you know.
50:37
Launch two hundreds of agents to find those data sellers.
50:45
All right, Doug, give us a hot take so we can wrap up because we've been going long enough. As tired as a market.
50:53
Yeah, I'm tired is the market.
50:55
My god, they're gonna clip this too, dude. I'm so it's so over.
50:59
We can't even see you man, the clip is gonna be fine.
51:04
I'll turn it hazard tunnel. That.
51:05
How's the new office, man? Is it looking good? Man.
51:08
We're getting these the Wi-Fi sucks. The Wi-Fi sucks, dude. AGI is fixed.
51:14
AGI is here when it can fix the Wi-Fi. Huh.
51:18
It found ten vulnerabilities that they publicly disclosed at CVE ten point level for Aruba Wi-Fi access points.
51:27
So maybe AGI can hack in everybody's Wi-Fi and fix it with the CVEs it finds.
51:32
Did you guys see the thing with the PHP vulnerability where somebody was proposing that you have AGI go and scan the internet for the vulnerable servers, log in, patch the server, and then lock it out?
51:47
They just need to spend five percent of training cost for that.
51:50
The five percent of rev just goes to friendly agents on the internet that help you when you're lost. Yeah.
51:57
Benevolent offensive cybersecurity models that find vulnerabilities and then patch them themselves once they find them. That's the future. All right.
52:08
Well, let's get creative on the upside here.
52:11
Underpromise, overdeliver.
52:12
Let's see if we can do that going forward. Good job, Anthropic. Nice job, guys. All right. We're done.
52:21
Daksha says, Doug, please lean into the camera and give us a smiling shot.
52:25
Smile for the algorithm, guys. Here we go. So beautiful. Looking good. Yeah.
52:33
I can make out your nose in this one. Your teeth too. There we go. We got a moment. Yeah, yeah.
52:44
Doug, what are you gonna be in what are you gonna be for Halloween? I have no idea actually. Can I log off? Just Halloween? Is this stuff going on? Can I log off?
52:56
I'll be in Paris, so you know, staying home because I otherwise I'm gonna die.
53:01
No, just he's gonna be a guy in a mask on the street with the moth telecon.
53:07
What am I doing at Halloween?
53:10
I'm probably gonna read Anthropic's S-1 because it's the biggest event ever.
53:14
So I need to read it again and again. Yeah.
53:17
Are you gonna if you have trick-or-treaters at your door, Jeremie, are you gonna hand them out Anthropic S-1 copies so they can study up at home? Yeah, exactly.
53:24
Kids, this is the future. Just read this.
53:28
No, the other way around, sorry.
53:30
Under what are you going as Halloween?
53:38
Any special costume with the fam?
53:40
No, I may I'll be on a flight to Singapore.
53:43
So I'm gonna be enjoying a nice seventeen hour flight to Singapore.
53:48
You should have you should I wanna dress up as seven percent ten year. That's my spooky thing. Yeah. Yeah. Yeah. Real fear. Yeah, yeah.
54:05
You can dress up as who's the new Fed chair? Kevin, what's his name?
54:10
Yeah, him as like the Grim Reaper. Yeah.
54:15
My daughter got a rocket costume.
54:16
She's gonna go as a rocket this year.
54:20
As a rocket emoji or just a costume?
54:23
It's just like a costume.
54:24
It's got a little bit of suspenders and then the rocket front and tail with the firing engine. Very funny.
54:31
I hope she doesn't explode when she doesn't get candy.
54:35
Or you just give her candy.
54:36
She's gonna get lots of candy.
54:38
Last year she went on a pizza slice.
54:39
I was the rest of the pizza.
54:40
So I held her and she was like the slice of pizza and I was the rest of the pizza. Steal that one, okay? That's a good idea. Yeah.
54:48
I think the wife would love it.
54:49
That costume on Amazon, I think. Yeah.
54:52
Well, anyways, happy Halloween, Spooky S-1.