1. Podcasts
  2. Uncapped | Benchmark Breakfast
  3. Benchmark Breakfast #1
Uncapped | Benchmark Breakfast

29 Oct 2026 · 56m · 20 chapters

Episode artwork

Benchmark Breakfast #1 | Jack and Ev

In short

The hosts launch a new “Benchmark Breakfast” format (no agenda, like their long Monday partner meetings) and use it to discuss AI markets and investing. Main focus is Anthropic’s S-1 and what it implies about AI unit economics, growth durability, and open vs closed inference. They also cover AI codegen’s “product-market fit” moment and its downstream infrastructure boom, plus emerging cyber risks from frontier models and “obliterated” (safety-removed) open-weight models.

Guests

Jack and Ev (the episode is “Jack and Ev”; no other guest names appear in the transcript).

Guest backgrounds

Both are Benchmark investors/partners who regularly run multi-hour Monday partner meetings and invest across AI, codegen, and infrastructure categories (they reference Benchmark portfolio companies throughout).

Key claims

Anthropic’s S-1 shows compute-to-revenue improving fast, with operating leverage (e.g., ~8% operating margin) and “money printing” economics. Codegen can drive broad, multi-derivative growth across the software stack. Open vs closed is a revenue-share split (open likely mid-single digits to ~10% of inference spend), with open not necessarily cheaper. Cyber risk is becoming concrete via internal model-assisted exploitation.

Notable examples

Anthropic revenue vs compute chart (Q1’25 ~0.4x; Q2’26 ~2.1x), 145% QoQ growth then deceleration, 80% consumption-based revenue. Codegen derivatives: Supermicro, Vertiv, Arista, Bloom Energy, Sandisk/Micron/Seagate; and venture examples like Fireworks, Base10, Modal, Superbase, ClickHouse, Railway, Cloudflare, Temporal, GitHub Actions, ChainGuard, Socket. Cyber examples include Claude disabling webcam indicators and a supermarket refrigeration attack; obliterated models used for hacking/defense.

Written by AI. May contain mistakes.

Chapters

0:39

Insights from Monday Meetings

Discussion about the informal and open nature of their Monday meetings.

“And I guess what we do on Mondays at Benchmark has basically been no agenda.”
2:05

Anthropic's S1 Filing Reaction

Exploring the implications of Anthropik's S1 filing and market reactions.

“So as somebody who can actually read an S1 of our mostly early stage group, what was your reaction?”
4:56

Revenue vs. Compute Expenses

Analyzing the relationship between Anthropic's revenue growth and compute expenses.

“Which is like, no, these are money printing machines.”
8:01

Growth Challenges and Market Concerns

Discussing growth rates and potential challenges for Anthropic in the market.

“the value that you can get as an individual is probably in like, at least in the thousands of dollars, if not tens of thousands of dollars.”
10:28

The Future of AI and Coding Revolution

Speculating on the future of AI programming and the evolution of coding practices.

“And it was basically like everybody who wanted a Zoom account got it.”
12:54

AI Market Trends and Investment Opportunities

Examining trends in the AI market and investment opportunities related to infrastructure.

“I mean, the other thing around that is all of the startups and markets around code, the right answer has basically just been bet on all of it.”
14:00

Impact of AI on Data Centers and Market Trends

Explore the various impacts of AI on data centers and the rising market trends in related sectors.

“But then people were like, shit, what else goes into an AI data center?”
16:48

Code Generation's Market Fit Moment

Discuss the significant advancements in code generation and its product market fit in the tech landscape.

“You start, you know, one year at 10, you go to 100, and then you go to like 300 or more.”
20:30

Durability in Tech Investments

Analyze the challenges in identifying durable tech investments amidst evolving trends.

“And some of the like more traditional like signs of durability and defensibility from like an enterprise context where it's like, oh, this is like three-year long-term contracts and all these things.”
25:08

AI Companies and Circular Revenue

Debate the implications of AI companies' circular revenue models and their impact on growth.

“And so you don't like when it just circles around.”
Show all 20 chapters
27:32

Open Source vs Closed Source in AI

Examine the arguments for and against open and closed source models in AI development.

“How do you feel about the closed versus open source stuff?”
28:00

Open vs Closed Source: The Debate Begins

Explore the contrasting arguments for open and closed source AI systems.

“and the price per token is very different.”
29:00

The Case for Closed Source AI

Understand the economic arguments favoring closed source AI models.

“open source API inference provider like a fireworks using Kimi K3 or any of these other models, US or Chinese open models.”
31:00

Arguments for Open Source AI

Dive into the advantages of open source AI systems and their economic implications.

“and co-design sort of their serving infrastructure with the actual models that they're training.”
36:20

The Future of AI: Synthesis and Market Dynamics

Discuss the potential market dynamics between open and closed source AI.

“Yeah, but then, you know, it's like the nice thing is we know the answer.”
40:40

Cyber Risks in the Age of AI

Examine the emerging cyber risks associated with AI technology.

“Like if it had been actually hacked from an external force.”
42:00

The Rising Threat of Cyber Attacks

Explore the evolving nature and personal impact of cyber attacks.

“It's a scary, you know, first exposure of real risk, I would say.”
44:25

Understanding Obliterated Models

Learn about the concept of obliterated models and their implications.

“And so obliterated models are literally open weight models that have no safety guards.”
46:55

The Slow Transition to Enhanced Security

Discuss the timeline and challenges of improving cybersecurity measures.

“the the hope i have is that you know like we were talking about you know in the cloud era it's like People bought Workday and Salesforce on these really elongated timelines.”
49:24

The Promise of Bio and Robotics

Examine the potential impact of advancements in bio and robotics fields.

“But a lot of that was because of just like infinite data.”
A man walking through an airport, asking his phone a question

Ask about this episode

Get the gist, explore an idea or ask about something said.

Claude

On Uncapped | Benchmark Breakfast, what are the key points of this episode?

Searching your listening

Transcript

Automatic transcript. May contain errors.

0:00I think I could joke about being just like a little bit bored of. It's like I interviewed every time.

0:05Ev Randle: Just shit on all the VCs you've interviewed. We could joke about that, though. Yeah. They're great. But I asked them all the questions and, you know. Learned all the lessons. Maybe we'd come back. But it was time for an update.

0:20All right. Well, this is the beginning of a new iteration of the pod. Where I guess, you know, at some point, I mean, I've learned a ton. But at some point you've interviewed a lot of VCs and you say, it's time to just talk to a VC that you talk to 10 times a day instead. But just changing it up a little bit. So now we're going to do this thing, Benchmark Breakfast, we're calling it, where we're kind of going to, I guess, I don't know, mock a little bit of the time that we spend on Mondays. And I guess what we do on Mondays at Benchmark has basically been no agenda. we get together as you know the five of us and we sit around talking about whatever came up that week or whatever we're thinking about and evan i were just saying well that would be a that'd be a fun new version and if i'm losing if i'm losing steam on the current you know the current podcast like let's let's take a new rep so i'm excited to be doing this 100 and i feel like whenever we tell

1:12Ev Randle: anyone um about our monday morning or just our monday partner meetings and the fact that they're typically five to six hours um the two i mean the two responses we get is like that is absolutely insane. It's a little long. It's like, you know, I, you know, people will be like, I can barely stand our hour long partner meeting. And then the second question is like, well, what's the agenda? And I think that's a really important thing about our Mondays. You review the pipeline. Yeah, yeah, yeah. Their agenda is like, you review a very strict list. And, you know, it feels like you're like a corporate vice president at Microsoft or something.

1:42Ev Randle: But our answer is always there is no agenda. We just think about, you know, it's a continuous conversation, as I think our partner, Eric says from one week to the next. And I think, you know, we thought it'd be fun to like, you know, open source or make public like the form factor of that conversation. So one thing we talked about, obviously, is big news is Anthropik dropped an S1, which is, you know, pretty critical to the industry. So as somebody who can actually read an S1 of our mostly early stage group, what was your reaction? Yeah, it was amazing timing because I feel like leading up to the S1 day, there was a lot of market turmoil.

2:24Ev Randle: There was a lot of carnage in the public AI trade, a lot of fun about Anthropic. Like, are they going to IPO? Are they going to delay it? And then they come out with this S1 and on its face, this is everything that you want from a technology business. And it's funny because most of the numbers in the S1 had been leaked at one point or another. Like there wasn't that much new information. You mostly could have stitched it together from, you know, if you believe the right things on Twitter. As you always say, Twitter is making us all dumber. But it was all mostly there. But it was like, you know, if you really were good at parsing through, it was all basically there.

2:57Ev Randle: But I think it was another thing entirely to have all the numbers put together in one place. And again, it was like on its face. What you want out of a technology business typically is you want immense growth and you want operating leverage. Operating leverage is the sense that you have this big upfront investment in R &D or sometimes with software businesses, you pay a ton for salespeople up front and they take some time to pay themselves back. But you invest all this money up front, you have fixed costs, and then your revenue grows so much and your incremental margins are so good that you become very profitable very quickly as you scale.

3:33Ev Randle: And we don't have to go through every number in stat. Obviously, it's all over X. It's all over the internet. But I think the most important visual from the S1, in my opinion, there's a chart that I'm looking at that you guys can't see. No, no, no. We'll put it up. We have the power of technology. That's true. We have the power of maybe even cloud. Maybe we'll use AI. But the chart is showing the relationship between their revenue and their total compute expenses per quarter. And so it's both the compute expenses for serving customers, but then also training. And in Q1 of 2025, Anthropic had$400 million-ish.

4:08Ev Randle: They do it in billions, and so it's only 0.4. So it's like$400 million-ish of revenue against$900 million of compute and infra expenses. And so that's like a 0.4x ratio where compute is larger than revenue. fast forward six quarters in q2 of 26 they had 11.6 billion of revenue against 5.6 billion of compute and infra expenses so a 2.1x ratio of revenue to compute it's like that like the entire concern that everyone had is like oh are the labs are all these ai companies you know selling a dollar for 90 cents yeah um you know or 50 cents like are they ever going to be able to turn a profit. Anthropic in six quarters did what, you know, it takes a lot of software and technology companies like a decade to prove out.

4:56Ev Randle: Yeah. Which is like, no, these are money printing machines. They're money printing machines. In a huge way. Yeah. And it's like the revenue just turned out that it's big enough to make up for all these fixed costs of training. And the other surprising thing, and you know, this has played out throughout inference and all sorts of areas of the AI stack, but the the price per gigawatt has just surprised everybody like i don't think anybody like the monetization yeah yeah yeah and just like what the value of the frontier tokens are that's just has just worked has just made it all work it's made it all absolutely work and you know i don't think it like i think there was a lot of contention around the quality of the frontier lab business model that was somewhat that was partially answered by the numbers in the s1 like you cannot say that Anthropics business model isn't working on a per unit basis or an aggregate basis.

5:45Ev Randle: They had 8 % gap operating margins in Q2 2026. So it's clearly working. I don't necessarily think that means that it's going to be a slam dunk out of the gates in terms of how public investors and the public markets value it. And if there's a ton of share price appreciation right out of the IPO, there's two reasons for that. The first is that there's still some battleground topics that people have uncertainty over that I think are somewhat fair. So the first is, you know, Anthropic in Q2 grew massively. They grew 145 % quarter over quarter. So it's like insane sequential growth, but then they didn't grow as much in Q3.

6:22Ev Randle: And there's a lot of, you know, pretty good reasons for this. But they grew 45 % quarter over quarter, which is still insane. But you know, it's like the deceleration from 145 % growth quarter over quarter to 45 % growth. people are like, wait a minute, like this doesn't, you know, with SaaS with like, if you go look at like the net new per quarter of like CrowdStrike or Datadog, it's just like consistent upward trajectory. There's very little volatility to the growth rate. And then 80 % of the revenue is consumption based. And so there's not like these, you know, if you had like, you know, Snowflake or, you know, Databricks or one of these, it's like a lot of that revenue is like multi-year long-term contracts.

7:00Ev Randle: You have more like fixed visibility into the future growth. One of the things that's been surprising to me, which I would not have guessed it, at least at this point, because it's like the last generation of models is pretty good now. But it still just seems that so much turns on who has the very best model at any given point in time. And the leading edge, even if it's by a little bit, and even if the last model could do most of the tasks, it just seems like it's worth a crazy amount of momentum. And we saw this go back and forth between the labs all year. and we saw it again sort of in the last month and then you saw it over the summer.

7:34That part is surprising to me that people will always seem to go for the highest intelligence even with some price insensitivity. You know what I think it probably is?

7:46Ev Randle: There's just so much surplus. And what I mean by that is whether you're paying$0 or$200 for like a Claude Max account or you're paying like next to nothing with, you know, some free service, the value that you can get as an individual is probably in like, at least in the thousands of dollars, if not tens of thousands of dollars. This is like a key thraboyism that there's like, you know, two task types. There's the ones where it like just needs to be completed and it needs to be right. Yeah. And you don't get any more bonus points for getting any better past a certain point. And like, you know, that's often in a company context talked about, you know, with regard to like what type of executive you should hire.

8:24Are you protecting downside or are you looking for uncapped upside? And like filing taxes, for example, you kind of either did it right or you didn't. At some point, I just have to believe those types of tasks will get saturated. But then if you think about writing code or doing sales or discovering drugs and biology or robotics, it's probably the majority of tasks people understand it as unbounded upside. And in an unbounded upside world, if that's the mindset, you're probably always going to buy the better thing. I mean, I do still think it's obvious that the sellers of tokens are doing really well.

9:03I still think that I believe it'll get there. But I don't think that in all cases, the buyers of tokens are proving to be that much more productive in all cases. There's still a lot of hope going on. Yeah. And, you know, obviously code got from, you know, it was actually a year ago. I think it was not being productive and people were buying a ton of it. Now I think it looks like it is totally productive. Yeah. But I still think we have a way to go until like, you know, Purina Farms can buy AI and we're sure that it's a better business now. And I don't know like how that's going to happen. You don't know about Purina Farms?

9:36I don't know about Purina Farms. I'm just guessing.

9:38Ev Randle: Chick-fil-A has an amazing engineering team. Yeah. And you think they're supercharged by AI? I bet they are. Everyone loves the Chick-fil-A. I don't know. A little robotics might help them to make it. Yeah. But so I think there's a little diffusion that needs to happen on the consumption of token side. It's 100 % true. And I think so then, so you have like weirdness and growth is I think one reason why people will be cautious about the stock potentially. There's debates. I think this is a core debate, which is how much cogen and related TAM. So like for cogen specifically, how long does the fusion take?

10:10Ev Randle: And then how much just cogen TAM is left? Yeah. And then what is the next thing? I mean, how do you think about the cogen, Tam? Like, you know, the analogy could be something like in Zerp, you know, time when everybody went remote. It was like Zoom got this absurd pull forward. And it was basically like everybody who wanted a Zoom account got it. And so Zoom grew ridiculously. And then it had maybe not full saturation, but like roughly it got saturated. and obviously at least in our corner of the world everybody who might want to use ai programming has been using ai programming for a while yeah because then you have market share trades you have those kinds of other things but you don't get 10x year over year from that forever yeah but then you probably have a lot of other places in the economy where it's not even close yeah so i don't know how you think about like the the code tam yeah that's it's funny because like i think of back to your saturation point of like things that are saturated with like a zoom call there's a pretty easy saturation point that's like can i get on do my headphones correct correctly but you could always have another coding agent you could always have it's like you you have like 30 agents going out once probably right all the time yeah all the time building stuff just building stuff yeah just just figuring out how to be mean to you in our in our partner meeting in chat but with the um but with zoom yeah it's like with zoom it's like there's a very clear ceiling and asymptote and i think the really interesting thing that we've seen with code is like i think there probably is some asymptote but the asymptote isn't like a crazy place and what i mean by that is like some of our newer companies like the companies that have been started kind of post opus 4 5 i'd say the way that they build software yes looks almost nothing it's like nothing it's nothing like how you built software in 2019 i met up with one of our founders that we just seated and And there's one other engineer on the team right now.

12:02But we went for a walk. And he's doing some stuff on his phone. What are you doing? He's like, oh, I have a few agents running. And it's going to be done by the time we're done and all this stuff. And I'm just like, it's just a different way to do it now.

12:16Ev Randle: And just like the idea of a software factory. And it's like a new organizational structure for building software products. And a lot of it is not even human driven anymore. So basically within Silicon Valley, there could still be many X to go. Yeah. And then I think it's like if you think that like, okay, you either have to adopt that method of building software products or get left in the dust. You assume that on some timeline, the rest of the economy starts to at least look more like that than not. And that's where I think that like I'm relatively – I don't know if you can go have like the 10X growth that you had again or something.

12:53Ev Randle: But in terms of the durability of GoGen, I think that you could continue to see like – Double every year for a while. It's insane. 100%. And I think there's enough to do that. I mean, the other thing around that is all of the startups and markets around code, the right answer has basically just been bet on all of it. And it's been funny because in a lot of these sort of, let's say, code-related spaces, you're like, ah, this can't possibly be that big, or there's three of them, or whatever. They just seem to keep going. And it seems like the market pull on some of the code adjacent areas is ridiculous.

13:29Ev Randle: Yeah. Which is one of the things we talked about earlier this week. 100%. It's funny. So like a lot of people, I think, know about the sort of like AI bottleneck trade that happened in the public markets. A lot of people probably participated in some version of that. But basically like after ChatGPT came out in November of 2022, there was like, you know, almost immediately AI demand first for ChatGPT and then other products. started skyrocketing and people were like, okay, well, LLM training and inference runs on GPU, so let's buy NVIDIA. NVIDIA goes up a ton. But then people were like, shit, what else goes into an AI data center?

14:03Ev Randle: What else will be impacted? What derivatives of GPUs are impacted? And so I have a list of these things that went through. So like 2023 servers and racks, Supermicro went up 10X, Dell went up a ton, cooling equipment, so a company called Vertiv went up a ton, 23 and 24, Arista basically doubled in both 23 and 24. That's crazy. In 25, power generation. Vistro went up 3.5x. Oclo in nuclear. People were pulling forward the future of nuclear. Bloom Energy went up 4x in 2025. Data center equipment. GE Vernova went up 2x in 25. And then 25 and 26, obviously, we've seen memory and storage. So Sandisk, Western Digital, Micron, Seagate all went up at least 3x, I think, in 2025.

14:48Ev Randle: and then now in 26 now the agents are now using more cpus again you have intel up 120 percent this year so it was like it's like falling an egg through a snake it's just like it's like a new derivative like every like every every every year or like there's like three new derivatives every year that you could have gotten like a multi-bagger return on in the public markets leopold spoke of this you know it's just like the profit the profit but then there was like so then in venture the version of this that happened is sort of like what what i think you could call like the code gen pmf derivative yeah so like obviously like in so so chat gbd comes out in november of 2022 but really this really kicks off in november of 25 when opus 45 comes out and supercharges cloud code that's right and that creates this mega product market fit moment for code gen and then like right after that you had like basically um and again this is sort of a parallel all these public categories where you had vibe coding infra yeah so you had like back end as a service and databases super base took off click house and our portfolios continue to accelerate like crazy um uh deployment hosts so like bursell um railway cloudflare um the inference clouds obviously we're large investors in fireworks which is growing like crazy base 10 modal data companies like mercore in the benchmark portfolio yeah code sandbox companies like modal and e2b are growing like crazy web search exa also benchmark portfolio company code review I'm glad you're getting all the plugs in as we go.

16:12Ev Randle: I know. I'm just kidding. But it's crazy. It's like if you have 10 ,000 X more code in the world, all of a sudden all of these derivatives of code generation also need to grow that much. And then obviously you have a very exciting – I know it's not public yet, but you have a very exciting investment in the CICD GitHub action space. Because, again, durable execution with Temporal, browser use APIs, and now supply chain security with folks like Chain Guard and Socket. Like all of these categories, every single one I just mentioned, have companies in them that are doing like the 10 to 100 to 300.

16:47Ev Randle: That's crazy. You start, you know, one year at 10, you go to 100, and then you go to like 300 or more. Which never used to happen. Never used to happen. And multiple are doing it in the same category at the same time. Yes. This is what was wild. I mean, you know, it's like Lovable and Replit, for example, which I've obviously done great. But it's like everyone's like, isn't this, you know, isn't this club code? Isn't this Codex? Yeah, it is. And they're going to do a version of it that is form-fitted for a different purpose. And the whole thing is so big that it's just all going to work. That's right.

17:18Which kind of goes to the Eric point of it's all going to work. And I think probably the modification on that is in some of these categories where the technology hits that hard, it's just so easy to underestimate it. Yeah, it really is.

17:34Ev Randle: I think like AI code generation, you can sort of abstract most of the value generation in AI today to code gen. It's probably the most impactful product market fit moment in the history of technology. Probably right. I'm trying to think what else would even... I mean, it's on the order of iPhone moment or something like that. That's right. Except for because so much of it is in the digital... Because the infrastructure was already there... It can just go so much wider, so much faster. It can go so much faster. Whereas like Uber and DoorDash and, you know, these companies that were built on the mobile internet that were in iPhones.

18:09Ev Randle: Yes. They were limited by the growth because they were. The growth of mobile phones that had internet. Yeah. Yeah. It's fascinating. Yeah. And, you know, it's funny like for like knock on effects of this. I think part of what's weird for founders and VCs is when you've got a thread like this that's pulling so hard, not just on code itself, but like you're saying, all of the like accoutrements around it that are supported. it and come before it and come after it and are just kind of all getting sort of yanked upwards it's like other categories that are very good it's like when the opportunity cost is people like well i can get this ridiculous ride over here right that has a big effect huge impact and it's weird because at once it's making it so that venture demand in these spaces is so obscene and And then LPs are obviously doing great and they're funding it.

19:01And it's all part of like one rational circus. But then you have these other categories and a founder is like, well, what the hell? I have a good business and it's hard to raise and all that thing. So it creates a weirdness.

19:16Ev Randle: Historically, I've spent a fair amount of time in growth investing. And as a growth investor, you keep all these historical precedents of the growth trajectories of all these amazing previous companies. and I'll go back and look and I think like data I mean this might be wrong but I'm pretty sure data dog went like 5 to 18 yeah and like 5 to 18 to like maybe 40 to 50 or something it's like now 5 to 18 it's like I don't know if you can raise I don't know if there's capital would you I don't know like would you look at a 5 to 18 or what would it take for you you would need you would need in this in this environment you would need to believe an 18 to 100 story for the next year or you would just need to think that they were doing something that was so difficult and unique that it just had insane growth durability forever after that but on the surface that used to be enough where everybody would be super interested if all you heard was it's a software company like 5 to 18.

20:05yeah whereas now if that's all you heard you'd be like well why wasn't it 5 to 40. that's

20:10Ev Randle: right you know which is insane it's so insane but it's all because of this code pull that's right 100 which you know and like a related topic that we've been talking about is that there's sort of two archetypes or there's almost like a continuum of like how levered you are to the trend and then do you have like the old hallmarks of like durability enterprise stickiness yeah um like some of the things that we loved about you know the the scions of the last era like data dog and crowd strike and you know even databricks even though they're not public um there seems to be some form of trade-off between like how levered you are to the AI, like basically the cogen pull or just like the AI trade in general.

20:57Ev Randle: And some of the like more traditional like signs of durability and defensibility from like an enterprise context where it's like, oh, this is like three-year long-term contracts and all these things. What do you think about sort of that trade-off and if you're attracted to one or the other or anything that you think about that? Well, this might be wrong. But for me, I'm in a mindset right now where durability is really hard to call. Like I think in the last era by let's say the late 20 teens, I think you had a good enough feel for cloud and enterprise at that point that you could not perfectly, but you could kind of name it when you're like, that's durable or that's going to get competed.

21:43That's sticky. That's got a high switching cost. Like we got to a place where I think you could make those calls somewhat correctly because you had too many examples that looked like it and you just knew how software worked. You also had situations where enterprises would look at software for years before they were buying and like, I know I'm going to buy Workday, but I'm not going to buy it yet. But when I do, I'm going to use it forever. You just had dynamics like that where right now, I don't know that durability is nearly as easy to identify. You know, it's like even take application software.

22:19I don't know, six months ago, everybody was kind of like, it's dead. At the moment, I think people are like, it's alive and it's good. And why should the labs do it? And there's all these edge cases to do. And they can actually be value-add to the intelligence tokens. And there's all these reasons to put your data in there and blah, blah, blah. and you need a UI. And, you know, that's great. And six months from now, people might flip again and be like, never mind, code got so good that that's all not true anymore and all these things. So calling the durability seems difficult to me. I think, you know, I was thinking about like, what is it about the 5 to 18 stuff?

22:58I think if you are trying to identify the next code, the next thing that's going to pull like crazy, I think that's probably the reason that people get really attracted. You know, like as an example, I think we all believe robotics will work. Yeah. We would all probably say it's not exactly working yet. We can't have a robot in here doing, you know, a bunch of things that, you know, a human could do. You can't really have one at home just yet. But I think we all believe we'll have one. Like you and I both love our Teslas. You like see the beginnings of it. And, you know, you go to these, you go to cool companies.

23:32Like if you visit like Pi or something like that, you're like, you know, You can see the future. Yeah. So it's going to come. And so you can bet ahead on things like that. Bio is another one where it seems like, you know, the smartest people at the lab seem to be saying it's going to work. There seems to be early indications that, you know, target identification and other things are possible that didn't used to be. And so I think that to me is more the spirit of investing at the moment is either on the trend or you're calling the next one. And obviously that returns are highest if you call the next one.

24:03Ev Randle: 100%. You're calling. Yeah. I feel like that is, it's funny, like as soon as you can sort of put into chat, you know, like, you know, for obviously the segment of like, what are all the derivative things? I went to chat. I'm like, well, we've seen these work. What other trends have worked? As soon as like chat knows the derivative trade. Everyone's talking about it. Everyone's talking about it. And the companies that you could invest in are probably already at like a billion plus valuation or something like that. Yes. And you can still make that call because you can believe that the market will still somehow just not appreciate how far it will go.

24:35Ev Randle: Yeah. A hundred percent. A hundred percent. And that's fine. It's super fine. Yeah. I think, I think on the, um, on some of the, the, like a final thing on the difference or the comparison between, um, even both categories that are growing and like, what is healthy growth versus, you know, riskier growth. I think that there's like two sort of, um, conflicting views. One is that like, well, if you want to see like a really healthy, like a sign of health for the AI ecosystem would be money that is not in the system coming into the system. And so you don't like when it just circles around. Yeah. So, well, I think there's two arguments, but I think the argument here is like, you know, okay, circular revenue.

25:16Ev Randle: And just like, if you're, if the only people buying AI stuff or AI companies, that is like, you know, it's just, it's more and more leverage. Like that's more and more leverage on the system because if it doesn't work out, then it all comes crashing down. Whereas like if large banks or large healthcare systems in like ohio and new york and florida are getting immense value in buying a ton of ai that is really healthy because that's that is like an independent non-ai oriented end market and it goes back to those are buyers of tokens and those are buyers of tokens which is ultimately what needs to happen here yes spending a lot of money to make these tokens yes they better be useful they better be useful and and like i think ai companies are much more willing to be risking their token spending or token buying relative to like a big healthcare system yes with you know eight percent profit margins they're not going to blow up their pnl well again you know if you're if you are an ai application company buying tokens yeah and then doing something to them and then giving them to somebody else somebody still needs to make people's better like at the end of the day somebody needs to turn the token into a better life in a way that you couldn't have done without the token or at least not as well that's right yeah but then the other so the other side of the argument that i think also has a lot of merit is like but look like the AI companies that are, you know, also buying a lot of these tokens, creating software factories, they're going to become the massive companies of the future because they're going to outcompete all the companies not doing this.

26:36Yeah.

26:36Ev Randle: And so like, this is like, when you look at Stripe, like Stripe didn't get, now they work with a lot of old kind of economy companies, but they didn't get big because they worked with old economy companies. They got big because they worked with DoorDash and Shopify, Instacart, Lyft and Lyft and like these and Uber, you know, a lot of the sort of like YC era internet native companies that got massive. And like, that's what, and then once they were massive, then there, it was much easier to attract old economy companies because they were so like, they're like, Hey, we just helped DoorDash, you know, kick the hell out of you.

27:07Ev Randle: It's like, you should probably use this as well. And so I think that like, I have no problem. I've, I don't really have any problem with AI companies buying AI, like, you know, AI tokens, like the circularity, because I'm like, look, like these are companies, they're just at the frontier. And just like internet native companies were at the frontier of the internet and older economy companies didn't really buy into what they were doing until it was too late and those companies were very large, I think the same thing's going to happen in AI. Yeah, that's probably right. Yeah. How do you feel about the closed versus open source stuff?

27:37We've talked about this a bunch.

27:38Ev Randle: Yeah. And obviously, we are exposed to a lot of the open source world through fireworks and other things. what we've seen in general I think has been a rise in open source this year. At the same time obviously the frontier labs have gotten massive and the price per token is very different. The use cases are similar but different. So anyway, what's your latest on open and closed stuff? Yeah, if you'll entertain me I feel like we probably have a similar synthesis of what we feel now. But I think what would be interesting to do is like maybe we steel man each side of the argument. Great. And then like that, I think once you have the steel man for each argument, because I think there's a sort of like a lot of like, there's a lot of talk online about open source versus closed source, which I think is sort of off track.

28:33Ev Randle: Right. From like the reality on the ground. So I think if you steel man both sides, you can actually understand like, okay, well, what's the synthesis that probably blends the two? Great. Steel man. So like I'll do closed source, you do open source. Yeah, you go first. So I think, so the steel man of the closed source argument, and this is, and what we mean by this is like, you know, why would a company, you know, get inference from a closed source frontier API rather than from an open source API inference provider like a fireworks using Kimi K3 or any of these other models, US or Chinese open models.

29:07Ev Randle: And people say, well, look, intelligence is fungible or it's like sort of a commodity. And if we take that seriously, that idea that like, you know, a unit of intelligence is this fungible thing, there's a lot of history and a lot of good literature about who wins those types of markets. And when you're optimizing for the cost of a fungible output. And it's oftentimes the winner is something we call a low cost producer. And so the defining feature of one of these low cost producers is that you are structurally as a company, you have a lower cost to deliver a comparable outcome. And so you can either charge lower margins or you can charge lower prices at the same margin as competitors or you can earn higher margins than your competitors at the same price.

29:50Ev Randle: And so two famous examples in tech. SpaceX is probably the easiest one to understand. SpaceX has the lowest cost to them in order to launch a kilogram into space in the world. No one can launch a kilogram into space as cheaply as SpaceX can. And that gives them a pseudo monopoly on launch, at least at the same level, unit level of capability. A previous one in computing is AWS. They had massive aggregated demand that created economies of scale. They're able to amortize a lot of investments over a massive customer base, which gave them lower cost relative to competitors in the CPU cloud business.

30:29Ev Randle: So for AI, the equivalent here would be how do you become the low cost producer of a unit of intelligence? So we could say this is like a given level of capability, latency, context length and reliability for some task. Like who can produce that task at a given level of capability, the lowest reliably? And I think proponents of the closed source argument would say the leading frontier labs have the best models in the world, many of which are not released to the public or not released via API. And they're using those models and pointing them at their infrastructure to improve their serving efficiency and co-design sort of their serving infrastructure with the actual models that they're training.

31:11Ev Randle: They have the most aggregated demand. They're the largest. And so they have economies of scale. Like the data center space that they're buying is cheaper for them on a per unit basis than anyone else because they're the largest purchasers of compute. and if all of these are true if you have economies of scale if you have like if we are in some weak form of rsi right now where the models are actually designing like you know actual inference efficiencies at any sort of rate that that is compounding that means that you can have a low cost producer of intelligence in form of in the form of an ai lab and that means that this argument that like open source equals cheaper is a misnomer and is not necessarily true because Because at the same level of margin, if this was true, a frontier lab at any point of the intelligence curve, like the capability curve, would be cheaper than an inference provider serving on an open source system.

32:04Ev Randle: So that's sort of the economic argument for closed source. And then I'll let you take open source. Okay. So, I mean, I'll start with the economic piece. And obviously, as you know, I mostly buy it. But economically, one difference is that because there's the model cost, and then you've got to do the inference on it. The model, in theory, has some training that's been amortized into the price. And so to the extent that the labs want to be profitable or have enough margin that they make up for it, they do need to do something to make up for the training costs. versus, you know, if you're using Kimi, you don't, you don't, the model's, the model is free.

32:44So there's no amortization there. And then obviously you still need to run the inference. And I probably, you know, it's probably right that, you know, on a per sort of inference unit basis, the labs are still going to have the advantage over the inference companies, but maybe in theory that's, that's made up. The other thing is, you know, do the models care to take their margins down to that level? And is that actually worth it? Or is that really expensive for them to do. And so, you know, the... Then they don't have as much money to train. They don't have as much money to train. And also it's like, you know, if they're really focused on the real frontier and they're like, you know what, use your open source for the taxes.

33:21Use it for recommending which, you know, iPhone you should buy and should you get the duo or not. And we're going to go solve biology. We're going to go do robotics. And it's just not worth it to go optimize two models ago to go do the basic stuff. And so there's basically some argument along the lines of we're going to get saturation on many tasks and as open source of the future is better than Opus 5.5 and Astra are today, that's going to be good enough for most tasks and the frontier labs are not going to care to go reduce their margins, optimized two miles ago. They just need to keep their eyes on the frontier.

34:04And so maybe there will be some space for that. But what I would say probably is more than the money pieces. And I wish I actually had, because to your point, the specific numbers on the margin are where it all turns on. And I don't have those on my hand. But I think more important is, do you want to own your own intelligence? And this is like the Alex Karp stuff of just like, you cannot trust as an enterprise you cannot trust any company to own all of your data yeah and to be fully reliant on any company and you need to own your own models you need to own your intelligence and i think there is an argument to be made there and i think for certain industries that trade is going to be worth it yeah and particularly as the open source models get good enough even if you're not saving any money you're in control and that's about something and then even like think about the

34:53Ev Randle: branding of like if you let's say you're all all your all your tokens are anthropic tokens and all of a sudden anthropic enters your space yeah and then your customers are like well yes you know like what model are you using yeah you're like 100 of our models that we're using is anthropical like why don't we just go to anthropic yeah if you're eli lily no special sauce if you're eli lily and all of your data lives in anthropic yeah and then anthropic's like hey by the way Like, BTW, we got some drugs coming. You're like, okay, well, I don't feel so good now. So I think there are some reasons that people won't want to do that.

35:25And then I think there's situations where you can post-train on open source, and that can be useful. And if you've got a lot of data, that data can outperform the extra intelligence that you might get from the next model. And so you'll see some of that too.

35:38Ev Randle: I feel like we hear that pretty consistently from founders that are using, that don't necessarily, aren't even selling an AI product, but just have a big product corpus and they're using AI throughout their products. where they're like, look, you can really like, if you really spend the resources to fine tune and you have really good trace data from the tasks that you're trying to do, like, yes, like maybe vanilla can be K3 or whatever isn't quite as good as whatever the model, but like you can actually get a really small model, like a really, really small open source model. And if you fine tune the hell out of it, like it gets to like, you know, a K3 or a Sonnet level performance and it is dirt cheap.

36:14Ev Randle: Yeah, it is dirt cheap. I think, yeah, we had a large public CEO come in and talk to us about that. It's like there's clearly all these use cases where that's true. Yeah, but then, you know, it's like the nice thing is we know the answer. The market's telling us the answer. And the answer is that the inference companies in combination are doing mid-single digits worth of total revenue. Yeah. That's a reasonable proxy for open source because that's where it's running. and that's whatever percent of the labs companies. And you can argue for that there'll be some mix shift but what we're basically seeing is that open source might be 5, it might be 12, it'll be some percentage like that unless there's some change of the overall spend and that's what'll be there.

36:58Ev Randle: So I think the synthesis between these two is that, again, shout out Eric Vishria, it all works. But also the way it all works is they're probably not even rivalrous to each other. I don't know if this would be the case, but also I'm sure the Frontier Labs, they're probably even happy about there being an open source ecosystem because it probably creates all, from an antitrust perspective, if it's like, oh, they're getting too powerful, it's like, no, there's a healthy open source ecosystem where people can also use. I think that's probably also healthy. I have no idea if that's an actual thing that they think about.

37:31Ev Randle: But right now, this will be excluding the hyperscalers. I'm not sure they think about it, but yeah. Yeah. But if you look at, so like right now, if you look at the, we'll exclude the hyperscalers because we have no idea how much open source inference they're doing. But if you look at, you know, just the inference clouds, as we've talked about, and then you look at just OpenAI and Anthropic. So this is very imperfect, but it gives us some, some okay proxy. By the end of this year, I think the main, the top three inference, three or four inference clouds will probably do like 6 billion of run rate-ish.

38:02Ev Randle: And then, you know, let's say OpenAI and Anthropic will do like$170 billion of run rate-ish if their growth kind of extrapolates from public disclosures. So that's like 3-ish percent thus far between the inference clouds and the top two frontier labs. If the growth continues - Yeah, and maybe if you add the clouds in, maybe it's double at most or something like that. That's right. So we can say like 5 % to 10 % in terms of revenue for open source versus closed. But if you have the growth that the labs have had and that people, I think, expect, Again, this isn't a guarantee, so this isn't a projection.

38:35Ev Randle: But if you do have the growth for three to four years on the frontier side, and the labs combined are at a trillion, and open source takes 10 % of that, that is like$100 billion of revenue. And the inference market has consolidated. There's really only three players. That's right. We have Fireworks. We have Base 10. We have Modal. And it's like, I mean, there's other ones that are good and all that stuff. But they really are the three champions of the market. And I think we're going to see more consolidation to those three than less. so you have like probably two or three companies at that time splitting 100 billion dollars of revenue which is incredible i mean it's incredible i mean it means one of those companies could be worth 100 billion or more probably both probably you know you have like a 40 billion dollar company if like even if it's 50 gross margin you know it's probably 200 billion dollar company yeah you have like hundreds of billions of enterprise value for open source and then you have the trillions in frontier labs i'm like everyone is very happy with that like that's that's huge like humongous companies all around.

39:30Ev Randle: And it's through the synthesis of like, open source will always on a revenue basis be some percentage of Frontier. And whatever that percentage is, there's so much room for the inference platforms to grow. By the way, and then we can move on from this, but it's like a couple years ago, I feel like it was, there was a worry of like, man, it seems like AI is either going to wake up the Terminator sand god and everything's going to be really bad. It's like just like not going to matter that much crypto all over again head fake and at the moment and you know this is why i'm just i'm happy to see the market reacting positively to anthropic it's like at the moment like unless you know rsi takes us into a weird place like we might actually actually be splitting the middle landing where like the labs are extremely important but they are not 60 trillion dollar companies where it is not useless technology that doesn't help anybody like that that that's a nice thing seems like the happy path it seems like that i feel like my odd my my p happy path is higher than it used to be and that's awesome amazing which is amazing that's amazing it's amazing um do you want to talk about um bio obviously you've you've got i feel like you've been really interested in bio yeah i have been i mean the i was gonna go to the dark place of cyber we can do cyber let's do cyber let's do cyber because like that would be that is one of the scary things and this this last week you know there was like the php thing and people are very afraid and understandably so it does feel like we are in a point in time where one of the risks that is most imminently being expressed is a cyber risk yeah where you know we we've spoken to some like prominent CEOs in the last few weeks who have told us that like they've pointed frontier models inward and said hack us and the model's like I'm gonna hack the shit out of you and it does And they said that day after day, they were getting company-ending vulnerabilities surfaced.

41:30Ev Randle: Yes. Like if it had been actually hacked from an external force. And these are great companies with great founders. Wonderful companies. And with great engineering teams. Like tech-forward companies. These aren't like old companies. And obviously, when you attack internally, you have crazy advantages that outside attackers don't have. You've got all the data. You've got the code base. It's just a different thing. But still, it seems like, you know, fear could be on the horizon and then, you know, you see all this stuff. So obviously, it's probably it's a good time to be a security company. Yeah.

42:02But it is also a scary. It's a scary, you know, first exposure of real risk, I would say.

42:07Ev Randle: 100%. Yeah. Yeah, and I think that the really scary thing is that cyber attacks, at least for the public, I feel like cyber attacks historically have been sort of an abstract thing. Where like at worst, your data was like your password for this website was hacked. And then you get like some notice in the mail and you like, you know, you change your password, but like nothing actually happens. Yeah. And recently, there's been a lot creepier and more things that feel really personally intrusive that are both now possible and have actually been happening. And so I pulled up a few examples where there was a researcher who used Claude to disable a webcam's recording indicator.

42:55Ev Randle: So started recording without like the light coming on and access the microphone and the firmware. That's scary. All those things like a French supermarket suffered an attack that manipulated refrigeration controls and actually spoiled food. So I think that the issue is that like now like the systems that are now that we thought were secure because these are open source systems that have existed for 15 years with humans looking to patch them and hack them in all sorts of ways. like these these critical systems to like firmware and hardware that's like tough overall internet like these are now getting exposed that that only have been able to be exposed via these ai models you can patch the software but can't really patch the hardware yeah and so it's like you know people are like oh like you know uh like update your smart tv and it's like what the fuck yeah my tv what's wrong with the tv like what like you know and they're like oh disconnect it from the internet yeah well that's really scary if you have a samsung frame smash it immediately Yeah, smash all of your smart devices in your home or disconnect from the internet.

43:54Ev Randle: So I feel like it just feels so much more widespread and intrusive than before, which is why it's now sort of like I definitely bought YubiKeys a couple of days ago. I want to go through the whole – actually, Andre Carpathia has a really good blog post on digital hygiene and just things you should do. And I feel like for the first time, given the parallelism and the scale that people can use and just like the extent to which these things that we're seeing as like strongholds like php yeah um are being hacked it's like okay this is this is for real yes um i think that the like really interesting thing so there's one other um piece of the cyber thing is this idea of obliterated models so an obliterated model is an open weight model that has its internal sort of safety guards removed and um to help help with its work because you know ultimately like when the hugging face incident happened, you know, they had some complaints where they're like, we had to use an open source model because we couldn't turn off, we couldn't get, you know, the frontier labs to turn off the safety guards in time to help us with the hack.

44:58Ev Randle: And so obliterated models are literally open weight models that have no safety guards. So it's very scary. And like, I think one of these is actually what some groups in China have been using, I think it's called Artex, have been using to, you know, commit some of these hacks globally. It's also an incredible defense tool for enterprises. Because you take an obliterated model, you do what that CEO told us, and you just say, hack us. Like, hack us to death. But, like, obviously, you're the one using it, and so nothing actually happens from it. And that is, I think, going to be an incredible tool, is you need these models that will...

45:33Ev Randle: Because guess what? The adversaries have obliterated models. The adversaries are not working with safety controls. You need to have that... You need to see the tools that they're using and use it on your own infrastructure. And I feel like we're going to hear a lot more about this idea of a Blu-rayed model. By the way, I think the tech world, which I think is mostly a good thing, at first I would say reacted negatively to pacing the frontier type stuff. And there's much more of just a let it rip mindset, which is very libertarian, tech forward good. I think that's basically a good thing. But this security dynamic is an instance where it not going too fast would be very helpful.

46:10Because you actually do want the defenses to penetrate before we have, you know, four generations forward of models. That's right. Because you do need things to get a little time to harden against this stuff.

46:21Ev Randle: It is. But this is the crazy thing is like, how long do we think it's going to take to harden all this stuff? It's going to take a long time. Like think about like, I don't know, think about like a hospital system in like West Virginia. Yeah. How long is it going to take for them to, because I think if you talk to some of these cybersecurity professionals, they're like, we basically have to re-architect. this is like the biggest code modernization project in all of history that's what we need cognition for that's right that's what we need that's what we need cognition for um you know but i mean i think like um i was going to try to do like an ad read there yeah cognition update your code yeah and security that's good but it's like um it's going to take a long time but i mean the the hope i have is that you know like we were talking about you know in the cloud era it's like People bought Workday and Salesforce on these really elongated timelines.

47:11They were slow. If you were talking to a government or a hospital or a law firm, it was like 36-month cycles on some of this stuff. That's not really how it's going right now on the purchasing of software side. I think people are pretty alert. So the hope would be that people see the security stuff for real. People are making it a top priority. And everybody reacts a lot faster than you think. I still don't think that that means it happens in three months. but maybe in 12 months people could do a lot. That'd be amazing.

47:40Ev Randle: I think people see the urgency. Yeah, hopefully. Maybe as a last topic, I think what's cool is the bio thing is interesting to me because I've kind of been of the view, I still am of the view, that all the software improvements that we're making are awesome but there is a ceiling to how much that can improve the world. And if you like touch some grass and you get out of San Francisco and you're like, what will actually make the world a better place? It's not that we're shipping software way faster and way more easily. It's things like people have better housing or people live longer or people live healthier or people have food or people have transportation.

48:23Like it's physical things. I just think there's a limit to what you can get when it's all sort of in the screen. And bio to me would be like one of the very best possible things if we're discovering drug targets and, you know, finding amazing new, you know, drugs quickly or at all. That would be, what are you laughing about? So anyway. Oh, you were on a good tirade. I was on a good tirade. So anyway, but I think that it is both really promising and it seems like it's on the horizon. Yeah. Same with robotics, by the way. Those are the two to me where you see the labs both talking about it, building teams around it, purchasing a huge amount of data around it.

49:07They're writing about it publicly. And so I think that that to me is likely an indicator that it's there. I mean, I haven't yet seen the labs be wrong when they say something's coming. That's right. Getting the timeline right is really difficult. With self-driving it took, now it's fully here, but it took a while. Yeah. And self-driving, Tesla's obviously dominant. Yeah. But a lot of that was because of just like infinite data. Right. And once you get enough data at these things, like maybe it just works.

49:32Ev Randle: I think that's and that's the there's this really weird sort of like inverse relationship between. I completely agree. I feel like you do you want like the you want the rest of America who doesn't have like an amazing, I think, view on AI, or at least a lot of people don't throughout the world and in America don't have an amazing view on AI. Like, do you want to change perception of AI? Like, solve disease, save people's lives, help people live longer, build more housing so housing prices are lower, like create physical abundance? That's what makes a better world. That's what makes a better world.

50:02Ev Randle: Yes. And the weird thing, though, is that like there's almost this inverse relationship with those things, the things that like when you actually touch the physical world could actually make the physical world a better place, and the things that have been able to get better very quickly from AI. And so if we take like, you know, like AI just tends to improve fastest when it can generate large amounts of trustworthy feedback cheaply. I think it's like a good way to extract it. So if you look at like code, you know, it's like this incredible area for this because it's extremely verifiable. Like, you know, you can do a unit test, you know if it works or not.

50:35Ev Randle: You can scale up experiments extremely cheaply. They're like infinitely scalable. It's all in the digital space. And so when you think about like cloud code or like how cursor improved its models very quickly, or even chat GPT, when they started getting feedback from users, all of these were in a digital realm where there was an API and customers could come on at the click of a button and start, you know, basically contributing into the Leviathan of training data for these labs. and in the real world robotics and now bio the tough thing is that like you you have real bottlenecks to that feedback loop like in robotics obviously like robots are kind of expensive like they're expensive to deploy expensive to maintain there isn't like a big customer base that can press a button i mean you could order a robot and you could start doing these things and pi has an api now and you could do the sunday beta program or any of these things but it isn't like chat gpt where it's like you can get 100 million people in a few months right via website yeah and so like that's just very different bio you also have natural rate limiters like the equivalent of like running a unit test in biology is a physical experiment yeah and like that you know like it could you know like and even that experiment might only answer like a small part of the question you actually care about and a lot of things do just take time they just take time like you know it's like i don't know physical experiment like even if you speed it up um you know making your own you have to parallelize that's what it is and i think so in bio it's like you can imagine a virtual sort of or like a real lab but where you're sort of you know kind of run somewhat autonomously or at massive scale yeah where you're just doing a ridiculous number of thoughtful you know trials at once 100 i think that could be like you know the time will hold you back but the parallelizing will still speed things up quite a lot yeah and but i think like this and this is i i know like you you have an investment that you recently did in in this general space for robotics um we've been very interested in things happening in bio as well but like the infrastructure to i mean we've obviously seen it with the ai labs as well like mercore as a business has grown to an unbelievable amount of scale yes because there's basically there's just no limit very unintuitively by the way everybody kind of thought that like the data thing would run out and they would buy the data and then they bought the data and now they can move on it's gone so like turns out it's just turned out to be wrong yeah reality has a surprising amount of detail as they say that's and there's just so many different things that these labs want to train on there's so many different data types that emerge.

52:50Ev Randle: And I think how much we've seen in the digital space, in the digital data space, we're going to see, I think you have to see more. Like Brendan, I don't know if he's the first one to come up with this, but Brendan Foodie at McCore says the internet was a one-time subsidy on the production of LLMs. And these physical spaces, they don't have as much of that one-time subsidy. They don't have the internet. They have some data. There's a lot of bio data. There's some robotics data, but like nowhere near. And you also need a ridiculous diversity of data, it turns out, to build these models well. 100%.

53:18So it's just not going to happen overnight.

53:20Ev Randle: But I think it's going to be like a civilization scale thing. Yes. Maybe economy scale, civilization scale is maybe a little dramatic. But to get these things where they need to be, I think it's going to be like Manhattan Project level scale. Yes. It'll be a bigger effort to make it actually happen. Yeah. But if you imagine what's really possible with good AI robotics, it's probably bigger than it's probably bigger than all the digital stuff so far if it actually works i can ask i mean even like if you think about the gdp impact of i mean maybe it could get scary because if you extend lifespans then you have like an even more gray population but we don't want to get too much in the macro yeah but like you know it's like you know not to get too personal with it but like again my dad got diagnosed with a very advanced form of cancer i look at like the survivor cohort rates yeah it was like you know eight percent by year two or whatever.

54:11Ev Randle: Yet immunotherapy drugs produced in the last 10 years have given him like a 10 plus year prognosis. Yes. Which would like, like, you know, if he got the disease 12 years ago, he's dead. Yeah. And now he's in remission. Right. And so like, like, and we are just at the start of probably a hundred to a thousand times that amount of things happening where it's like every single disease area, every single cancer, like, you know, we've solved everything else from AI. And when we solve bio through all the scale of investment we're going to need to make it's going to make probably a much much larger impact than any of these other things well i think that was enough for this one yeah this was lovely this is great what are you going to do for halloween with the family uh we're gonna we're gonna we're gonna go trick-or-treating in an extremely dense neighborhood well i'm gonna just try not to lose the kids that's great that's always my goal is it nsf in the on the peninsula on the peninsula yeah yeah yeah not which you're welcome to come to Well, Brett and I are going to be in San Francisco.

55:08Ev Randle: Why don't you guys come down? We might come down. We'll talk about it. And then do we get a duo review or is that next episode? The duo is unbelievable. It's the duo. Do you have the duo? Yeah. The duo is unbelievable. With the open. Wow. It's incredible. You liking it? So this is day two? This is, yeah, I'm 24 hours in. And you love it. I'm obsessed with it. Amazing. But next episode, let's see if I've returned it. Yeah, we'll see if I've returned it. Yeah. All right. I'll see you in a minute. Thanks, Jack. Goodbye. Bye. Bye.

From the publisher

Benchmark Breakfast is a recurring conversation between Benchmark partners Jack Altman and Ev Randle. It’s a window into the firm’s Monday no-agenda partner meetings, where five partners sit around and talk about whatever came up that week.

In this inaugural episode, Jack and Ev dig into the Anthropic S-1: what the revenue-to-compute chart actually shows, why growth volatility and consumption-based revenue complicate the public market story, and whether code gen is the most impactful PMF moment in the history of technology. They steelman open source vs. closed source, run the numbers on what the inference cloud market actually looks like, and get into AI’s emerging cybersecurity threat — including “abliterated models,” open-weight models with safeguards removed that adversaries are already using. They close on why bio and robotics may ultimately matter more than all the software gains combined, and what Manhattan Project-scale investment in the physical world could actually unlock.


TIMESTAMPS:

(0:00) Intro

(0:45) Introducing Benchmark Breakfast

(2:09) Anthropic S-1 drops

(3:59) The revenue vs. compute chart

(6:19) Growth volatility and consumption-based revenue

(7:11) The race for the best model

(10:21) Code gen saturation and software factories

(13:39) The AI-bottleneck trade

(17:39) Code gen as the most impactful PMF in history

(21:27) Durability is hard to call

(23:36) Investing in the trend vs. calling the next one

(26:09) Circular revenue vs. outside customers 

(28:15) Open source vs. closed source steelman

(34:25) Own your intelligence

(36:29) Open source as a % of frontier spend

(40:52) AI and cybersecurity

(44:35) Abliterated models

(48:08) Bio, robotics, and physical abundance

(51:52) Why physical feedback loops are slower

(54:12) Ev’s dad and the promise of bio

(55:13) Halloween plans and the iPhone Duo


LINKS:

https://x.com/jaltma

https://x.com/everettrandle

More from Uncapped | Benchmark Breakfast

All episodes