Big Technology Podcast · · 54 min
Why The Anthropic IPO Will Succeed — With Aaron Levie
With Aaron Levie
ChorusBring the market's conversation to your AI. Learn more
Summary
Aaron Levie joins Alex Kantrowitz to argue that Anthropic's IPO can succeed, mainly on the strength of enterprise demand for frontier models. He cites the S1 figures: customers spending $100k or more rose from about 950 to 11,900 in a year, revenue grew roughly 12x year over year to about $17B, and inference margins improved sharply from $0.40 to $1.72 per compute dollar. He answers the bear case that cheap open-weight models and routing tools will commoditise frontier labs with a Jevons-style argument: frontier plus cheap models makes agentic work economic, so spend on both rises. Levie extends the case to agents that remove the limit on knowledge-work output, and to large new token markets in cybersecurity, CAD and product engineering, each worth roughly a trillion dollars in economic activity. He concedes that margins, commoditisation risk and the pace of diffusion remain open, and he dismisses overcapacity and duration-mismatch concerns for now.
Key takeaways
- Aaron Levie (bullish on Anthropic): Frontier AI demand is very strong and broadening; Anthropic's growing enterprise spend shows real, widespread adoption.
- Aaron Levie (bullish on Anthropic): Anthropic's latest model, Fable 5.2, is strikingly capable: it designed a home stereo from the ground up.
- Aaron Levie (bullish on Anthropic): Anthropic now earns more per compute dollar than it spends, so its frontier business looks attractive and cash-generative.
- Aaron Levie (bullish on Anthropic): Cheap open-weight models and routing will raise total spend on frontier models, not commoditise Anthropic.
- Aaron Levie (bullish on Alphabet): Google's Gemini is a surprise, with a crazy release cadence and models getting much better.
What was said, by company and theme
13 views from 1 speaker. Every quote is checked against the episode's transcript (how).
Anthropic
Across the shows: 134 speakers in other episodes, 230 bullish and 114 bearish calls, past 30 days
Where everyone stands on Anthropic
Aaron Levie · BoxBullish
Frontier AI demand is very strong and broadening; Anthropic's growing enterprise spend shows real, widespread adoption.
Levie is bullish on frontier AI demand, with a market of perhaps half a dozen investable companies at the frontier, including Anthropic and OpenAI. He does not give a view on the share price or the right market cap. The evidence he cites is the S1 metric Kantrowitz raised: Anthropic's customers spending $100k or more rose from about 950 to 11,900 over the past year. Levie reads that breadth as proof that AI is broadly useful, since companies are 'voting with their wallets'. He also cites revenue growing roughly 12x year over year (Q3 2025 $1.4B to Q3 2026 $17B), but says nobody should expect another 12x in the next twelve months. He argues AI is at the early inflection point of adoption and diffusion. Frontier token spend could be 10x higher and total economy-wide token use 100x higher as agents spread, though that diffusion takes years and needs better software form factors and change management. He flags open questions on long-term inference margins and open-weight price competition.
“But when you have, in the tens of thousands of companies spending 6 or 7 figures on AI, like it's working. Like this is like really happening.”at 7:11
Aaron Levie · BoxBullish
Anthropic's latest model, Fable 5.2, is strikingly capable: it designed a home stereo from the ground up.
Levie cites Fable 5.2, which he calls Anthropic's latest frontier model, as evidence that models now do real engineering and design work. One example he saw online has the model build a home stereo system from the ground up, doing the circuit design, the physical design and the manufacturing plans over a long autonomous run. He also says frontier models now show much better reasoning and judgment, and understand what the user is trying to accomplish. In his own prototyping work, a year ago models needed heavy prompting and produced janky output, but now they produce results that feel uncanny. He uses this to argue that product and engineering work will be transformed.
“Fable 5.2 built like a home stereo system from the ground up... It did the physical design. It built the manufacturing plans.”at 10:02
Aaron Levie · BoxBullish
Anthropic now earns more per compute dollar than it spends, so its frontier business looks attractive and cash-generative.
Kantrowitz cited the S1 figure that Anthropic earned $0.40 for every compute dollar in Q1 2025 and $1.72 by the later period. Levie reads this as proof that frontier AI businesses can make money despite very expensive training. Inference has become very profitable, and training is a smaller share of total compute than it was about a year and a half ago. That shift makes these businesses look attractive and cash-generative. He says critics who thought this would never happen should update their views. He admits the bet depends on the margins holding and the models not becoming fully commoditised.
“Now, all of a sudden, fast forward today, inference very profitable, training sort of less of the total percentage of compute... these businesses actually look pretty attractive. They generate cash. It works.”at 23:54
Aaron Levie · BoxBullish
Cheap open-weight models and routing will raise total spend on frontier models, not commoditise Anthropic.
Kantrowitz put the bear case that routing tools like OpenRouter could send work to cheap models and leave frontier labs commoditised. Levie says a year or more of evidence already points the other way. Frontier models plan, orchestrate and make judgment calls, while cheap open-weight models process large volumes of data. Customers end up spending on both, so frontier spend rises. He gives a hypothetical: an agentic audit costs $100,000 on frontier-only, but $25,000 with frontier plus cheap models, against a clearing price of $35,000. That makes the audit economic for the first time, so total usage grows. This is his Jevons paradox argument: cheaper efficient models expand demand, and Anthropic's own spend goes up in that scenario.
“You're actually spending more on Frontier than in the AB scenario where you didn't even have open weights. Like this is the Jevan's paradox part”at 26:34
Alphabet
Across the shows: 74 speakers in other episodes, 64 bullish and 51 bearish calls, past 30 days
Where everyone stands on Alphabet
Aaron Levie · BoxBullish
Google's Gemini is a surprise, with a crazy release cadence and models getting much better.
Levie says Google has been a surprise for many people. Its model release cadence has been very fast, and Google is in a flywheel that he says is unquestionable. He points to benchmarks and real-world scenarios showing Gemini models getting much better, and says the latest Gemini model is incredibly strong, comparable to the Anthropic model he praised. He treats Gemini as one of the frontier labs competing at the top of the market.
“I think Google is slightly the surprise for a lot of folks. They've been shipping... the release cadence for them has been crazy... They are clearly in a cycle... they're in a flywheel right now.”at 29:40
Cybersecurity
Where everyone stands on Cybersecurity
Aaron Levie · Box
Agents will review every line of code and log in future, growing cyber token use about 100x.
Levie argues cybersecurity is a large token-consuming use case. Every line of code will be reviewed by other agents. Infrastructure will be constantly reviewed, and agents will go through every log and event in any connected system, including critical infrastructure, IoT devices and cars. He estimates this area will see probably 100x token growth from current levels. The argument is that cyber is economically valuable enough for labs to train models on it, which then improves the models.
“every single line of code in the future will be reviewed by other agents. Our infrastructure will be constantly sort of reviewed by agents.”at 10:29
Aaron Levie · Box
Agentic AI could make cybersecurity two to three times larger, with AI-assisted security teams worth hundreds of billions.
Levie cites his own exposure to cyber through Box and his investments and startup watching. He says the public cybersecurity companies (CrowdStrike, Palo Alto Networks, Okta, Fortinet, Zscaler and others) total about $500B in market cap, nearly all pre-AI revenue. He argues agentic cyber should make the software category two to three times larger. Augmenting every security operations team could support hundreds of billions in cyber inference revenue. He argues that niche-looking markets become very large when measured against total global spend, and that AI could capture 10-20% of that spend. He rejects Kantrowitz's narrowing-tasks argument.
“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.”at 43:11
Enterprise AI adoption & AI software
Where everyone stands on Enterprise AI adoption & AI software
Aaron Levie · Box
Agents break the link between knowledge-work output and human time, so usage can be uncapped and token spend will become pervasive.
Levie argues that for forty years knowledge-work output was limited by how fast people could type, design or code. Agents break that limit. A business can now set a goal and a budget, such as $20,000 for a research project, and get the output of a thousand people's work with one person starting it. He says this was not technically possible two years ago, because models were too error-prone for it. Reliability has now improved enough, so organisations will spend on agents without a cap. For example, a company will pay $5 million to secure its enterprise if the spend reduces risk. He expects token use to become a pervasive operating expense within about three years, and says the slope of usage growth is still unclear.
“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.”at 14:59
Aaron Levie · Box
AI could do engineering and CAD design as well as code, a roughly $1 trillion market within a couple of years.
Levie says engineering is probably a trillion-dollar TAM if AI can do strong CAD work. He expects models to reach the level of geometry and component design they now reach in JavaScript or Python within a couple of years, once that training is done. Product design, which he says looks almost identical to software development, is one of the areas that will consume many more tokens, across every physical goods company.
“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.”at 20:40
Frontier AI labs & model race
Where everyone stands on Frontier AI labs & model race
Aaron Levie · Box
Domain post-training costs only millions per domain, so models should keep improving across commercial use cases.
Levie pushes back on the idea that narrowing returns make expansion uneconomic. He says post-training on specific domains is far cheaper than pretraining. You can make a lot of progress with millions of dollars of RL data and compute in a domain, and coding has so far absorbed about 90% of post-training effort. Spending billions across cyber, bio, legal, manufacturing and design would meaningfully improve models in each of those areas. He calls math work a lab competition with less direct monetization.
“it's not like hundreds of billions of dollars of post training. ... You can make a lot of progress with millions of dollars of RL data and compute going into post training on particular domains.”at 40:09
Aaron Levie · Box
Frontier labs can capture a sliver of multi-trillion-dollar TAMs as direct model revenue.
Levie argues that each domain he lists (cyber, bio, legal, manufacturing, product design/CAD, code) has at least a trillion dollars of global economic activity behind it. If frontier models make each of those areas 2-3x more productive, he thinks labs can capture a sliver of that as direct revenue. He says this is about frontier labs generally, not Anthropic specifically. He cites public cybersecurity companies (about $500B combined market cap) as a sense of scale for one of these markets. He is bullish on the revenue opportunity and treats it as a matter of productivity gains flowing into the labs' pricing.
“Like every TAM I just mentioned has at least a trillion dollars of like global economic activity going into it. ... I think you can capture a sliver of that as direct revenue to these models.”at 41:39
AI infrastructure & data centers
Where everyone stands on AI infrastructure & data centers
Aaron Levie · Box
Inference customers keep hitting compute limits; no sign yet of underutilized AI compute capacity.
Levie checks the overcapacity bear case against what he sees from companies deploying inference. Friends and customers keep running into compute limits as their first constraint. He says he is not yet seeing the surplus capacity that a glut scenario requires. From his position he only sees demand moving up and to the right. He grants that other scenarios are plausible and should be planned for, but this is the evidence he is using now.
“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 ... we have so much compute capacity that's underutilized. That hasn't occurred yet.”at 46:17
AI bubble & tech valuations
Where everyone stands on AI bubble & tech valuations
Aaron Levie · Box
AI spending commitments are fine if AI revenue grows as expected; the timing mismatch is an opportunity.
Kantrowitz argues that AI companies' commitments may come due before enterprises can change their workflows. Levie replies that the commitments are justified if the technology generates hundreds of billions in revenue a year, and everything depends on that. He says he sees the utility of these models as very high. The real debate is how quickly the technology diffuses through the economy. Blowups along the way are likely, as in every buildout. Some people call the gap a duration mismatch, but he sees it as an opportunity. He also says he has not put money into the bets that carry capital-allocation risk, which makes the debate more of a hobby for him.
“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.”at 47:14
Across the shows
What other episodes Chorus covered said about the same companies in the 30 days up to this one.
Topics
- AI agents in cyber defense
- AI for engineering and CAD
- AI spend vs revenue timing
- AI-driven cyber market expansion
- Gemini model progress
- agent-designed enterprise workflows
- agents and uncapped token usage
- cheaper models and frontier demand
- compute capacity utilization
- domain post-training
- enterprise demand for frontier models
- frontier lab revenue capture
- frontier model capability
- inference profitability
More from Big Technology Podcast
- Why The Anthropic IPO Will Fail — With Gary Marcus · November 17, 2026
- How Humans Will Work Alongside AI Agents — With Mike Cannon-Brookes · November 17, 2026
- Storm Clouds Ahead For Apple?, Debating Amazon’s Bee Frontier, AI as UI · November 11, 2026
- Will the Democrats’ Big Win Stop the AI Buildout? — With Teddy Schleifer · November 5, 2026
- This AI Expert Argues We're Underestimating AI's Productivity Impact — With Ethan Mollick · October 28, 2026
Ask your AI about this episode
Chorus gives Claude, ChatGPT and other assistants every view from this episode and every other show Chorus follows. Ask who agrees, who changed their mind, and what else was said about the same companies.
Chorus reports what speakers said. It isn't investment advice. How Chorus works · About Chorus