Ev Randle

1. Key Themes

Anthropic's S-1 Settled the "Are Labs Money-Printing Machines?" Debate

Ev Randle walks through the S-1's most important chart, revenue versus total compute expense (serving plus training), and shows the economics flipped in six quarters. The prior fear was that labs were selling a dollar of revenue for $0.90 of cost. Randle: "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." 00:04:23 This compares to Q1 2025: "400 million-ish of revenue against 900 million of compute and infra expenses. And so that's like a 0.4x ratio." 00:04:00 His conclusion: "Anthropic in six quarters did what... it takes a lot of software and technology companies like a decade to prove out, which is like, 'No, these are money-printing machines.'" 00:04:44 He also cites "8% operat- like GAAP operating margins in Q2 2026." 00:05:26 Jack Altman adds that the surprise was how valuable frontier tokens turned out to be: "what the value of the frontier tokens are has just made it all work." 00:05:20

Why Public Investors May Still Hesitate: Lumpy Growth and Consumption Revenue

Randle flags two legitimate objections to a smooth IPO pop. First, growth deceleration: "they grew 145% quarter-over-quarter... But then they didn't grow as much in Q3... they grew 45% quarter-over-quarter, which is still which is still insane." 00:06:11 Second, revenue quality compared with SaaS comps like CrowdStrike and Datadog: "80% of the revenue is consumption-based... a lot of that revenue was like multi-year, long-term contracts... You had more like fixed visibility into the future growth." 00:06:46 The core debate he identifies: "how long does diffusion take? And then how much is code gen TAM is left? And then what is the next thing?" 00:09:51

The Best Model Wins Because Most High-Value Tasks Have Unbounded Upside

Altman is surprised that buyers keep paying a premium for the single best model, even when the last generation can do most tasks. He explains it with a Keith Rabois framework: "there's like two task types. There's the ones where it like just needs to be completed and it needs to be right, and you don't get any more bonus points for getting any better past a certain point." 00:08:07 He contrasts these with "writing code or doing sales or discovering drugs in biology or robotics... it's probably the majority of tasks, people understand it as unbounded upside. And in an unbounded upside world... you're probably always going to buy the better thing." 00:08:36 Randle's complementary view is that consumer surplus is so large that price barely matters: "the value that you can get as an individual is probably in like the at least in the thousands of dollars, if not tens of thousands of dollars." 00:07:56 Altman caveats that on the buyer side, "there's still a lot of hope going on" 00:09:15 and token buyers outside tech still need to prove productivity.

The Code-Gen PMF Derivative Trade, Mirroring the Public "AI Bottleneck" Trade

Randle maps a public-market pattern onto venture. In public markets, after ChatGPT: "2023, servers and rack, Supermicro went up 10x... Vertiv... Arista basically doubled in both '23 and '24... Vistra went up 3.5x, Oklo... Bloom Energy went up 4x in 2025... GE Vernova went up 2x... SanDisk, Western Digital, Micron, Seagate, all went up at least 3x... Intel up 120% this year." 00:13:53 Altman's gloss: "It's like following an egg through a snake." 00:14:55 The venture version was kicked off when "in November of '25 when Opus 4.5 comes out and supercharges Claude Code. And that creates this mega-product market fit moment for code gen." 00:15:20 The resulting derivative categories: backend/databases (Supabase, ClickHouse), deployment hosts (Vercel, Railway, Cloudflare), inference clouds (Fireworks, BaseTen, Modal), data (Mercor), sandboxes (Modal, E2B), web search (Exa), CI/CD, durable execution (Temporal), browser-use APIs, and supply-chain security (Chainguard, Socket). Randle's summary: "if you have 10,000x more code in the world, all of a sudden like all of these derivatives of code generation also need to grow that much." 00:16:12

Growth Benchmarks Have Reset: 5-to-18 Is Now a Hard Raise

The code pull has recalibrated what counts as exciting. Randle recalls Datadog's early trajectory, "5 to 18... maybe 40 to 50", and asks whether it would be investable now. Altman: "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." 00:19:44 He adds: "if all you heard was it's a software company going 5 to 18... now... you'd be like, 'Well, why wasn't it 5 to 40?'" 00:19:58 Randle notes multiple companies per category are hitting "10 to 100 to 300... which never used to happen." 00:16:34 The flip side: other good categories are starved of attention. Altman: "a founder is like, 'Well, what the hell? I have a good business and it's hard to raise.'" 00:19:07

Durability Is Harder to Call in the AI Era, So Investing Means Riding the Trend or Calling the Next One

Altman contrasts the late-2010s, when "you could kind of name it when you're like, 'That's durable,' or 'That's going to get competed. That's sticky,'" 00:21:33 with today: "right now, I don't know that durability is nearly as easy to identify." 00:21:57 Application software sentiment whipsaws: "six months ago, everybody was kind of like, 'It's dead.' At the moment, I think people are like, 'It's alive.'" 00:22:19 His conclusion: "the spirit of investing at the moment is either on the trend or you're calling the next one. And obviously, the returns are highest if you call the next one." 00:23:37 Randle adds that once a derivative is obvious enough to be listed by a chatbot, "the companies that you could invest in are probably already at like a billion-plus valuation." 00:24:22

Open vs. Closed Models: Steelmanned, Then Synthesized as "It All Works"

Randle steelmans closed source on low-cost-producer economics. He gives SpaceX ("no one can launch a kilogram into space as cheaply as SpaceX can") and AWS (economies of scale) as precedents, then argues labs have "the best models in the world, many of which are not released to the public," plus co-designed serving infrastructure and the biggest compute purchasing power. 00:29:28 Altman steelmans open source: no training amortization ("The model was free" 00:32:39), labs won't chase margins on "two models ago" 00:33:32, and enterprise control: "you cannot trust any company to own all of your data... you need to own your intelligence." 00:34:27 The market data: "the inference companies, in combination, are doing mid-single digits worth of total revenue." 00:36:29 Randle's synthesis is that the two are not rivalrous, with open source worth 5-10% of frontier revenue. 00:38:25

AI Cyber Risk Has Become Concrete, and Hardening Takes Time

Altman: "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 going to hack the shit out of you.'" 00:41:10 Randle adds they were "getting company-ending vulnerabilities surfaced" day after day 00:41:24. He cites a webcam recording indicator disabled with Claude and a French supermarket attack that "manipulated refrigeration controls and actually spoiled food." 00:42:55 The problem is structural: "these critical systems to like firmware, and hardware... that only have been able to be exposed via these AI models." 00:43:20 Altman's hope: because software buying cycles are faster than the 36-month cloud-era norm, "in 12 months, people could do a lot." 00:47:36

Physical-World AI (Bio, Robotics) Is the Bigger Prize, but Feedback Loops Are the Bottleneck

Altman: "there is a ceiling to how much [software] can improve the world... It's things like people have better housing, or people live longer, or people live healthier." 00:48:00 Randle frames why code improved first: "AI just tends to improve fastest when it can generate large amounts of trustworthy feedback cheaply." 00:50:03 Robotics and bio face real bottlenecks: "the equivalent of like running a unit test in biology is a physical experiment." 00:51:28 Altman's signal: "I haven't yet seen the labs be wrong when they say something's coming. Getting the timeline right is really difficult." 00:49:08

The "Happy Path" for AI Is Likelier Than It Looked

Altman: a couple of years ago the worry was either "Terminator" or "crypto all over again, head fake." Now: "we might actually be splitting the middle landing, where like the labs are extremely important, but they are not $60 trillion companies, where it is not useless technology that doesn't help anybody." 00:39:58 Randle: "It seems like the happy path." 00:40:32

Circular vs. Independent Token Demand: A Healthy-Growth Debate

Randle presents two sides. The worry: "if the only people buying AI stuff are AI companies... that's more and more leverage on the system." The healthy sign is "large banks, or large healthcare systems in like Ohio, New York, and Florida" getting value. 00:25:06 The counter is the Stripe analogy: "Stripe didn't get big because they worked with old economy companies. They got big because they worked with DoorDash and Shopify and Instacart and Lyft," after which old-economy firms followed. 00:26:35

2. Contrarian Perspectives

The Biggest Software Ceiling Is Real: Physical Progress Matters More Than Shipping Code Faster

While much of tech celebrates software acceleration, Altman argues the bounded-ness of the screen. Altman: "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." 00:47:52 He adds, "If you like touch some grass and you get out of San Francisco... 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." 00:48:03 His practical conclusion is to bet on bio and robotics, because "I haven't yet seen the labs be wrong when they say something's coming." 00:49:08

Circular AI-on-AI Token Spending Is Fine, Even Optimal

Conventional wisdom says AI revenue from AI companies is fragile. Randle disagrees: "I don't really have any problem with AI companies buying AI tokens and like the circularity, because... they're just at the frontier. And just like internet native companies were, you know, at the frontier of the internet and older economy companies didn't really buy into what they were doing until it was too late." 00:27:09 The Stripe example supports it: it scaled on internet-native customers first, then won old-economy customers once those references existed. 00:26:35

Open Source Is Not "Cheaper"; Frontier Labs May Be the Low-Cost Producers of Intelligence

The popular assumption is that open source equals cheap. Randle's steelman inverts it: "this argument that like open source equals cheaper is a misnomer and is not necessarily true, because at the same level of margin... a frontier lab at any point of the intelligence curve... would be cheaper than an inference provider serving on an open-source system." 00:31:31 Altman agrees partially ("on a per sort of inference unit basis, the labs are still going to have the advantage") 00:32:59 and argues control, not price, is open source's real value. The empirical backstop is that open source is only mid-single-digit percent of revenue. 00:36:29

Fixed-Cost Frontier Economics Are Not the Problem; Revenue Scale Absorbs Them

The prevailing narrative was that training costs doom lab economics. Altman: "it's like the revenue just turned out that it's big enough to make up for all these fixed costs of training." 00:05:01 Randle shows this with the 2.1x revenue-to-compute ratio 00:04:23 and treats the deceleration from 145% to 45% as a lesser concern than the economics.

Today's Rapid Software Purchasing Cycles Could Compress Cyber Remediation to ~12 Months

Most security thinkers view hardening legacy systems as a decade-scale problem. Altman is more optimistic: "it's like people bought... Workday and Salesforce on these really elongated timelines... 36-month cycles... That's not really how it's going right now on the purchasing of software side... maybe in 12 months, people could do a lot." 00:46:59 Randle describes the scale of the challenge ("the biggest code modernization project in all of history") 00:46:28, but both see urgency.

3. Companies Identified

Anthropic

Frontier AI lab behind Claude. Discussed because its S-1 revealed a revenue-to-compute ratio of 2.1x and 8% GAAP operating margins. Randle: "Anthropic in six quarters did what, you know, it takes a lot of software and technology companies like a decade to prove out." 00:04:44 Also: "They had 8% operat- like GAAP operating margins in Q2 2026." 00:05:26

OpenAI

Frontier lab and Anthropic's main competitor. Included in Randle's open vs. closed sizing: "let's say OpenAI and Anthropic will do like 170 billion of run rate um ish." 00:38:02 Also mentioned for ChatGPT's role in kicking off AI demand. 00:13:53

Fireworks AI

Inference cloud serving open-source models. A Benchmark portfolio company. Randle: "we're large investors in Fireworks, which is growing like crazy." 00:15:43 He identifies it as one of three consolidating champions: "we have Fireworks, we have BaseTen, we have Modal... they really are like the three champions of the market." 00:38:50

BaseTen

Inference platform. Named among the three dominant inference clouds. Altman: "They're dominating. That's right." 00:38:56

Modal

Inference and code-sandbox cloud. Randle lists it under both inference clouds and "code sandbox companies like Modal and E2B are growing like crazy." 00:16:02

ClickHouse

Database company in the Benchmark portfolio. Randle: "ClickHouse in our portfolio has continued to accelerate like crazy." 00:15:43

Supabase

Backend-as-a-service/database provider. Randle: "So Supabase took off." 00:15:43

Vercel

Deployment/hosting platform for vibe-coded apps. Named in the deployment-host group. 00:15:43

Railway

Deployment host. Named in the deployment-host group. 00:15:43

Cloudflare

Deployment/infrastructure provider. Named among deployment hosts benefiting from code gen. 00:15:43

Mercor

Data/expert-labor marketplace for AI labs, in the Benchmark portfolio. Randle: "Mercor as a business has grown to an unbelievable amount of scale because there's basically... there's just no limit." 00:52:13 Altman: "Everybody kind of thought that like the data thing would run out... It's just turned out to be wrong." 00:52:34

E2B

Code sandbox company. Randle: "code sandbox companies like Modal and E2B are growing like crazy." 00:16:02

Exa

Web search for AI agents, a Benchmark portfolio company. Randle: "web search, Exa, also a Benchmark portfolio company." 00:16:02

Temporal

Durable execution platform. Randle: "durable execution with Temporal." 00:16:30

Chainguard

Supply chain security. Named among fast-growing code-derivative categories. 00:16:34

Socket

Supply chain security. Named with Chainguard. 00:16:34

Cognition

AI software engineering company (Devin). Randle and Altman on the cyber-remediation theme: "That's what we need Cognition for." 00:46:44 Randle: "Cognition, update your code and security." 00:46:55

Lovable

AI app-building startup. Altman: "Lovable and Replit, for example, which have obviously done great... it's 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." 00:17:01

Replit

AI coding/app platform. Cited with Lovable as examples of "code-adjacent" success despite "isn't this Claude Code?" skepticism. 00:17:01

Claude Code

Anthropic's coding agent. Randle: "in November of '25 when Opus 4.5 comes out and supercharges Claude Code. And that creates this mega-product market fit moment for code gen." 00:15:20

Cursor

AI code editor. Randle cites "how Cursor improved its models very quickly" through user feedback loops. 00:50:47

Codex

OpenAI's coding agent. Mentioned as a comparison ("Isn't this Codex?"). 00:17:08

SpaceX

Cited as the canonical low-cost producer. Randle: "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." 00:29:46

AWS

Cited as a low-cost producer through scale. Randle: "They had massive, aggregated demand that created economies of scale." 00:30:01

Stripe

Payments company. Randle: "Stripe didn't get big because they worked with old economy companies. They got big because they worked with DoorDash and Shopify and Instacart." 00:26:36

Nvidia

Start of the public AI bottleneck trade. Randle: "let's buy Nvidia. So like Nvidia goes up a ton." 00:13:53

Supermicro

Server/rack maker. "Supermicro went up 10x" in 2023. 00:14:12

Dell

Server provider. "Dell went up a ton." 00:14:12

Vertiv

Cooling equipment. "A company called Vertiv went up a ton." 00:14:12

Arista

Networking. "Arista basically doubled in both '23 and '24." 00:14:12

Vistra

Power generation. "Vistra went up 3.5x." 00:14:26

Oklo

Nuclear. "people were pulling forward the future of nuclear." 00:14:26

Bloom Energy

Power generation. "Bloom Energy went up 4x in 2025." 00:14:26

GE Vernova

Data center/power equipment. "GE Vernova went up 2x in 2025." 00:14:26

SanDisk, Western Digital, Micron, Seagate

Memory and storage. "all went up at least 3x." 00:14:45

Intel

CPU maker. "now that agents are now using more CPUs again, you have Intel up 120% this year." 00:14:45

Zoom

Analogy for demand pull-forward and saturation. Altman: "Zoom got this absurd pull-forward... everybody who wanted a Zoom account got it." 00:10:26

CrowdStrike

Datadog and CrowdStrike are cited as "scions of the last era" with consistent net-new trajectories. 00:06:41 Randle: "it's just like consistent upward trajectory, there's very little volatility to the growth rate." 00:06:42

Datadog

Example of a 5-to-18 growth trajectory. Randle: "pretty sure Datadog went like 5 to 18." 00:19:16

Databricks

Cited with Snowflake as multi-year contract businesses and an "era scion." 00:20:35

Snowflake

Cited as a multi-year-contract comparison. 00:06:46

Workday and Salesforce

Altman's example of slow cloud-era enterprise buying cycles: "'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.'" 00:21:57

Physical Intelligence

Robotics company, referenced as "Physical." Altman: "you go to these you go to cool companies like If you visit like Physical or something like that, you're like, this... you can see the future." 00:23:16 Randle: "Physical has an API now." 00:51:14

Tesla

Altman: "you and I both love our Teslas and you like see the beginnings of it." 00:23:16 On self-driving: "Tesla is obviously dominant... because of just like infinite data." 00:49:23

Palantir

Referenced via Alex Karp's thesis: "you cannot trust, as an enterprise, you cannot trust any company to own all of your data." 00:34:27

Kimi (Moonshot)

Chinese open-weight model, used as an example. Randle: "an open-source API inference provider, like a Fireworks using Kimi k3." 00:28:44

Hugging Face

Cited for an incident in which they needed an open-source model because frontier labs couldn't turn off safeguards in time. 00:44:57

Eli Lilly (hypothetical)

Altman's hypothetical about data ownership: "if you're Eli Lilly and all of your data lives in Anthropic, and then Anthropic's like, 'Hey, by the way... we got some drugs coming.'" 00:35:13

Uber, DoorDash, Instacart, Lyft, Shopify

Internet-native companies that Stripe scaled with. 00:26:35 Randle also cites Uber and DoorDash as mobile-internet companies limited by smartphone adoption. 00:18:05

Samsung (Frame TV)

Altman's joke about smart-device vulnerability: "if you have a Samsung Frame, smash it immediately." 00:43:47

Chick-fil-A and Purina Farms

Banter examples of non-tech companies. Randle: "Chick-fil-A has an amazing engineering team." 00:09:38

Yubico (YubiKeys)

Hardware security keys. Randle: "I definitely bought YubiKeys a couple days ago." 00:44:04

Artex

Randle names a group "in China" using abliterated models: "I think it's called Artex... to commit some of these hacks globally." 00:44:57

4. People Identified

Benchmark investor; grew up in growth investing. His analysis of the S-1, the derivative trade, and the open/closed steelman drives much of the episode. He also shares a personal story: "my dad got diagnosed with a very advanced form of cancer... immunotherapy drugs produced in the last 10 years have given him like a 10-plus year prognosis, which... if he got the disease 12 years ago, he's dead. And now, he's in remission." 00:53:59

Jack Altman

Benchmark investor and host of the Uncapped podcast. He contributes the unbounded-upside task framework, the Zoom saturation analogy, and the physical-world-over-software thesis. He also notes a founder he just seeded runs agents from his phone: "he's got like you know, he's doing some stuff on his phone... 'I just I have like a few agents running and it's going to be done by the time we're done.'" 00:11:54

Eric Vishria

Benchmark partner. His "it all works" philosophy is referenced repeatedly. Randle: "shout-out Eric Vishria, um it all works." 00:36:58 Randle also attributes the "continuous conversation from one week to the next" description of Mondays to him. 00:01:34

Keith Rabois

Investor and operator. Altman credits him with the two-task-type framework: "This is like a Keith Rabois-ism that there's like, you know, two task types." 00:08:07

Leopold Aschenbrenner

Referenced by first name. Altman on the public AI-bottleneck trade: "Leopold spoke of this... He's the prophet." 00:15:07

Alex Karp

Palantir CEO. Altman invokes "the Alex Karp stuff" on owning your own intelligence. 00:34:26

Andrej Karpathy

AI researcher. Randle: "Andrej Karpathy has a really good blog post on digital hygiene." 00:44:05

Brendan Foody

Mercor founder. Randle: "Brendan Foody at Mercor says, 'The internet was a one-time subsidy on the production of LLMs.'" 00:52:45

5. Operating Insights

Build Software with a "Software Factory": Agents as the Default Org Structure

Randle describes how newer companies work: "some of our newer companies, like the companies that have been started kind of post-Opus 4.5... the way that they build software looks almost nothing like how you build software in 2019." 00:11:29 The organizational form is "like a new organizational structure for building software products and like a lot of it's not even human driven anymore." 00:12:19 Altman's founder example: one other engineer, with agents running while on a walk. 00:12:12 And "you either have to adopt that method of building software products or get left in the dust." 00:12:32

Point Frontier Models at Your Own Systems, Including Safety-Removed Ones

Randle's defensive playbook: "take an abliterated model, you do what that CEO told us, and you just say, 'Hack us. Like hack us to death.'... the adversaries have abliterated models. The adversaries are not working with safety controls. You need to have that." 00:45:18 Practically, CEOs who tried this saw "company-ending vulnerabilities surfaced" day after day. 00:41:24

Post-Train Small Open Models on Your Trace Data to Get Frontier-Level Quality at Near-Zero Cost

Randle on founder reports: "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... you can actually get a really small open-source model, and if you fine-tune the hell out of it, it gets to like a k3 or a Sonnet-level performance, and it is dirt cheap." 00:35:38 Altman adds that "if you've got a lot of data, that data can outperform... the extra intelligence that you might get from the next model." 00:35:26

Don't Build Your Entire Product on a Single Lab's Tokens If That Lab Can Enter Your Market

Randle: "let's say you're all all your tokens are Anthropic tokens, and all of a sudden, Anthropic enters your space. And then your customers are like, 'what model are you using?' 'Anthropic.' Well, then why don't we just go to Anthropic? You have no special sauce." 00:34:52 Altman's Eli Lilly example reinforces the data-leakage side of it. 00:35:12

Run an Agenda-Free, Long-Form Partner Meeting

Benchmark's operating cadence: 5-6 hour Monday meetings with no agenda. Randle: "The agenda is like you review a very strict list... But our answer is always, 'There is no agenda.'" 00:01:34 The meetings function as a "continuous conversation... from one week to the next," generating the cross-pollination visible in this episode (e.g., Altman raising CI/CD and Randle raising derivative maps).

6. Overlooked Insights

"The Internet Was a One-Time Subsidy" Implies Physical-World AI Has No Free Training Corpus, Making Data Collection Infrastructure a Giant Market

Mentioned quickly through Brendan Foody's line, but it reframes where durable value accrues. Randle: "Brendan Foody at Mercor says, 'The internet was a one-time subsidy on the production of LLMs.' And in these physical spaces, they don't have as much of that one-time subsidy, they don't have the internet." 00:52:45 Altman adds that "you also need a ridiculous diversity of data, it turns out, to build these models well" 00:53:14 and calls the effort "Manhattan Project-level scale." 00:53:30 The implied investable thesis: picks-and-shovels data/feedback infrastructure for robotics and bio could resemble what Mercor did for digital data, and the earlier assumption that "the data thing would run out" 00:52:34 was wrong.

"Reality Has a Surprising Amount of Detail" Plus Abliterated Models Hints at a Whole New Defensive Tool Category

The abliterated-model discussion is dropped in as a side note ("we're going to hear a lot more about this idea of abliterated models") 00:45:46, but it is a specific unmet-need signal. Enterprises need models with no safety guards to red-team themselves at the capability level adversaries already have, yet Randle notes frontier labs could not turn off safeguards fast enough during a real incident. 00:44:57 A product that supplies controlled, authorized, safety-off offensive security models to defenders sits in the gap between lab policy and attacker reality.