第041期 - 紧急：深入解读 Anthropic 的 S-1（代币经济学/市场）

All right, guys. Welcome. Let’s talk about Anthropic’s S-1.

Once-in-a-lifetime event. Biggest day in human history.

Certainly in the history of capitalism, right, Joey? No hyperbole there.

There’s zero hyperbole. What an amazing IPO it’ll be. How capital-efficient this growth curve was, how fast it happened. You can’t ask for more.

They’re going to come in as, like, the sixth most valuable company in the world from the start of trading in the IPO.

It’ll be up there, I think. Tech companies, I think, cover 7 of the top 10.

Yeah.

So this and OpenAI will be trillion-plus.

1. Profitability

Okay, what do you guys want to talk about for profitability?

I think that’s a good place to start. They kind of give the compute margin numbers, but I think even on the non-GAAP basis, they already have 10%+ margins and have been profitable for 2 quarters. Really solid. What we’re seeing on the gross-margin line over the past 3 years—the scaling of compute and inference efficiencies—is amazing. What they’re able to spend and increase in the training splits is really impressive.

So, 85% inference gross margins, 41% compute margins, and extremely capital-efficient sales, marketing, and partnerships. Another thing: a lot of software and tech companies spend a ton to acquire a customer. Their net dollar retention number is incredible. A customer who spent $1 at this time last year is now spending $9. In Q1, that was 500%, so a customer spending $1 was spending $5.

That all falls straight to the bottom line. There’s not much you have to do to upsell the usage-based side of that. Really solid headline numbers all around the business model.

2. The Datacenter View

Yeah, I would say it’s really interesting to me because, if you talk to folks in the data center market—I’m the data center bro here—folks have been looking forward to this IPO for one big reason: Anthropic has the biggest compute demand of anyone alongside OpenAI. They’re on par with hyperscalers, and they can’t deal directly with the folks actually building the data centers. One reason might be that they have too small a team to build; the other is mostly financial. They can’t fund a data center on Anthropic credit.

We’ve seen the market kind of hating on Anthropic and OpenAI, like naked exposure, whether it be on the data center side or the GPU side, over the last few weeks. So, I think it’s going to be really reassuring for folks to see the actual financials. The first thing I wanted to say is, talk to anyone in the data center market: everyone was just looking forward to this IPO. “Dude, whenever these guys are going to IPO, I’m going to sell so many megawatts to these guys. It’s all going to be so easy.”

Obviously, that’s contingent on Anthropic being investment grade. When you look at some of these numbers—how much cash they have on the books, over $100 billion, how much they’re going to have in addition to what they’re raising, and the promises they have from Google and others—these guys are going to have over $200 billion. As Joey said, that’s a company that, in the last 3 years, despite being very unprofitable, barely burned any cash.

I think it’s extremely reassuring. Also, the compute commitments frankly aren’t that large, which clearly shows they’ve struck some good, fairly flexible deals with folks like Amazon and Google. Instead of having to commit to gigantic deals like Stargate, this huge $300 billion agreement for 4.5 gigawatts, they’ve managed to do something more efficient on the commitment side. I think that’s overall great news for AI infrastructure as a whole market.

3. Compute Commitments

Jeremie, can you walk us through? I think a lot of people see the headline—$500 billion and $600 billion of compute commitments. Can you walk us through why that’s not that large?

Yeah.

No, it’s a good question. People ask that, and when you break it down, it’s not that much.

Yeah, for sure. If you look at recent deals that have been done, such as Anthropic and Nscale, let’s take this one. The compute per-gigawatt commitment is $100 billion. There’s a world where half a trillion dollars would be just 5 gigawatts, which is really not enough for Anthropic to serve its roadmap. What they have in terms of gigawatts contracted is materially higher than that, so their effective ratio of gigawatts contracted relative to commitments is really low.

Right now, the compute they have is under 5 gigawatts. That’s going to go up materially as we get into the end of next year. Plenty of compute to accelerate revenue growth, cut the cost of models, and keep finding efficiencies. It’s pretty clear all of that is going to trigger an acceleration of revenue. These guys are going to be in the multiple hundreds of billions of annual revenue next year.

So, it doesn’t feel that complicated for them to pay half a trillion dollars, given the revenue they’re making, their free cash flow profile, and strong balance sheets. Overall, it’s very reassuring, especially when you compare it to the first freakout about compute commitments last year, which was OpenAI. People were saying, “Dude, you’re making $20 billion annualized, and you have $1.4 trillion”—right? I think we’re in much more reasonable ratios here. $80 billion, half a trillion in commitments: it’s a lot, but there’s $200 billion on the books post-IPO.

Overall, they’ve done a mix of long-term deals like Nscale, but also more flexible short-term deals with hyperscalers and folks like SpaceX, which they can cancel effectively anytime. Though, overall, 30 years ago today—can I—

4. Incremental Margins

Greatest day in the history of capitalism?

I actually just made that up in terms of the history of capitalism, but it felt right. Look, it’s a big business, dude. It is a big business. I remember once upon a time, a $1 billion run rate was the grown-up number, and they obviously did a lot bigger than that.

I think the thing I thought was best on that is incremental margins: 90% gross margin and 40% EBIT. That’s a best-in-class business. 40% EBIT is as good as it gets, if not better than the hyperscalers, if they get there, with gross margins at 90% incremental. It’s very clear. It’s on a very large base. I think it’s as clean as you possibly could want it to be.

We’ve been saying this for a long time: there are obviously really great incremental margins. There’s real revenue. When you sell it, you’re not selling a dollar for $1.10; you’re selling a dollar for 10 cents. That’s margin, right? The biggest bears have always believed there’s heavy subsidization, and it’s very clear there isn’t. I don’t know what else you want to say.

I also think the fact that Ed Zitron hasn’t tweeted is probably the most interesting fact of all. Him calling Jensen a crook last week was pretty based. You’re watching the speedrun of a short seller in real time. We’ll see. I hope he’s mentally okay. It might be a hard day to be Ed.

Yeah, so inference is a good business model, I guess.

Yeah, it’s a great business model.

Yeah. Controversial.

5. Customer Concentration

Controversial: people actually want these tokens. How about customer concentration? There was a comment in there saying that the top 5 customers represent 25% of revenue now, and then the CSP commitments, the marketplaces, represent another 50%. So, what’s the leftover 25%? I mean, it’s just a really big base, so the leftover 25% is still many billions.

I’m confused about what you mean. I think you think that it’s either/or, mutually exclusive. 58% is the total distribution. That’s the total distribution, and then 25% can look through that, if that makes sense. They don’t add together. The entire top 10 customer could be 100% on the CSP.

You think that some of the top 5 or top 10 customers are actually going through the CSP marketplaces. It implies that the numbers show that, right? To be clear, I really want to make sure we get this straight on the podcast, Jordan: 58% of distribution is their sources, right? But I don’t think when they’re talking about their net customer, they’re talking about the CSP as the net customer. They’re talking about the look-through. I’m assuming there’s some kind of pass-through, meaning that when Anthropic gets billed, they get the customer who purchases through the distribution platform, aka Bedrock. Right?

Okay, let’s just cut this part out where I’m being stupid, then, and not talk about customer concentration.

Well, no, we can talk about the 5%.

No, no, we should talk, Jordan.

No, I mean, dude, the fact that you didn’t know is important. People don’t know that the top 5 are doing 25%, which is pretty good. Honestly, most companies are more concentrated.

6. Can Anyone Catch Up?

Yeah, Doug, I think the point he’s making is that the 48% of revenue from the CSP marketplaces doesn’t include any of the top 5 customers.

Okay, great. Surprise.

Which makes sense.

Yeah, that kind of makes sense. Why would you—

Those customers would buy through that channel?

So then you have the CSP marketplace, which we know is going toward your traditional buyers: large AWS accounts and large Google Cloud accounts buying through that channel, a little more Fortune 500. Then you have another 25%, and there’s 27% left over. I think the point is that it’s still rather concentrated. If we look at alternative data sets, the top 1% of customers make up 70% to 80% of the panel. They said the top 5% of Claude Code users at these end companies are spending $18,000 per user per year. Your non-Claude Code user spends less than $75 a month.

You’re talking to 3, maybe 4, of the top 5%. Come on.

I was wondering: there are 250,000 customers on a trailing 12-month basis. I know we’re in the million-plus group—the 1,900, one of 1,900. We’ve got to be getting close at some point.

We’re one of 1,900, yeah.

At some point, we’ll be in that 250,000, I would think.

Yeah, at some point.

Yeah.

So, if Dylan’s listening to this, that’s not a KPI.

It’s not a result; it’s the result of some other KPIs being enforced. Not that we have KPIs.

7. Top Customers

Okay, but let’s talk about the actual customers they named: Customers A, B, C, and D. Customer A started at 43% of their revenue, went below 10%, and is now back up over 10%. We have our estimates for who those people are, but let’s just say—

You can’t look at 2023 numbers, I don’t think. There were customers at OpenAI that were over 10% in 2023 who are nothing today.

I don’t know about the 2023 numbers. We know some of those large customers today. There was one, I think, 10% customer, and, you know—

Who do you think it is? Let’s speculate. My meme was that it’s Kimi, MiniMax, and GM. They were 5% to 7%.

Yeah, yeah, yeah. A 5% customer.

To be clear, I don’t know if their KYC is good enough to completely cut them off. I’m sure someone at Moonshot is using Claude Code. I think I saw on a benchmark—I forget which one, maybe DeepSeek—they were using Claude Code as the final verifier to check all the results. So they’re using it, obviously, but maybe not for distillation.

Yeah.

So who do you think the top 5 are? Let’s speculate.

We know Meta was the one that rationalized spend in Q3—or, not rationalized spend; they moved from an external provider to an internal one. The tokens are still there from an industry perspective. I think Microsoft is one. Cursor we’ve talked about as being another one.

NVIDIA, hedge funds like Jane Street, HRT.

Do you think NVIDIA is that big of a spender?

Yeah, definitely. I think they’re probably the 43% that went lower. Who else was using this much in 2023?

My understanding is they were really late to getting access to Claude. They were a cloud adopter in 2024.

I don’t think NVIDIA was using it in 2023. That’s my impression. I might be wrong, but I feel like Jane Street is probably the most rational guess. Or maybe you can argue Google or something. I don’t know.

That would be crazy. I know Jane Street is producing a ton of cash flow right now, but they already have such big commitments with OpenAI. It’d be crazy if they were also a top 5 customer and spending so much on Anthropic as well. But it’s totally possible. There are only so many people who can be spending tens of billions of dollars on these tokens.

Actually, no—I was going to say, I don’t know. It feels like we’re going to figure out who it is over time, and I bet it’ll be someone really obvious that we just didn’t think about. The point of clarification for me is that these top customers are not the CSP marketplaces. If they’re using Bedrock, Foundry, or Vertex, they have a direct relationship with Anthropic through the CSP marketplace. I think that’s pretty interesting, too, because if people have that direct relationship with Anthropic and make purchase commitments ahead of time, the way they do with an AWS MSA, that’s a big lock-in for using those models.

I made another chart. They said that in Q3, 20% of revenue was from models released before April. That’s Opus 4.6 and 4.5. They said 80% was from the latest model, so people move to the latest models. But I’m shocked that 20% of revenue is still coming from models that are 3 months old. How long does it take somebody to get off the old models and onto the new ones? I’d think a Fortune 500 company buying through Bedrock, which has to certify something, would have that lag time, whereas a startup would be more nimble. Both of these things point to the models being really sticky.

If we’re saying Microsoft is a top 5 customer, then they’re probably the one spending on all the Opus 4.5. I think that’s via GitHub Copilot, too—maybe Microsoft 365 Copilot or something like the Cursor or Perplexity equivalent, where they don’t pass the relationship through to the end user. Their API key is doing all the spending. Microsoft has the Foundry business, where they do pass through the relationship, and then the GitHub Copilot stuff, where they consume so many tokens. If they had more granular data, you could probably see Microsoft’s spend be high at the beginning, dip down when they have all these OpenAI options, and then come back up now that they’re more motivated to serve Anthropic and aren’t tied so closely to OpenAI. We’ll see.

8. New Funding

Okay, what else is hot about this S-1?

Compute commitments. Jeremie, talk about the $70 billion MPA.

Sure. They’re essentially going to get even more money than they already have, which is mostly good news. A bunch of those commitments are already known if you look through Google’s and Amazon’s filings, but they’ve signed a bunch of large agreements. For example, the $40 billion with Google has multiple milestones. Google has a $20 billion commitment, and there’s another $20 billion that will be additional cash for Anthropic based on milestones, such as going public. Effectively, all of that is additional funding, which means these guys will have more money on the books and it’ll be easier to pay for stuff.

They also got the Broadcom one, which is $42 billion, enabling them to buy a lot more TPUs. Overall, the 2027 compute roadmap de-risks a lot of what we said in an article and on a podcast a few weeks ago about the data center buildout. Most of what they’re going to get next year is de-risked, and it’s going to be huge. The implication is that they’ll have all that compute next year. Can they accelerate revenue enough? I think that’s still the question. But maybe one thing people weren’t aware of was how much they have on the books right now. Now we know, and it tells you at least they’re not going to go under.

Yeah. If rates go up, they’ll make even more money, baby.

Hell yeah. If rates are 5% a year, they’re making $5 billion a year.

That gets us to almost $1 trillion, Doug.

9. Underpromise, Overdeliver

Yeah, dude. Just a little bit of the way there. Honestly, the thing I thought was most interesting in the whole S-1 is the stuff they put in. I think they’ve been pretty consistent about this. Dario’s been saying this for a long time: “We never want to overextend ourselves.” And then essentially saying, “We’ll never miss your expectations,” is an insane thing to put in the S-1.

Like, that is insane. I think this is the most nutty thing they could possibly do. Then everyone is really worried about Q4 because of what they’ve been telling the bankers for ARR, and you’re like, “They just told you they’re never going to miss your expectations.”

They’re like, “We’ve talked to all these—” Oh, I think Joey’s frozen. “We’ve talked to all these people and all this stuff. You should never be worth one more dollar than what you’re actually worth. And so we’ll underpromise and overdeliver.” I was like, “Oh my God, dude.”

Yeah, that’s a core tenet of the business: underpromise and overdeliver, Doug.

Yeah, I guess that maybe that’s in their fit interview, dude. It’s crazy.

All of this talk about—okay, wait a second. So this is a core principle of the business. 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 has a 10% chance of causing the death of humanity. This is underpromising.

Yeah. Apparently, according to Dario—

It can get worse.

It can get way worse. It’s the worst—you saw—did you guys see the head of cybersecurity? He’s like, “If you think it’s bad, it’s going to get way worse.”

I’m laughing. Scary, man.

10. Elon and Neoclouds

It is actually scary. This is going to be our last secure podcast. I was freaking out last week with all of the cybersecurity stuff, and then there was interesting stuff over the weekend, as Elon and Chamath are calling for the death of all neoclouds. Then we have this counterweight to all of that, which is, “We’ll save it with capitalism.” I guess that’s the thesis of today.

I don’t know. The money looks good to me. I think Elon is smart, and he knows when he has an advantage. He’s trying really hard to lower his cost of capital at SpaceX, and he is a neocloud dude. He’s saying that while being the largest neocloud in the world, so, yeah, it’s kind of kooky.

I would also say our view—and I think folks are starting to be really worried about this—is that there may be an inversion. Right now, data centers are more constrained than chips, but maybe by the end of next year it’ll be the other way around. Then you have this guy who says, “Hey, I’m single-handedly going to build up to 20 gigawatts, maybe just 10,” you know? So he’s going to make sure he has the chips, right?

That’s the game of Elon, dude.

One of the ways to make sure others don’t buy the chips is to make them go broke.

Yeah, just break the markets. They can’t fund all of them for you.

I mean, that’s happening right now. There’s essentially a liquidity run on all the neocloud debt, so anyone who wants to place new neocloud debt isn’t able to. If he presses his advantage—if he presses his foot to their throat and they essentially never get funding—that’s all his opportunity, right? Then everyone’s like, “Dude, I want to invest. I want to invest. I want to do this SPV.” SpaceX is the only game in town, and he’s like, “Great. That’s a better cost of capital for me,” right? Elon is a competitor.

Yeah, exactly. And as he said, he wants to do half a trillion dollars to a trillion dollars of ARR next year. The less these guys can fund their stuff, the more he can build it himself and score deals at $50 million a megawatt a year, and just make a whole lot of money.

Yeah. Jeremie, how real do you think it is that $14-an-hour GB300 contracts that are on demand and can be canceled any month just keep going because Anthropic keeps growing and can’t bring enough compute online, so they just keep it?

There are many things that go into understanding how sustainable this is. One of them is basic supply and demand. If demand continues to outperform supply growth—we’ve said very publicly that we think supply growth is going to be huge—but what if ARR and overall demand also outperform that? Then that could augment the premium. That’s a realistic scenario, because who else has as much capacity available in 2027?

If the equilibrium stays the same, there’s probably still availability to sell. And obviously, if demand starts to fade, things get more complicated. There are a few other things to look at. We talked about compute commitments. The beauty of SpaceX contracts is that they’re much more balance-sheet efficient, or compute-commitment efficient, than signing a six-year deal. You saw at scale, at around $100 billion per gigawatt—that’s likely going up, perhaps to $120 billion or more. This guy is actually at $50 billion a year, but for 90 days, so the commitment is extremely small. It’s like $15 billion a gigawatt. It’s more expensive, so margins go down, but for your books, and what people are looking at, it’s actually quite efficient.

The last thing is that the drivers of this market remain these small companies that don’t yet have the full trust of creditors and lenders. It’s going to be hard for them to fund an insane amount of supply. Even for hyperscalers, it’s not that easy. We saw tiny cracks at Meta, for example, when spreads went up. This environment is obviously very positive for companies that can take a lot of risk. These guys are basically taking more risk than anyone else by equity-funding all of that stuff. Some do as well, and they’re building stuff without a customer.

11. Compute Efficiency

Yeah. Let me throw this chart on screen. Joey, what’s your take when you saw this chart they put up about compute efficiency?

I think it’s great. Even looking just at inference compute efficiency, the revenue per megawatt in inference has gone from around $16 million last year to $60 million-plus currently. It tells even more of the story of how great a business model inference is. This is the entire fleet, so it shows what you need to do to continue investing in new models.

This is great. Obviously, in Q3 it went down a little bit, with some of the new SpaceX compute coming online, a little less monetization via Meta distillation, and some additional safety compute. But I think this really tells the story: AI model businesses are really solid, and they scale extremely well. To Doug’s point, incremental margins are amazing.

Yeah. Let me read the line from the S-1 that jumped out at me most about compute efficiency. They said, “Our compute efficiency is demonstrated by revenue that grew 1,368% year over year for the 3 months ended June 30, compared to compute and infrastructure expenses that grew 300% during the same period.”

Although, to be fair, how much of that is Anthropic being geniuses versus actual market efficiency being higher? We run inference. What do you think, Jordan? We’re seeing GB being 50 times better than Hopper, for example. Isn’t that just a reflection of the silicon and systems market overall?

Yeah, the new stuff is better. Maybe one of the other quotes I can read from the S-1 is related to their fungibility across the fleet, which is fascinating. I think they have a unique advantage in picking the right accelerator for the right job: TPU, Trainium, Nvidia GPU, and AMD GPU are all online, whereas other people are catching up. If you want to realize the true benefits of new hardware coming online, which is always more performant than the previous generation, you need the software to be able to use it. Who has the best software for writing code for this stuff? Probably the company with the top engineers and the best model.

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 bring on the latest and greatest. They have a track record of being able to adopt whatever the best silicon is for the job.

I think we also maybe underestimated the models, to be honest with you. We were in the camp where we thought maybe the moat would never be breached, but that kernel optimization would be a harder problem. We’re watching models reverse-engineer binaries, which was the gold standard for gaming. Reverse-engineering games was allegedly impossible because of all the anti-cheat stuff. Clearly, it isn’t.

Maybe we also underestimated the fact that Claude is good enough that Fable 5.5 can probably port things. That makes the AMD-versus-Nvidia question not matter as much as it used to, because the same model that can do Navier–Stokes can also program a GPU to make the economics work out. By pushing to be multi-chip, they were the first to really do it, and I think that strategy paid off for them. OpenAI was very much an Nvidia-first shop, at almost anyone else’s expense, and I think that’s been part of Anthropic’s catch-up story: supply availability. They continue to be really smart and intelligent about supply availability. They want vendors that aren’t just TSMC.

They want to make sure every aspect is diversified so no party has power over them, and they have power over everyone else. That’s just good business and strategy, and it definitely played out. The model is the differentiator.

Yeah. I view it as a race where you have to start running at some point. Some people took off earlier than Anthropic, like OpenAI or even Google DeepMind. They made a bunch of commitments, spent a bunch of money, and brought on compute. Maybe now that compute is less efficient than Anthropic’s because they don’t have the bloat of as many H100s from 2022, when they weren’t doing anything or building any compute.

But, conversely, if you’re a startup that’s getting off the ground with a billion dollars of funding in the last 3 months, you’re not profitable. You don’t have the data flywheel, and you’re missing all of this compute. Anthropic has a bunch of advantages over somebody who’s trying to get started right now.

I think the dynamic is that they struck it perfectly. They have both this financial flywheel and all the research and data that put them so far ahead of somebody else. Not only are they ahead, they’re also running faster. I don’t know how anybody catches up from here. There are kind of no moats left in terms of other companies having researchers, compute, or data that Anthropic doesn’t have. I think they’re clearly in first or second place on any of those dimensions.

You have to make a case for any of these new labs catching them in specific industries where they have great researchers and great data. But I don’t know how anybody possibly builds more compute than them at this rate. As far as compute being an advantage, they’re ahead of anybody else, except maybe Meta or Google have a flywheel that’s better than the Claude API as a way to fund this buildout with their ads business. But even that, I’m questioning as Claude goes into next year and starts to grow so quickly.

Yeah, I think the only places where someone has the potential to win are in multimodal, where I think Anthropic has always been nearsighted. There’s no other way to put it. It’s just not as good as, let’s say, Astra for vision.

Maybe you could argue that we love Aqua Voice, for example. Really niche, specialized applications can maybe take tiny slivers, but I think the big pie in the sky is definitely a proprietary VLM. I’m sure they’ll let little things loose, but in the same way that they can just vibe-code a new model, they can catch up if they ever had the time or attention.

Yeah. Well, let’s look at the domain. Let’s say inference efficiency or systems engineering on hardware is an important domain. They point Claude at it, spend some time on it, and come out on top, even though others were ahead of them. Now they’re going to build a chip. Is their CPU-derivative chip going to be better than Jalapeño [?]? Is it going to be better than what all the startups can come up with?

They’ve got Claude working for them. That’s a bit of an advantage. They’re classifying everybody else so they can’t work on designing chips using their models. So what’s next? Can they point that at voice-to-voice interaction models? Can they point that at the decision models where we saw Jev [?]? Can they point that at bio and start building cancer drugs?

I think the thesis that AI research and software engineering are fundamental, and that they come first before you win any of the other domains, seems to be playing out exactly as their thesis was.

I don’t see any reason to doubt that at this point. They’re going to have more money, more compute, more researchers, and more data than everybody else going into next year.

100%.

12. Surprises and Runway

I’ve got 2 questions for you, Joey. First, you had this great piece maybe 6 months ago, and I love the chart. That was so good. Great job. You built this model of Anthropic’s financials, and it was super accurate. I’m curious: how much of a surprise were the numbers, relative to what we had? And the second one: any disappointments for you overall on Anthropic as a company, as a business?

Yeah, I think we had the margin story right. Obviously, Q3 was a little tougher number internally for both OpenAI and Anthropic, but I think we had a pretty good understanding of the business model. So there wasn’t anything massive. I think most stuff got leaked, too, so there wasn’t a massive surprise here today.

It’s so clean. Everything has been leaked, but the big thing is, it’s like seeing all the different parts of the picture, and then you see the whole picture and you’re like—

You could nitpick a number in there, like any numerical disclosure. I don’t think you could really go in and nitpick or say, “Oh, that’s not really—”

Maybe you can be really particular about off-balance-sheet items or the debt commitments to that.

Yeah.

We all knew that, too. It’s nothing crazy, dude.

Jeremie, what was the second question?

Sorry. Any disappointments with regards to the business overall in this one?

No. If you’re of the view that every employee at an enterprise is going to pick up their spend, or even that the median looks like what the 75th percentile is today, they don’t have to get up to the 90th or 99th percentile user. There’s a ton of runway here just in enterprise work.

The really interesting number I shared earlier is that the top 5% of users are at $18,000 per year, and the non-Claude Code users are at under $75 a month. We’re at 0.1% of Fortune 500 OpEx budgets today. That’s under $8 a month per employee, or under $8 to under $50 per employee.

Joey, is it time to apply the 0.85x ARR-per-cohort SaaS number? So 900% is just going to be 0.85 for the rest of the cohort’s life? Sorry, that’s some real brain-dead SaaS-bubble stuff. I don’t know if you remember that.

No, if you look at the cohorts, I think the other big thing as we look toward 2027 is that we can talk about cyber and traditional white-collar work. Anthropic is obviously extremely coding-based, but there are a lot of other verticals these cohorts are going to rip through.

We’re going through the first Q4 traditional enterprise IT budgeting cycle for AI. That’s the point: it’s such a small percentage today of Fortune 500 and Global 2000 companies. Most people aren’t a top 1% customer.

What inning are we in, Joey?

I like Doug’s comment today: we’re still paving the foundation, the parking lot. We’re painting the lines. The asphalt—we’re that early.

The park has not been built. What are we doing?

Yeah.

They’re still grading. They’ve got the dirt.

Oh, I was dying when you guys said the cement.

And the best part of this, Jordan, is we’re already at like $120 billion of ARR combined between the 2 frontier labs.

If that’s where we’re at today—

You know, maybe one day they can get to the enterprise SaaS revenue penetration of Microsoft Teams or Salesforce and start monetizing more than $8 a month per user.

Yeah.

Right. And no, that’s—

That’s a big thing. Some people might say that goes to Microsoft 365 Copilot, but I’m—

There are suspicions—no, I mean, there will always be demand for frontier intelligence. As we look toward next year, I think the pieces are in place to have a really strong 2027. We’ll look at 2028 when we get there, but it’s very early innings.

13. Market Reaction

Beautiful. Doug, what’s your take on the market reaction when this came out? How does this affect other things? We talked a little bit about neoclouds and debt generally. What do you think happens to everybody else when they come public?

I would think, on the cleanest quarter—or, not the cleanest, a pretty clean S-1—the reaction in semis, and all the AI beneficiaries, was that the land-buildout companies were the number-one winners today. The data center side is really starting to pick up the bid.

But it wasn’t as strong as what we’ve been seeing all year. I think the AI trade is exhausted. There’s no other way to put it. The AI stocks are tired, but we just got a refresh of Q4 budgets. Who knows?

They’re so tired, dude. They’ve traveled so far.

Bro, this is going to go viral on Twitter. You’re going to see Taiwan pick up your face in a video analysis.

They’re not going to see it, man. He’s just like Batman with that light coming behind him. We just talked to a shadow of Doug now.

14. Holiday Usage

Well done. Maybe the one letdown I can come up with is that they talked about the low usage over the holidays.

So, for Q4, what sort of revenue impact are we expecting Christmas vacation to have on Claude Code usage?

They said it was a 20% decline compared with the 2 weeks prior. On the month of December, let’s call it a 10% cut. I imagine it’ll be less pronounced this year.

Why less? I think it gets worse, man. Maybe we can see a Polymarket for this one. The line says, “Consistent with this pattern, for the 2 weeks ended January 2, 2026, total customer spend declined by approximately 20% compared with the prior 2 weeks ended December 19.” I feel like the mix has gone toward consumption over the all-in products, and I think a lot of people log off over the Christmas holidays. They just don’t use it. That’s what’s going to happen.

I think Claude Code is more part of the average user’s daily routine now, even if they’re working 2 or 4 hours a day, than it was at this point before.

Jeremie, are you going to be on your phone Claude coding while you’re having Christmas holidays?

You have Muse now, bro. Muse is for this. Muse is actually taking vacation dollars out of Claude Code, man. So, yeah, I don’t know. I’m not sure. I use Perplexity Computer now. Maybe I’m a weirdo.

Oh, breaking news.

This is a call for frontier labs: please make the experience of using AI models on the phone, continuing on the computer, and then going back to the phone better. Shout-out to Perplexity; they do really well.

I didn’t see that as a material risk in their S-1. Jeremie moving to Perplexity—I’m surprised they mentioned that one.

Yeah, and you’re going to see it in the Q4 ARR. It might go down.

Big partner dropped off.

They’re going to call it another one-off after Meta cuts. We had a one-off: Meta left to use its own models, and somebody named Jeremie Eliahou said something else is gone. What models do you choose, though? Are you still using Claude, or have you gone for some Codex models through Perplexity?

You just tell it, “Use whatever state-of-the-art model, subagent, max effort,” and let it figure it out. Just use the best stuff.

Generally, I think the orchestrator is Fable, though, by Fable 5.1.

Okay. Yeah. So, Fable 5—you know, Codex hasn’t even been considered here. We obviously saw Opus 5.5 come in, and that had a big impact. I love that model. Fable is still on 5.1. What happens when they just keep ripping that? Do we see that before the next version of Astra? I’ve got to expect so, right?

Yeah, we’ll see 5.5 through the IPO. You have to juice the first-quarter numbers.

What happens when they release Fable 5.5 fast mode at $500 output tokens? It’ll be a good one.

They don’t like to disappoint expectations, is what I’m thinking.

They’re going to release a faster tier in Q4, too, with 5.5.

Promise, overdeliver.

Underpromise, overdeliver. Yeah.

Dude, the numbers people are saying for Q4 are so stupidly low. They’re not believable.

They’re actually going to make a new class of model: Gigafal [?].

15. Is OpenAI Cooked?

Gigafal [?]. The unreleased models are coming out, honestly. Okay, let’s talk about what we haven’t talked about: OpenAI. Is OpenAI kind of cooked? Astra is their best. Maybe their base model improves or something like that. But where is OpenAI? Opus 5.5 accelerating is pretty significant. It comes at someone’s direct expense, and at least within the SemiAnalysis spend, it’s very clear who it is: Astra. What do we think about OpenAI? They’re caught in the middle of everything. They’re in the middle between consumer, on Gemini and Meta, and Anthropic and open-source models for enterprise. They’re literally number two in everything, which in markets is the single worst place to be.

I mean—

Yeah, I kind of said it earlier when I was talking through Anthropic’s justification for why they have a lead and why they have a moat. I think OpenAI got running a little quickly, and they might be a little bit behind now. They’re certainly behind in ARR. They land some counterpunches in model capability, and then Anthropic comes beyond them. If people make significant prepayment commitments to Anthropic, that’s tough for OpenAI. But I can’t see all of this and not be bullish on both of them. Isn’t there a need for multiple models from multiple providers for different things? I still use OpenAI models for lots of stuff that I don’t think Claude is good enough at. I find them more reliable for data analysis, for example. So, I don’t know. I’m not an OpenAI bear by any means on this result.

I feel like every day someone gets the vibe that, hey, today Claude is great; today GPT is terrible. Believe it.

His first tweet is, “Hey, is Opus 5.5 dumber?” Genius.

Yeah.

I don’t know. I feel like every week we have some big user shifting from one to another because the model feels dumber for whatever reason. So, if both are good, I don’t know why OpenAI would be done. The lowest-cost producer with the highest margin is Anthropic. You see this in semiconductors, too: number two has less share and worse margins. OpenAI isn’t done. Obviously, 90% incremental gross margins tell you they can figure out a business. They can put their heads down and figure it out.

Pick up a few hundred billion dollars of scraps.

Yeah, there’s money to be made for them. But I think their positioning, their competitive positioning, is worse because if you’re in the lead, you charge a premium margin and get a premium product. It signals that you’re a higher-quality, luxury product. OpenAI will now always be competing with Anthropic on price, and that’s such a shitty place to be. We’ve seen that one of the big takeaways is that AI is about best-in-breed. If OpenAI isn’t best-in-breed, then they’re competing on price, and that’s a worse place to be.

Yeah, overall, I agree. One pushback I’m curious what you guys think about: I’ve heard that OpenAI is much better at enterprise sales right now. They’ve been more successful at signing a lot of contracts with these, call them boomer, Fortune 500 enterprises. So maybe that’s one bull case for OpenAI: as enterprise spend starts to pick up, as Jordan was saying, it’s such a small portion today if you exclude the superpower users. Maybe that’s good for OpenAI as the better enterprise sales platform.

Yeah, OpenAI has been very bullish on its Q4 enterprise ramp. Bedrock will ramp a ton for them, and they’re going on Vertex as well. It’s a different business model. They don’t have the third-party share. They haven’t had that; it was majority consumer up until a few months ago. They’re back on the right track. There’s an argument that we’re ping-ponging between model share going forward, based on who’s best. If you’re signing these 1- or 3-year enterprise deals, you’ll see less of that, and switching costs will be a little higher. There’s also been a lot of news in the last week about them aggressively discounting against Anthropic accounts to win some of that business. They’re willing to fight for it, and they’ve taken enterprise a lot more seriously since July with the release of 5.6. They executed well and had a really good Q3. Their margins don’t look anything like Anthropic’s, but they’ll get there.

They have just too much consumer.

16. Government Revenue

Yeah, they have a lot of consumer subscriptions. You have to support all the free users, too. If only 6% of the billion weekly active users convert to paying, that’s tough. They’ve also had very little of this usage-based API revenue we talk about monetizing at $100 million per megawatt plus.

How about government as a proxy for what’s going on there? They said something about government. Let me read the passage.

It’s sub-1%, right?

Yeah, sub-1%. “Our customers include U.S. and non-U.S. government agencies. Revenue from government agency contracts represented in aggregate approximately 1% of our revenue in each of the periods in this prospectus.”

That’s a huge market for everyone.

There are a lot of seats you can sell in government.

Let’s size this relative to all cloud or something like that.

Yeah, look at GovCloud or FedRAMP, or any of those programs. But then just look at the number of Microsoft 365 seats out there.

How many Microsoft 365 commercial seats are there? Hundreds of millions. If you can collect $30 a month from a lot of people, regardless, you can sell to them. Microsoft has been selling these lower tiers to Walmart cashiers for enterprise use.

I think they’ve proven that it’s worth more than $30 a month, though, even on the—

Exactly. But you have to start somewhere. The general middle- and back-office roles—and even front-office roles—aren’t going to start out spending anywhere close to what coding would cost, even as they ramp up. A lot of enterprises don’t let people use frontier models; if you’re in a certain job type, you have to use Sonnet.

17. Wrap-Up

All right, Doug, give us a hot take so we can wrap up. We’ve been going long.

I’m tired, brother. I don’t have any more hot takes. I’m as tired as the market.

Oh my God. They’re going to clip this, too, dude. I’m so—It’s so over.

We can’t even see you, man. The clip is going to be fine. You’ve got to fix that. How’s the new office, man? Is it looking good?

It’s nice, man. We’re getting used to it. The Wi-Fi sucks.

The Wi-Fi sucks, dude. AGI is here when it can fix the Wi-Fi. It found 10 vulnerabilities that were publicly disclosed at a CVSS 10.0 level for Aruba Wi-Fi access points. Maybe AGI can hack into everybody’s Wi-Fi and fix it with the CVEs it finds.

Did you guys see that thing with the PHP vulnerability where somebody was proposing that—

—you have AGI scan the internet for vulnerable servers, log in, patch the server, and then log out? They just need to spend 5% of the training cost on that. Five percent of revenue just goes to friendly agents on the internet that help you when you’re lost.

Yeah. Benevolent offensive cybersecurity models that find vulnerabilities and then patch them themselves. That’s the future.

All right. Let’s get creative on the upside here. Underpromise, overdeliver. Let’s see if we can do that going forward. Good job, Anthropic. Good job, guys. All right, we’re done. Dax says, “Doug, please lean into the camera and give us a smile for the algorithm, guys.” Here we go.

So beautiful.

Oh, Doug is looking good. I can make out your nose in this one. Your teeth, too. I won’t—

It’s for the Taiwan videos.

We’ve got a moment. Yeah. What are you going to be? Sorry.

What are you going to be for Halloween?

I have no idea. Actually, I haven’t thought that far ahead.

What are you going to be for Halloween?

Is this still going on? Can I look off?

Yeah.

I’ll be in Paris, so I’ll be staying home. Otherwise, I’m going to die.

He’s going to be a guy in a mask on this trip.

What am I doing for Halloween? I’m probably going to read Anthropic’s S-1 because it’s the biggest event ever. I need to read it again and again.

If you have trick-or-treaters at your door, Jeremie, are you going to hand out copies of Anthropic’s S-1 so they can study up at home?

Yeah, exactly. I mean, kids, this is the future.

Underpromise, overdeliver. Oh no, the other way around. Sorry. Joey, what are you going as for Halloween? Any special costume with the family?

No, I’ll be on a flight to Singapore.

Jeez.

I’m going to be enjoying a nice 17-hour flight to Singapore.

I want to dress up as a 7% 10-year. That’s my spooky costume.

I would love that.

Yeah. Everyone should be scared. That’s real fear.

Yeah. Who’s the new Fed chair? Kevin—what’s his name? Yeah, him. He’s like the Grim Reaper.

My daughter got a rocket costume. She’s going to go as a rocket this year.

As a rocket emoji, or just a costume?

It’s just a costume. It’s got a little bit of suspenders, and then the rocket front and tail with the flames.

I hope she doesn’t explode when she doesn’t get candy. Or you can just give her candy.

She’s going to get lots of candy. She was a pizza slice, and I was the rest of the pizza. So I held her, and she was the slice of pizza and all.

Oh, that’s so sweet.

Joey, you’ve got to steal that one.

Okay.

That’s a good idea.

Yeah.

It is a good one. You can just buy that costume on Amazon, I think.

Okay. Well, anyways, happy Halloween. Have a spooky one.