Apr 30, 2026 · 1h 28m · news
Anthropic Raises $45B but Falls Short on Compute & Thoma Bravo Hand Back Medallia Keys to Creditors · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Harry Stebbings, Rory O'Driscoll, and Jason Lemkin analyze the seismic shifts in the technology landscape, detailing how autonomous AI agents are disrupting enterprise SaaS valuations, the high capital intensity of scaling foundational AI models, and the compounding liquidity challenges facing private equity and venture capital markets.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 10.1% of the talking time here. How this is scored →
Harry as informed peer 4.1 Guest teaching 4.6 Guest disagreement 3.2 Harry pushing back 3.5
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 24:42 Rory rejects Sam Altman's compute thesis
Rory forcefully attacks Sam Altman's claim that compute equals revenue, calling it a 'fucking stupid statement' that confuses correlation with causation.
Hardest push from Harry ▶ 14:00 Harry challenges Canva's IPO narrativeHarry directly confronts Jason on a logical contradiction, asking how Canva can have a wildly successful IPO while simultaneously being useless to AI agents.
Biggest teaching moment ▶ 42:49 Jason corrects Harry on Pluralsight precedentJason gently corrects Harry's assertion that Medallia was the first multi-billion dollar private equity software wipeout by reminding him of Pluralsight's earlier collapse.
Harry holds his own ▶ 1:09:09 Rory details YC's market-making strategyRory demonstrates expert command of venture dynamics by explaining how Gary Tan's memo acts as a strategic move to preserve market trust across YC's 25% seed share.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Analyzing OpenAI's Missed Targets and Performance Lag | 3 | 4 | 3 | 2 | Harry sets up the premise surrounding OpenAI's missed revenue targets, but Rory immediately reframes the situation as backward-looking news already well known to AI insiders. | |
| The Rise of AI Agents and API Preferences | 1 | 6 | 2 | 1 | Jason explains his thesis that AI agents rather than human preferences will dictate LLM and API selection, declaring his return to OpenAI because his agents prefer its API. | |
| The Agentic API Grader and Winning the Agent Wars | 3 | 5 | 4 | 3 | Jason details his weekend project building an agentic API grader and dismisses legacy sales tools like Marketo as obsolete, while Rory acts as co-interrogator. | |
| The Impact of AI on Multi-Year Enterprise SaaS Value | 5 | 5 | 3 | 6 | Harry offers explicit pushback by citing a prior guest's defense of multi-year enterprise contracts, prompting Jason and Rory to reframe multi-year deals as merely deferred churn. | |
| Canva, Rippling, and the End of SaaS Pixie Dust | 5 | 4 | 3 | 6 | Harry presses Jason on an apparent logical contradiction regarding Canva's strong IPO prospects versus its lack of utility to AI agents. | |
| Anthropic's Mega-Round and the Astounding Cost of Compute | 4 | 6 | 2 | 2 | Harry frames Anthropic's mega-round, after which Rory delivers a detailed breakdown of the massive CapEx intensity required to expand compute capacity. | |
| Debunking the 'Compute Equals Revenue' Thesis | 4 | 4 | 6 | 3 | Rory exhibits strong combativeness by aggressively rejecting Sam Altman's 'compute equals revenue' claim as a foolish confusion of correlation and causation. | |
| The Hyperscaler Bundling Play and Specialized Chips | 3 | 5 | 2 | 2 | Jason and Rory break down how hyperscalers use custom chips and bundled equity investments to bypass Nvidia's high gross margins. | |
| Investment Showdown: Google vs. NVIDIA | 4 | 3 | 5 | 4 | Rory explicitly rejects Harry's quickfire stock pick premise as flawed before picking Google, while Jason forcefully counters with Nvidia. | |
| Geopolitics of AI: China's Block on Meta's Manus Deal | 3 | 4 | 3 | 2 | The guests analyze the geopolitical fallout from China blocking Meta's acquisition of Manus, exploring employee visas and tech transfer leverage. | |
| The Collapse of Medallia and the Danger of Overpaying | 5 | 6 | 4 | 3 | Jason gently corrects Harry's assertion that Medallia was the first big PE software collapse, after which Rory explains how overpaying destroyed equity value. | |
| LP Portfolio Fallout and the Loss of 'Plan B' Exits | 4 | 5 | 2 | 3 | Rory explains how PE write-offs threaten LP risk appetite and eliminate the safety-net acquisition route for venture-backed SaaS businesses. | |
| The $1B IPO Standard and Founder Key Hand-offs | 4 | 6 | 2 | 3 | Rory and Jason explain why IPO bars now require $400M-$1B in ARR, predicting an era where founders hand over company keys to peers. | |
| Is the Classic B2B Software Thesis Broken? | 5 | 5 | 5 | 6 | Harry pushes back on the doom narrative by highlighting how depressing the collapse of exit and B2B markets sounds, prompting guests to defend power-law returns. | |
| Gary Tan, Y Combinator, and the Fight Against 'Bullshit ARR' | 4 | 5 | 2 | 2 | Jason touches on Gary Tan's memo regarding inflated seed ARR, while Rory analyzes YC's strategic move to act as a self-regulating market maker. | |
| Thrive Eternal and Investing in AI-Resistant Assets | 6 | 5 | 4 | 5 | Rory directly refutes Harry's attempt to compare Thrive Eternal to Sequoia's evergreen fund, explaining that Thrive is targeting physical, AI-resistant assets. | |
| The Economics of Sports: US Franchise Moats vs. European Relegation | 6 | 4 | 3 | 6 | Harry poses a thoughtful bear case for sports media rights in an AI world, which Rory counters by highlighting the financial moats of US franchises versus European relegation. | |
| Evaluating Retail Venture Funds like Robinhood & AngelList | 4 | 3 | 2 | 4 | Rory admits to placing $500 into AngelList's VC product to test the flow, leading Harry to tease his thriftiness before Rory breaks down retail allocation math. | |
| Debating AngelList's High Management Fees and Venture Return Drags | 5 | 4 | 3 | 3 | Rory breaks down VC fee drag math, explaining that AngelList's 3.6% management fee is comparable to traditional 2-and-20 VC models when carry is included. | |
| The Negative Signaling of Top Funds Passing on Follow-On Rounds | 4 | 3 | 4 | 5 | Harry confesses to panic-selling his Figma position down 40% mid-show, prompting Rory to lecture him on public market investment discipline. |
Statements from this episode (51)
O'Driscoll: Ten mega-exits can cover total LP expectations for VC
“It's entirely plausible that 10 super big exits cover the entire nut from the LP perspective, such that it's still a good business.”
O'Driscoll: VC returns are mostly generated in rare hype years
“The dirty little secret adventure I get is how much of your money you make in that one year in 10 when everybody buys the dream.”
Lemkin: AI agents will choose software vendors and models
“More and more the agent is going to choose what models and just what vendors we use.”
Lemkin: AI compute spending may decouple from revenue growth
“It's possible you look back and see that as the first disconnect from compute equals revenue.”
O'Driscoll: OpenAI lost market share to Anthropic in late 2024
“And there's no doubt in the back half of last year, OpenAI failed at the first part of that job. They didn't build great models. And as a result, their traction relative to entropics declined markedly. Their market share declined markedly, right?”
Lemkin: Anthropic took OpenAI's market share by dominating AI coding
“The rate of growth was incredible. And some of that was market share, right? It wasn't all AI. Some of it was market share and it stole market share. And Elon was clear about this, that Anthropic had quote, something special in coding, which underestimates how…”
Lemkin: Anthropic's Claude holds no competitive advantage for most workflows
“As workflows expand to do everything, as agents do more, they will pick the LLM, and I see no competitive advantage to Claude for most workflows.”
Lemkin: Internal AI agents natively choose OpenAI's API over competitors
“Our agents, our AI VP of marketing, AI VP customers, our applications, their font, they love OpenAI.”
Lemkin: AI agents will manage most software workflows by 2027
“20 late 20 26 into 2027 is most workflows are going to be managed by A.I. Agents, not A.I. Agents working autonomously, not crazy open claws blowing up our Mac minis, but running everything.”
Lemkin: Stripe received sole A+ grade in agentic API benchmark
“Over, over the weekend my nine nine six project was I built an agentic API grader where we went, I just had Claude, OpenAI, and Gemini together take the top 120 APIs and grade which ones they thought were the best, which tools I thought, 11 Labs everything on …”
Lemkin: AI agents render Marketo, Outreach, and Salesloft worthless
“They're not going to recommend Marketo for your agent to do marketing automation, ok? In fact, it mocked Marketo outreach and sales loft as tools useless to agents. Ok, this API grid. They said there is no place in the, an agent will never send an email throug…”
Lemkin: Agent benchmark ranked Anthropic ahead of OpenAI and Gemini
“In fact, the order is Anthropic. And so the greater it graded Anthropic just above OpenAI and then Gemini was just down here.”
Lemkin: Atlassian and Monday threatened as AI agents bypass project tools
“The public markets have the right idea, but the wrong direction. The public markets think vibe coding and Claude are their threat. No, the threat is what the agents pick. And actually, if you look at overall, the markets are, they almost get it right. They're …”
Lemkin: Multi-year contracts hide deferred churn and mediocre SaaS leadership
“Churn that is churn that is deferred still exists, and it is where the rent to CEO and the mediocre hide.”
Lemkin: Workday's standard enterprise contract terms effectively total eight years
“Workday does three year contracts up front and five year renewal. So the average Workday customer effectively signs up for an eight year contract. Three and five is their standard term.”
O'Driscoll: Analysts will soon press Benioff on Salesforce AI agent API calls
“And I'm willing to bet in a quarter or two, someone's going to be asking Benioff for how many calls to your agent, your headless API did you get?”
Lemkin: AI agents will not use Canva, Jira, or Confluence
“I don't think any chance an agent is going to use Can that I don't think any, there's any chance an agent is going to use Jira or Confluence unless it's forced to, like it has no need for these products.”
O'Driscoll: The SaaS valuation premium expired in 2025
“And the truth is SAS pixie dust credit expired in 2025.”
O'Driscoll: Canva and Rippling won't fetch 30x revenue premiums
“Good SaaS companies that aren't AI first are going to trade at fair value, which means if you've created value, you'll get value. And I think Canva and Rippling have both created enormous value. So they'll get value, but what they won't get is that stupid, 30 …”
O'Driscoll: OpenAI has compute but flawed models; Anthropic lacks compute
“OpenAI got one job right, they have enough compute, and they got the model wrong, so that's why they're in trouble. And Entropic did it the exact opposite way, right? They got the model perfect. In fact, they may have over-succeeded, and as a result of that, t…”
O'Driscoll: AI foundation models require $4-$5 of CapEx per $1 revenue
“And the capital intensity for every dollar of run rate revenue, it probably takes four or five dollars of CapEx to support that, right?”
O'Driscoll: Compute without a good model yields no revenue, citing Grok
“No compute equals no revenue, but compute and a shitty model also equals no revenue. See Grok for details, right?”
Lemkin: Salesforce bill jumped to $22,000 while seats dropped to three
“Our Salesforce bill went up from 12,000 to 22,000 dollars a year, and our seats went down from 10 to two plus one.”
O'Driscoll: AI agent adoption will sustain strong medium-term compute demand
“The big picture trend is over the next, as agents kick off, the demand for compute over the medium term will be there.”
O'Driscoll: GPUs account for 50-55% of total datacenter buildout Capex
“GPU spend is roughly 50, 55% of total capex on any build out.”
O'Driscoll: Nvidia makes $14B in raw profit per gigawatt of datacenter compute
“And NVIDIA's gross margins are 70%, which means fourteen billion of that per gig is raw profit to NVIDIA.”
O'Driscoll: Zero chance VCs return capital from blocked Manus deal
“The investors who've gotten their capital out, the chances of them having to or being willing to return that capital are zero.”
O'Driscoll: Manus-style cross-border tech acquisitions will never happen again
“What it really is doing is it's less about getting this thing back than it's preventing it from ever happening again. That's the first, last, and only one of these deals that anyone will do, right?”
Lemkin: California will pass billionaire tax and wealth exodus will continue
“I think the California will pass its billionaire tax and the Exodus will continue.”
Lemkin: Coupa, Zendesk, and Anaplan may fail to repay their debt
“Coupa, New Relic, Anaplan, even Zendesk, Avalara, Smartsheet they all look like they may not be able to fully repay their debt”
O'Driscoll: Private equity losses will reduce LP appetite for venture risk
“If the safe part of the business takes some significant hits, it, you know, it's definitely going to reduce the appetite for risk.”
O'Driscoll: Private equity will no longer provide 3x revenue fallback exits
“You can't build a company big enough to go public. Strategics don't care. So you can sell this thing for three X revenues to fill in the PE from. That's not going to be true going forward.”
O'Driscoll: Early VCs must increase portfolio company count as hit rates drop
“From a portfolio construction at the early stage, early broadly defined, you just have to have a higher end count because your probability of getting one right is lower.”
Lemkin: Blocked SaaS founders will increasingly pass company control to peer CEOs
“This is a micro trend that I think is going to accelerate this year is founders giving the keys to their friends, not completely quitting, like eight months after an accelerator that didn't work out, but I mean, it's just, I'm at 40,000,050 1,000,020 million, …”
O'Driscoll: Operators will roll up mid-market SaaS into 30% EBITDA compounders
“You're some guy who's a mid-career operator who's willing to take the pain is going to say, I got this. I'll take these five software companies all broadly speaking in the systems management space. We'll put them together. We'll run them like a hard ass. We'll…”
Lemkin: Only one portfolio startup received a PE offer this year
“In my own portfolio, I've only seen one soft offer this year. Okay. But it was from a PE firm that was exactly that. A startup at scale that is not growing at astronomic rates, but growing at really good rates that is clearly AI enhanced in the AI category got…”
Lemkin: A $100 billion startup exit will happen within a year
“The big exits will solve it, right? The whizzes and the, I mean, we thought whizz was big. Now we have cursor. Now I'm going to win the bet of a hundred billion dollar exit in the next year.”
Lemkin: $100M+ startup reports three different ARR metrics monthly
“And for example, I've got one investment I made that's north of nine figures in revenue. I get three different ARR numbers each month. Three different definitions.”
O'Driscoll: YC guidelines will become the seed revenue standard
“I predict if it sticks, the shorthand version of the seed stages will be, does this conform to the Y Combinator revenue guidelines, right?”
O'Driscoll: Y Combinator holds 25% seed market share
“Y Combinator has a dominant market share in the seed market, right? 25%.”
O'Driscoll: Sequoia evergreen thesis will prove right over 10-20 years
“Though I think over 10 and 20 years, their analysis will still be correct. If you build enduring companies, even in the public markets, the compounding will happen.”
Lemkin: Ryan Smith's Utah Jazz investment quadrupled, outpacing software returns
“When Ryan Smith sold Qualtrics, I think he made about a billion dollars. After 20 years or so and I believe that billion, most of it went back into the jazz, and it has quadrupled... He's up three billion on the jazz or something like that versus the 20 years …”
O'Driscoll: US sports leagues are antitrust-exempt oligopolies with zero accountability
“Europe, the alleged socialist capital of the world, has a far more performance oriented sports culture than America, where it's a nasty little oligopoly. I mean, the NFL and all the American things have been constructed partly because they're the only three bu…”
O'Driscoll invests $500 in retail VC product for Anthropic and SpaceX exposure
“I felt this morning, I put the literally the lowest amount possible in the VC product, so I am now an individual investor in Anthropix, SpaceX, and OpenAI, and even as we speak, I'm adding the logos to our website, right?”
Lemkin: AngelList charges a 3.61% annual management fee for USVC fund
“So AngelList charges 3.61% a year to manage this fund, right?”
O'Driscoll: Successful VC funds face a 4% to 5% fee drag between gross and net
“The average venture investor is charging two percent, And then, 20% of the profits, which typically turns into, if you're successful, a four or five percent drag between gross and net, right?”
O'Driscoll: VC requires 25%+ gross returns to justify 2&20 fees over public markets
“The only reason you can pay in the long term, two percent to VCs and 20% of the profit is because the gross returns have to be high enough, 25% plus, that the net return is still, you know, 20%, which is so far above the Ibbotson small cop return of 11, 12% th…”
Lemkin: Top funds passing on growth-round super pro rata is a terrible signal
“When a top fund doesn't go all in, On an investment, it's such a bad signal. Not only is it bad if the tier, if the, if Andreessen does your seed and doesn't lead your array, that's the classic discussion we could have done on 20 VC in 2015. Right. But the sub…”
Stebbings: Sold Figma position at a 40% loss
“While you guys were doing my Skydio, I just sold Figma, 40% down.”