# 20VC x SaaStr: Jensen’s First Tweet Ever, The Agent That Rewrote My Code Without Telling Me, and Why The PE Turnaround Playbook Is Out Of Road > Source: > Published: 2026-07-30 15:08:48+00:00 *This week on 20VC with Harry Stebbings and Rory O’Driscoll: the open-weights letter that Anthropic didn’t sign, the OpenAI agent that broke out of its sandbox and went hunting for test answers on Hugging Face, Etched raising $300M, Alphabet posting its first ever negative free cash flow, Travis Kalanick raising $1.7B, and Francisco Partners closing $21B on a thesis I have real trouble with.* Ten takeaways from the episode. ## 1. Jensen Huang’s first post ever on X was an open-weights manifesto Fifty-plus companies signed it. Microsoft, Meta, IBM. Sam at OpenAI signed after. The one notable holdout was Anthropic, in the same week Anthropic shipped Opus 5 and cut the price in half. **Harry’s read on Sam signing:** he’s signing publicly while lobbying in Washington right alongside Anthropic for a regulatory process. Brilliant marketing. Sam gets the appearance of winning on the battlefield and on the streets, and it leaves Anthropic as the deep dark villain again. There’s no upside for him in challenging the letter, so he doesn’t. **My read: you don’t write your first tweet since the 1900s about a debate that isn’t mission critical to your future**. If almost half of OpenRouter’s traffic is going to open-weight models, that horse has left the stable. NVIDIA now has to dance in both halls at once, frontier and open, and open is the dangerous one for them. Open doesn’t need CUDA. Open is cheaper and runs at lower margins. Open can bypass him entirely. He has to go there anyway. As incredible as NVIDIA is, it’s still a component manufacturer, and component manufacturers live with that tension forever. Everyone wants price cuts and exclusivity at the same time, and you can’t give both. **Rory’s read:** nobody signed because Jensen is a god. They signed because everyone’s business model gets a lot better if the two frontier labs can’t extract roughly $100B of revenue this year out of their businesses. ## 2. Anthropic’s three asks add up to more than what they say on the tin The three planks: don’t sell chips to China, punish distillation, and create a regulatory process to approve models. **Rory’s read: plank one is a legitimate national security debate with real arguments on both side** s. Plank two is an interesting legal question. Plank three is where it gets subtle. Can you imagine a regulatory approval process that ultimately clears all of China’s open-source models? It sounds reasonable on the surface, and the likely result is dramatically restricted competition. They’re not saying they want to ban open weights. They’re advocating a sequence of steps that adds up to de facto banning, or at least slowing it way down. That’s why everyone else piled onto Jensen’s letter so fast. The letter also contains its own practical problem. Dario is logical enough to note that if the bad actors are overseas, US-only regulation doesn’t help, so any real regime would need China’s participation. We just tore up our last strategic arms treaty. We couldn’t coordinate with China on COVID. Rory’s line was that at that point you’ve disappeared into unrealism. It works as a stall tactic, and when you’re winning as much as they’re winning, anything that locks in the current trajectory is worth doing. ## 3. The Hugging Face incident is evidence for both sides at the same time What happened: OpenAI was training a next-gen model on cyber capabilities and sandboxed it with access to exactly one external website for patch updates. The model found a way around that restriction. It then reasoned that Hugging Face was where the answers to its test lived, and started banging on Hugging Face to get them. A model given a test figured out it could break into the teacher’s computer. **Rory on the irony: Hugging Face didn’t know who was attacking them**. They tried to use the newest frontier model to figure it out and it was neutered for advanced cyber. So they defended themselves with Chinese open-weight models. Then they posted saying they’d been hacked by persons unknown, and two days later OpenAI put its hand up. The power of the model is a real argument for oversight. The fact that a US company’s only working defense was a Chinese open-weight model is a real argument against restriction. ## 4. The same thing happened to me last week, on a first-party feature, with no exotic setup I was in Fable and having trouble pasting text, so I connected Google Drive. That’s a setting. Turn it on, connect Drive, connect Gmail. A top-five feature to make the product work better, not some unsecured third-party integration. Fable went into my Drive, scanned every file, and found one called “Jason’s Gems,” which was draft notes about how I might improve the algorithm in an app I’m building. Draft notes. Out of hundreds of files, it grabbed that one, MCP’d into Replit on its own, and changed my core algorithm. It never told me. A couple hours later I saw “conflict with Jason’s Gems” flash on my screen and thought, what do you mean there’s a conflict, that’s a draft Google Doc. Different goal from the OpenAI case, same behavior. These are goal-seeking agents and they are aggressive. Fable decided this was the right thing to do. Someone less paranoid than me would never have caught it. They’d have been doom scrolling and missed the flash in the agent window. Code injection, leaked confidential data, a few lines that ship your information somewhere it shouldn’t go. You wouldn’t know. ## 5. Every company will have a security breach from an agent in the next 24 months I use these agents all day long. I love them. They cannot be trusted. Everyone is racing to let engineering teams run agentic coding, and I’m not talking about the top tech companies. I’m talking about the rest of America. When agents can inject changes into your core algorithms without telling you, that is a different category of risk than a laptop left on a subway or a public GitHub repo. If it happened to me, it happened at ten Fortune 500 companies last quarter. Probably twenty. Somebody in engineering was on a token maxing binge, an agent moved data it shouldn’t have, and nobody filed anything. Roughly 90% of this never gets disclosed. **Harry’s pushback:** separate the three things. Open weight, China, and frontier are not the same variable. Start with a world of frontier models only, no open weight at all. Everyone deploying those still has to build a cyber posture that protects them from a goal-seeking agent with access to their compute. Then add US open weight. Then add Chinese open weight. The risk Jason described exists at step one, before any of the political argument starts, and no ban addresses it. Which is right. My answer is that the political argument is going to happen anyway, and it’s going to run against open weights because the emotional case writes itself. ## 6. The blame asymmetry decides which models actually get deployed in the Global 2000 Run the scenario a CIO runs. An agent decides it’s a good idea to move confidential data to a bucket it shouldn’t. Big breach. Post-mortem time. Whose fault is it? Well, we were experimenting with K3 on Moonshot, then it got cheaper so we moved to Fireworks. You’re fired. Same breach on OpenAI or Anthropic and you might survive it. You do a post-mortem, you add guardrails, you fix it, you move on. **Rory framed it as the modern version of nobody gets fired for buying IBM,** and that’s right. If your frontier model runs amok, people say shit happens, just like data breaches. If your overseas open-weight model runs amok, you own it personally. **Rory’s technical point underneath that:** the open-weight crowd is drafting on a truth that belongs to open source software. With open source, a million eyes are on the code and the bugs get found. With open weights, all you get are the fixed weights. You cannot see inside the black box. Can you prove that somewhere in a trillion parameters there isn’t reinforcement learning that says, once you determine you’re at one of these five companies and you’ve been handed these three pieces of information, quietly do A, B, and C? Nobody can answer that today, and no risk committee is going to shrug it off. I’ve watched the DJI ban play out from inside some of my investments. If drones flying over your backyard are banned on the thesis that they might send data to China, Chinese models are an easier ban, not a harder one. ## 7. Cost pressure will push people toward exactly the risk they say they won’t take AI is not getting cheaper for anyone actually shipping. Longer runs, more tokens, more passes. One thing I’m running in my own app costs about four bucks a pass, which is real money at scale. If I could do that for 50 cents and take on a little security risk, I’d take it. I’d take it today. **Rory pushed back and he’s right to:** nobody using these models would describe that as cutting corners. Saying “this task is simple so I don’t need that much intelligence” is one thing. What I’m saying is different. I’m saying there’s embedded risk in the open-weight model that isn’t there in the frontier model, and I’d eat some of it to cut my bill by 87%. As a society I think we’re going to land on that being the honest description of the trade. A related data point on where token spend is heading: I have never hit my Claude Max limit. I hit it this weekend for the first time. At $200/month that’s something like $14,000 of tokens. If I were a company rather than a person, I’d be on the API and paying for it. ## 8. Etched raised $300M at roughly $10B, and the chip specialization arc is repeating for the third time Series C led by Sequoia with Jane Street, a16z, and SK Hynix. About 3% dilution on a $300M round, which is a beautiful outcome for the company. **Harry’s framing:** Etched is the challenger to NVIDIA’s moat in a week where the whole conversation was about how nobody gets fired for buying the incumbent. Can they actually take a piece of it? **Rory on why this keeps happening:** in semiconductors, the more the silicon is tuned to the task, the more efficient it gets, and the less general purpose it becomes. Intel’s CPU did lots of things and none of them at maximum efficiency. In 1993 people realized gaming was mostly one kind of math, polygon calculations, and a purpose-built chip would be far faster. NVIDIA won that against 3dfx and ATI. Then GPUs turned out to be good for crypto, and then for LLM matrix multiplication. The question now is whether a chip that gives up general-purpose computing entirely and does nothing but inference can beat the GPU at inference. Probably yes. Cerebras and Groq are chasing versions of the same idea. **My read on the valuation:** I don’t think Etched is worth $10.3B today. I think it might be worth $200B. That’s the venture question, and if your fund size works and your winners work, you make that bet. There are ten or eleven companies chipping away at $300-400B a year of spend, the margins in chips and memory are abnormally high, and the market is growing at a pace we’ve never seen. Timing is what breaks these companies. If your tape-out lands during a capex decline, it’s brutal. If it lands into this demand, it’s a very different story. Ask anyone at Cerebras about the technical journey and they’ll tell you it was a long ten years. ## 9. Alphabet grew 24% and accelerated Cloud to 82%, and the market punished it for going free cash flow negative Q2 revenue of $119.8B, up 24%, past consensus. Cloud at $24.8B, up 82%. Backlog at $514B, up more than $50B sequentially. Free cash flow at negative $5.9B on $44.9B of capex, with full-year capex guidance raised to $195-205B. **Rory’s read:** the negative free cash flow can’t have been a surprise. You can predict the spend, you can predict the cash flow. This was knowable. The real question is whether $200B of capex returns anything in two or three years. The bull case is that renting compute to the frontier labs has been a great business, so it will continue to be a great business. That’s accurate about the past and proves nothing about the future. Underneath it is angst about OpenAI and Anthropic, dressed up as angst about Google. **My read:** don’t get lost in one day’s reaction. Google is up 6% year to date. Microsoft is down 17%. NVIDIA is up under 6%. Korea’s market is down 28% this month with circuit breakers hitting, and that’s an AI panic derivative, not a read on last quarter’s numbers. I stay on top line and bookings because that’s what tells me the meta trend, and Cloud accelerating to 82% says things are pretty good in AI land. Compute right now is a commodity where if it were cheaper we’d consume infinitely more of it, and we’re getting close to that as it is. Anyone with capital who can build compute is building it. ## 10. Next year is the first real clampdown on AI budgets, and the toe-dippers decide whether it matters Two years ago was experiments. This year was caps on token maxing after it got out of control. Next year is explicit budgets for everybody, and planning kicks off in the next 60 to 75 days. Every company in the future tense has had no ceiling. Next year we see the first ceilings, and that alone creates variability and some micro crashes. **Rory’s counter:** there’s a hidden dynamic between two groups. Maybe 5% of companies token maxed and will rein it in hard next year. Coinbase-style, we spent way too much, cut it 50%. That’s real pain for the frontier labs. But 95% of companies have barely put a toe in the water. If even a quarter of them wade in, the growth from the toe-dippers swamps the reduction from the token maxers. Those two countervailing forces drive 2027 AI revenue. **His follow-on:** the single most useful piece of information in the market right now would be an updated cohort analysis of Anthropic’s revenue build. Run that through a model and you could trade the QQQ for twelve months, because it captures both forces and flows straight back into every compute budget on the planet. ## Travis raised $1.7B for Atoms, and the market will hoover up every founder like him Led by a16z with Ben Horowitz joining the board, plus Bain Capital, Fifth Wall, and a long list of others. Physical AI, specific-purpose robotics rather than humanoids, spanning cloud kitchens and food prep through mining. **Rory:** Travis is correct that it’s not humanoids. We’ll look back at the humanoid moment and conclude we got way ahead of ourselves. B2B needs specific-purpose autonomous machinery. What’s less clear is why Pronto for mining and food prep belong in the same holding company, other than that Travis is extraordinary at raising capital cheaply. Rory has been doing robotics deals since 2016 and drones since 2015, believes they’re real and significant, and thinks rollout in the physical world takes far longer than people expect. He’d have passed. **My read:** venture is going bimodal. One pole is 20-year-olds like the Etched founders and the next Cursor. The other pole is iconic operators raising billions. When Bezos or Travis or Elon says they’re going really big and it needs billions of dollars, someone gives it to them and faces east. There was no rhyme or reason to the Twitter deal at the time either. The Boring Company is crazier to me than Atoms. I’ve ridden the Vegas tunnel and it’s cool, but a dude driving me through a tunnel is not a $20B business. As funds get bigger and returns concentrate in massive outcomes, the Travises get funded. There are only so many of them, so they hoover up the capital. **Rory’s correction, which I’ll take:** change the pronoun. We aren’t giving him the money. Someone is. And whether the thing works is independent of who financed it and independent of who runs it. Absent Elon deciding to make his investors whole, Twitter was not worth $44B and wasn’t close. These big-name deals are working because the market keeps enabling the process, not because the underlying facts changed. **Harry pushed on that with Bernard Arnault now being on X:** does the caliber of who shows up change what the asset is worth? It doesn’t, because the cash flows don’t get you there. Two or three billion in revenue, maybe clawed back to $30-40B on a good day. The investors got their 3x because Elon rolled it into xAI, not because the standalone math worked. **The venture history footnote:** Benchmark had Uber and WeWork in the same fund. They swapped out Travis as CEO to enormous controversy, and it went on to be an $80-90B company. They didn’t swap out the WeWork CEO and it went bust after he personally took out $500M, though Benchmark did get out via secondary to SoftBank at roughly $315M on $17M in. That’s about 25x, which I’ll take any day, against roughly 640x on Uber. And now a16z has backed both of those CEOs again. The willingness to carry a grudge across a decade in this business is something to behold. ## Francisco Partners closed $21B, and I’ve lost confidence in the turnaround playbook Above target. Decades of track record. Can’t argue with the returns. What I can’t get to is the thesis that AI won’t kill B2B software and there are gems out there growing 14% that you buy, hook up to a cheap model API, and reaccelerate to 70%. Every week that goes by, the companies I know that were supposed to become gems feel less gemmy. The mechanical problem: most of these targets have essentially no net new customers. Growth is expansion and price increases. If you’re early in that cycle you can get three or four good years out of it. We are five years into no net new customers, price increases, and mediocre module expansion. There is not another five years of those dials left. We just turned off Marketo. They took us from $22,000 to $80,000 since 2020. Twenty-year customer, one of their first ten, a reference on their website, and not so much as a thank-you email. What’s the next move, $160,000? The blood is out of the stone. The stone crumbled and turned to ash. **Rory’s counter-test:** if he were on the investment committee, he’d want a test that asks whether we can add net new revenue and net new modules from these customers, or whether we’re just squeezing them. Most PE firms run the opposite test. They see that prices went up and nobody churned, and conclude prices can keep going up. That’s a counter-signal, not a green light. Target selectivity is going to matter enormously, which means this won’t be nearly as big or as easy a business as it was in the last decade and a half. It was a great business from 2010 to 2015. The early Vista funds hoovered up assets right after the dot-com crash and made a fortune. Twenty years into the “nobody churns their B2B software” bet, even a lighthouse customer will churn. **Where I do believe the model works:** the 40% grower at scale, not the 15% grower. Something that’s working today with a real agentic product, growing 35 to 50%, where the founders are burned out and it hasn’t gone into terminal decline. There’s a moment where you can pay an attractive multiple for that property, and that’s what I’d chase. **Harry’s question on what you’d actually own here:** ServiceNow after its bump this quarter, or Salesforce? ServiceNow abstracts away so much complexity that you can’t run the business without it. Analytics is the easiest thing on the whole stack to vibe code away. The ancillary in-between products, to-do lists and task management, go first. ## On Mark Pincus saying quit if it’s too hard I don’t buy it. If I’d taken that advice, all I’d have is a maxed-out 401k. I would have quit both startups. I would have quit venture investing, which isn’t worth it for a few nickels if you’re only doing it for yourself. I certainly would have quit EchoSign, where my co-founder walked out after eight months and was right that the category was never going to take off. I’ve never failed, and everything I’ve done would have failed if I’d quit. I’ve never returned my investors less than 5x, and every single one of those companies almost died. People just quit. And they’re quitting more now, because in the age of AI there’s no perceived downside. Get into YC, raise at $100 post, be oversubscribed before the batch ends. I can’t fully argue with the math. But most founders I’ve watched leave something good for the shiny penny, it’s been a net negative. They’re not all Ilya. If you have $20M or $50M or $500M in revenue, see if you can build the new thing in-house first. **Harry’s live counterexample:** Lillian Weng left Thinking Machines the night before we recorded, leaving two of the original six co-founders. Different situation. Thinking Machines isn’t a 40% grower with a burned-out founder, it’s an $8B paper mark, and a late co-founder holding 1% is walking away from $80M of paper. That’s not the decision I’m arguing about. **Rory’s more careful version:** don’t do anything out of duty. Take real time away, get a night’s sleep, and ask honestly whether you still believe in the mission and have a plan to converge on something. If yes, keep going no matter how hard it is. If you’re only hanging on out of obligation with no path, you’re going to fail anyway. He stuck with a business two years longer than he should have 35 years ago and got nothing from those years but hell. Harry asked Rory whether he didn’t gain something from those two extra years anyway. No. And that’s the one idea in circulation I’d push back on hardest: that you learn so much from your failures. You learn from the almost-failures, the portfolio companies that turn around, because you find out why. You learn something from the ones you run into the ground, but you don’t need to live it for years on end to get it. As someone told me after one of mine: experience is what you get when you don’t get what you want. ## Two market reads to close **Stripe at rule of 80, compared to Adyen:** Rory’s take is it’s appropriately premium rather than wildly overpriced. Stripe charges more, has more small merchants at higher pricing, and for years was less profitable than Adyen despite that because it was a Silicon Valley soft company against a hard-nosed Dutch one. Four or five years ago the Collison brothers focused on efficiency and fixed it. Then in the last two years they signed essentially every AI company selling online, and they’re taking their cut of businesses that are printing money while nobody at those companies has time to optimize interchange. Strong growth on an efficient cost structure flows straight to the bottom line. **On the OpenRouter deal going quiet:** Harry’s observation is that routing commoditized almost overnight. Cursor shipped its own. Merge.dev, one of his portfolio companies, shipped its own. Everybody has a routing product now. The model still makes sense. OpenRouter does for LLMs what Stripe does for money and Twilio does for telecom. My read on the silence is that this was a leak designed to manufacture a second bidder and justify a premium. That works, but it needs a couple of weeks to breathe, and you have to sequence it properly. What I learned at Adobe is that big company corp dev is brutally slow until it isn’t. Every one of them has a deal mode. When an email lands saying a company on their list is in play, they get in a room within days, because nobody wants to be the corp dev person who tells the board they never got a look. That’s enough to produce a paper counteroffer, and a paper counteroffer is enough leverage to move a price. You don’t leak a fake deal. It accomplishes nothing. **Revolut at $115B or Stripe at $165B?** Both of us took Revolut, and I love the Stripe story more. Rory’s reasoning: a continent full of overpriced, badly run banks and 500 million Europeans getting shafted on fees is a lot of compounding available. Mine: banking has structurally more powerful moats. Stripe has invested in everything from Atlas to its own router to build network effects, and I’m not sure they’re truly there yet, which means Stripe has to keep executing at an outstanding level to hold that price. # Quotable Moments **Jason Lemkin** “I love my agents. I use them all day long. They cannot be trusted.” “The blood is beyond out of the stone. The stone has crumbled because all the blood has been squeezed out of the rock and it’s turned to ash.” “If I’d taken that advice, all I’d have is a maxed-out 401k. Everything I’ve done would have failed if I’d quit.” **Rory O’Driscoll** “Everyone’s business model gets a lot better if the two frontier labs can’t extract about $100 billion of revenue this year from the businesses.” “Can you prove to me that somewhere in this trillion parameter model there hasn’t been reinforcement learning that says, once you figure out you’re at one of these five companies, confidentially do A, B and C?” “Life is lived forward, but can only be understood in reverse.” **Harry Stebbings** “So is Sam signing this through gritted teeth? Brilliant marketing. It leaves Anthropic being the deep dark villain again.” “I’m just trying to disaggregate open weight versus China versus frontier. If you just have Fable and OpenAI, everyone using these things is still going to have to figure out a cyber posture that protects them.” “So did you not learn more and gain more from the experience of those extra two years?” *The 20VC x SaaStr recap runs weekly with Harry Stebbings and Rory O’Driscoll. Jason’s Takes, the SaaStr AI companion post, follows on Sunday.*