20VC x SaaStr: “I Don’t Buy Dario Anymore”, Mythos Withheld Over Zero-Days, Meta’s Muse Spark Gets Back in the Game, SpaceX at $2T on 108x Revenue, and Why 60% Agents Are the Slow Death Spiral for Public SaaS

With Harry Stebbings, Jason Lemkin, and Rory O’Driscoll

The team is back!! Anthropic unveiled Mythos and then refused to release it. Why? Because it can autonomously discover thousands of zero-day vulnerabilities, including ones sitting in code bases for years, without anyone asking it to. The market reaction: cyber stocks tumbled. Which is backwards. Every security vendor on earth is about to have to buy this capability to survive.

Meta debuted Muse Spark, the first model out of Super Intelligence Labs under Alex Wang. It’s not the best model. Not close. But it’s good enough to prove the $14 billion Scale AI acquisition wasn’t a disaster. Meta is back in the game, and they’re pivoting toward closed source, which changes the ecosystem for everyone who built on Llama.

SpaceX filed for a $2 trillion IPO on $18.5 billion of revenue. That’s 108x. The most expensive IPO at scale of all time. As Rory put it, “The Elon discount rate is zero and the Elon probability of failure rate is zero to get to $2 trillion.”

And in the background of all of it: Jason said the thing out loud. “I don’t buy Dario anymore.” Not as a CEO. As the boy who cries wolf. Meanwhile, public software stocks keep sliding because almost everyone is shipping 60% solutions that customers will not pay for. ServiceNow. HubSpot. Figma. The question every one of them has to answer: can you charge for the agent you just launched? If not, you are in a slow death spiral, and nobody on those earnings calls wants to say it.

Here are the key takeaways.

Top Takeaways

#1. Mythos Is the Rifle vs. Machine Gun Moment for Cybersecurity

Mythos doesn’t just find vulnerabilities when prompted. It finds them autonomously. Point it at a code base and it surfaces zero-days on its own, including ones sitting dormant for years. Some of the Twitter counter-argument was that older models could find the same bugs if you steered them correctly. True. And totally beside the point.

Rory’s framing was the cleanest in the episode: “It’s the difference between a rifle and a machine gun. Both of them can kill someone. But one shoots one bullet and then you stop and reload, and the other just spews bullets out. In the First World War, we all tragically learned that machine guns are… it might be the same thing, but quantity makes a huge difference.”

The speed at which Mythos processes and reasons across large code bases means it will find more bullets. It will shoot more bullets. The Twitter cynicism that “it’s not that different” misses the quantum step in real capability.

Jason’s extension: AI is enabling every single breach possible, every security hole to be found. Not a subset of the hottest companies. Not folks trying to attack OpenAI APIs. Everyone. The MyFitnessPal Cali acquisition got breached two days after close, 3.2 million records stolen, because there was no authentication on Firebase. Superbase has had to deal with the same class of issue. As more databases are built by AI, and more apps are built by AI, the number of issues is going to explode. If Mythos lets bad actors find every site the second it launches with any PII and steal it, we’re entering a transition era where security gets worse before it gets better.

#2. Cyber Stocks Going Down on Mythos News Makes No Sense

Security stocks fell on the Mythos reveal. The logic doesn’t hold. What Mythos actually says is that the go-forward security process will require using Anthropic or another code model to check your code before you deploy, to find these vulnerabilities before someone else does.

Rory’s framing: “If the other side now have machine guns, then you’ve got to build tanks. What security is may change. The vendors who step up and meet the challenge will triumph and the ones who don’t will fall away.”

And the six-month delay matters. As Rory pointed out: “If Anthropic say that their model now allows anyone to find any vulnerabilities and are going to withhold it for six months, that means that in six months and one day every bad guy on the planet is going to be pinging your code and trying to find the bad bits.”

That’s the opposite of a bear signal for cybersecurity. That’s every enterprise security budget getting bigger. The defenses have to match the offense.

#3. “I Don’t Buy Dario Anymore”. The Doomer Fatigue

Jason said the part out loud: “He may well be the second greatest founder of all time behind Elon. Five years to $30 billion. The greatest rage startup of all time. But I am just so burned out of the boy who cries wolf. Every job’s going to be destroyed. Everything is insecure. Enough already. I’ve rotated back to Team Sam.”

What specifically? The endless doom. 80% of jobs destroyed in two years. No programmers by next week. Every podcast. Every interview. Every keynote. The message stops landing after the eleventh repetition.

“Show me something that’s inspiring. His message is uninspiring. That’s the problem.”

Rory pushed back thoughtfully. The doom warnings may be wrong, but they’re sincerely held. And there’s value in the grandiosity. “Half of Silicon Valley is running around thinking they’re inventing the next thing after the atom bomb. I don’t think we’re going to unemploy 50% of white collar workers. I think it’s madness. But that overwrought intensity has allowed them to build a culture where it had no churn, gave them mission clarity, and given them a $30 billion revenue line and a possibly trillion-dollar market cap.”

The Mars analogy landed. “The minute we had to file an S1, someone said you might have to go to prison if you say things wrong. We said we’re not going to Mars. We’re going to the moon. The cynical but incorrect approach would have been: I don’t think they’re going to Mars, therefore I’m not going to do the deal. The more evolved approach is: I don’t think they’re going to Mars, but I do think the Mars vision over 20, 30 years is a rallying cry that allows them to do amazing things in the short term.”

Jason’s ask: what leaders like Dario and Sam need now isn’t more warnings. It’s a positive vision. The Mark Andreessen deflation-and-abundance pitch. The Vinod “it’s going to be okay” pitch. A little inspiration about the good. Not the fluff. Just a glimpse of what we’re building toward.

Harry’s landing on it: “Don’t look at what we say, look at what we do. Ignore the people ringing their hands. You’re not buying the guilt. You’re buying the revenue. And the revenue is amazing.”

#4. The Oppenheimer Moment: Ship the Bomb, Skip the Crybaby Act

Rory closed the Dario section with the best metaphor of the episode. Founders all want their Oppenheimer moment. They want to be Vishnu, destroyer of worlds. They all reference the book, they all channel their atomic bomb.

“My favorite moment in that movie was when Harry Truman says, ‘Get that crybaby out of the White House.’ In the end, Harry correctly says, ‘I dropped the bomb.’ The equivalent is if a whole bunch of people are fired, Jamie Dimon will fire them. He doesn’t need you ringing your hands with guilt, Dario. History won’t say, Robert Oppenheimer, you killed all those people. History will say, you built the bomb, and I dropped it.”

Get over the guilt. Ship the product in a methodical fashion. Be careful. But it’s not stoppable, and other people are going to own the problem.

#5. Trainium Is a $20 Billion Business. Nvidia Lost 10% of Its Revenue.

Jassy disclosed Trainium is now a $20 billion annualized business growing triple digits, with Uber as one of its biggest customers. Trainium is nearly sold out. And Mythos was largely trained on Trainium.

Rory’s precise framing matters. Amazon doesn’t have a merchant silicon business competing directly with Nvidia. Amazon is not buying Nvidia chips for its own internal use, and instead uses its own chips to offer cloud hosting, inference, and training services. With a $200 billion capex budget, about half of which is chips, they buy their own where they can and Nvidia where they can’t.

Anthropic running on Trainium isn’t a statement that Trainium is better. It’s a statement that Anthropic is consuming Amazon compute, and some of that compute happens to be Trainium.

But $20 billion is still $20 billion that didn’t go to Nvidia. That’s about 10% of Nvidia’s potential revenue. Not a mega competitor. Just an in-house bundle at meaningful scale. Nvidia may have its own Fortnite moment on a relative basis. But as Jason noted, “no one is friends with his frenemies and partners better than Jensen. No one’s played this game in the history of mankind of being kind to everybody, pulling back, being right, understanding the dynamics, and still winning.”

And the Amazon-Nvidia relationship is famously difficult. Amazon is the most motivated hyperscaler to not buy from Nvidia. At the margin, 10% is material. But compute is scarce, chips are sold out, Nvidia is at $194, and we’re up against the constraint limit rather than anything else.

#6. Anthropic Going After Lovable May Be Real, Even If It’s Just Maiming

Claude Code is directly competitive with Cursor. And now Anthropic appears to be coming for the Lovable/Replit segment with announced screenshots of a vibe-coding product. Harry said “launched.” Jason corrected him: “You sure it was launched?” Harry: “Well, they announced it.” Rory: “Harry’s a media guy. He thinks when things are announced that they’re real. Jason is a software guy. He actually thinks you have to ship product.”

Eric from Bolt said everyone knew this was coming. The CTO of Lovable knew. It was just a question of when.

Jason’s take on competitive threat: if this were 52 episodes ago, the answer would be “it’s not important enough. They’re going to have to get into databases and hosting and identity management and OAuth and end-user support. It’s a whole bunch of things culturally they don’t want to do.” But the pace at Anthropic is so intense that on a whiteboard, it’s hard not to want to grab a couple billion of extra revenue from vibe coding. It’s 30 days of work for them.

The competitive reality: they don’t have to replace Lovable or Base44 at the prosumer level. They might just go halfway and target the more technical product teams. That nerdier segment is the highest-ROI category in Replit and Lovable. Taking that slice alone is enough to maim them. And maiming hurts.

#7. The 60% Solution Problem Is Why Public Software Stocks Keep Sliding

This was the most important segment of the episode and Jason was the clearest on it.

“Almost everybody’s building a 60% solution. Make is an example. You look at what Claude did or prompting did last November, and you spent the last four to six months building something that’s kind of similar to what these products were like five to six months ago. You can’t afford the tokens. You’re worried about the cost. You can’t build all the features. You’re trying to use cheap models to bring cost down. And you end up with Make.”

Here’s the meta problem for the incumbents. A 60% solution has to be free. You have to include it with your base charge. If you could charge 60% of what Claude or Lovable or Replit charges, that’d be great. But you can’t. A 60% product has to be given away.

HubSpot launching its next group of AI agents is the perfect case. If they’re only 60% as good as standalone solutions, HubSpot cannot charge for them. You can’t get away with charging another 20, 40, 60 grand to a HubSpot customer for an agent that’s fine in isolation.

“When I meet with internal product teams at large companies at scale, they are so effing proud of themselves. They show me their agent. They show me their vibe coding thing. And if I didn’t use any other products, I’d think they were great too. Or if this was 51 weeks ago, I would think these products were great. They’re so insular at their 2,000 person company on their fancy campus. And they’re failing even as they’re proud of themselves, because the market will not pay. They will use your 60% solution but they’re not going to pay for it.”

ServiceNow has the biggest moat. But if its agents are only 60% as good, it’s in a slow death spiral. “Check the box” does not work with agents. The check-the-box feature cannot be monetized in the AI era.

Rory’s crystallization: “Unless you have a product that’s good enough to charge for independently, you won’t have revenue acceleration of any meaningful scale. And if you don’t have revenue reacceleration, you’re in a different valuation bucket.” The gut-level test: can you charge for it? Yes, you’re in the growth bucket. No, you’re in the tragic value bucket.

#8. Moats Keep Customers In. So What. They Don’t Pull New Ones In.

Jason’s line cut through the entire mature-SaaS debate: “The problem with moats is they keep your customers in, but they don’t lure any new ones in. No one’s excited to cross the moat except the folks that want to breach the castle walls. The whole moat discussion is at the edge of moronic. Hooray, you signed a five-year ServiceNow deal. So what? That doesn’t bring in anybody new. It doesn’t bring in agentic revenue. It just means I’m trapped. Prisoners don’t create growth other than at the margin.”

The models have radically accelerated their power since December. Opus 4.5 and what’s coming is a step change. Anybody building to last year’s spec slowly cannot compete. It’s too furious. And the moat protects the prisoners while the new customers go elsewhere.

Would Jason buy ServiceNow? No. Wouldn’t buy any of them. Looking for products that are more than a 60% solution. The world is moving too fast.

#9. Public Markets Don’t Have Access to Growth on the Software Side

Rory’s zoom-out: “The weird thing right now is the public markets don’t have access to the growth side of software. The trade is sell SaaS, buy semis. Which effectively means sell SaaS, buy AI.” What the public markets don’t have yet is AI-native companies, starting with Anthropic and OpenAI.

“You’re comparing the practical values of owning Salesforce with the mythical values of owning this company that’s growing 10x where you’ve never seen the actual financials. No one’s seen GAAP financials, but oh my god it’s amazing.”

Until five or six foundation model and AI-native companies go public, the public market cannot price this correctly. Once they do, investors can choose: Anthropic growing 5x at 30x revenue with huge losses and big SBC, or boring Salesforce growing at 10% with 30% operating margins and no SBC loss. At that point, the stocks will find their real level.

Until then, public SaaS has one path up: achieve growth. Which means passing the Jason test. Build an agent customers will actually pay for.

#10. Wix Did the Buyback and the Stock Fell 23%

Wix spent $1.6 billion to buy back nearly 30% of its stock at $92. The stock fell 23% that week.

Rory’s read: “It’s kind of like an inverse Bill Gurley. Instead of selling stock and then it goes up, you buy stock and then it goes down. That’s like the IPO premium but the other way. Oh my god that hurts even more. Talk about leaving money on the table.”

Wix did one thing brilliantly. They shipped a product (Base44) that got them beyond a 60% solution to 9-figure revenue. The argument here isn’t that the buyback was the wrong instinct. It’s that they should have kept their capital in reserve for six or twelve months more, either to buy another AI product or invest behind their AI product.

Salesforce did the bigger version. $25 billion of debt to buy back stock. On a spreadsheet it looks brilliant. In the short term, the market shrugged. Rory’s flag: “One of the big advantages you have as a public company with cash flow positive is financial flexibility. I wouldn’t trade that for anything.”

If the AI wave rolls on without a blip and you don’t have a growth story, your stock will be cheap in two years. No hurry to buy it now. If there is a blip, the guy with a public currency and $25 billion in cold hard cash can buy five big things in private land. “That’s why I would have kept my $25 billion in my back pocket. You got to win the war, not the quarter.” Come to daddy, I got money.

#11. Muse Spark Is Meta Back in the Game. That’s Worth $14 Billion on Its Own.

Muse Spark isn’t the best model. Reviews said it’s decent, not quite as good as the others, but good on things Scale AI was working on a year ago and weaker on newer areas the advanced labs have been developing. Makes sense. They implemented a year-ago state of the art.

Rory: “If you decide as Mr. Zuckerberg that you want to be in this game, then you achieved your mission. You spent $14 billion and you’re back in the game. Now you got to move up the league table. Had you done all this and done a Llama 4 which was disappointing where people are like it’s not even credible, then you’d have felt like a fool. And you don’t.”

One underappreciated shift: Meta is moving much more closed source. Llama was the American version of open source. Now they’re pivoting. That has significant implications across the ecosystem.

Jason’s strategic read: “If I’m Zuck, I don’t want to be stuck buying tokens from Anthropic or OpenAI. I’m Meta. I have the dominant consumer advertising platforms. I have the dominant social networks. I sure better own this. It turns out it is not a commodity. This is core to our existential existence just like it is for Google. And I don’t want to wither.”

The scorecard on Meta big bets: $1 billion on Instagram, win $100B+ (maybe $400B). $70 billion on Reality Labs/Meta, still lose it all. $14 billion on this with a clear win. Somewhere in the middle of those outcomes is a great portfolio. And remember: a lot of folks thought AI would kill Google Search and maim Facebook. Today Google Search is a better business than it has ever been, and so are Facebook and Instagram. This is not a time to retreat. Get the lance out, go straight into battle.

#12. OpenAI’s Path to $100 Billion in Ads Is Plausible. It’s Also Not Enough.

OpenAI is projecting $2.5 billion in ads in 2026, $11B in 2027, $25B in 2028, $53B in 2029. The pilot hit $100 million annualized in six weeks with 600 advertisers.

Rory’s math: “The funny thing is I’m mentally giving them 100% credit for getting that. OpenAI is typically not unaggressive in its projections, so let’s assume this is progressively aggressive. This is going to sound really awful, but $100 billion is amazing. It’s not enough relative to the market cap. In the context of a total trillion dollars of ads with Meta at 300, Google at 200, Amazon at 100, and TV has to eat too, that’s a realistic high-end estimate. You need another hundred billion plus from your enterprise business. Consumer alone ain’t going to be enough to feed this beast, because your competitor who’s all in on enterprise is already at 30.”

Jason’s take on execution: “This is our job. Every week we’re going to iterate. We’re going to improve it. It is mathematically possible. It’s not as aspirational as the enterprise stuff. This is doable. And if it comes up short a year or two like Elon, it would suck for the IPO, but it’s not fatal. Because it is achievable, it will be achieved.”

#13. The Denise Dresser Memo Is the Enterprise Playbook Coming to AI

Denise Dresser (President and CRO at OpenAI, formerly CEO of Slack) leaked a memo this week claiming Anthropic is overstating revenue and that OpenAI has capacity while Anthropic is out of it.

Jason’s first read: “It seemed like a flashback to something Marc Benioff might write. More appropriate for Salesforce than for OpenAI.” But on reflection: “This is the exact type of messaging you want to win traditional enterprise customers. I’m pretty bullish on OpenAI in the enterprise because all this enterprise DNA they have, which probably didn’t help a snail’s inch the last 12 months, is going to matter more and more when the models are so powerful and big enterprises are making decisions between a couple of top brands.”

Last year’s playbook was “ask your developer.” Anthropic won that game. Developers picked Opus for their apps, and the revenue followed. Classic Twilio mantra. It worked until it didn’t.

Going forward: OpenAI doubling in size with enterprise selling motions will work better and better in 2027. The traditional playbook wins when you go deep into the Fortune 2000.

#14. Token Budgets and the CIO Coming Back Into the Picture

Aaron Levie had a tweet this week about his road show meetings with CIOs that both Jason and Rory flagged as must-reads. CIOs are token-maxing: creating fixed dollar budgets for AI tokens for the coming year and making departments fight it out per project.

Jason: “This is where the game is going to change again. When most of the budget is essentially rogue, when it’s developer teams picking Anthropic almost universally last year, that’s one thing. When the CIO takes control again of how many tokens across a large enterprise, that’s a very different calculus of which vendor I choose. And if the CIO prefers to buy OpenAI because it’s got a more traditional sales motion and it’s packaged better, then the enterprise is standardizing on OpenAI for 2028.”

Rory’s implication: “If you are correct, the people who should go into guidance counseling are Microsoft and OpenAI, because they need to get their relationship back together again. Microsoft is the dominant path to every single enterprise in the world. They need couples therapy. That’s an indulgence OpenAI can no longer afford.”

Microsoft can go top down. The other guys stole the march on developer love. If enterprise is two-thirds of the game, Microsoft is the key, and OpenAI plus Microsoft aligned is a very hard combination to beat.

#15. Enterprise Might Be Two-Thirds of the AI Market. That’s a Total Flip.

Rory’s biggest aha of the episode. Historically, consumer has been the bigger half. Google’s consumer business and ads is two-thirds of the value, cloud is one-third. Facebook is pure consumer. The big money has been in consumer.

“But consumers fundamentally when I go home want to buy Netflix. When I go to work I want to buy intelligence. It may well be that enterprise is two-thirds of the ball game in AI and consumer is one-third or less. Which is the flip of the last time. Three years ago when ChatGPT exploded, you would have thought consumer was the launching point. But if enterprise is the better place, then yeah, you’re going to have a good consumer business, but the ball game from a customer’s perspective is two-thirds enterprise, one-third consumer. Which is the mirror opposite of the internet.”

This ties to the persona question. Consumers want supportive answers. Enterprises want harsh critique and decisions. Research already shows younger users prefer OpenAI because it’s more emotionally supportive. In business, the opposite is true. Investors don’t want support. They want “this is a stupid deal.” Whether the same underlying model can serve both is an open question. The tuning for each may need to diverge.

#16. Aaron Levie Is a Walking Investment Insight. And If Box Can’t Reaccelerate…

Rory on Aaron: “He’s in that rare combination of grounded enough in terms of 15 years calling on CIOs to really know what they think, and at the same time young enough and flexible enough to really understand what AI is doing. It’s literally like an investment insight. My colleague even said after he spoke at our annual meeting: it was like an investment memo spewed out on enterprise AI.”

Jason’s edge on it: “If he can’t reaccelerate Box with his incredible depth of like 10 out of 10, what hope is there for so many other leaders, so many other unicorns? If he can’t get Box to 20 to 30% growth, I’m giving up on the rest of the world.”

If the best and most insightful enterprise software CEO in the game can’t pull his own company back to high growth, the prognosis for the dozens of mid-cap SaaS companies behind him is grim.

#17. SpaceX at 108x Revenue Is the Elon Discount Rate Made Literal

$18.5 billion in revenue. $5 billion loss, mostly from xAI acquisition accounting. $2 trillion target valuation. That’s 108x revenue. No company has ever gone public at that multiple at this scale.

Rory’s caveat on the accounting: “Pooled accounting doesn’t exist anymore where you retroactively recast the financials as if the companies were together. So that’s only the loss from when xAI was acquired. We actually don’t know the actual run rate.”

And then the valuation frame: “The case will obviously be all the future things. You take these adjacent TAMs. If you give them a 100% probability of happening and a 100% probability of them happening right now with no NPV because it takes five years to make it happen, then you probably get to $2 trillion. If on the other hand, you apply a probability of it not happening and a time value of money, you get to a lower number. The Elon believers are saying, these are the future things, and I ascribe 100% probability of success. It’s basically the Elon discount rate. The Elon discount rate is zero. And the Elon probability of failure rate is zero to get to $2 trillion.”

The existing business is real. A near-monopoly on cost-effective launch, with prices potentially coming down with the next-generation rocket. Starlink is an amazing business. The xAI $250 billion acquisition is harder to defend, but it’s done. The IPO will print. The question is what the multiple looks like 18 months after pricing when the new initiatives (direct-to-cellular, data centers in space) either prove out or don’t.

#18. Everyone Wants to Be Small by Choice

AppLovin: $4.5 million in revenue per employee with 898 people. Rory on the caveat: “Revenue per employee is not a useful metric as a one-shot-fits-all. Cursor does $4 million per employee and Salesforce does $700K. Well, Salesforce has 30% operating margins and Cursor is losing a lot of money because they spend a whole ton on tokens. You can compare companies in the same business on efficiency, but one shot fits all doesn’t cut it.”

That said, Jason’s directional point is correct: everyone wants to be small by choice. Not because the economics are forced, but because leaner is better. AI is an enabler. It lets your best engineers do more. It lets you get rid of SDRs. The core test Harry tweeted: knowing what you now know about a person, would you hire them again? Jason’s version: would you rather replace them with an agent?

“I’d rather have an agent than a mediocre person. Everyone thinks that. Everyone. Not everyone says it out loud.”

In 18 months, everyone will know how to build a good agent. Prompt engineers went from hot to irrelevant in 24 months. Agent engineering is on the same trajectory. The skill isn’t building the agent. It’s deciding which humans to replace with one.

Rory’s reformulation as a metric: “It’s not fair to say to a SaaS company, hey, AppLovin does $4.5 million per head and yours does $500K. But what is fair: last year you were at $500K. This year you better be at $600K because Jason’s telling you. And next year maybe you got to be at $800K. If you’re not making progress on that metric as a software company, you are not with the program.”

#19. Thoma Bravo Shutting Down Growth Equity Is a Signal About Legacy SaaS

Thoma Bravo shutting down its growth equity arm matters. Their core playbook: buy control positions in software companies with leverage, add on other companies, sell to another PE buyer. That’s 90% of the money. Growth equity (non-control, minority positions in late-stage high-growth companies) is a different business.

Rory: “In a period like right now where your core business is threatened, the first rule of threaten is you retreat to the core. PE guys regularly come into growth venture at the top of markets thinking this looks easy and regularly retreat when they discover it’s hard.”

On the underlying private companies (Coupa, Anaplan, Medallia): these are mature vanilla SaaS at single-digit growth. Public equivalents trade at 2-4x revenue. If you bought at 10x with leverage, the math is terrifying. Apply public market comps and deduct debt, you get little or no enterprise value. Big losses in some PE funds are very possible.

The counter-argument: private ownership means acute clarity. Replace management. Buy AI assets. Bolt them on. Make the transformation to AI-first that public markets won’t let you make. If you can add a 100% agent that customers will pay for, you earn your carry. If you can’t, and the underlying software markets don’t bounce back, it’s a train wreck.

Jason’s tragedy framing: “If you have 10,000 or 50,000 or 100,000 reasonably happy customers, and you’ve had 15 months to build them an agentic product with access to the LLMs and 50 of your best engineers, and you didn’t take advantage of your installed base for real, this is a missed opportunity. It’s tragic. Most teams are so mediocre and so paralyzed. People are asking me what they should do with their agents. They should be showing me their agents and asking for constructive criticism. People are paralyzed with fear. They don’t want to work twice as hard as they used to and they don’t know what to build.”

Adobe-era lesson from Jason: when he was a VP at Adobe during the cloud transition, there were 100 surplus amazing engineers. By design or by accident. Sometimes on the back half of their career. Not as raised as the top Replit or Cursor engineers. But great. It’s a crying shame you can’t take a team of six at Coupa or Anaplan, build the world’s best agent, and ship it to your 50,000 customers. The base is there. The opportunity is there. Stebbings/Driscoll/Lemkin AI consulting ($2M a year, checkered shirts, checkered ties, coming in to teach you how to get to a 60% solution by year-end) is not going to get them there.

#20. IPO Order: SpaceX, Anthropic, OpenAI. CFOs and CEOs Cannot Be on Different Pages.

Jason’s call: SpaceX first (already filed), then Anthropic (just added the Novartis CEO to the board, classic pre-IPO move, along with finding an audit committee chair), then OpenAI.

Will Sarah Friar be CFO of OpenAI when they IPO? Rory’s frame: “CEOs and CFOs have to be wildly aligned, and the CFO should probably report to the CEO. Right now the CFO reports to the president VJ, which seems anomalous. Rather than doing yes-or-no on whether she’s in or out, what I’d say is: the IPO is going to happen, this team is going to make it happen, they need absolute sync and probably a more traditional reporting structure. Stop the leaking, stay in sync. Those are going to be the two people on the road show. They should be able to finish each other’s sentences. You should be able to put them in separate rooms like a police interrogation and each should say exactly the same thing.”

Jason’s nuance: there’s often a case for islands of stability in your management team even if they’re not perfect. Every turnover takes a toll. “Sarah is so experienced that things work. Workday finally works. We finally got all these things to work. I got 99 problems. This is not one I want to tackle.” But the daylight cannot be public. If an executive is truly arguing with the CEO in the media, or pulling a Dario-circa-OpenAI move of going directly to the board with “I’ll only stay if I report to the board,” they have to go. No matter how good they are.

And Jason’s advice to any VP who gets this far on the pod: “Whatever you do, do not reach out to your VCs to say there’s problems with your CEO. You’re losing your job. And it doesn’t matter if you’re right. What are they going to do, fire the CEO over you? Zero. Unless there’s fraud, zero. You’re gone. Maybe in three months, maybe that afternoon. Just resign. Resign with grace.”

The Market Keeps Bifurcating Ever More

The market is splitting into two buckets. AI-native companies that can’t go public fast enough. And legacy software companies getting repriced against a backdrop of 60% solutions and flat growth. The public markets know this. The trade right now is sell SaaS, buy semis. It’s a proxy for buying AI exposure when you can’t buy the AI-native companies directly.

Mythos is a preview. Anthropic built something so capable it felt irresponsible to ship. That capability will exist in the wild within 12 months regardless. Every software vendor needs to plan for a world where AI can break and fix everything, and where the only moat is velocity.

SpaceX, Anthropic, and OpenAI will be three of the largest IPOs in history. And every public B2B company trading below 5x revenue is either a value trap or a comeback story in waiting. Which one depends entirely on whether they can build an agent customers will actually pay for.

Speed. Ship. Charge for what you build. If you can’t charge for it, you don’t have a product. You have a line item.

Quotable Moments

Jason Lemkin

“I don’t buy Dario’s endless doom anymore. He may well be the second greatest founder of all time behind Elon. Five years to $30 billion. The greatest rage startup of all time. But I am just so burned out of the boy who cries wolf. Every job’s going to be destroyed. Everything is insecure. Enough already. I’ve rotated back to Team Sam.”

“When I meet with internal product teams at large companies at scale, they are so effing proud of themselves. They show me their agent. If I didn’t use any other products, I’d think they were great too. Or if this was 51 weeks ago. They’re so insular at their 2,000 person company on their fancy campus. And they’re failing even as they’re proud of themselves, because the market will not pay for a 60% solution.”

“The problem with moats is they keep your customers in, but they don’t lure any new ones in. No one’s excited to cross the moat except the folks that want to breach the castle walls. Prisoners don’t create growth.”

Harry Stebbings

“Harry, you’re a media guy. You think when things are announced that they’re real. Jason is a software guy. He actually thinks you have to ship product. No, it’s all about the announcement. Surely you’ve seen that in the last few days.”

“Jason, you said he’s not inspiring enough. What would make you happy? What do you think he should say?”

“Aren’t you loving this, Rory? Jason is like, 10% is material. We don’t need to spend all over 10% is material, right?”

Rory O’Driscoll

“It’s the difference between a rifle and a machine gun. Both of them can kill someone. But one shoots one bullet and then you stop and reload, and the other just spews bullets out. In the First World War, we all tragically learned that machine guns are… it might be the same thing, but quantity makes a huge difference.”

“Can you charge for it? That’s the test. If yes, then you’re on the increasing value scale. If no, then you’re in the tragic value bucket, doing hard things that would make Dario and Sam cry. SBC reduction. Headcount reduction. At best you build a mini version of IBM/CA. Nothing magical will ever happen to a high single-digit growth tech company that’s kicking off cash.”

“The Elon discount rate is zero and the Elon probability of failure rate is zero to get to $2 trillion. If you put a more conservative number in both of those, you probably end up in a different place. You still have the upside. But are you getting paid for the risk?”

This post is part of the ongoing 20VC x SaaStr collaboration with Harry Stebbings and Rory O’Driscoll.