Benchmark Breakfast #1 | Jack and Ev

Anthropic

Across the shows: 69 speakers in other episodes, 73 bullish and 55 bearish calls, past 30 days

• Ev Randle · BenchmarkNeutral

  Anthropic's IPO may not pop on day one despite strong numbers, given growth deceleration and consumption revenue.

  The S-1 is strong on its face, with immense growth and operating leverage. But Ev says it is not a slam dunk at the open because of two open debates. First, Anthropic grew 145% quarter over quarter in Q2 and about 45% in Q3, a sharp deceleration. Second, roughly 80% of revenue is consumption-based, unlike SaaS peers such as CrowdStrike or Datadog, which have steadier growth and multiyear contracts. He calls these concerns somewhat fair even though the business works.

  “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 in the public markets value it and if there's a ton of share price appreciation right out of the IPO.”at 5:57

Fireworks AI

Across the shows: 1 speaker in other episodes, 1 bullish and 0 bearish calls, past 30 days

• Ev Randle · BenchmarkBullish

  Fireworks, an inference cloud in Benchmark's portfolio, is growing very fast as AI coding derivatives scale.

  Ev places Fireworks among the inference clouds that have grown with the code-gen wave. He says Benchmark is a large investor and that the company is growing like crazy. The argument is qualitative, with no revenue or customer figures.

  “The inference clouds, obviously, we're large investors in Fireworks, is growing like crazy.”at 15:54

Tesla

Across the shows: 22 speakers in other episodes, 16 bullish and 19 bearish calls, past 30 days

• Jack Altman · BenchmarkBullish

  Tesla is dominant in self-driving, driven largely by its large data advantage.

  Jack says self-driving took a while to arrive, but it is now fully here. Tesla is obviously dominant, and he attributes much of that to infinite data. He suggests that once these systems get enough data, they may simply work, which he uses as a parallel for robotics timing.

  “self driving, Tesla's obviously dominant. But a lot of that was because of just infinite data.”at 49:23

Enterprise AI adoption & AI software

• Jack Altman · Benchmark

  Token buyers are not yet proven more productive in all cases; broad enterprise diffusion still has a long way to go.

  Token sellers are clearly doing well, but Jack doubts buyers are proving much more productive across the board. Coding was unproductive a year ago and now looks productive, but much of the economy has not yet adopted AI. He says there is still a long way to go before ordinary businesses can buy AI productively. He still believes it will get there, so this is a cautious view about the pace of diffusion rather than a rejection of it.

  “I don't think that in all cases, the buyers of tokens are proving to be that much more productive in all cases.”at 9:04

• Jack Altman · Benchmark

  AI coding adoption is near saturation among likely users, so 10x growth will fade; the runway is elsewhere in the economy.

  Jack compares AI coding to Zoom during the pandemic. Everyone who wanted the product got it in a burst, growth was extreme, and then it saturated. Anyone in the Benchmark world who might use AI programming already does, so from here growth comes from market-share shifts, not 10x years. The larger opportunity is in the many places in the economy where adoption is far behind, so coding growth should slow even as AI broadens.

  “everybody who might want to use AI programming has been using AI programming for a while... you don't get 10x year over year from that forever.”at 10:38

• Ev Randle · Benchmark

  AI code generation is the most impactful product-market-fit moment in tech history and could keep doubling yearly.

  The breakout moment came when Opus 4.5 in November 2025 supercharged Claude Code. That created a wave of derivative companies (Supabase, Clickhouse, Vercel, Modal, E2B, Exa, Chainguard, Temporal) growing from 10 to 100 to 300. As code volume rises, derivative products must grow with it. Ev sees agent-run software factories spreading to the rest of the economy, so durable code-gen demand could keep doubling each year even if the 10x years end. He calls it on the order of the iPhone moment.

  “It's probably the most impactful product market fit moment in the history of technology.”at 17:36

• Ev Randle · Benchmark

  AI companies buying tokens from labs is acceptable; they will become the large companies, as internet natives did.

  Ev concedes that if only AI companies buy AI tokens, that is more leverage in the system, and that buyers in banks or health systems would be healthier. But his view is that AI-native companies are at the frontier and will become massive by out-competing firms that do not build software factories. He compares this to Stripe, which got big by serving DoorDash, Shopify and Instacart before old-economy firms adopted it. He says he has no problem with circularity.

  “I don't really have any problem with AI companies buying AI tokens, like the circularity, because I'm like, look, these are companies.”at 27:06

Open-source AI models

• Jack Altman · Benchmark

  For some industries, owning open models is worth it even without savings, as open models get good enough.

  Enterprises cannot fully trust any vendor with all of their data, so they want to own their intelligence, a theme Jack links to Alex Karp's argument. For certain industries that control will justify the trade-off. As open-source models improve, a company gets control even if it saves no money. Post-training an open model on proprietary data can also outperform a newer closed model, and founders say they hear this often. He expects open models to be good enough for most tasks, while frontier labs focus on the frontier.

  “I think there is an argument to be made there. I think for certain industries, that trade is going to be worth it. And particularly as the open source models get good enough, even if you're not saving any money, you're in control.”at 34:30

• Jack Altman · Benchmark

  Open source likely stays mid-single-digit to ~12% of AI revenue unless total spend changes.

  Jack treats the market as the answer. Inference companies together are doing mid-single-digit percent of total revenue, which he takes as a reasonable proxy for open-source usage because that is where open models run. He says mix shift could move this, but the trend suggests open source at around 5% to maybe 12% of the total unless overall AI spending changes. He frames this as the synthesis of the open-source and frontier-lab views.

  “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.”at 36:26

• Ev Randle · Benchmark

  Open source stays about 5-10% of frontier-lab AI revenue, a roughly $100B pool if growth holds.

  Ev uses the top three or four inference clouds (about $6B run-rate by year-end) plus OpenAI and Anthropic (about $170B run-rate if their disclosed growth extrapolates) as an imperfect proxy. That gives open-source inference roughly 3% of revenue today, which could about double with the hyperscalers included, so he calls it 5-10% of open versus closed revenue. He says it is not a projection, but if frontier labs keep their growth for three to four years and reach a combined $1T, open source at 10% would be about $100B. He expects consolidation into roughly three inference champions (Fireworks, Baseten, Modal) splitting that pool. He also argues fine-tuned small open models can reach Sonnet-level performance at very low cost on specific tasks.

  “So we can say 5% to 10% in terms of revenue for open source versus closed.”at 38:11

• Ev Randle · Benchmark

  Uncensored open-weight 'obliterated' models will become key enterprise tools for defending against AI hacking.

  Ev describes obliterated models as open-weight models with safety guards removed. He says adversaries, including groups in China, are already using them for hacks, and that Hugging Face reportedly needed them to run an incident response that frontier labs' safety guards would have blocked. Because the adversaries' tools have no safety controls, he argues enterprises need the same kind of model to red-team their own infrastructure, which makes it an incredible defense tool. He expects to hear much more about this.

  “It's also an incredible defense tool for enterprises.”at 45:15

Software & SaaS

• Jack Altman · Benchmark

  The growth bar for software companies has risen sharply; a 5-to-18 growth story no longer clears a raise.

  Ev cites Datadog's historical path from about 5x to 18x revenue, then up to 40 to 50. Jack says a 5-to-18 company used to draw interest, but now investors ask why it isn't 5-to-40. Being credible today requires believing in an 18-to-100 story or a company doing something so hard that growth durability is unusually high. He ties this to the current code-driven pull on growth.

  “that used to be enough where everybody would be super interested if all you heard was it's a software company going five to 18.”at 19:59

• Jack Altman · Benchmark

  Durability of application software is now hard to call; sentiment has flipped within months as models improve.

  In the last cycle, investors could tell durable software from competed-away software because the patterns were familiar and enterprise buying was predictable. Now that is harder. Six months ago application software was seen as dead, and now it is seen as alive. The labs could move into these applications, but applications add value through edge cases, data and UI. Jack expects sentiment could flip again if code gets good enough to make those advantages disappear.

  “I 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.”at 22:19

• Jack Altman · Benchmark

  Enterprise software buying is faster now than the slow, 36-month cloud-era cycles.

  Jack compares the cloud era, when customers bought Workday and Salesforce on long timelines and government, hospital and law-firm buyers took about 36 months, with today's software purchasing. He says it's not really like that now, and people are alert. He hopes buyers will treat security as a top priority and react faster than expected. He still doesn't think that means it happens in three months.

  “people bought Workday and Salesforce on these really elongated timelines. And they were slow... That's not really how it's going right now on the purchasing of software side.”at 47:04

Robotics & autonomous vehicles

• Jack Altman · Benchmark

  Robotics will work and become common, so investors can bet ahead of it even though it is not working yet.

  Robots cannot yet do much of what a human can, and no one has one at home. Jack still believes they will arrive, pointing to the early signs of autonomy in Tesla vehicles and to demos at robotics startups. He concludes that this is the kind of trend investors can bet on ahead of time. No date is given.

  “I think we all believe robotics will work... But I think we all believe we'll have one.”at 23:09

• Ev Randle · Benchmark

  Robotics faces slow, costly data feedback loops, so physical AI needs Manhattan-Project-scale effort.

  Ev says robots are expensive to deploy and maintain, and there is no large customer base that can press a button to generate data. Unlike ChatGPT, which reached 100 million users in months through a website, robotics feedback is slow. Labs got the internet as a one-time data subsidy for LLMs, while physical domains lack that and need a wide diversity of data. He cites Brendan Foody at Mercor for this framing. He concludes physical AI won't happen overnight and will need investment on the scale of the Manhattan Project.

  “robots are expensive. They're expensive to deploy, expensive to maintain. There isn't a big customer base that can press a button.”at 51:10

• Jack Altman · Benchmark

  Good AI robotics could be bigger than all digital AI so far, if it actually works.

  Jack argues the physical world is where AI makes life better: housing, longevity, health, food and transport. He sees robotics and bio as the two areas where the labs are talking about the technology, building teams around it and buying large amounts of data, which he takes as a sign it is coming. He has not seen the labs be wrong when they say something is coming, though timing is hard, as self-driving showed. Ev says Jack has recently invested in robotics, so Jack has a stake in this space.

  “But if you imagine what's really possible with good AI robotics, it's probably bigger than all the digital stuff so far, if it actually works.”at 53:38

Pharma & biotech

• Jack Altman · Benchmark

  AI for biology is likely to work; early signs show target identification now possible that wasn't before.

  Jack says the smartest people at the labs believe AI for biology will work. Early indications show that target identification and other tasks are now possible when they were not before. He cites this as one of the areas where the next big bet is forming, alongside robotics. The evidence is anecdotal.

  “the smartest people at the lab seem to be saying it's going to work. There seems to be early indications that target identification and other things are possible that didn't used to be.”at 23:41

• Jack Altman · Benchmark

  AI-driven bio could be among the best outcomes, as labs build teams and buy data for it.

  Jack argues software improvements have a ceiling because the world's biggest gains are physical: housing, longevity, health, food and transport. Bio would be one of the best outcomes if AI finds drug targets and new drugs quickly. He points to labs talking publicly about bio, building teams around it and buying large amounts of data as evidence it is on the horizon. He calls bio and robotics the two areas where this is most visible.

  “bio, to me, would be one of the very best possible things if we're discovering drug targets and finding amazing new drugs quickly.”at 48:31

• Ev Randle · Benchmark

  AI-driven bio progress will likely have a much larger impact than other AI gains.

  Ev uses his father's cancer as an example. Recent immunotherapy drugs gave him a ten-plus-year prognosis and put him in remission, where a diagnosis twelve years earlier would have been fatal. He argues we are only at the start of perhaps one hundred to one thousand times that much progress across every disease area. Once AI helps solve biology, with the investment that will require, the impact should be much larger than from the other AI applications so far. The argument is partly personal but grounded in the survival-rate shift and the scale of the opportunity.

  “And we are just at the start of probably one hundred to one thousand times that amount of things happening, it's like every single disease area, every single cancer.”at 54:24

Cybersecurity

• Jack Altman · Benchmark

  AI-driven cyber risk is imminent, so it is a good time to be a security company.

  Jack says cyber risk is the most imminently expressed AI risk. Recent CEO conversations described pointing frontier models at their own systems and finding them breakable. He notes that attacking from inside has advantages an outside attacker lacks, since insiders have the data and code base. Still, he argues this is a good time for security companies and a scary first real exposure. He expects enterprises, governments and hospitals to move faster than the old 36-month software cycles, though not within three months. He hopes that in about twelve months people can do a lot.

  “obviously, it's probably a good time to be a security company.”at 41:58

• Ev Randle · Benchmark

  AI-enabled attacks now reach personal devices and critical systems; hardware can't be patched easily.

  Ev says companies with strong engineering teams are finding company-ending vulnerabilities day after day when probed with AI. Cyberattacks used to mean a leaked password. Now AI makes them creepy and intrusive. He cites a researcher who used Claude to disable a webcam's recording light and access the microphone, and a French supermarket whose refrigeration controls were manipulated. Systems that were thought secure, such as decades-old open-source code, firmware and hardware, are now exposed. Software can be patched, but hardware largely can't, so smart devices may need to be disconnected. He says hardening will take a long time, possibly the biggest code modernization project in history, which is why the threat feels real now.

  “you can patch the software, but can't really patch the hardware.”at 43:32

Frontier AI labs & model race

• Jack Altman · Benchmark

  Customers pay for the leading-edge model even at higher prices, so frontier leadership carries outsized momentum and pricing power.

  Labs have swapped the lead all year, and momentum follows whoever has the best model, even by a small margin and even when the older model handles most tasks. Users seem to pay for top intelligence despite price sensitivity because individual value is in the thousands or tens of thousands of dollars, and that surplus is large. For unbounded-upside tasks such as coding, sales, drug discovery and robotics, users will keep buying the better model. Jack also notes that price per gigawatt and the monetization of frontier tokens have surprised everyone.

  “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.”at 7:18

AI infrastructure & data centers

• Ev Randle · Benchmark

  The AI supply-chain bottleneck trade keeps producing new derivative stocks that can multibag each year.

  After ChatGPT, investors bought Nvidia, then the derivatives: Super Micro, Dell, Vertiv, Arista, Vistra, Oklo, Bloom Energy, GE Vernova, memory and storage names, and now Intel as agents use CPUs again. Each year brought new derivatives that produced multibagger returns. Ev's evidence is largely a list of past returns, and he expects the pattern to continue.

  “Or like, there's like three new derivatives every year that you could have gotten, like, a multibagger return on in the public markets.”at 14:59

Venture capital & startups

• Jack Altman · Benchmark

  Venture demand for AI code-adjacent startups is so intense that other good businesses now struggle to raise.

  The code-adjacent AI wave has made venture demand in those spaces extremely strong, and LPs are doing well and funding it. Jack calls it one rational circus. The result is an opportunity-cost squeeze: investors can get a ridiculous ride in AI, so founders of solid non-AI businesses find it hard to raise capital.

  “it's making it so that venture demand in these spaces is so obscene. And then LPs are obviously doing great and they're funding it.”at 18:54

AI pricing & unit economics

• Jack Altman · Benchmark

  Frontier labs likely keep a per-inference cost edge over inference providers, though training amortization complicates it.

  Frontier labs must recover training costs through their prices, while an open model is free to use, so it carries no amortization. Inference still has to be run either way. Jack thinks labs still have the advantage per unit of inference over inference companies, but that gap could be offset by other factors. He says he mostly buys Ev's economic case for closed labs, but he doesn't have the margin numbers to be sure.

  “And it's probably right that on a per inference unit basis, the labs are still going to have the advantage over the inference companies.”at 32:46

AI bubble & tech valuations

• Jack Altman · Benchmark

  Frontier AI labs are very important but unlikely to be $6 trillion companies; a middle outcome is the happy path.

  Jack says the earlier fear was either a Terminator-style AI takeover or that AI wouldn't matter, like crypto. He points to the market's positive reaction to Anthropic as evidence that the outcome is landing in the middle: the labs are extremely important, but not trillions-of-dollars companies, and the technology is useful. He calls this the happy path and says his odds of it have risen. The argument rests on the market's reaction and his view of AI's usefulness, with the caveat that RSI could change things.

  “where the labs are extremely important, but they are not [$6 trillion] companies”at 40:19

AI: overall outlook

• Ev Randle · Benchmark

  AI improves fastest where cheap verifiable feedback exists, so physical-world gains lag digital ones.

  Ev argues AI improves fastest when it can generate large amounts of trustworthy feedback cheaply. Code is the ideal case because unit tests verify it and experiments scale in the digital world. Cursor and Claude Code improved quickly from user feedback, and every such interaction feeds training data. Biology and robotics have real bottlenecks: physical experiments are slow, and a single experiment may answer only a small part of the question. He sees an inverse relationship: the things that most improve the physical world are the ones AI improves most slowly.

  “AI just tends to improve fastest when it can generate large amounts of trustworthy feedback cheaply.”at 50:18