Anthropic IPO前に知っておくべきこと — Tanay Jaipuria(Wing VC)

⚡️Anthropic: Know Before You IPO — Tanay Jaipuria, Wing VC

Latent Space: The AI Engineer Podcast53分
Anthropic IPO前に知っておくべきこと — Tanay Jaipuria(Wing VC)

Anthropic IPO前に知っておくべきこと — Tanay Jaipuria(Wing VC)

Latent Space: The AI Engineer Podcast

0:0053:37

要約

Latent Spaceが、Wing VCのTanay Jaipuriaを迎えてAnthropicのS1(上場目論見書)を読み解いた回。冒頭でswixが、想定時価総額1.3〜1.4兆ドル、約7倍の応募超過、10月の年換算売上810億ドルなど直近の動向を整理する。対談では、キャッシュ効率の高さ、学習・研究に偏った計算資源配分、モデルの鮮度への依存リスク、NeoCloud/NeoLabの見方、ライフサイエンスの位置づけが論じられた。

この回のポイント

  • ●冒頭でswixは、Anthropicの想定価格帯は70〜78ドル、時価総額は1.3〜1.4兆ドルと述べ、10月のランレートは810億ドル、年末1000億ドルの見通しを紹介した。
  • ●Tanayは最大の驚きはキャッシュ効率だと指摘する。調達した約1400億ドルのうち約1320億ドルがバランスシートに残っており、Uberなど過去の上場企業より少ない資金消費でこの規模に達したと述べた。
  • ●効率の理由として、コーディングとエンタープライズ中心で無料ユーザー向け推論が少ないことを挙げる。粗利率は報告上85%で、無料枠の推論は計算支出の約5〜6%にとどまる。
  • ●計算資源の約70%が学習・研究に使われ、推論は約30%。外部データへの支出は直近四半期に急増しており、RL環境型のポストトレーニングが効いている兆候だと解釈している。
  • ●売上の約80%は直近6か月に出たモデルから生じ、APIの従量課金が中心のため、モデル競争で数か月遅れるだけでシェアと売上を失うリスクがあるとTanayは述べる。
  • ●約9000億ドルの計算資源コミットメントに対し、ライフサイエンスなど自社でコントロールできる計算の使い道が、需要拡散が遅れた場合の保険になるという見方が示された。

章立て

  1. IPO直前のアップデート

    swixが価格帯、時価総額、応募状況、10月ランレート、ロードショーの論点(Cowork、チップ、利益率)とモデル発表の影響を整理する。

  2. S1を読む理由と売上の数字

    Tanayが上場企業のS1分析を書く動機を語り、ランレートの算出方法や月次成長率の見立てを議論する。

  3. 驚異的なキャッシュ効率

    調達額に対し手元資金がほとんど減っていない点を、AirbnbやUberなどと比較。OpenAIとの違いや、無料ユーザー推論の少なさも説明される。

  4. 成長鈍化と次の需要源

    Q2の急拡大からQ3の減速、トークンの使い込みが終わる懸念、コーディングに続くセキュリティや科学の可能性を論じる。

  5. 計算配分とデータ支出、著作権和解

    学習・研究が計算の70%を占める点、データ支出の急増、15億ドルの和解とBooks3の背景に触れる。

  6. モデルの鮮度とチャネル依存

    売上の約半分がチャネルパートナー経由で、利用の80%が直近6か月のモデルという構造が、競争上のリスクになると指摘する。

  7. IPOのタイミングとNeoCloud/NeoLab

    早期上場の利点、エコシステムの流動性、NeoCloudとNeoLabの見方、オープンモデル推論の利益率を議論する。

  8. ライフサイエンスと締めの展望

    9000億ドルの計算コミットメントとライフサイエンスの役割、月次成長率が注目点になるという結びで終わる。

解説記事

Latent Spaceのswixが、Wing VCのTanay Jaipuriaを招き、Anthropicの上場目論見書(S1)の中身とIPOの意味を語った回だ。冒頭ではswix自身が、上場前日までの最新状況を整理している。以下は番組内の発言に基づく要点である。

IPO直前の状況

swixによれば、価格帯は70〜78ドルで、想定時価総額は1.3〜1.4兆ドル。約7倍の応募超過で、80ドル前後に決まり、初日の上昇を含めると1.4〜1.6兆ドルのレンジになりうるという。10月のランレート(年換算売上)は810億ドルと新たに開示され、年末に1000億ドルという見通しは維持されていると述べた。ロードショーの論点としては、Claude Coworkなどコーディング以外への展開、NVIDIA・AMD・TPU・Trainiumにまたがるチップ戦略、利益率(調整後で19%から15%へ低下、長期では35%超を目標)の三つが挙げられた。OpenAIやGeminiの新モデルが出ても、ストーリーは大きく変わらなかったという認識だ。

驚きは成長ではなくキャッシュ効率

Tanayは、ランレートの急伸は事前の報道で想定内だったと言い、最大の驚きはキャッシュ面だったと述べる。調達した約1400億ドルのうち、約1320億ドルがバランスシートに残っている。上場時点の売上ランレート1ドルあたりの資金消費は、近年の上場企業ではAirbnbくらいしか近くなく、絶対額でもUberより少ないと指摘した。swixは、OpenAIはSnowflakeやCoreWeave、Cerebrasに近い水準で、はるかに非効率だという見方を示している。

効率の背景として、Tanayはコーディングとエンタープライズに集中し、チャネルパートナー経由で配信している点を挙げる。無料ユーザー向けの推論が少なく、計算支出のうち無料枠に使われるのは約5〜6%にすぎない。報告上の粗利率は85%だ。一方で、計算資源の約70%が学習・研究に使われ、推論は約30%という配分だった。推論自体が非常に効率化していると、Tanayは見ている。

成長の持続性とモデルの鮮度リスク

Q2に年換算で約250億ドルを積み増した後、Q3は四半期成長率が100%超から40〜45%程度に下がった。Tanayは、これが「トークン使い込み」とAnthropicの一時的な優位の反動なのか、という問いが市場にあると説明する。オープンウェイトモデルの台頭やモデル間のルーティングも逆風になりうる。ただし技術の普及はまだ浅く、リスクはそこまで大きくないという立場だ。

コーディングに次ぐ需要源としては、カスタマーサポートや法務、そしてセキュリティが候補に挙がった。さらに長期では、科学やライフサイエンスが語られている。

もう一つの重要点は、利用の約80%が直近6か月のモデルから生じ、売上の大半が従量課金という構造だ。数か月でも遅れれば、API層では簡単に乗り換えられてシェアを失う。自社アプリの利用を増やすほど定着性が高まると、Tanayは述べる。なお売上の約44%はチャネルパートナー経由だという。

データ支出とNeoCloud・NeoLab

データへの外部支出は、従来は年10億ドル弱と推定されていたが、直近四半期で急増し、年換算で20〜25億ドルに近いとTanayは試算する。RL環境を使ったポストトレーニングが成果を上げている兆候だと解釈した。また、書籍の著作権和解(15億ドル)にも話が及び、swixはBooks3というデータセットの経緯を紹介している。

NeoCloudとNeoLabについては、平均的にはNeoCloudのほうが有利だが、NeoLabのうち一、二社が極めて大きくなれば後者のバスケットが上回りうる、というのがTanayの見立てだ。swixは、計算資源を確保したNeoLabが活用先を失うとNeoCloudになる、と補足した。オープンモデルの推論事業については、価格競争が激しく、素のトークンでは長期的に40%超の利益率は難しいと予想し、ポストトレーニングなどへの展開が進むとみている。

9000億ドルの計算コミットメントと科学

Tanayは、約9000億ドルの計算コミットメントが懸念材料だと述べる。来年は700〜800億ドル、その後は1000〜1200億ドルの支出になりうる。需要が追いつかない場合の備えとして、ライフサイエンスや新素材など、自社が主導して計算を使える領域に意味があるという。顧客企業の導入時期に左右されない使い道になるためだ。

まとめ

今回の議論から、Anthropicの強みは成長率そのものより、資金効率とエンタープライズ中心の収益構造にあることがうかがえる。一方で、モデルの鮮度への依存と巨額の計算コミットメントが主要なリスクとして語られた。日本の読者は、上場後も月次の売上成長率と新モデルの競争力が焦点になる、という番組の見方を押さえておくとよい。なお、発言は出演者の見解や推計であり、S1の数値そのものとは区別して読む必要がある。

文字起こし(英語・自動生成)

Good morning. Today we're doing our last Anthropic IPO podcast before the IPO and TBD on whether or not we'll actually continue covering it after the IPO because that becomes the domain of finance podcast. But I think it's still obviously very relevant to a lot of listeners, particularly those of you that work at Anthropic, but also those of you that work at OpenAI and everyone else connected in the ecosystem. them. The IPO Roadshow has been going on for a couple of weeks, and we know a bit more than we used to before. And we covered a lot of it in the Jam and Ball podcast. That went really well. Thank you for all the positive feedback. And I think I try to pull in people that I respect and look up to in terms of the crossover technology and finance, because I mean, I think that is, if But if anything, my career has set me up to do that. So we have a price for the Anthropic filing, I guess, between $70 to $78, which is an implied valuation of $1.3 to $1.4 trillion.

It looks like it is roughly seven times oversubscribed and probably will price in a range of $80-ish. And then there's also Ford contracts or crypto contracts on speculating where the final pop is going to end up. If it looks anything like the SpaceX IPO, there will be a small pop. It will be about 19% to 20% on the first day. Maybe it would go up to $100. And so that is in that range of, let's call it 1.4 to 1.6 trillion. And that represents roughly 38 times trailing 12-month revenue, 20 times current revenue, and obviously what for revenue is is just dependent on what your projections are. It's still anticipated to – so October run rate was $81 billion. That was also a new disclosure. We'll still project it to end the year at $100 billion. And what really people are thinking about is like, well, where does it go from here, right?

Like what is the market pricing? This is obviously the highest multiple that anyone's ever priced at, especially at the size. And, you know, the ranges for the future, like, future years, like 2030 or 2028, really range from, like, $400 billion to $1 trillion in revenue. And so that is a very wide range. But you kind of have to do this kind of math in order to underwrite long-term investments in antibiotic. And so you can see by people looking ahead, you know, I think 11, 12,000 customers spend at least $100,000 per year and per customer, but once they land in a company, they also expand nine times in their future years. So there is NDR, a very good net dollar retention. Usually state-of-the-art is, let's say, 150%, and this is 900%. So really the discussion about the roadshow has focused on three things.

One, which is Cloud Cowork. It is basically the next phase as far as the market narrative is. How well are you extending outside of cloud code? I think there is probably going to be a different form factor, I would say, than cloud code work because we did the first cloud code work part with Felix, if you scroll back in our archive. And it's a form factor that's there, but I think it's not very cloud native. and definitely the sort of autonomous teammate that they are pitching for their S1 that we covered in the first podcast, that is definitely going to have to be cloud-native. That's cloud-duet. That is the instinct. That is the grok bots. That is the cognition bots. All those things are like everyone is trending towards that direction, and I don't see why cloud would not do that, and probably that is going to be something different than cloud-cowork. The second one is chips. Enthlopic wants to run across NVIDIA, AMD, TPUs, Tranium, and they also released information

about how they were easily able to migrate much, much more easily from NVIDIA to AMD. And that is obviously part of the whole RSI narrative that people are concerned about. But I think that obviously just makes a lot of business sense for them that, like, hey, if I'm bottlenecked on NVIDIA, I'm sorry, I have to do this, right? And by the way, it's verifiable, and by the way, it's one of the core tasks that is easy for LLabs to do. Third is margins. So we have a little bit more in terms of disclosures on margins. We know that margin went down from 19% to 15% using adjusted margin. It went from a gap operating profit to a slight gap loss, and that is just intentional and not a big deal. Enthopic is projecting that it can get above 35% in the long run. So that would be a very big deal because then it would be very profitable. And that's something that, again, as we discussed in this podcast and the previous one with Chapman,

that is just unheard of for a large tech company of this size IPO. Usually you're unprofitable, like deeply unprofitable, like negative 100% profits for, I don't know, three, four years before you even start to turn it around. And that's good. I mean, this is growth with profit, which never happens. I think the final thing that happened this week is OpenAI's model launch. As we already established, model launches are incredibly, incredibly important to the viability of the frontier labs. The shelf life of a frontier model is roughly six months. And if you don't release the next model at the frontier, then obviously the usage is not going to come to you. and it is very sensitive to benchmarks and vibe checks. And so OpenAI shipped GPT 6.2 and Fable, Anthropic shipped Fable 5.2. And obviously Fable 5.2 is still on top, and I think the vibe checks are mostly holding out that Fable is on top.

So there's that, and then there's the efficiency frontier, which OpenAI does claim. I think people are a little bit more mixed into practice in terms of there's benchmarks and then there's just like impractical use. And I would say that basically both model launches have not really changed caps either way. If you're a cloud person, you continue to be cloud. If you're happy with OpenAI, you continue to be OpenAI. Those key risks for the IPO that may have dramatically changed the story, either for the positive or negative, haven't really changed. There was no Fable 5.5 that people were anticipating. and Gemini shipped, Gemini 4, that did not affect the narrative, even though Gemini Carbon did very, very well. And that's about it. So those key risks are behind us for the Anthropic IPO. Anthropic is still the top model lab in the world, and it is IPOing tomorrow. And it is going to be a blockbuster IPO, and that's really good for the ecosystem because a failed Anthropic IPO would have read through really badly

together with the NVIDIA drama that is also happening at the same time. So that is the overall update in the last week. You are mostly prepped. This is our podcast that we recorded to go through some finer details in the S1. And I hope you enjoy, and good luck tomorrow. Okay, we're in the remote studio. This is another in our series covering the Anthopic IPO with Taneja Puria. Welcome. Thank you for having me, Swix. Long-time listener of the podcast, and I'm excited to be on. Tane, how come you're writing S1 breakdowns? Yeah, no, I think in general, I guess, maybe for the last six or seven years, I've been writing about various things, like more on long-form and on Substack, and one of the topics I've been writing about is just like breakdowns of interesting companies as they go public, and I think the S1 is usually the time when as an outsider, you get a look into all aspects of the business.

And the reason I enjoy reading them so much is that you can kind of get a sense of the story of the business so far. And I feel like in every business, there are some interesting lessons that people take away about business in general, about specific kinds of businesses. And so I used to follow them from that perspective. And then obviously with the last year or so, some of the ones we've had go like with SpaceX and now with the Antropocas one. And these are like, you know, truly generational defining companies. And so I think it's been even more fun to look into those. And I think they do give you a sense of the current state of those companies specifically, but then also the market much more broadly. And so the Anthropics one kind of gives you like a full view into everything from, you know, the idea of like, you know, the model development, inference, that everyone's talking about like inference as this, you know, D-next big, huge market. It gives you a sense of all, you know, the current state of all the compute needs and like NeoCloud and then also what's to come in the future

with things like science and some of those things. And so I think it's obviously like a category-defining company and one that's still relatively early in incidents, but it's kind of crazy that, you know, in five or six years, it's created this much value so quickly and it's something we've never really seen before. across any company, really. Yeah, for sure. And, you know, I'm happy to dive right in there. You have a lot of charts that have been prepared, and I can go wherever you want to take this. Headline numbers, you know, they're ramping revenue a lot. It's really interesting, basically up like 13, 14X from same quarter last year. I think the run rate numbers just leaked for October, which we knew was something that they had in store, and they kind of left blank in the S1. But now they're reporting 80 billion run rate, right? Yeah, so I think it's 80. Divide by 4 is 20, yeah. Yeah, and I think my guess is the way they calculated that 80

is this kind of more like on a monthly basis. And so I think the exact definition they might use is actually 28 days times 13 plus monthly subscription revenue. And so it's interesting because I think it's being, even on Twitter, I think the reactions to that will mix where there are some people saying, oh, this is like a very good, very positive number and signals reacceleration, whereas some are saying that it'll mean that they likely will fall short of the 100 exit run rate that they've been talking about. I think it's hard to tell, but I think it probably is in that kind of 8% to 10% month over month growth, which I think is, I don't know if it's a recaleration, but it's at least roughly in line with what they did in the last few months. It might be slightly up is my guess, but hot to fully know at this point because they haven't broken down Q3 and stuff by month. Yeah, yeah, totally, totally. 80% consumption, 20% subscriptions, which is kind of funny. It's a pretty public secret how much of a discount the subscriptions get. So actually people are incentivized to make this 80% consumption base lower.

but you have to sort of figure out the subscription system. And, you know, number of customers over 100K is up eight times to 12K. Yeah, you know, like, I'll let you drive. Like, let's assume that people have looked at the headline stuff and they know about the profitability. What are the major topics that you think are the main messages that people should take away? Yeah, I think at various points across the last few months, there have been all these leaks about roughly their run rates and stuff. So I don't think, to me, it's kind of like those things were not as much of a surprise to anyone that's kind of been following this in terms of what to expect based on what has been publicly reported even before the S1. I think perhaps the biggest surprise to people was just around the cash flow of the business and in a positive way. So they raised like, you know, $140 billion of cash and literally like they still have on the balance sheet like $132 billion or so.

And so they've bought like under $10 billion to get to where they have. And I think it's like one, no, I don't think other than, you know, obviously the people that actually knew the numbers from being on the inside, I don't think anyone expected that the cash war chest that they had on their balance sheet was actually that large just based on, you know, based on expectations around the burn and the margins and things of the business. And so it's actually been like a very, very, very efficient business. And I kind of had just this quick analysis around, you know, given the scale they've gone to, right, which is, you know, roughly 70 billion at the time of IPO if you do like the quarterly times four what the capital consumed in terms of like the cash burn for a dollar of kind of revenue run rate at the time of the IPO And there are very few businesses of any meaningful scale really that we seen go public over the last five seven years that are anywhere close Like, I think Airbnb is probably the only one that kind of comes close

on a revenue run rate basis in terms of the cash efficiency in terms of the burn. And so I think this was something that was quite surprising. I think even from, like, an absolute dollar basis, Like, I think Anthropic has literally burned, like, less dollars than Uber did to get, like, at the time when they went public, which, you know, given what they've created with this is, you know, pretty remarkable. And it's like, you know, the cash flow, and I was thinking about it, it's like, like, Meta is still burning, like, $10 billion on their Reality Labs initiatives every, like, two quarters of a year. I think it's now down to, like, $16 billion a year, so maybe they burn, like, $8 billion. But they're still burning every year way more. That's still as third. Yeah, then, like, you know, this company that got off the ground and, like, built, like, you know, basically built, like, Long Horizon agents and, like, you know, was the first of so many of these things. Wait, I expected them to take it to zero. Like, what the hell? No, yeah, they're still losing, like, $12 million in the photo. What are they doing? Yeah. Yeah. What?

It's, like, they're burning, like, a million dollars per spectacle for glasses or something. Like, what the fuck? Yeah. that's great that's crazy so i expected so my my actual pitch for msl was look the uh they were burning something like 20 billion uh per year uh on on on on vr and you just cut that 20 and reallocate to msl uh and you you have a new lab right like yeah and which is they now have it it's just that his pockets are so deep that he can just do both yeah he was like i'll do that but maybe i won't do the cut fully and like i'll slowly do the cut yeah i mean like clearly like he's enjoying it he's loving life uh whatever but like uh you know these glasses are nothing compared to the news right um yeah so uh it's so funny oh man i can't believe they're still burning five billion a year. Five billion per quarter. Yeah. Come on. All right. So

that's the cash burn. I think that's something that I feel like is underrated. So the one number that you're not seeing on here is where's OpenAI, right? Because you don't have the OpenAI S1. It is way more inefficient. It is probably closer to the Snowflake, Cerebris, CoreWeave range. There have been unofficial charts. So this was not a completely new fact if people were paying close to the leaked S1, where people were already drawing charts like... Anthropic was, let's say, doing... was known to be doing like this as like a projection. And then OpenAI was kind of like way deeper. Way, yeah. Something like that. And so this is a known chart. We just didn't know how big it was. Yeah, I think the magnitude was impact. Yeah, yeah. So I think I attribute this to a couple of things, right? Anthropic started later.

So they basically skipped GPT-123 because that was done on OpenAI's, you know, that R&D. Yeah, it was done on their dime. It's the same founding team. So, like, you know, they already paid the tuition there, and now they come out and they can just kind of repeat that work from a clean room perspective. But, you know, what else? Why is Anthopics so efficient? I think that is a really, really under-examined thing. Partially, I think, it's just straight up, there are now unit economics profitable, which is cool, but they also didn't burn that much to get there, which is super unusual, basically unheard of in tech. Yeah, and if you go to the next, the chart right below around kind of gross monitoring, I think, and this also probably contrasts something with OpenAI, which is that because Anthropic focus ends up focusing on coding and on enterprise maybe more generally and more just like, let's have the best models

and both on let's distribute primarily through channel partners, but then also let's leverage them for compute wherever possible. And they didn't have this, because they weren't trying to do consumer, they didn't really have this like massive consumer base that they had to support with, like, you know, mostly free inference. The economics are, like, very, very positive for them on kind of, like, you know, on supporting, like, their paid users. And so I think the most common number, like, you know, is, like, this kind of idea of the 85% gross margins that they report. And the way to think about that is that essentially, like, as a percent of revenue, their, like, inference on paid users is about 15%. And there's a chart further below that kind of breaks that down. And so they reported these 85% gross margins on essentially paid inference. But then one thing that a lot of these companies obviously have been doing is that there's these free-tier users. And it's been common to put this into sales and marketing spend because you could argue that it's like, oh, it's how we attract users to a platform and then we monetize them if they become paid.

And so they put that free-tier inference in sales and marketing spend. and so you could be like oh maybe that should go in cogs and but i think the point is that entropic because you know they primarily relied on enterprise and on um coding claude is not as big as like you know chat gpt is uh from a margin perspective and maybe like a cost uh cashbone perspective that actually is a positive for you because you don't have this huge user base you have to support on on the free plan and so they only spend like you know only about like 4% of margin points or about like 5-6% of their compute spend ends up going on like this free user inference. And so I think that's kind of like the other aspect of it. I imagine when the numbers come out for OpenAI, that's going to be significantly higher just given the size of that user base. Yeah. Do we, is 4% is small, right, for any feature? I don't know how you look at your other S1s and investing. Yeah. usually features like people just kind of treat this as marketing right like this is not marketing

expense yeah exactly and so obviously this is specifically on the compute usage and so it's hot like there aren't that many companies that have like you know that are public that have like compute like free user with compute plans but i think essentially like they're not using up a ton of their compute on serving these free users and so most of it is either being used for training and research or is being used to support these paid users which i think like you know if you if you think about like you know remember open ai gathering rate of sora uh one kind of reason for that i think was just that if one it was obviously not really monetizing because it didn't intend to but it was such a use it was such a large use of compute that it was like you probably have more productive uses of that especially given that they weren't really trying to monetize it into some kind of like you know like ai video gen platform truly um yeah so i think you see that with with than Tropic. I think there was a case where, you know, Sam has always wanted to build a social network for OpenAI. And what do you do if you're not Twitter, you're not Facebook or Instagram?

You do AI TikTok, right? And that was a reasonable bet and it didn't work out. So I don't know. I think it was worth a try. I actually like, I'm kind of like an OpenAI defender here in terms of, yeah, you can do coding in B2B SaaS or enterprise, but like, you know, AI is supposed to be a bit more broad and creative and weird than just, you know, we help you code better. And so I, you know, I defend like the deep minds and the opening eyes for doing things that are not, you know, just work. Yeah. And I think it's one of those where I think that's the, and Prophec has worked partially because of its focus, right? And if you see where they've gone through, that's helped them a lot. But as you think about one big question that people have, and you kind of start to see it in some of the numbers, is that, hey, like Q2 was this like generational quarter for y'all, where y'all added like, you know, like 25 billion of like, you know, like on an annualized basis of like run rate revenue.

And already you see like a decline in Q3, you bought like a massive slowdown where you went from like, you know, over a hundred percent quarter over quarter growth to like 40%, 45% or so quarter over quarter growth. But then also you ended up now that it was, people are like, well, was that just a quarter of like mass token maxing and peak kind of like entropic share? Cause no one else had caught up on coding and on like a good, like, you know, like reasoning model, like a long horizon reasoning model. And you just basically captured all of that, that market available. and now everyone's kind of caught up. Now open-weight models are strong. Token maxing is ending. Everyone's talking about routing across models, and so you're going to suffer from that without maybe that next leg to go on. And I do think that is partially one of the risks. I don't think it's as big as that in the sense that we're still so underpenetrated in terms of just diffusion of this core technology that's already available to us. But I do think there is this big question of is there going to be the next thing off the coding

that is like actually the size of this off-coding where someone, you can imagine like some large base of people are spending like $8,000 a month or $10,000 a month on kind of tokens off that kind of for that kind of use case. To me, I think what the, if you think about just like your product market fit of various segments, like obviously I think coding is number one. I think customer support, legal have probably been some of the others. I think the one that seems to be coming, and there have been various debates on X about, like, you know, what percent of token use next year is going to be around security. I think that could be another one where, in some ways, these labs have kind of created the problem, but then also offer the solution where now attackers have access to all these models and can kind of go ham on these. And I was trying to see some of this play out, right, with some of the PHP stuff, and then just, like, more and more of these paths happening and, you know, some of these, like, the models that are like ablated and like

that allow you to kind of really use these and so I can't I see that as like a promising area in terms of like you know what's that what's another like massive source where like a lot of tokens can be used but then I think what they talk about a lot for long term is around like science and life sciences and you know let's maybe use these to start doing things, drug discovery and all of those things. And they touch on some of the wet lab space they have and those things as well. Right, right. We'll leave the wet lab stuff to the end because obviously that's newer and less. We honestly just have less details about it. Yeah. What other, okay, so we've talked about a lot of the headline financial things. What really surprised you? What else? Yeah, I mean, I personally was pretty surprised by, in some ways, the share of the compute that was on training and research still.

Like, you know, basically they're using about 70% of their compute on training and research, and only like 30% is going towards inference. And I think we're surprised both. One, that just means that, you know, the inference is very efficient. So like, you know. It was going up, right? It was 50 at the start of last year. Exactly. And so I think they're really squeezing inference efficiency out, which to me, I mean, if you think about it, right, the business of the scale of Antropic, if their quote unquote inference costs are only like, call it like $10 billion this year, maybe like $10 to $12 billion a year, that's kind of crazy if you think about it. And it does make you question a little bit about some of the sizes of these build-out malls. Now, again, we're still very early in aggregate, but the fact that they've gone to this scale and they're only spending $10-15 billion a year on inference to support that size, I think that was surprising to me. I think the other piece of it was that they didn't technically do this, but they vaguely gave you these numbers that let you break out the spending on data.

And even that to me was surprising in that it was until this past quarter, which they only kind of published through the amended filing like a week or so ago, my best estimate on it was there was like about like 900 million or a billion of run rate spend on data, which I thought was actually quite small, given that you consider how many of these companies are there in Silicon Valley selling like various forms of data, like our own environments of all these different niches, and the scale that like, you know, that so many of them are at. And so that was surprising to me. I think the one thing you did see is that there was, at least on a quarterly basis, there was a big jump. I think they spent about million or so in the June to September quarter on data And so it seems like they have very recently ramped up their spending quite a lot And so today you could argue they now spending close to like two two and a half billion a year run rate on data And it's interesting because now it feels like there's been this renewed interest in all these like buying operational data from companies and like these data brokers that are helping get like real company data onto people.

To me, it's a sign that this kind of RL environment paradigm we've been in and just post-training and using RL to keep scaling across every one of these domains is working and they're seeing good results from it and they're continuing to ramp up their spend on that across all these forms. And so I thought that was interesting. Well, and roughly the spend on data has more or less been, you know, like five to six percent of their spend on compute over time. Although it kind of did have this jump, as I mentioned, the last quarter from like, and so it went from like four percent compute to six percent. But I think that's another thing to watch. Yeah, that's something I've been I've been checking a little bit. Do you think that scaling it by percent compute is the right way to do it? I don't think so. I mean, I think like one is just like, what's the absolute number? And like, let's talk about that. I think, but if you think about it, if you think about it, like it's ultimately a core input into training, maybe percent of training compute could be a good, like percent of compute that's going on training and research.

Well, the other thing that's not captured in this is that this is the spend on external vendors that I, my guess is that it's also sizable, just like, or not my guess, but there is a sizable synthetic data effort internally as well to synthetically generate data. that is probably just getting captured in training and research compute as well and so that is the other aspect they're also buying books and destroying them yeah and so I think this was the, they had this separate line item for one of the quarters in the S1 for the settlement that they paid 1.5 billion? yeah exactly I didn't touch on it but I think it's called the bot settlement or something And so in one of the quarters, there's a separate line item in the income statement for specifically that. And the thing I found funny with that whole thing was that the issue was not that they – or the things that they ended up getting fined for, I believe, was not the ones that they actually bought, but that they torrented a bunch as well.

And so it was not that you actually used them. You bought some and you used it. There was more that there was some that you didn't pay for, which you used. If they just bought all of them, I think they may not have to pay the settlement. So I have a tiny bit of context about this. This is actually probably from Books 3. I don't know if you know the context because this is kind of deep lore. I can cover it briefly for people interested. So basically GPT 1 and 2 was trained on this data set of Books 1 and 2. And it's literally called Books 1 and 2. You can go look it up. These are sort of copyright clear books. Now, a random person, I forget his name now. I think his name is Sean as well. And just from nowhere, just an open source AI person goes like, hey, Books 1 and 2 are a really good idea. Let's keep going. And so he creates a very, very illegal data set of actual books, including Harry Potter and whatever, and calls it Books 3.

You can call things whatever you want, But clearly, you know, books one and two were like official copyright clear. Books three was not. Books three, but however, books three was like, it's like pretty widely cited or used in the model trading circuit. So I don't know if Anthropic was exactly caught by books three, but it's going to be something like a books three, where it was included by somebody in some corpus that was just like, okay, like, you know, researchers just got used to trading on this for like an academic project where it didn't really matter. But now it matters for Anthropic. And so this decision costs $1.5 billion. Yeah. And I'm sure after that settlement, they're going to be a lot more careful and they have much better prominence on everything else that goes into their training sets. I mean, so that was the whole discussion that I think should have gone to the Supreme Court. It did not, I think. Yeah, I don't think it went to the Supreme Court. But it should be transformative use. We literally call it Transformers.

I'm allowed to go and read Harry Potter, and maybe I have a perfect memory, and I can go somewhere else and write down word-for-word Harry Potter from memory. Is my brain illegal? That's effectively what we're doing. Just because there's machines involved and you can view LLMs as databases, maybe that's a bit more questionable there. but I do think that there's some amount of dual standard or double standard when it comes to treating humans versus treating AI differently. Anyway, fun questions for now. So yeah, I think the other thing that I always press people on is like, what percent is cloud code? Do we care? We used to have these disclosures of like, oh, cloud code is like 1 billion in revenue. cloud code is 2 billion in revenue the last update we got by the way was 2 billion in February or March of this year and then they stopped reporting so when people stop reporting

usually you assume the number is bad but the truth is that cloud code as a CLI stops being a thing it basically just got adopted in the cloud tag and cloud code work now 3 out of the 4 products that they list on the S1 is like cloud code based things so arguably cloud code is responsible for i don't know like 50 billion in revenue so i i don't know if that matters i you know it's just one of those things that like people debate yeah yeah i think like the other thing i i would have liked to see but they did not disclose which i don't fault them for not doing so is that what percent of actual revenue is coming from you know like these first party applications that you've built on top like the true like you know like like like the actual like you know the end of it like versus like you know resellers yeah or just or just anyone using your if they could use it directly from you but like like a hobby using you and their like vn api is pretty

different than a consumer using like cloud code or cloud code or something where you kind of have like one more level of uh um abstraction that you kind of build which allows you to like you know arguably have a little bit more of uh retention over time and the reason i say that is like a few things so one is so we do know for example that about half their revenue is coming from channel partners i think that's actually now going down slightly it was on the way up i think it's now down to 44 it's coming from these channel partners and that they roughly are paying if you think about like the big cost right it's pretty much compute which we've touched on people which again is as a percent of revenue just keeps going down because they are getting more and more efficient i think the revenue per employee is now up to like six million per year from like a from a million or something. And then the third is this kind of the sales and marketing, but most of it is not like, you know, brand building sales marketing. It's like the fees that they need channel partners, which is about 30% of the revenue,

30 to 35% of the revenue that those channel partners are getting for them. But I think the other reason this is important is that they disclose that about 80% of their, most of the model usage, or most of their revenue comes from models that were released in the last 80% of the usage comes from models released in the last six months. And I think what that, the dynamic that kind of ends up creating, especially when you add to that fact that most of their revenue comes from usage-based plans or from usage-based pricing as opposed to like, you know, subscriptions is that you can't fall behind at all, or you kind of are putting like a large percent of your revenue at risk. So like if like, and I think it would be hard for this to happen, imagine, like, for whatever reason, like, they just fall behind for three months. And, you know, they're like, you know, Gemini's new model is very good. OpenAI's new model is very good. And Entropics is just, like, slightly behind for a reason. It's like you're still using Opus 55 is the best that they have,

and the others are further ahead. You could find yourself in a scenario where because people are just kind of using you at the API layer in some system where they can easily just swap you out, that you have lost a lot of share suddenly and lost a lot of revenue. And so I think that's like a tricky place to be in. So from a company strategy perspective, I think, you know, the more they can get people to use your apps, like if you're using, you know, Claude, you're not going to switch of like, or switch everything over to chat to you. You might have a bunch of scheduled tasks and everything. You won't do that switch if like, you know, it's behind for a few months or something. But if it's true, the more it's just API driven usage, the more risk there is of just people, you know, just like the revenue can go away more quickly as well. And so I do think that that is one of the things where you're in this game where you just always have to stay, and not necessarily at the frontier, but also just at the, like, on every point of the price performance that kind of matters from a usage and workload perspective.

Yeah, amazing. I think that's a really good recap of the state of things. Okay, so it's been a couple weeks since the S1. A lot has changed. We are now tracking, what is it, like 50-something percent probability of IPO. How critical of a moment are we in, whether this happens or not, right? If they don't IPO, is it the end of the world? If they do and it still flops, is that the end of the world? or is there no end of the world and actually just keep on the fundamentals screw that forget the stock market? I think that my guess is that it would probably, it definitely would be better for them to even go out sooner rather than later, even if that means like, you know, 10, 20, 30%, like a lower valuation in that like, you know, indefinitely delaying it might actually lead to more fear and uncertainty than just like, you know, going out at like 1.25 instead of 1.5

or whatever that ends up being. I think the other reason this ends up, probably the area this also ends up mattering in is just that so much of liquidity in the ecosystem rests on, you know, the IPOs of some of these companies. SpaceX is the one that has gone out. And I think a lot of the continued funding of these very large rounds, like, you know, whether it's like these Neo Labs and some of the other, like, you know, like the compute build out and ecosystem stuff also just relies on, Hey, like these things actually going public and providing liquidity to the venture community and then to the RLPs, which then allows them to invest back in the next funds of the VCs that then allows them to invest in the next set of companies. And so I do think from that perspective as well, it is, it is important. Now, I don't think it's going to be the end of the world if it happens in January instead of November. But I do think like, you know, like it probably from an aggregate perspective and, I wonder what the Entropic team thinks, but it might actually be better to go out now, like, you know, like before the end of November, as an example, than maybe wait for the new year, even if that does mean a slight, like, kind of discount on the valuation.

Yeah, because I do think that the fundamentals of this business is strong, so I don't think, I still think it will be the first of the, like, it will be, it will be easier for them potentially to go out than it would an OpenAI. I think if you're, if you're entropic, you probably still want to go out in this, this year you have where you're still growing, like, you know, at least like on a year over year basis, like this enormous amount, because I do think what, what the what the risk for them is that you could have like some month where you have like some negative month over month growth, and then that could actually freak people out a lot. and so you want to go out while you're still growing this year at 5-10-8% a month and already people aren't doing this as though you went 10-100 I think that is true and they likely will go 10-100 but there's also the, well, in the last 3-4-5 months the growth is declining but it still is growing at an amazing phase for the scale of stats

you still want to get out in that period I think. Yeah, you know, like, look, there's the short term, like, you know, is it this month, is it the next month? Like, look, if it's within the next, like, six months, like, we're, you know, we're obviously it's like rounding error. When employees, I think, you know, I talk to developers and engineers and researchers all the time, and they're asking me, like, look, okay, it's 1.5-ish, and so now they're looking for the next, like, 2x, the next 5x, right? Like, is it 3 trillion? Is it 7.5 trillion? Is it 10? And that really the big unknown They had this big jump in the past year plus What is the ongoing prospects right And so that the other thing I really want to emphasize which is like they're trying to also tell that story about what's the long-term growth. It's basically agents. They call it virtual collaborators in the S1. You know, and that's a bet that you have to take, right? Like obviously the S1 is not going to tell you

anything beyond like recent financial performance and maybe some risk discussion. but that's about it. Yeah. Yeah, I think this idea, I think this question that people have been debating around what is the right amount over time for a company to spend on kind of tokens and any, because I think it's like, you know, people will have a view on that of like how big is this kind of spend on tokens going to be and then they'll have a view on like what is a tropic share of that going to be and then like how is that going to monetize based on where you are in the frontier or like, you know, the profitability of that. I think other than stats that was interesting from there, S1 was that only today, only about half the Fortune 500 spend over 100K a year on Enthropic. So I was just wondering, like, what are the other two? Like, how are there, like, still 250 companies in the Fortune 500 that aren't spending 100K a year on Enthropic? Like, that's kind of crazy to me. And then the other one was that, like, off the ones that are spending this over 100K, so off that first set, Enthropic is under 0.1% of their, like, aggregate expenses.

And so it's like, you know, you would think that everyone has different views on this, but what is tokens spend as a percent of OpEx or of expenses for these companies over time going to be? And it's obviously going to vary depending on the kind of company you're in. But my guess is that at least it's going to be like, you know, in the few percentage points. And even that could be like, you know, undercounting. And I'm talking about like three, five years from now. And so there's probably like a 10 to 30x if you just think about it on just what you're doing now, even without that much footer, you don't need, like, you know, the intelligence to get smart. It's more maybe just the things around it and the actual, like, you know, this idea of the diffusion gap or the capability overhang that people talk a lot about. I think as that gets addressed, I think this token usage should just continue to grow. But I think then the other questions become, like, okay, like, what is the, if most of the usage, you don't need these, like, the frontier models for, will more of that go to open-width? Will more of that, will the margins of those be low because they're much more price competitive versus like the frontier models where, you know,

people are willing to pay higher prices and stuff. But those might only apply to a smaller set of the use cases than like, you know, everything that's being used in the enterprise. Yeah, yeah, that's great. Cool. I'm going to play a little bit of like a read around the broader ecosystem and then we'll end with the wet lab stuff and the science stuff. So are you more bullish or more bearish NeoLabs and NeoClouds? You can sort of interpret that either way. I think that at the current pricing of the NeoClouds, post some of this dips and stuff, probably more bullish on those ones. If you were just to take the average one, like how that would turn out, I think if you were to take a basket of the neoclouds versus the neolabs, I think what's going to happen is that you may end up having one or two of these neolabs becoming extremely, extremely significant companies.

And so I think that basket may end up outperforming the neocloud basket. But if you're just taking your average one, you definitely are better off with the neoclouds, I think. And I think the other thing is many of these neolabs will end up being the neoclouds because they've already bought the compute. and if they can't find the productive use for it, they will end up... Exactly, yeah, yeah. Because most of them, I think, ended up, you know, ended up basically, like, securing, like, long-term compute in one way or the other. And so now they could end up, if they don't have a productive use for it, then they'll end up being the supplier of it. Yeah, I mean, that's the, you know, that's the failure mode of a NeoCloud, or a NeoLab is to become a NeoCloud, that's for sure. Yeah. But I do think, like, just in terms of the profitability of inference, we now have a couple data points of how profitable Anthropic inference is, how profitable DeepSeq inference is, because they released the DeepSeq inference stack, and they said it's something like 80-something percent. Yeah. So I think that should make all NeoClouds just generally more interesting as an investment.

Because not only do we know that Anthropic is profitable, but also the state-of-the-art stack with DeepSeek is open source. And so you can just buy a rack of GB200s and go nuts. And obviously it's not that simple. But, you know, the payback period for a GPU is not 10 years. It's not four years. It's not, maybe not even three years. It might be a year. Yeah. Yeah. Yeah. SpaceX on their earnings call said that they're seeing their sub one year paybacks on their build outs, right? Which is kind of crazy. So then, like, why are people skeptical? Like, why are people, like, why is, if that is actually true. So this is probably not true. And the reason is the debt markets are the smartest people in the world, right? They do by far the most due diligence, and they are hiking interest rates on all these guys. The whole reason we're raising equity is because it carries debt. I don't know. To me, I think one aspect of this is that obviously we are in this moment in time where definitely there is this imbalance of supply and demand.

I think people just have a different view in terms of when that does end up clearing. I also do think that if you look at, like, you know, Anthropics inference numbers, I don't think that at steady state or even right now, for example, like the open source inference market where it's like you're just serving inference on the models that are open source, like, you know, you see different numbers thrown around for the at-scale players there, like the fireworks, the base stands of the world. But I don't think, you know, I don't think at steady state, these things end up close to, like, what the entropic numbers do, for inferences, because, like, that model is available to everyone, and someone will price, will undercut you on price to kind of just get share. And as you kind of see this playing out, I think Open Router is just, like, fascinating to look at. Like, you can go look at, you know, go look at, like, Akimi K3 or, you know, GLM 5.3. Like, it doesn't really matter what model you look at. And you will see, like, the set of players, like, the official, you know, the lab itself

that made that model, like, Akimi, and, like, the price of they charge at and there are like five and ten plus like players now that are basically like at like 80 on like basically 20 of that price or even 10 of that price but they are like undercutting the you know the the what the price that like the deep seat themselves might give it at or that like a mini max or whatever might give it out that they're like undercutting them by like 10x and these are like smaller players but i do think you're going to continue to see this trend of like, oh, if you try to charge too much, especially on the open source model and depending on the licensing of that model requirements, like there used to be like some new player, like, oh, I have like, I'll just capture share now and I'll run it at like 20% gross margins or even 10% as opposed to, and just like capture that share. And so I do think it'll be interesting to see how some of that plays out and settles out. My guess is that I don't expect any open source inference, open source model inference provider to have at least like maybe over at least like 40% margin long term

on like the vanilla open source tokens, which I think is partly the reason why, you know, you see many of them pushing into things like post-training and helping customize models because then that becomes a slightly different, more lock-in, slightly different model you're serving and you can maybe, you know, you can benefit from like as you optimize it further and maybe charge higher margins, have higher margins on those. Awesome. So I just want to wrap it up with the wet lab and the sciences work. Dario has been pretty open about his personal relationship with AI for science. We haven't seen that much yet. What are people, first of all, what was disclosed and what are people talking about there? Yeah, I mean, so maybe for a step back, I think the other thing is that, you know, So, like, the property obviously has signed, like, these, like, $900 billion of compute commitments, which has been, you know, one of the kind of, like, eye-popping figures in the S1 that has some people worried. And I think, to me, the question for seeing that becomes, like, you know, they have, like, a rough schedule.

And so, let's say that, you know, next year they might end up having a compute expense of $70 to $80 billion the following year onwards at RAM, so, like, more like the $100 to $120 billion range. And so, the question becomes, like, what happens, given that, you know, these are, like, mostly contractual, other than SpaceX one. What happens if you don't have enough demand to serve it? Which is, you know, like, it's obviously not a great scenario, but it could happen. And the reason it could happen is just that it doesn't have to do with anything around even just competition and things. It could just be that diffusion takes longer, and so you don't have, it's just, you know, beyond coding, that next thing isn't using as much tokens as you don't have that next thing. And so I think the life science of stuff is important because it provides, It's like some use of that compute that is more in anthropic zone control where you can focus your models and then, hey, like, let's, we will be the primary user of our own models and help, you know, you can partner with companies, obviously.

But, like, you would be then looking for these novel discoveries. And it doesn't have to be just black sciences, right? It could be any kind of thing in science, new materials, new things like that. And so I think it's really important there because it then means that at least you have an avenue where you can direct compute to, which has another way of being productive that isn't dependent on timelines of enterprises and others to adopt your technology. And so if this diffusion gap ends up being wider or takes longer, you have this other business segment where you can direct some of the compute to. And so I think it's a pause from that perspective. beyond just like, you know, just further, the other obviously big use of computers, RSI as well, and just make the models better and better. But I think that may not still solve the customer demand piece, whereas like having like these life sciences, that's, and I think they also touched on robotics a little bit, but not as much as life sciences. Like having those then also puts you in a different,

having some of these other things that are expanding to the physical world, but then also things that may be a bit more in your control where you are operating these to try to find, like, make novel discoveries and that you, the US Alpha, you know, potentially with Thama Pontnos taking through the pipeline. Yeah, awesome. All right, any parting thoughts, messages for people, founders, you know, talking to Wing, whatever? No part of that. I mean, I think it's going to be very interesting to see how this plays out, and I'm curious to see what the next couple of months look like, and I think until they go public the number that everyone's going to be watching very closely is just like the month over month the actual revenue growth for them oh man yeah it's going to be which is why it's better for them to do it sooner or later otherwise it's like they have to like somehow like encourage everyone to token max for the next yeah I mean in that case they need to release Fable 5.5 yeah and you know whatever is beyond that but yeah I mean

that would obviously be a driver of more attention and usage and that would that is a lever that they have to pull. But regardless, thank you for taking us through. I'm really just trying to give people a good grounding on what they should know going into this. I think a lot of engineers are wondering if they should join, if they should cash out. Same thing for just the general broad rest of the market because Anthropic is probably one of the three IPOs that are happening that are the leading frontier labs in this sort of one-year period. And I do think it really matters for the OpenAI IPO, which if Anthropics fails in whatever shape or form, OpenAI is not going to do it. And that sucks for multiple reasons. Yeah, so pretty pivotal moment. We'll see. Thank you for coming. Yeah, thank you for having me. It was great to discuss. And yeah, I love the stuff you put out.

and so we'll continue to listen and excited to follow along. Yeah, and same for your work. Everyone go read his substack as well as deploy codes. We're actually starting a Ford Deployed Engineering Track podcast, and we're launching it with Ode, which we're pretty excited about. Regardless, the AI is here, and we have to deploy it, right? Whether or not things are public or not, we have to deploy it. Exactly. All right. Thanks, Jenny. Awesome. All right. Thanks. See you next week. Bye.

番組の概要欄(原文)

Watch the full show on YouTube: https://www.youtube.com/watch?v=wX8nDtGIfYA

出典・この記事について

番組
Latent Space: The AI Engineer Podcast(Latent.Space)
元の回
⚡️Anthropic: Know Before You IPO — Tanay Jaipuria, Wing VC
配信日
要約作成日
公式サイト
www.latent.space/podcast

この要約・解説記事は、元の音声を文字起こしした内容をもとに AI(Claude)が日本語で作成したものです。 固有名詞や数値に誤りを含む場合があるため、正確な内容は元の音声でご確認ください。制作方法について

X でシェアSpotify で聴くApple Podcasts で聴く

関連エピソード