
Anthropic の IPO はなぜ成功するのか――Box CEO Aaron Levie が語る強気シナリオ
Why The Anthropic IPO Will Succeed — With Aaron Levie

Anthropic の IPO はなぜ成功するのか――Box CEO Aaron Levie が語る強気シナリオ
Big Technology Podcast
要約
Box CEO の Aaron Levie が、Anthropic の IPO に対する強気の見方と企業向け AI の将来を語った回。コーディング以外にもエージェントの用途が広がりトークン消費が大きく伸びる、安価なモデルの普及はむしろフロンティアモデルの支出を増やす、という論点を示した。一方で債務、成長鈍化、投資回収の時間差といった懐疑派の指摘にも応じ、需要側の実感から楽観を保ちつつ、シナリオが外れる可能性は認めた。
この回のポイント
- ●Levie は投資助言ではないと断った上で、コーディングが先行しているが、エージェントがコンピューターやツールを使えるようになるため、法務・製造・サイバーなど他分野へ波及するとの見方を示した。
- ●エージェントの登場で、知識労働の「投入した時間」と「得られる価値」が初めて切り離され、予算と計算資源しだいで仕事量を伸ばせるようになると述べた。
- ●S-1 の指標として、年間10万ドル以上を支出する顧客が約950社から約1万1900社に増えた点や、計算資源1ドルあたりの収益が0.40ドルから1.72ドルになった点を、需要の広がりと推論の収益性の裏付けとして挙げた。
- ●安価なオープンウェイトモデルの併用は、ジェボンズのパラドックスによりこれまで採算が合わなかった業務を自動化可能にし、オーケストレーションを担うフロンティアモデルへの総支出も増えると主張した。
- ●懐疑派の三点(債務、トークン最大化後の成長鈍化、収穫逓減)に対し、債務は成長が続くかどうかしだいと整理し、トークン最大化は一時的な現象、ポストトレーニングのコストは過大視されていると反論した。
- ●投資回収の時間差で資金が尽きる可能性は否定せず、計算資源の不足を訴える顧客が多い現状から、自分は時間差より機会の大きさを見ていると述べた。
章立て
強気シナリオの前提と導入
Levie が投資助言ではないと断りつつ、フロンティア AI 市場は強く、普及はまだ初期段階だという立場を示す。
コーディングから他分野への波及
S-1 の顧客数の伸びを手がかりに、製造・サイバー・ライフサイエンス・営業・映像制作などでトークン消費が増える具体例を語る。
知識労働の制約がなくなる
エージェントにより時間と価値が切り離され、予算しだいで仕事量を拡大できるという論点と、ポッドキャスト編集の例を通じた信頼性の課題を話す。
推論の収益性とモデルの安価化
計算資源あたりの収益の改善と、安価なモデルの併用がフロンティアモデルの支出も増やすというジェボンズのパラドックス論を展開する。
最新モデルの実感と論点の整理
Fable 5.2 や Gemini を使った製品プロトタイピングの実感と、司会による強気論の要約を行う。
懐疑派への反論
債務、トークン最大化、収穫逓減という三つの批判に応じ、各分野の市場規模から収益機会を試算する。
投資回収の時間差リスクと Box の取り組み
資金が尽きるリスクを認めつつ需要の強さを強調し、Box Architect などの新発表を紹介する。
解説記事
Box CEO の Aaron Levie が Big Technology Podcast に出演し、Anthropic の IPO に対する強気の見方と、企業向け AI の今後を語った。前回の回で懐疑派の Gary Marcus が登場したのに対する、対になる楽観論の回である。Levie は冒頭で、自分は投資助言をする立場になく、適正な株価や時価総額には言及できないと断った。その上で、フロンティア AI の市場そのものは今後も強いと主張している。
コーディングの先に広がる需要
Levie は、現在の応用 AI の主役はコーディングで、OpenAI と Anthropic の急成長の大きな部分を担っていると見る。ただし「コーディング以外ではまだ機能していない」という見方は、あらゆる業務がコードに近づいている点を見落としていると述べた。エージェントがコンピューターを使い、ソフトウェアを書き、ツールを操れるようになれば、エンジニアが得てきた恩恵が他の領域にも移るという理屈である。
具体例として挙げたのは製造、サイバーセキュリティ、ライフサイエンス、営業、映像制作だ。製品設計や CAD は将来トークンで生成される、あらゆるログをエージェントが見張る、調査エージェントの群れ(エージェントスウォーム)が仮説検証を並行して進める、といった未来像を語った。営業担当が商談前に顧客を調べさせるような地味な用途でも、数万社規模に広がれば運用費の大きな一部になるという。番組では、S-1 に載った年間10万ドル以上の支出顧客が約950社から約1万1900社へ増えた点も、普及が実際に進んでいる証拠として取り上げられた。
知識労働の「物理的な制約」が外れる
Levie が最も強調したのは、知識労働の歴史で初めて、投入する時間とエネルギーが生み出す価値から切り離されるという点だ。これまでの成果は、人数、時間、使えるコンピューターの台数に比例していた。今後は、予算と計算資源をどれだけ割くかが唯一の制約になる。1人がエージェント群に目標と予算を与え、千人分の調査に相当する作業を任せる、という使い方である。その結果、需要には上限がないように見え、減速の時期は自分にも分からないと語った。
ただし信頼性の課題は残る。司会はポッドキャスト編集で、息継ぎの音だけを除く作業を AI がまだできないと例を挙げた。Levie は、そうした小さな例は学習データ(RL データ)が少ないニッチな課題だからだと説明する。市場規模が大きい領域、たとえば CAD やサイバー防御には、各ラボが重点的に学習させ、周辺のスタートアップがデータを供給するため、改善は早いという見立てだ。
推論の収益化と、安価なモデルの影響
S-1 では、計算資源1ドルあたりの収益が2025年初めの0.40ドルから1.72ドルに改善している。Levie は、モデルの訓練は資本集約的だが、複製しにくい製品を持てば需要に対して十分に収益化できると指摘し、この数字を見方を改める材料だとした。ただし商品化(コモディティ化)が進まないことが前提だとも付け加えた。
フロンティアモデルが安価なモデルやルーティングサービスに置き換えられるという懸念には、ジェボンズのパラドックス(効率が上がると総消費が増える現象)で応じた。難しい判断や全体の調整は最先端モデルが担い、大量のログや文書の処理は安価なオープンウェイトモデルに任せる分業になる。コストが10万ドルから2万5000ドルに下がれば、これまで誰も実行しなかった業務が初めて自動化され、両方の支出が伸びるというのが Levie の主張だ。
懐疑論への応答
司会は Marcus の三つの論点、すなわち巨額の将来債務、トークン最大化による成長の鈍化、収穫逓減を挙げた。Levie の回答は次のとおり。債務は、年数千億ドルの収益が見込めるなら妥当な投資で、そうでなければ破綻するという前提しだいの話だとした。トークン最大化は一時的な流行にすぎないと見ている。収穫逓減については、ポストトレーニングには数百万ドル規模のデータと計算でも大きな進歩が得られ、これまでコード中心だった投資を、サイバー、生物、法務、製造、設計に広げる余地があるとした。サイバーセキュリティ企業の時価総額の合計が約5000億ドルであることを挙げ、エージェント型サイバーによってその市場が2〜3倍になっても不思議はないと述べている。各分野の世界の支出の1割や2割を AI が取りうるという試算も示した。
投資回収の時間差で資金が尽きるリスクについては、起こりうると認めた。ただし、Levie の周囲の顧客が計算資源の不足に直面している現状から、自分は時間差を機会の大きさとして見ているという。懐疑論には、AI の有用性を過小評価した前提のものが多いとも指摘した。
Box の取り組み
Box が発表した Architect は、契約管理システムなどの業務プロセスを、Box の既存部品(ワークフローエンジン、ノーコードアプリ、ファイルシステム、セキュリティ機能)だけで自動設計するエージェントだ。コードを書いて自前で運用させる形ではない点を強調していた。
まとめ
本回は、AI 需要の大半がコーディング由来である現状に対し、次の成長要因を整理した楽観論として読める。鍵となる仮定は、エージェントによる用途拡大が実際に続くことと、フロンティアモデルが商品化されずに価値を保つことの二つだ。Levie 自身も、時間差のリスクや市場の過熱を否定してはいない。日本の読者にとっては、AI の導入効果を人数や時間ではなく予算と計算資源で考える視点と、安価なモデルとの併用で採算が合う業務が増えるという視点が参考になる。ただし数字や予測は発言者の見方であり、S-1 の数値は番組内の紹介にもとづく点には留意したい。
文字起こし(英語・自動生成)
Will Anthropics run on the public markets live up to the massive expectations that traders are setting for it? Our next guest says yes. An Optimist view on the Anthropic IPO coming up right after this. How do you price intelligence as a service? How do you charge for delivering end-to-end work, not just access? Those questions remain front and center for AI founders and operators. And getting to the right fit business model, whether that's selling usage credits, outcomes, seats, or a hybrid combination requires constant iteration. Chargebee is the billing and monetization infrastructure that powers just that for ambitious, fast-scaling AI businesses. Visit chargebee.com slash Alex to learn more. That's C-H-A-R-G-E-B-E-E dot com slash Alex. Welcome to Big Technology Podcast, a show for cool-headed and nuanced conversation of the tech world and beyond. We continue our coverage of the Anthropic IPO and the company's entry to the public market
with an optimist view on the company and its prospects. And we're joined today by Vox CEO Aaron Levy, who's going to tell us why he believes it has what it takes to meet expectations. Aaron, great to see you. Welcome to the show. Good to be here. Thanks for having me. So I will note we are recording this before the Anthropic IPO, but you are a believer in the company. And a couple days ago, we had Gary Marcus, who is not a believer in the company. So we're juxtaposing the optimistic view and the pessimistic view. Is he not a believer in the company, or does he just think the expectations are inflated? I think a little bit of both, but by and large, he thinks the expectations are inflated and is not worth, the company is not worth investing in. Now, we're not an investment advice podcast, but you work in Enterprise AI, which is in basically the bullseye for Anthropic and Anthropics World. So I think you're someone who's a little bit more positive about the company and the arena that it's entering. So we'd love to just hear you make the bull case for why Anthropic will thrive on the public markets.
Yeah, so obviously there's like nine disclaimers that I need to provide as legal advice. I'm not a financial advisor. I don't make investment advice. I probably can't speak to the specific, like, price that anybody should pay for Anthropoc shares, to be clear, especially because, you know, we'll see how it ends up trading. I think what I can make the case for is why I believe that you're still going to have – you're going to have a very strong market for frontier AI, and there's, you know, less than, you know, half a dozen companies you could even invest in in that zone and then probably even half that number truly at the frontier, so, you know, OpenAI, Anthropic, et cetera. And what we're seeing from the customer base and what we likely expect to be on the horizon of the use of these frontier models, all of that leads me to be very bullish on the market broadly. Again, like, no clue on exactly what the right market cap should be and at what price, you know, people should be paying for these shares.
But in general, I think we are still in the very early part of the kind of inflection point of AI adoption and diffusion. And when you look at the kind of token consumption going into really any kind of company that's using these models at or sort of nearly at the frontier, you can extrapolate that as that sort of diffuses across the economy, these businesses will only be much larger from at least a top-line revenue standpoint. And obviously long-term questions on what's the right kind of margin structure for inference, how does the open weights competition sort of bring down the price of inference. But for where I see things right now where we tend to see the kind of enterprise adoption, we're still in the very early phases of the overall diffusion of AI in the economy. Now, you route a lot of workloads to these AI models. And you were talking about the token consumption, and I think that's really important because, you know, we've seen the charts of Anthropics.
Revenue is pretty extraordinary. I mean, just to give you an example, if you look at Q3 2025, the company was $1.4 billion in revenue. Q3 2026, $17 billion in revenue. Yeah, just a standard 12X. at tens of billions of dollars scale. So, yeah, I mean, incredibly unprecedented. Now, I don't think anybody should think that there's another 12X in the next 12 months. So like that, we should probably sort of tamper expectations on that front. But what's interesting is the way I generally think about it is, coding is sort of the dominant sort of use case and applied AI use case right now, and that's driving, I think, a large portion of that sort of acceleration from both anthropic and open AI. And when you start to see the shape of a bunch of these adjacent sort of areas of AI adoption,
could be legal work, manufacturing, any kind of, you know, anything in sort of broad knowledge work that, you know, we're kind of doing on computers, coding is starting to have a broader impact across all of those other spheres of work. And so everybody sort of says, well, you know, yes, it's working in coding, but nowhere else. And the thing that that sort of misses is that everything is becoming sort of code adjacent because once these agents can use computers and they can write sort of software to execute tasks and they can use these tools, you know, very effectively, then actually the same sort of benefits that engineers have been getting from AI in the coding domain, and that will actually translate to a bunch of other areas. And so when you add up not only just the continued progress and adoption we're going to see just even in coding, but then you layer in cybersecurity. You layer in IT automation. You layer in, you know, broad areas of sales and marketing and legal work. You look at manufacturing, product design and development.
These are all areas where we have the opportunity to expend vastly more tokens on these workflows than is sort of happening today. Some of that is the models just need to get better and better. And then a lot of it is actually the diffusion of these models into the economy needs the right kinds of software form factors. They need to be implemented properly. You need to kind of, like, help with the change management of companies. But that's all, like, we are off to the races on all of that. That is just a multi-year kind of like, you know, half a decade, decade-long kind of journey that you have to go through for technology adoption. But again, you can easily sort of squint and see that you're at least at the frontier. Easily, we could be spending 10 times more on frontier tokens. And then overall tokens in the economy, we could be, you know, we could be using 100 times more just on all of the sort of economic activity that happens with what we can now do with agents. Yeah, and Anthropix numbers that have come out on the S1 or it's released in the S1 are wild. Speaking of that diffusion, right, I think what was September 2025 in the quarter ending then,
950 customers spending $100,000 or more over the past year from then. Now it's $11,900. So that diffusion is really happening. Yes. And I think that should also, that actually is a really interesting metric because it does show, So, like, you can't do that metric if you don't have fairly broad-based sort of utility being generated from AI. Like, it was one thing when it was, like, a few hundred companies spending a few hundred thousand or a million dollars. But when you have, you know, in the tens of thousands of companies spending six or seven figures on AI, like, it's working. Like, this is, like, really happening. The adoption is actually occurring. Companies are deciding to sort of, you know, they're voting with their wallets that this is actually highly useful. and so now imagine that number is not 10,000 but it's 50,000 and imagine the metric of people paying a million or 10 million is now 10x larger than it is today. You start to see sort of the shape of how these become very big businesses. Yeah, and I'm going to get into the rebuttals in a bit
so we're not going to just totally glaze an anthropic. I prefer to do that. I can just be purely the optimist. Okay, so we'll go back and forth in a bit but I do think it's worth seizing on something that you said and seeing if you can help us flush it out a bit, which is that it's not just going to be coding. It's going to be a variety of different industries that will use tokens from companies like Anthropic. And some of these industries that you mentioned, like manufacturing, it's harder to wrap your head around how they're going to be heavy users of tokens. Coding makes sense, right? Coding is code. You break down tokens. You code a lot. You check your work. You build more. You build competing prototypes. You see what works. You know, you could see almost like an explosion of tokens that might be never-ending with that. But the other industries are harder to see. So talk through some of the industries that you listed and why you think they're going to be using lots of tokens from a company like Anthropic. Yeah, I would say that the kind of capability that engineers are getting
is now starting to diffuse into these other domains. So if you take, you know, manufacturing as an example, you know, there's obviously many sort of stages of manufacturing. Those are the design process. The product design process absolutely looks almost identical to the software development, you know, sort of, you know, space that we've been seeing. So you will have, you know, imagine every single CPG, consumer electronics, any physical product company. It could be, you know, people building buildings to cars to, you know, our consumer products that we use every day. every single one of those sort of, you know, product designs, blueprints, renderings, CAD files, every single one of those things will be generated by tokens in the future. And so, and exactly to your point about prototyping, we will do that across every sphere of manufacturing and any of the physical sort of world and goods that we're building. And you just got, you know, as of obviously this recording, Fable 5.2 just came out,
And you've already seen examples online of, like, complete new ways to render sort of physical, you know, spatial kind of experiences around these products. And one example was Fable 5.2 built, like, a home stereo system from the ground up. And I forget how many, you know, kind of hours it was running for, but it basically did the entire kind of, you know, circuit design. It did the physical design. It built the manufacturing plans. So now you can imagine, like, every single thing that we build in the world will have a substantial amount of tokens, you know, being applied to us. That's one example. Cybersecurity, another great example. It's obviously, you know, highly code adjacent, but you can imagine every single line of code in the future will be reviewed by other agents. Our infrastructure will be constantly sort of reviewed by agents. we will be looking at every log and event that ever happens in any of our systems, and agents will be reviewing those logs.
And this could be for not just like software applications, but this will be for critical infrastructure. This will be for IoT devices. This will be, again, for cars. Anything that is connected to the Internet, every single log that those systems generate, eventually agents are going to review in some fashion. So that's all of the cyber domain that now has, again, probably 100x token growth from here that's associated with it. And then you look at what's starting to happen with kind of this rough sketch of like an agent swarm where you basically, you know, sort of tell an agent, you give an agent a goal, and it can go and sort of fan out and then duplicate itself and go and do some kind of domain of work where there's a lot of research, there's testing hypotheses, it has to kind of compare notes. That will be massive token consumption in areas like life sciences and bio-research. You'll imagine a future where you'll probably have agents running constantly on your medical records because it'll be attempting to triage information from either new research that's in the market
or something that your doctor just learned, or when you go to a visit, it's going to look through your entire medical history, every single visit that you have. we're going to have a lot of tokens in kind of these very kind of research sensitive areas of knowledge work so even kind of quote unquote somewhat pedestrian use cases like if you're a sales rep and you're going in to meet with a customer you're going to have an agent swarm go and do every single piece of research possible on that customer, their org chart, all of the different touch points you've had with that customer, the pipeline that you've generated, everything from your CRM system And what's going to happen is you'll be 5% or 10% or 20% more effective in your sales process because of what agents are going and doing to kind of prepare you for that. Those are all the kinds of use cases that we've only scratched the surface of, like maybe like dozens or hundreds of companies like in Silicon Valley are already operating that way. Now imagine you have that across tens of thousands or hundreds of thousands of companies. This will become just a more meaningful part of our operating kind of expense as an organization
is all of these agents that are running in the background or being deployed by users or workflows that kick them off. That's what this looks like. Yeah, it's amazing how much the agent story plays into this because a year ago, we'd be sitting down here, if you told me this agent story, I would say that's a fantasy story. Now these things are starting to work a lot more. Yeah, and you've had this interesting benchmark of the podcast editing, right? Oh, yes. It's gotten very good. And it's gotten very good. And you can see these examples online of like more autonomous video editing as an example Again kind of fable 5 recency bias but there great examples of basically using this kind of tool called Blender to generate 3D animations So you imagine a future where, you know, if you are Pixar or the equivalent, you know, anywhere else, you're going to be using AI to generate, you know, any kind of, you know, piece of entertainment in the future. And so what will happen? You're going to see an explosion of that kind of content. It's not like we're just going to create five movies a year in 10 years from now.
You're going to have 1,000 movies a year that are all of Pixar quality today, and they'll take on any genre that you want, and they'll all have – some will be in movie theaters, some will just go straight to Netflix. This is the kind of era of abundance that we're about to enter, which will just mean that, again, tokens are just a more consistent part of our operating expense that we have inside of our businesses. Right, and for the listener or the viewer, if you have an agent going out and doing the work, they can just basically burn through tokens, you know, for better or worse. Yeah. Because instead of you just putting the prompt in and getting something back, they're just, like, out there for hours maybe just, like, using tokens to go do things. And then, obviously, Anthropic will make money every time one of those tokens are used. Yeah, I mean, the entire stack, you know, makes money on that. The thing that you – I think it's not intuitive to us yet. it will be like totally pervasive, you know, like three years from now. It's the first time in kind of knowledge work history where you have a complete decoupling
of my time and energy into the system and my value out of the system. So basically for, you know, 40 plus years in knowledge work, you know, since we've had computers, like basically I can only generate value as fast as I can, you know, type on a computer and write a document or type code into an IDE and generate code, or maybe if I'm doing graphic design, how quickly I can kind of move my mouse and be able to go design things. All of knowledge work was effectively rate-limited by that, which meant that the amount of output, of digital output that we had as a society was roughly correlated with how many people are there, how much time do they have, How many computers do they have access to? How much software has been sort of distributed to those computers? Like all these various kind of parameters. We don't have those limits anymore. All of a sudden, your only limiter is how much compute do I want to apply to this problem?
First time in history, in knowledge work. So I can go and say, you know what? I would like to do the equivalent of 1,000 people typing on those computers, doing research in some life sciences field, but I want to do that with one person kicking that agent swarm off, and I'm going to give it a goal and I'm going to give it a certain budget, you know, $20,000 to do this particular, you know, research in this domain of life sciences and come back when you've either accomplished a goal or failed or whatever. Like, you've never been able to have that level of sort of accelerated deployment of work. And this wasn't even possible technically two years ago. The models just couldn't do it. So the reason why we couldn't imagine it was because it was like, it was so shoddy. It was so prone to failure. Either the hallucination or just like the rate of error that the models would have would be so high that it wasn't even like you wouldn't even know when it would occur. Like would it be 2026 or would it be 2040, you know, because like we're just never able, like the scaling laws don't hold or we don't have enough RL data to be able to pull it off.
It just turns out that we did, and it's working. And so now all of a sudden you have this decoupling. I can spend a certain amount of money and I can get useful work done by these models or these agents that now I'm no longer sort of constrained by the laws of physics that I used to be constrained by. That's why you are going to see basically uncapped level of usage of these systems because it'll just be like you'll have a business problem that shows up. I have to go secure my enterprise. Okay, well, like it costs $5 million. Go for it. And then you're just like, okay, well, I spent the $5 million, and now I'm X percent more secure. We couldn't do that before. It didn't exist. So that's why you're going to start to see even, you know, like I just don't know when the slope slows down because we're only now, again, sort of just experiencing the full potential of these use cases. But it has to be reliable in order to work there. Yeah. And the reliability is, let's just give one example. Like for my audio editing, I'd like, you know, give it some parameters. Like, I'd like for the bots to be able to, like, take – so, for instance, when I'm on the mic and you just took a breath,
and I'd like it to be able to remove when you breathe. So, basically, like, we can have clean audio on the other side. Yeah, I'm sorry for breathing too much. No, I'm not a heavy breather. We welcome your breathing, but audio sounds better when we just remove that track. So, this is a real-life example, right? Right. And so, I had been handing the files over to – and you know what? it cannot distinguish between you breathing and saying, yeah, what you just did. So then it will just either delete all, yes, keep all breaths. It just can't figure that out. It can't transcribe transcripts. And so I want to, obviously, it would be so useful for me if it was able to do this. So I want to say what you're telling us today, that's definitely going to happen in places like, I mean, think about, that's a small example. If I'm an architect and I can't make these distinctions. Well, it's a small example, but unfortunately, sometimes the small examples are actually the reason why it's not happening yet.
So there's just probably not a lot of RL data about removing the breaths from podcasts. So unfortunately, you're just not. Now it's going to listen to this conversation and then they'll get working on it. Your TAM, I think, is the TAM for this is like 10,000 people, and the amount of tokens that are being spent on that particular problem is measured at $10 million a year, versus the TAM for doing cyber protections of an enterprise and being able to fan out agents and figure out all the different sort of chaining events you need to do to be able to either create a zero day or find a hack. That's incredibly economically valuable. And so, lo and behold, the labs go and train their models on just that problem. And then there's an ecosystem of startups that emerge to go and build data for that problem. And then they spend money on that data, and then it gets even better. And then all of a sudden, like, the models just get 100 times better at that particular task. And then the economy sort of adapts to that and starts to spend money on those tokens. So, you know, first of all, I actually think it's a fantastic benchmark.
You should, like, create the breeding benchmark, and we should, like, release it and track all the models. But, like, you know, unfortunately for you, that just happens to be a slightly more niche, you know, kind of problem. And as versions of that, you know, become more valuable to the AI, you know, labs, I think it will solve those kinds of problems. But, you know, I would just say that might not extrapolate perfectly to other domains. versus, like, engineering, as an example, is probably, you know, a trillion-dollar TAM if you could, like, begin to really do, like, incredible CAD work with AI. So, lo and behold, again, you're going to probably have AI models get better and better at being able to do, like, you know, perfect geometry and sort of measurements and underlying componentry of these kinds of systems. Like, we're probably just a couple years away from the models, you know, being able to do that as well as they can do, you know, JavaScript or Python. And that just has to get trained into these systems.
Okay. One more thing about this and we can move on. Doesn't this all come down to tool use, like making the perfect geometrical shape tool use, right? Yes. And so is honestly, like, determining what a breadth sounds like. Because you could, like, put it into a transcription tool. And if it produces yes, then you keep it. If it produces nothing, that's a breadth. Yeah. You remove it. So shouldn't this stuff be generalizable? I think you actually see surprises of that generalization that's happening. So there'll be a lot of scenarios where, and I don't have maybe a perfect example off the top of my head, but you hear from the lab researchers, which is like, okay, all we did in this training run was this, we get better at this one particular area of code. But then it turned out that because we did that, this other tool use capability opened up and got unlocked because there was enough kind of similarity in what those workflows look like. So you will, that will just happen. Like, I'm sure if we had a 100 trillion parameter model, like, it might just be smart enough to figure out that, like, what you meant by remove breathing, you know,
was to look at this particular waveform and, like, it sort of, you know, did some trial and error and then realized and then listened. Like, it just, you know, we have not trained in maybe, like, the superset of all audio engineering into these models. So even if you don't do hyper-specific breathing, maybe it does just solve that in a year or two from now. But I guess I think you can kind of squint and just see the curve, and you're like, yeah, these things will just get better and better at computer use. Now, I'll steer you to another bullish area about Anthropic is the token efficiency. Yeah. And there's another interesting chart in the S1. And they were spent, for every dollar they spent on compute in Q1 2025, so last year, at the beginning of last year, they were making 40 cents. Now for every dollar of compute, they're spending, they're making $1.72. So they're actually making money on the compute build-up. Yeah, which, and I think it's actually, I think it's a really important thing to underscore.
There has been a certain community or view that that would never happen. I do think that's a moment that people should kind of update their views on this, which is, yes, we all agree. Training of these models is capital intensive, very unprofitable, very expensive. But when you do sort of have a unique good at the end of that that has some amount of hard to replicate kind of qualities, you can monetize that very, very successfully with the amount of demand that there is. And that's what that metric is. So what that shows you is, you know, a year and a half ago, higher proportion of compute spent on the training side, less was going into inference, and inference wasn't maybe particularly profitable at that point. Now, all of a sudden, fast forward today, inference very profitable, training sort of less of the total percentage of compute, and then all of a sudden, wow, these businesses actually look pretty attractive. They generate cash. It works.
Now, you have to obviously bet that that continues to sustain and that they don't get kind of like completely, you know, commoditized. But I'm, I would be, you know, I'm more in the camp of like, like, no, like the frontier still remains very attractive at the moment. Okay. And that basically opens up this very interesting question of let's say all this is true, right? And AI can do all these different tasks and agentic use cases use a tremendous amount of tokens. there's been a discussion that maybe the frontier, and you've brought it up, the frontier is valuable. The discussion has been maybe the frontier is actually not that valuable. Maybe routing tools like OpenRouter will, like, basically, you know, people will learn how to do these things with the frontier models, and they'll use a routing tool to say, okay, I want you to do it now, but just use a much more efficient model. And then, therefore, if you're an anthropic, you're commoditized, and you're on the outside looking in. Your thoughts? Well, first of all, the good news is we actually have enough. We have probably at least a year of evidence of what that should look like.
Like there's been enough sort of we've had pretty good open weights models. Maybe we're like nine months into that. We've had routing for maybe about that long. We've had open router for a couple years. And generally speaking, what I'd say is the data is showing that kind of everything is rising. It's a rising tide lifting all boats kind of scenario. where basically you kind of use the frontier to figure out the next use case that now becomes possible. For the long tail of work in that use case where you're just cranking through tokens, you're just like, I got to process every document. So that way the model can make a decision about something going on in this audit that the company has. But to do that, I need all of the underlying structured data from these documents. This is sort of a use case that would show up in the Vox world. You might use Frontier to plan, to orchestrate, to make decisions, to use judgment on the whole task.
And then you might use OpenWeights or equivalent, maybe from Meta or Gemini, that's really, really cheap and efficient. And guess what? You're spending on both of these platforms or kind of ecosystems. You're actually spending more on Frontier than in the A-B scenario where you didn't even have OpenWeights. Like, this is the Jevons paradox part, which is by making the kind of efficient models affordable, and you can now actually successfully automate that task. It's, like, affordable to do that because, like, in one scenario in this hypothetical example, it costs, like, $100,000 to agentically audit that deal or that company. In the other scenario where you have Frontier Plus cheap, then you can do it for $25,000, and the clearing price for doing it is, let's say, like, $35,000. So all of a sudden, you would do it for the first time, where you just weren't spending tokens on that at all. So there's a lot of, there's far more of those in the economy than anybody realizes.
And so what I think you'll see is the frontier will often be used as kind of the orchestrator. It's the thing determining the tool use. It's the thing that does the computer use. A great example would be like let say you designed a cyber agent and the cyber agent job was to go look at all of the risks inside of an enterprise you probably want whatever the most frontier is let say Mythos 5 as an example or the latest from Astra. You might want that as you're like, this is the cyber researcher. But then eventually you have to go process tens of thousands, hundreds of thousands, millions of log files to kind of give you some kind of indication as to, like, is that log event like a risky, you know, event or not. Maybe you use JEV or something like that or use some other kind of decision model or something that's super, you know, very, very token efficient. But then that cyber agent is sort of directing the task and then reviewing data
that those other models are kind of, you know, doing some summarization or classification of. That is exactly how you're going to see AI get deployed in these kind of complex scenarios, is frontier model for decision-making, reasoning, the orchestration, and then whatever allows you to process large amounts of information and data. And then guess what? Again, now you're spending on both. Both ecosystems win. Open Router and its ilk win, as well as OpenAI, maybe the latest from Gemini or Grok or Anthropic end up getting the Frontier tokens. So your Anthropic spend is going up? In that scenario, 100%. But in that scenario or gross, like overall? Oh, gross, for sure. Yeah. Because you're just doing way more. You're deploying these systems across now everything. Talk a little bit about Fable 5.2. I've obviously seen examples. I've read the benchmarks. but your company is always, I think, pressure testing these models to see what they can do in real work.
So what have you seen on the Fable 5.2 side? Because everyone is talking about this model. Yeah, so Fable 5.2 and just even expand it up. Anthropics latest frontier model. Anthropics latest frontier model. Even the Gemini model, Carbon, is incredibly strong. And I think Google is slightly the surprise for a lot of folks. They've been shipping. I mean, oh my goodness, the release cadence for them has been crazy. What's gotten into them? The release cadence is crazy. They are clearly in a cycle of, they're in a flywheel right now. And, like, that's unquestionable. And I think you're seeing in the benchmarks and even the real scenarios, like the models are getting way better from Gemini. Fable 5.2, fantastic. Obviously, there's rumors about what will eventually, you know, kind of emerge from OpenAI, maybe even as of whenever this comes out. We'll see updates there. But basically, you know, it's all of the latest crop of models, you know, generally have roughly a similar feel, which is just the reasoning levels have gone up.
Obviously, the coding capabilities are way higher. You know, people are, you're finding way more nuance in the judgment that the models are making. You know, they kind of start to really sort of feel like, okay, they kind of understand what you're really trying to accomplish. It's a little freaky. It's a little freaky. I do a lot of product prototyping with, you know, whatever the frontier happens to be. And, like, you know, a year ago, I was really having to prompt these things. Like, I was like, okay, here's a bunch of research. Here's a bunch of information. Here's what I'd like you to go prototype. and then it prototyped something and, like, it's, like, nothing was, nothing's, like, particularly uncanny about what it came up with. The UI tends to be kind of janky still. Like, that's, like, six months or a year ago. Today, you know, it's, like, it's very astonishing. If you looked at my computer, I have, like, you know, five or ten tabs, two to three apps open, and it's just usually, like, building some prototype in something that we're trying to do,
and I'm just, like, you know, trying lots of things. And what's uncanny is, like, the ability, you know, you kind of brought up the generalization, you know, kind of point. And I'm sorry it's not happening in the breathing benchmark yet. It's happening in software design. Okay. You can give these models a task like, oh, you know, design a new enterprise experience for XYZ thing that will do that as a box. And I'll say, study, you know, these 20 companies of how they thought about it. The 20 companies might not have anything to do with box, just, like, general industry patterns. and it'll come back and it'll absolutely do something that no product manager could ever have done because no product manager could have read the documentation of 25 different companies and then correlate like interesting trends that are happening outside of our industry and pulling those into our experience so we could like deliver something that was a totally different outcome. We're now there. Like it's like happening and like that will completely change. For those that adopt the tools, it'll completely change the shape of how you build software
or how you build technology in the future. Because you have a super, you basically have, you know, I don't, I sort of freely use this term just because I don't really care about the actual definition, but you do have effectively a super intelligence now that just will go and do an infinite set of research that you could never have accomplished before, that will just be, you know, like the super researcher and design whatever product or experience or whatever you're trying to do in a way that you could never have accomplished just with the number of hours you have in a day or the amount of ingenuity that we tend to have. Okay, so let me see if I can sum up why you think the Anthropic IPO or Anthropic is going to work. If it wasn't clear. The models are getting better. As they get better, they can take on more tasks outside of coding, things like architecture, gaming, production, et cetera, film production, et cetera, cyber. While they're doing that, they're going to be agentic. So they're going to be operating basically outside of you just having a prompt, prompt, prompt, and that will be a lot to token spend. And frontier models are still going to be useful even if,
and they might be more useful if standard models end up taking on some of the tasks. So if you put it all together, you have a company at the center of that, like an anthropic, and then they're just well-positioned for what comes next. Yep. Okay, so. Very succinct. I probably should have said that just way faster. No, no, this was good. It was good to tuck it out. Okay, you pulled it out. Yeah, okay. All right. So we're going to take a break. When we come back, I want to talk about some of the rebuttals and some of the dangers of believing too much in Anthropik, according to some of the critics. And we'll also talk about your latest BoxWorks news and then maybe get into the frontier of the technology. All right. Let's do that right after this. This episode is brought to you by LinkedIn Pulse. One thing I've learned covering technology is that the pace of change is relentless. Just when you think you've wrapped your head around one major shift, another comes along. And keeping up isn't just about following the headlines. It's about understanding what's actually changing and what those changes mean for businesses, the workforce and the economy.
Now, that's where LinkedIn Pulse comes in. LinkedIn Pulse provides unique insights on work and opportunity, informed by data from 1.4 billion members, 71 million companies and 42,000 skills. These insights help us understand how the era of new work is taking shape as businesses evolve and the skills people need change. And what's especially useful is that LinkedIn Pulse isn't just trying to predict what might happen next. It's helping us understand what's happening right now, which is essential when everything is moving this quickly. So explore LinkedIn Pulse at pulse.linkedin.com. This episode is brought to you by Intelligence AI. Is your investment in AI coding tools helping your team ship better software? Intelligence connects your AI spending to code quality and engineering outcomes, from individual sessions and pull requests to entire projects. You can see where AI saves time, where it creates rework, and how to improve.
And its model router selects the right model for each task, based on quality, speed, and cost. See how Intelligence AI helps enterprises optimize AI-powered delivery with connected visibility into cost, quality, and outcomes at intelligence.ai. I know planning for the future isn't always the most exciting thing to think about, but it's one of those things that can really matter in the long run. I've always been a fan of anything that makes planning for the future a little easier. Ethos is a great resource for exploring life insurance options without making the process more complicated than it needs to be. Ethos makes the process fast and easy, and it's 100% online. You can get a quote in seconds, apply in minutes, and even get same-day coverage. There's no medical exam. You just answer a few health questions online. You can get up to $3 million in coverage with some policies as low as $30 a month. Take 10 minutes to get covered today with life insurance through Ethos.
Get your free quote at ethos.com slash bigtech. That's E-T-H-O-S dot com slash big tech. Application times may vary. Rates may vary. Everyone's talking about AI, but there's a big difference between running a promising pilot and actually making AI work across an enterprise. That's why we're launching the Big Technology Guide to Series presented by DXV Technology, available on YouTube. Throughout this series, we're tackling some of the biggest questions facing businesses today. How do you take enterprise AI from pilot to production? How do you navigate an increasingly complex cybersecurity landscape? What's the real opportunity with AI agents? And how do you calculate the return on your AI investments? On my Alex Cantor's YouTube channel, you can hear from experts like DXP Technologies, Chief Digital Information Officer, Russell Jukes. When you ask a question to somebody in your company, do you trust them? And what made you trust them? Do you trust but verify every single time? Russell and I discussed the challenges and opportunities of building production-grade AI
and what it really takes to move beyond the hype. Check out the Big Technology Guide To Series presented by DXC Technology on my Alex Kantrowicz YouTube channel today. And we're back here on Big Technology Podcast with Aaron Levy, the CEO of Vox. It's always great to speak with you, Aaron. Thanks for coming on the show. Always fun, always fun. Very quickly, the rebuttals to why Anthropik will necessarily do well. I'll just give you the big three that we got from Gary Marcus earlier this week. and you can give us your perspective on them. The first one is that they have a very big valuation. They have a lot of debt or a lot of promises to spend money that they may never be able to fulfill. $900 billion in future liabilities. I think 80% of them, they can't get out. Then there's the token maxing thing. They were making their revenue from Q1 to Q2 grew 147% and then token maxing slowed down. The revenue still grew 47% from Q2 to Q3, but that was growth slowdown.
And then the last thing is diminishing returns. To be able to get those increases in capability that you're talking about, they have to spend much more money, and it's applicable in much narrower fields as opposed to being applicable to everybody. So if you put that all together, you might say, hmm, is this really a company, despite all the positive things we talked about in the first half, that you want to get behind? Yeah. Your thoughts. Should I do each one or the whole thing? I'll just let you take it. First of all, I really enjoy Gary's content, and I actually enjoy all of the, you know, I don't know what the term everybody's agreed on, but, you know, the critical ecosystem. I think it's good to pressure both sides of this. I'm obviously going to be, like, a super optimist, and so you need to make sure that some people are couching that. Gary said similar things about you and suggested one day we might get you both together on the show or at one of our events and have a little bit of a debate.
Yeah, for sure. So I like all sides of the debate and I've taken the other side in certain kind of bubbly situations where I felt like, I'm not sure this is going to work out. But AI is just, I think, a very different kind of moment in our technological kind of progress, you know, timeline. So, you know, there's a few kind of maybe separate points in there. On the debt thing, I think I would just have to put that in the category of, like, if you believe what we talked about just now, then the debt thing is just like, yeah, that's just like how, that's just how buying the infrastructure works. And so, like, you know, hard to parse, like, what is the right amount of commitments Anthropics should have. If you believe that the thing is going to generate hundreds of billions of dollars, you know, a year in revenue, then those commitments are fine. And if you don't believe that will happen, then obviously that will blow up the company. So, like, everything is sort of downstream from whether you just think that the model, that, you know,
that the growth, you know, continues to sustain at some level. It doesn't have to sustain at the same percentage. Okay, so let's mostly focus on probably the second and third points. I think the – I'm not sure the token maxing point is that relevant. I think we've moved from token maxing, you know, like that was like a two-week window where like people thought it was fashionable to like have leaderboards and incentivize token maxing. I don't think that is like the core thing that is responsible for why Anthropic is, you know, an 80-plus billion dollar, you know, kind of revenue run rate company. So I'm not sure, like, that is sort of maybe a phenomenon that I don't think is, like, that meaningful. On the narrower and narrower tasks, I'm not exactly inclined to believe that's true. So I spend a lot of time in the cyberspace both because Box is a platform We spend an incredible amount of energy working in cyber And then both in investments and just knowing startups I watched the cyber ecosystem
And it's an example where you might have concluded cybersecurity is like a niche task, I think, in your third example is like, okay, like the task is smaller and smaller, smaller and smaller. Well, I would say basically like the cyber, okay, I might call it a niche task, but it might just be like you'd run another, you know, $100 billion pre-train, and it helps, like, you know, a handful of biologists with. Yeah. So you really need to, like, come up with, like, and as they talk about this cure, you really need to come up with a cure for Alzheimer's for that to be worth. Yeah, well, so first of all, I'm not 100% sure about the $100 billion pre-train. Like, that doesn't appear to be the cost of that. Right, that's probably exaggerated. Yeah, exactly. And the post-training side, which is where, like, a lot of the RL data shows up, First of all, like, you know, this stuff is not, it's not like hundreds of billions of dollars of post-training. Like, you can, you know, make a lot of progress. You know, I, you know, pay attention to some of the post-training startups. You can make a lot of progress with millions of dollars of RL data and compute going into post-training on particular domains.
And so I would actually say that up until, you know, this summer, coding was kind of the dominant conversation that we had. So I would probably guess that like 90% of our post-training energy went into code tasks. You layer on cyber, you layer on bio, you layer on legal, you layer on manufacturing, layer on product design and, you know, CAD. you layer on. Math actually is maybe the one that is maybe more exotic. I'm sure there's downstream benefits we get from all the math work, but you probably monetize that the least. That appears to be more like a fun competition between the labs. But all these other commercial use cases, unquestionably you could spend billions, not tens of billions maybe. You don't need to spend tens of billions right now, but right now you could spend billions of dollars on getting all of this RL data from every single one of those spaces. And I think the models would improve just meaningfully on every one of those dimensions
as that kind of rolls through the models. And each of those TAMs are measured in the trillions of dollars. Like every TAM I just mentioned has at least a trillion dollars of like global economic activity going into it. And so if you can make each of those TAMs, you know, 50% more productive or 100% more productive or 300% more productive, could you capture 20%, 10% or 20% of each of those TAMs? I don't see why that would be impossible. So then you're like, okay, so, you know, $10, $20 trillion of TAM where the, that doesn't have to be anthropics specifically, but Frontier Labs, let's say the Frontier Labs make each of those TAMs, you know, two to three X more productive in terms of just like the actual raw, you know, productivity gains of those TAMs, I think you can capture a sliver of that as direct revenue to these models. Like the world spends, you know, just even public,
the public market cap of cybersecurity companies in total is probably, I don't know, like half a trillion dollars or something like that. Like, you'd have to, like, you know, peel out, like, what you think Microsoft's TAM is and Google's, or sorry, you know, spend and Google. But you look at CrowdStrike, Palo Alto Networks, Microsoft, you know, Okta, Fortinet, you know, Checkpoint, Sentinel One, Zscaler, all of that. Like, all of that adds up to probably about half a trillion, again, when you added Microsoft's and Google's. That's all nearly pre-AI revenue, okay? Just on cyber. So now you layer on agentic cyber, and there's just no reason why that category isn't two to three times larger as an end result of just the software part of that category. Now imagine the kind of human augmentation of every single security team, every single cyber operations group. All of a sudden, you could easily sort of underwrite that
as hundreds of billions of dollars in revenue just in cyber inference. Again, go industry by industry. What appeared to be small niche kind of markets all of a sudden are very, very large because you measure them against the total spend globally on that area of work, and you say could AI capture, again, like 10% or 20% of that whole area of work over the next five or 10 years, and I think it's not unreasonable to think that would happen. Let me give you one more counterpoint, which is that there could be a duration mismatch here, which is that basically there's been so much money invested in these companies and all these industries that we're talking about are going to take a long time to change. And they might not be ready to change or might not be able to change before the bills come due. And in fact, OpenAI's last round is a good illustration of this. There was talk about this round. They were going out. There were stories about investors having indigestion. They can't possibly put more money into these companies anymore and they needed to be paid out. And, in fact, they went out looking for like $30 billion.
They ended up with $11 billion, maybe a little bit more. But I think the round is still open. It's still open. Yeah, it's still open. I'm not, I don't know what we would conclude yet from that. Right. But at a lower valuation also than they were seeking. And, in fact, Anthropic was supposed to go out. I think a trillion dollars is like we're doing pretty well here. Okay. But you have to admit that it's a potential issue that this stuff just doesn't pay off and the money runs out. Oh, so like, I don't know what percentage I'd peg it at, but like 100% that could happen. Well, like, I probably shouldn't have used a number in that. Let me not give you a number, but my percentage. It is absolutely possible that that could occur. So I'd like to, you know, sort of retract any number from that sentence because that could be confused, and your AI model will get very confused. I know. Yeah, it will start to pry its brain as it gets through the breathing exercise. Exactly, exactly. I should have just breathed all over that when I said it. So, like, that is truly an accurate skeptical take.
And it's like, I don't know if that will happen. You know, from what I see from the usage, just even internal at Box and our own productivity, and when I go talk to customers and when I kind of, you know, kind of try and calibrate, like, where are we? Like, how much am I seeing in Silicon Valley versus the rest of the world? It's ironic because some people see that as the duration mismatch. I see that as the opportunity ahead. Yes. And so then you're just sort of like, okay, well, fine, maybe, like, we're off by a year or two. But, like, when I go look at, like, any friend that is deploying inference and they are, like, always running into compute limits as the first thing that they run into, I'm like, I'm not yet seeing the part where sort of we have so much compute capacity that's underutilized. Like, that hasn't occurred yet. So, I'm, like, I totally believe that that is a plausible scenario. You know, I, like, if I were having to run, like, a massive capital allocation organization,
like, sure, I might be more stressed about that. You know, I'm not having to make that decision. Like, in our part of the world, we're only seeing things that are up and to the right in terms of the overall demand for AI. But I totally believe that there could be other scenarios, and we should, like, do scenario planning for all the possible outcomes. But I think there's – I think some of when I hear those scenarios, they sometimes come from the perspective that underestimates the utility that these models are generating. And so I'd believe it a little bit more if there was first a deep recognition for the raw utility and then there was just a conversation about the mismatch of duration. but usually it sort of like starts with the premise that these things are, you know, these things aren't actually that useful, and so then there's going to be a duration mismatch, and that's the part where I'm like, no, actually, like, these things are actually insanely useful, and so then it's just like, okay, how quickly does this diffuse? And that's, I think, a super interesting debate, and like, will there be blow-ups in the meantime?
Sure, of course, there is in every build-out, but anyway, I enjoy both sides of this, And because, you know, again, because I don't have to, like, make massive capital allocation bets, you know, as an investor. Like, I think it's a little bit more of a, you know, kind of like just a hobby topic. So I can enjoy debating it a little bit more. Okay. And we should also mention that you're getting involved in a lot of the things that we've discussed today. So, for instance, you just had your big annual event, BoxWorks. A couple of big announcements there, including the Box Architect, where you're going to give companies the ability to build these agentic tools, and Box Industries, where you're working with some of the industries that we talked about to make AI applicable for them, including banking, public sector, media, and entertainment. Yep. Yeah. So for us, I mean, part of my excitement is certainly driven by our own internal sort of like what we're seeing now be possible.
So Architect is a great example of this. We have had a box agent capability for very like formally a year. We had AI sort of like basic summarization and kind of quick responses to questions like for the past couple of years. but our thing that really most represents like the power of agentic capability is about a year old and what that agent does is it goes and you ask it a question, it'll review your documents, it'll answer questions for you, it'll process data for you and then you can drop that agent into workflows and so you can design a workflow and the agent can go and sort of say, okay, I'm looking at this loan document, I want to summarize it for the decision that the loan officer is going to make and so then you can design a workflow that agents kind of drop into and allow you to augment or accelerate those workflows. That was great. With Fable and Astra, we started to sort of see another kind of like interest,
like, you know, kind of new step functioning capability and it goes back to like that kind of computer and tool use that we talked about earlier where we started to see that like these models are now at a point where you don't have to just deploy the agent into the workflow. You might be able to have the agent design the entire workflow. And if you gave an agent all of the tools in Box, our workflow engine, our no-code app builder, our file system, our security features, let's say each of those were underlying tools, and you could describe to the agent roughly how somebody thinks about these tools and how they get configured, could you use the deep domain expertise of these models in, you know, legal work or loan work or, you know, banking transactions or insurance because they're being trained more and more on these domains. Could you use that level of knowledge plus tool use to help you actually design your entire process? And so, you know, with a bunch of testing and iteration, we realized that, like, yes,
we're on the cusp of that becoming really, really possible. So we announced a new agent in our platform called Architect. And what Architect does is exactly that. It architects full solutions, and then eventually we'll kind of run and optimize those solutions for you as well. So for the first time ever in our category, you can effectively have an autonomous platform that manages, automates, administers the workflow around enterprise content. So you come to Architect and you would say, design me a contract management system. And it'll just go build you the contract management system. With an important key thing, it's not actually like, you know, writing code and then you have to host it and you have to maintain it. It's all using the box underlying primitives. But now it can design a contract management system for you that will now effectively autonomously improve in the future. And it'll learn from your workflows. Yes. And so now imagine that for any workflow in your enterprise that deals with content, that's effectively what we've built out.
Okay, so when you were at our event, we were in the Big Technology Summit, the Big Technology AI Summit, I'll get the name right. We were in the wake of the SpaceX IPO. Yeah. And you had mentioned that you had bought some SpaceX shares. Yes. Do you still hold them, and are you happy with that decision? Yes, I still hold them. I'm happy about it. But I want to, like, again, like, back in the category of financial advice, like, I don't do much trading. It was a token, like, participation in SpaceX. So this is not like, you know, like, wow, because of, you know, I made some financial decision. You should trust me on that. Like, I did it because I want to participate in SpaceX. And are you going to do the same with Anthropoc? Yeah, sure, sure. Okay. Yeah. Aaron, really, it's always great to speak with you. We joke that every time we talk, every time we schedule something to talk, there's a new twist in turn. I don't know. I always say it can't get crazier or more interesting from here, but it always does. So thank you again for spending time with us and helping us unpack these pretty complex topics in a very accessible way.
So it was really great to see you. Thank you. All right. Thank you, Aaron. Thank you, everybody, for listening and watching. And we'll see you next time on Big Technology Podcast. Thank you.
番組の概要欄(原文)
Aaron Levie is the CEO of Box. Aaron joins Big Technology to discuss the bull case for Anthropic’s IPO and the future of enterprise AI. Tune in to hear why he believes AI agents will drive massive demand, how adoption could expand beyond coding, and why cheaper models could benefit the frontier labs. We also cover AI’s reliability challenges, the risk that revenue won’t arrive before the bills come due, and how Box is putting agents to work. Hit play for a clear-eyed discussion of AI’s business potential and the assumptions behind the optimism. --- Enjoying Big Technology Podcast? Please rate us five stars ⭐⭐⭐⭐⭐ in your podcast app of choice. Want a discount for Big Technology on Substack + Discord? Here’s 25% off for the first year: https://www.bigtechnology.com/subscribe?coupon=0843016b Learn more about your ad choices. Visit megaphone.fm/adchoices
出典・この記事について
- 番組
- Big Technology Podcast(Alex Kantrowitz)
- 元の回
- Why The Anthropic IPO Will Succeed — With Aaron Levie
- 配信日
- 要約作成日
- 公式サイト
- www.bigtechnology.com
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
関連エピソード

機械が「主体性」を持つために必要なこと(マイケル・トマセロ)
Machine Learning Street Talk (MLST)

Workday のプロダクト責任者 Max Wessel が語る、生成UI・信頼できるAI・ポリマス型組織(Workday Rising 2026 ライブ収録)
AI and the Future of Work: Artificial Intelligence in the Workplace, Business, Ethics, HR, and IT for AI Enthusiasts, Leaders and Academics

超知能には物理世界が必要だ:Arda共同創業者Bob McGrewが語るAIと製造業
No Priors: Artificial Intelligence | Technology | Startups

Runwayはいかにして AIエージェントでクリエイティブ業務を変革しているか
The Artificial Intelligence Show

MCP、MHS、そしてオートメーションピラミッドの終わり?
Industrial AI Podcast

Orbと永続スレッド、そしてAmpの次の一手(Changelog Interviews #691)
Changelog Master Feed