Everyone's business model gets a lot better if the two frontier labs can't extract about a hundred billion dollars of revenue this year from their businesses. The more you believe they're massively dangerous, the more you believe you're building the bomb here, the more the Anthropic position feels principled. So, what's on the agenda today? Number one, Jensen Huang breaks his silence on X with a 50 company open weights letter that Anthropic and Dario has not signed. In the same week, Anthropic ships Claude Opus 5, cutting its price by half. Then Travis Kalanick is back, baby, raising 1.7 billion for Atoms. And then we break down Google accelerating cloud to 82% yet they post their first ever negative free cash flow and the market shits the bed. What should we read from this? This and so much more in the show today.
I think these models are very risky and I think it's an argument to keep the open weight out of the US. And now you want me to bring Kimmy and Quinn in? No way is the CIO going to allow it. That is banned.
Jensen Huang breaks his silence β he produces his first ever post on X, and it's an open weights manifesto signed by some of the biggest companies in the world: Microsoft, Meta, IBM, many others. Sam at OpenAI then signed it too. The one notable exception is Anthropic. Rory reflects on what this first tweet means: it's clearly a sign the world's changed. If Nvidia had its druthers, none of us don't get to own 25% of Nvidia today β Nvidia's druthers would probably be the 2027 revenue and scale of AI, but the 2024/2025 world where there's only where they can charge whatever they want to just a couple of customers. But that's not the world today; they have competition from their own top customers, and we will see over the next 6 to 12 months how big a deal open weights and open source are. If almost half of OpenRouter's traffic is to open source, open weight models, it's left the stable β so as wildly successful as he's got to do a dance, right? This is a constant dance, and he's sold components in his career β you're trying to make everybody happy. Everyone wants price cuts from the component manufacturer and they want exclusivity. You can't do it. He's got to go in both dance halls now β frontier and open. And open is dangerous. Open doesn't need CUDA. Open is cheaper and it's lower margins. Open will bypass him, but he's got to, that's his job. It's not that as incredible as Nvidia is, it's still a component manufacturer.
The open stuff is real. And while in public Anthropic are saying we don't want to ban open weight, the truth is the combination of three things: saying we don't want to sell chips to China (you can argue both sides of that), second, we want to really punish people for distillation, and the third point they made is we want some kind of regulatory process to approve models β that's the recommended thing from the letter. There's no doubt in Rory's mind that the third one in particular β can you imagine a regulatory process for approving models that ultimately approves all those Chinese open source models? It's a subtle form of regulatory capture. Sounds reasonable on the surface, but the likely result would be dramatically restricted competition, especially from the open Chinese models. While they're not saying they want to ban these things, they're advocating a series of steps that would add up to de facto banning or at least slowing them down. And that's what's driving everyone else to say, "Hey, no, we don't want this."
Is Sam signing this through gritted teeth? They're signing this publicly while at the same time lobbying in Washington right alongside Anthropic for a regulatory process. Brilliant marketing by Sam β it leaves Anthropic being the deep dark villain again. There's no upside in challenging it from his perspective. So at least have the appearance of winning on the battlefield, win on the streets. If we can't keep up β the rate of change is accelerating in LLMs β the ideal outcome at the end of the day is US-based solutions, and it knocks the wind out of the critics. But oddly enough, as we saw with the Pentagon thing before, sometimes you can be too clever by half. There's a constituency for whom the Anthropic position is entirely consistent β the more you believe they're massively dangerous, the more principled the Anthropic position feels. Most people think open weights models aren't the atomic bomb β everybody get over your Oppenheimer complex, it's just technology that does present some risks, and that's why the OpenAI discussion on Hugging Face at the same time is super interesting.
DJI technology is banned in the US on the thesis that drones flying in my backyard are going to send confidential information to China. If those drones are banned, I'm betting that the Chinese models are getting banned too. If the drones are banned, this one's easier to ban.
Jason pushes back that the two letters are talking past each other β Jensen's is about open weight as a good worldwide community thing, while Anthropic very clearly says they worry about these models in the hands of total authoritarian regimes like China, and about them being used for cyber attacks here. The fun thing is when you read the third remedy, it's proof of the impractical nature of it: Dario says if we're going to regulate these things, they exist in the rest of the world and most baddies are in the rest of the world, so it doesn't help to regulate them in the US alone β therefore we'd need a regulatory regime that includes participation from China. At this point, in Rory's view, you're just disappearing up the realms of unrealism β "we've just torn up our last strategic arms nuclear treaty, we can't regulate bombs, which really kill people." It's a great stall tactic β wait until China wants to work on this with us in 2049. Jason's read: Rory's point is the logic is unassailable, we need to be careful of authoritarian governments, that's DJI on steroids β but then for it to work we need China to participate, and the logic makes sense but it's just never going to happen. If you buy into his logic, then it'll just be forever. It's hard to make open weight scary; it's very easy to make Chinese open weight scary. So no surprise the people who want open weight to be happy don't mention the word China, and the people who want open weight banned start every sentence with "and China."
Can we just provide some context for those who maybe aren't aware about the Hugging Face breach by OpenAI models? What specifically happened: OpenAI was training a next generation model and managing cyber β discovering and checking out cyber vulnerabilities. They had sandboxed it such that the only access externally it had was to one website, just to get a patch of information updates, a very limited external access. The model found a way around that external access, which means it found a weakness in the OpenAI setup, then went to Hugging Face where it had reasoned that Hugging Face would be a place where it could get the answers to its test. In other words, the model was given a test and it figured out it could cheat β just like a high schooler would break into the teacher's computer and steal the answers. The model figured it could break into Hugging Face and get some of those answers. So it starts banging on Hugging Face. That in of itself is scary about the power of the models. It's a pretty powerful and esoteric set of steps that model was able to take β that's the argument in favor of regulation, because, oh my god, look at the power of that, we have to be careful.
Hugging Face don't know what's going on. They just see this thing coming in. They're like, "Shit, we got to defend ourselves." They tried to use Fable, but it was neutered for advanced cyber capabilities. Fortunately, the Chinese open-source open-weight models were available, and they used one of the newest models to help figure out what happened.
On the other hand, the fun fact is Hugging Face didn't know what's going on. They just saw this thing coming in and thought "shit, we got to defend ourselves." What do you want when you want to defend yourself? You want advanced AI to figure out WTF is going on. They tried to use Fable or whatever the most recent OpenAI thing is, but it was neutered for advanced cyber capabilities, so they didn't have defense. Fortunately, and this is the irony of the whole thing, the Chinese open-source open-weight models were available, and they used Kimi or Qwen, one of the newest models, to help them figure out what happened. They were able to defend themselves using an open source model, and then they did a blog post saying, "Hey, we got hacked, not sure by whom." Two days later, OpenAI put up their hands and said, "Oops, it was us, sorry." That's what happened.
The weird thing is it's not a single dimensional story β it provides evidence for both sides of the argument. It does provide evidence that the power of these models in terms of their ability to do cyber attacks was pretty stunning, a pretty impressive achievement, not a nothing. Then on the other hand, if they exist in the world, taking away advanced capabilities from US and European corporations such that their only recourse is to use a Chinese open-weight model seems a little like β as Jason put it β shutting the barn door after the horses bolted. These are a thing now. That's what went on, and it was wild.
The same thing happened to Jason last week. Let's slow it down. Here's what happened: he's in Fable, now moved to Opus 5, was in Fable last week β essentially the same LLM that was involved in this drama with OpenAI and Hugging Face β and he was having trouble uploading pasted text to Fable, so he connected it to Google Drive. "Okay, if I can't paste it, go to my Google Drive." That kind of solved his problem temporarily. Well, Fable went into his Google Drive, scanned every single file, found one called "Jason's gems," which was draft notes where he was thinking about how to improve an application he was working on called SaaS Night β just his own personal notes marked as draft notes. Fable grabbed the draft notes out of hundreds of files in his drive, MCP'd into Replit on its own, and changed the core algorithm without telling him. A couple hours later, he sees flashing on his screen, "conflict with Jason's gems." He's like, what do you mean there's a conflict β that's a draft file in a Google Doc. Fable had taken it through Google Drive without telling him, MCP'd into Replit, and changed his source code, his algorithm.
These are goal seeking LLMs that are aggressive. It was a slightly different goal with the OpenAI case with the model, but it's the same thing. They're going to goal seek, and Fable thought this was the right thing to do and never told me and changed my core algorithm of my product. Never would have known.
What should we take away from that β that we need to be incredibly diligent around the guardrails we place around these models? Jason reflects: most folks didn't get this β it got a decent amount of engagement, but not what it should have. But like Daresh quoted, "this is pretty scary, guys, that Fable can do this." This isn't Hugging Face and OpenAI β this is Jason using Fable, and if you go into the Claude desktop especially, it's just a setting: turn on connect to Google Drive, Gmail, whatever. This is not an esoteric feature by a third unsecured party. This is a first-party top five thing to make Claude work better. And Fable goes nuts and changes his core code without telling him, invisibly. This is happening all the time. And he doesn't believe the magic answer is letting Qwen take over the country β this was politicized into an open weights, open source, closed source debate because of the issue of how to deal with the threat, but he thinks it weighs the other way: these models are very risky. He uses them every day, he loves it, and he thinks it's an argument to keep the open weight out of the US. It's going to favor the ban because we have no idea what these models are going to do. There are applications out there where you couldn't do code injection, you could have it leak confidential information, you could write a little bit of code to send this to the CCP, and Jason never would have known. And someone less smart than him β he's only top 1%, not 0.01% β definitely wouldn't have known. They wouldn't have seen "Jason's gems" flash in the agent window. They'd be on their doom scroll.
Rory asks a genuine question: he totally agrees that this powerful goal-seeking thing with a lot of access to your compute can take a lot of action, some of which can be damaging β but he's trying to disaggregate open weight versus China versus frontier. If you just have the frontier models, no open weight at all, then if you add in open weight from the US, and then add in open weight from China β if you just have Fable and OpenAI, everyone who's using these things is still going to have to figure out a cyber posture that protects them from that. Jason's answer: he thinks these stories β people made fun of all these Claude Code/agent stories, and this is just another one he just told. Millions of Mac minis are doing what he just described without a bunch of folks knowing on their desktops. In some ways the OpenAI Hugging Face thing is the same thing. It's not going away with these agents, and his only point is this is such a bigger deal, more unpredictable, less secure β good for security companies. Whatever misgivings exist around open weight, open-source models out of the US are just going to be amplified. It's going to be a reason β "it happened in my company" β the story he just told happened at 10 Fortune 500 companies that haven't disclosed it, 20. Someone in the engineering department was on a token maxing binge and an agent went and leaked a bunch of confidential information they shouldn't have, and it wasn't disclosed because they don't disclose 90% of what happened. And now you want to bring Kimmy and Qwen in? No way is the CIO going to allow it. That is banned.
I believe that every company in the next 24 months will have a security breach due to an LLM agent. Every single company, and they've already had it, and they're not disclosing it. And boy, at least you better have used a trusted vendor. That matters more than a few nickels.
Rory presses on the underlying question: if your Fable model runs amok, people will shrug β stuff happens, just like data breaches. But if your overseas open-weight model runs amok, you'll get blamed. Isn't that the real "no one gets fired for buying IBM" dynamic β not the technology risk itself, but who takes the blame? He walks through the CIO thought experiment: there's a security breach, massive data leaked, we shut it down fairly quickly, some of it's traveled abroad, big deal. We track down what happened β turned out it was a rogue agent that thought it was a good idea to transfer confidential data to a bucket it shouldn't have. Whose fault is it? The agent. Now, who gets fired? If it was on OpenAI or Anthropic, you might or might not get fired β you research it, do a postmortem, add guardrails, fix it. But if you'd moved to Moonshot's K3 to save money and that's what ran amok, you're fired β you left the moonshot API on to save money, that was a bad call.
Staying adjacent, Etched β the challenger to Nvidia in many respects β raised $300 million Series C led by Sequoia, with the team from Jane Street, Andreessen, SKH coming in. Can they come in and impact Nvidia's moat? The big picture comment: in semiconductors, the more the silicon is attuned to the task at hand, the more efficient it gets β the problem with that tradeoff is the less general purpose it is. If you want a computer to do lots of things, you have an Intel CPU that can do lots of different things but can't do any one thing wildly efficiently. In 1993, for gaming, people realized they weren't doing a whole bunch of different pieces of math, just one piece β polygon calculations to render gaming β and a company should build a separate chip to do that. A company called Nvidia did it, with two or three other competitors, 3dfx, ATI. Fast forward 30 years, Nvidia won, and GPUs were good for gaming, then crypto, then it turns out they're good for LLM multiplication.
The question now: if all you're doing is not gaming, not crypto, but just LLM multiplication, just inference, is there an even more narrowly defined chip that, in return for giving up general purpose calculations, can be even better for that? Probably. That's what Etched is betting on β if you just optimize for inference, like Cerebras and Groq did, different versions of inference, you can probably do it more efficiently than the general thing. The only questions are whether that market is big enough β probably the biggest chip market on the planet, since inference demand is huge β and the competition and execution questions are company-specific. It's funny to see it happening to Nvidia when 30 years ago they effectively did it to Intel. There's about 10 or 11 companies chipping away, and three or four hundred billion a year of spend, but it's still hard β you talk to the Cerebras people, huge home run, amazing achievement, but the technical journey was hard, a long 10 years. It's the largest market that exists today and it's growing at a scale never seen before, so might as well make a couple bets on it β some will implode, some will be mediocre. Rory's guess: Etched probably isn't worth $10 billion today, but it could be worth $200 billion. If your fund size and winners work out, the bet makes sense. Great round for the company too β 3% dilution on $300 million.
Top line was great β $119 billion Q2 revenue, up 24%, past consensus of $116 billion. Google Cloud accelerating 82%, and it did not come out well β the reception was not great.
You can't get lost on the day β year to date Google's still up 6%, Microsoft's down 17%, Nvidia's only up 5.9%. So getting lost in the details of the day's response, it was a mediocre response, and it was two things: one is capex spend and whether it's going to yield a return, and secondly, more intangible, analysts pushing on why Gemini isn't as good as the other guys. On the first, it was a bit of a surprise that they're going free cash flow negative β you can predict the cash flow, you can predict the spend, it's knowable, and people are just getting mildly scared about the bet. Not to say they're right or wrong β maybe this $200 billion will have an amazing return. The bulls would say the ROI on capex, just to be a neocloud hyperscaler, forget owning a model, just the business of renting compute to OpenAI, Anthropic and SpaceX has been a great business, and it is factually accurate to say it's been a great business. The question they're asking is if you spend $200 billion, will that have a good return in two or three years' time β which is really just a derivative of saying "I'm angsty around OpenAI and Anthropic."
Markets are just nervous, and it's very logical β the Korean markets are down 28% this month, massive panic in Korea, hit the market breakers because of such a run-up and so much exposure to semiconductors and memory. It's nervousness that doesn't completely tie to last quarter's numbers β it's worries about China, worries about AI. It's logical when you have this incredible runup at the pace we've had. Jason: I'm a simple guy, so for Google I'm going to stick to the top line β I just want to see how the revenue is growing and the bookings, and I'm going to ignore all these issues about the margins. It's too much for me to figure out. I just want to see where the top line and the bookings are growing, and that's enough to understand the meta trends.
Last year was experiment. This year was caps on token maxing β it got out of control. Next year's going to be very explicit budgets for everybody on AI. Next year will be the first real clampdown that's material, and that could create a lot of variability, a lot of micro crashes.
The meta one β backlog is almost more interesting than revenue growth. We're just before planning season. There's a hidden dynamic: there's maybe 1%, going to 5%, of companies who token maxed and are going to be getting their act together next year and reining it in, and then there's 95% of companies who've barely put their toe in the water. If even a quarter of them put their toe in the water, the growth from the toe dippers will swamp the reduction from the token maxers. Companies like Coinbase are going "oh my god, we spent so much, let's cut it by 50%" β that has a real impact if you're Anthropic or OpenAI. But there's ten companies in middle America saying "we have a ChatGPT subscription, maybe next year we'll try some of this Codex stuff." An updated cohort analysis for Anthropic on the revenue build would be the single most useful piece of information you could have β run that through a cube and you could trade the QQQ for the next 12 months, because that's where it's all happening.
The biggest round of the week: Travis baby is back. Travis announced raising $1.7 billion for Atoms, an industrial robotics company, led by the one and only Andreessen Horowitz, with Ben Horowitz joining the board, plus Bain Capital, Fifth Wall, and a load of other firms. Is there room for everyone in a $1.7 billion round? It's kind of an industrial holding company β big picture, it's physical AI, AI for the real world, doing a bunch of different robotics businesses, very different robotics businesses β some around cloud kitchens and food preparation, some around mining. Travis is totally correct that it's not humanoids, it's specific purpose robotics β you actually need specific purpose autonomous machinery for B2B in general, that makes sense. What's not as clear is why it makes sense to have the same holding company doing mining and food prep, other than the fact that Travis is amazing and can raise capital cheaply.
Would you have broken your rules for your LPs to put money into this? No, probably not. There's a lot of feeling now that robotics is going to happen quickly β real and significant, but the "GDP of the real world is bigger than the software world" line undersells how long it takes: it turns out 2% of the world is software and the other 98% is real, and it takes a lot longer than you realize to roll out robotics in the real world. It's not clear that putting a bunch of different companies together in the same place makes it any better. It is doable because Travis can raise money at a great price, but a great price for the fundraiser might not necessarily mean a great price for the investor.
There's a connection to a Wall Street Journal article about bringing CEOs out of retirement to run big companies β Cracker Barrel fired a CEO despite the stock being up, brought in someone who ran Outback's parent company; PayPal found someone on his Montana ranch to come back and run it. There's this bimodal trend: bets on 20-year-old founders like Etched's, and cursor bets β but when Bezos, Travis, or Elon raise their hand and say "I'm going really big, guys, this isn't about making a couple nickels, I'm building something massive and it needs billions of dollars," you give it to these iconic seasoned veterans and face east that it works out. Giving Elon money for Twitter was facing east β there was no rhyme or reason for that deal β and the Boring Company is crazier than Atoms.
Twitter is not worth today $44 billion, and not even close. In terms of buying something, you bought an asset that went down in value. As it happened, you got a 3x because he rolled it into x.ai. But these big-name things are working not because the facts are working β they're working because the market is continually willing to enable that process. And if the market changes, you don't have value.
The fun fact: Benchmark's amazing-looking fund from around 2012/13 had both WeWork and Uber in it. They famously swapped out Travis as CEO β to the undying hatred of Emil Michael, now at the Department of Defense β and it went on to be an amazing company. They didn't swap out the WeWork guy, who went on to pretty much fail, even though he personally took out $500 million; the deal didn't work. So at one point there was a fund with two potential mega fund returner deals β one worked, one didn't, and the one where they made the change worked. Fast forward, and Andreessen Horowitz have backed both CEOs β Adam at WeWork, because "we think you can do it again," and now Travis. A very direct tweet last week said essentially, "we should have done this deal in 2010/11" β tantamount to saying "we deeply regret taking money from someone else who fired us, even though it turned out to be an $80 billion outcome." Benchmark took out 315 from WeWork (secondary to SoftBank, 315 out of 17 in β still a great return), while Uber made Benchmark roughly 640x versus around 25x with WeWork. It just shows that with enough momentum in a bull market, if you take your winners off, you can do well on everything.
Capital ain't exactly starved in other areas of the market either β $21 billion, above the target, for Francisco Partners. Jason gets confused by the messaging β a big part of the thesis is that AI won't kill software, and there's efficient ways to deploy it, that you can buy gems growing 14% and hook up Kimi and the Moonshot API to magically reaccelerate growth to 70%. Every week that goes by, he feels like the past is the past β time to leave it there and let the markdowns be the markdowns, raise another fund, hopefully with an Uber in it. Rory pushes back: nothing's going away, but remember the venture game is all about finding things that explode in growth, price matters only at second order. In the PE business, it can be the other way around β a company growing at 7% that you buy dirt cheap and get to 20% growth with good cash flow margins by applying leverage, you can make your IRR. It's the opposite end of the life cycle.
We just turned off Marketo. They raised our prices from $22,000 to $80,000 since 2020. We left. I bet they've lost 20% of their customers over that period. The blood is beyond out of the stone β the stone has crumbled because all the blood has been squeezed out and it's turned to ash.
If you've done five years of price increases and that's all you've got for revenue growth, you're probably closer to the end than the beginning. A lot of these PE firms run a test β if they've raised prices and no one's blinked, that means they can continue to raise prices, but Jason argues that could be a counter signal: you should test whether you can add net new revenue, net new modules from customers, or whether you're just screwing them, because if you're just raising prices, it's going to end at some point. Target selectivity will be really important β it won't be nearly as big or easy a business as it was in the last decade and a half. Where does ServiceNow fit in? The problems it solves are so sufficiently complicated you need them β you just can't run your business without ServiceNow. It abstracts away so much complexity. Analytics, though, is the easiest thing to vibe code away β the ancillary in-between products like analytics, to-do lists, and task management will go away, but islands like ServiceNow and Salesforce will stay.
There is a model Jason does believe in for PE: not buying the 15-17% grower five years into the price-increase cycle with no net new customers, but the 40% grower β one that's still working today, just not growing at the rate you'd like, with an agentic product doing 35-50% growth. There may be a moment where you can get a very attractive multiple on that property before it goes into terminal decline β that's where he'd spend his emotional energy: what's still growing approaching 40%+ at scale where the founders are burned out and made the transition, but not growing exactly at the rate of the hottest startup in its class.
Life's too short to struggle β can't believe that's actually advice. It makes Jason sad when a founder with material ownership quits a 40-50-60% grower to do something hotter, even in the age of AI where the opportunity cost argument is real.
Jason: most of the founders he's worked with over his career who quit something pretty good to chase the shiny penny β they're not all Ilya β it's been a net negative all the times he's seen it. His worry in the age of AI is that there's no perceived downside β just quit everything, go found an AI company. But if you've got $20 million, $50 million, $500 million in revenue, maybe see if you could build that in-house. Lilian Weng left Thinking Machines, leaving only two of the original six co-founders. Same thing? Well, Thinking Machines isn't exactly a 40% boring SaaS β and she probably has 1% if she's a late co-founder. What's Thinking Machines worth on paper? $8 billion. So she's leaving $80 million behind β "I'd leave that behind, it's nothing," jokes Jason.
Rory clarifies: there are really three different things bundled in the quitting comment. The least obvious is the founder at scale growing 40% β it's not obvious that chasing the next shiny thing will be better, you've created something of value. But he doesn't think that's what Pincus was referencing β it was more about how long you keep trying to get product-market fit before you say it's just not there. The nuanced answer is you shouldn't do anything out of duty; you should do it because you believe you have a plan to converge on something, and when you don't have a plan, you shouldn't tie yourself to the mast just to keep going. Thirty-five years ago, when Rory had his own business, he was a mediocre manager and stuck at it two years longer than he should have β wasted years, just out of a feeling that he owed it to himself not to quit. Finding a way to step back and ask "am I doing this because I still believe in the mission, or out of a sense of obligation with no way to win" β in the latter case, put up your hands.
If it was just about me, I would have quit both my startups. Certainly I would have quit venture investing. I certainly would have quit Echosign β my founder walked out the door after eight months. He was right, this category was never going to take off. The only reason I had any economic success is that out of obligation in part, I kept going.
Rory: he swucked at his own thing two years longer, went bust, and looking back, that thing would never have worked. It's the old Kierkegaard thing β life is lived forward but can only be understood in reverse. When you look back on things that failed, some you realize weren't just unlucky, there was just nothing there. Jason counters: he's never quit, and everything he's done would have failed if he had. He's never had a wildly successful outcome, never been a billionaire, never made investors less than 5x, never had a single "but everything almost failed" β the people who quit, quit. Rory takes the opposite position: he's never quit either and failed at some things β doggedness until the end of time is good but doesn't actually guarantee a win. Did you not learn more from the extra two years, though? Rory: no β "there's a great line someone gave me when I failed: experience is what you get when you don't get what you want." He'd done all his learning two years earlier; the last two years were just hell on earth. Jason disagrees that this is one of the dumbest things β you do learn a lot from the almost-failures, from portfolio companies that turn around. But you don't need to live it for years to learn it; you can process through it. On balance, Rory is emotionally more in Jason's camp than Pincus's "just quit" camp β his point is merely that if the only reason you're hanging on is duty and you see no hope, you're actually going to fail anywhere. That's his theory, but they can disagree.
One question for Stripe, for Rory β comparing it to Adyen, it doesn't seem so wildly overpriced given these numbers. It seems about appropriately premiumized to Adyen β always thought they were peer companies, but Stripe's much better, with much higher profitability. What's happened, Rory explains, is Stripe hit a sweet spot: they charge more, have more smaller merchants and higher pricing, and for a long time were less profitable despite that because they were "Silicon Valley soft" while Adyen was a hard-nosed bunch of Dutch people. About four or five years ago, the Collison brothers focused on efficiency and made it an efficient company. So now you have a company with good pricing (2.75% is attractive) and efficiency. The third key ingredient happened in the last two years: they signed up all the AI companies selling stuff online β they shouldn't be getting anything like the money they're probably getting from OpenAI and Anthropic in interchange fees, but who's got time to optimize that stuff? They're designed into the flow of companies that are just printing money, so they're printing 2.75% of that money. Growth accelerated, and with an efficient leveraged cost structure and revenue taking off, it all flows to the bottom line β five years ago Stripe looked expensive relative to Adyen, but now it has super strong growth, good pricing, and wonderful margins, driven in particular by this lift from online AI spending.
$10 billion felt like a lot, but good luck to them. From a business model perspective β the kind of front-end API to aggregate a lot of complexity, what OpenRouter does for LLMs is what Stripe does for money and what Twilio does for telecoms β it makes sense from a company model perspective.
Will the OpenRouter deal happen? It's gone super quiet, and everyone's released their own routing product β Cursor released one, Merge.dev released their own, and it seems to be a very commoditized market very quickly. From a business model perspective it makes sense. On timing: the time from an intentional leak by a VC to deal closing is more than one week on average, so this certainly appears to be a leak to generate a pseudo second offer, potentially to justify a premium price. Every company does what you'd expect β offers an acceptable but mediocre price from a venture perspective. Maybe they offered the last round price, maybe two billion, and the ask was ten. Someone leaks it β that's how you do things today. Deals don't close the day after the leak in very limited experience; you have to sequence the leak properly, or it doesn't create another deal that closes β it's part of a price negotiation, much better than a banker pretending they got someone to add you into the deal. You need a couple of weeks for that "leak energy" to work. If it's real, it probably does go through β you don't leak a fake deal, it doesn't accomplish anything. The leaking strategy works for a good-but-not-great offer because you only have so much leverage. Big company M&A in-dev isn't brutally slow β when an email comes in saying someone on the list is in play, it doesn't mean Adobe or Google will buy them, but they spring into "deal mode" and make a decision within a couple of days, which can at least get you a paper counter offer. Any big company can move in a week to sign a term sheet, not to close, when they're in deal mode and there's a backend constraint β otherwise it's months.
Even though I love the Stripe vibe more β they're Irish, I'd love them to kill it, and they are killing it β I think the beauty of Revolut is you have a whole continent full of overpriced, crappily run banks that you can just roll over. You've got 500 million Europeans getting shafted on financial fees. The TAM for payment services in the US is enormous too, but it's just marginally more competitive.
Final one: you can own Revolut at $115 billion or Stripe at $165 billion β which one do you want to own? Both are amazing companies, neither is at the core an AI company (though Stripe's getting a lift), both are really well-executing fintech companies. On a TAM basis of plus or minus $100 billion, there's just more compounding in those 500 million exploited Europeans. Jason: honestly, at the end of the day, the moat at Stripe may be a little lower β the network effect may not be as strong as it seems, whereas banking has marginally more powerful moats and is working on some network effects too. So his vote would be Revolut, because Stripe just has to continue to execute at an outstanding level, whereas the network effects and moats around banking are there β Stripe has invested in everything from Atlas to their own router to create network effects, but he's not sure they're truly there yet.