Summary

OpenAI got warrants for up to 10% of AMD at a penny — effectively free warrants for the privilege of being sold to. Rory O’Driscoll’s decode: Nvidia is strong, so it gets equity in OpenAI for supplying chips; AMD is weak, so it hands its own equity to OpenAI for the right to be a supplier. “Paul Graham was right. Sam Altman understands power.” AMD’s stock price went up 30%+ and $60B against warrants worth $30-40B — but they only land if the chips ship and get bought.
The structural frame worth trading on: this is Wintel replaying. OpenAI is the new Microsoft (it has the users), Nvidia is Intel, AMD is running “exactly the same shtick” as second source thirty years later, and Microsoft is playing IBM — the incumbent that “set this viper in motion.” The open question for Redmond: “did we just create a monster?”
The only profit pool in AI is Nvidia’s 50% margin, and everyone knows it. Jason, an ex-components seller: the normal rule is “cost plus twenty percent,” so Nvidia’s margin is “the number one place I’d like to attack if I could.” Jensen is running an “elaborate dance” — conceding a little share politely to keep ~90% — and the cautionary comp is memory: if GPUs ever trade like Samsung and Micron, “that business looks very sad.”
Venture math is being stress-tested at the entry price, not the trend. Naveen Rao’s $1B at $5B pre is defensible (“proven person, hard problem — hard problems are getting rewarded”), but Harry’s line holds: there will be Amazons in this wave, “Amazon was priced as Amazon was” — these are not. Rory on comps-based pricing: “I wanna bludgeon them to death.” Nothing changes “until Mr. Market delivers a sad lesson,” and with a16z expecting $40B+ back from Databricks, that lesson isn’t imminent.
The mid-tier exit window is quietly shut: PE isn’t calling. Snyk at ~$300M ARR growing 26% sits just under the IPO bar (median IPO this year: a stunning $931M run rate), pencils to “mid twos” billions on a Netskope comp versus a $7.2B last round — and Jason’s three comparable portfolio companies have had zero PE offers. “Crickets. Crickets from the PEs.” Rory’s board playbook: equity-for-growth regrants, profitability, and an AI second act — “a plan that doesn’t rely on the kindness of Thoma Bravo.”
Vercel ($300M at $9.3B) and Supabase are “Captain Obvious bets” — the default hosting and Postgres layers of an exploding app wave — and not suicide rounds, because revenue multiples are holding flat as scale 10x’s. The kingmaking debate resolved on Harvey vs LawGora: Harvey had Sequoia, capital and mind share, “then LawGora came in from Sweden, for God’s sake, and killed it… if you have two people being kings, there can’t be a king.” But in capital-hungry AI categories, “king-making becomes a prophecy because you just bow out.”
SPACs are back and Polymarket is being legitimized — the sign of the times. Chamath’s new terms are “almost legit” (sponsor promote only above a 50% uptick), though SPACs still legally pump where IPOs can’t, and “a well-run IPO beats it by a head.” NYSE-owner ICE putting $2B at $9B into a company that was “essentially illegal last year” is, per Jason, “quietly the craziest story” of the week.
Vibe-coding traffic has plateaued — and that’s healthy. Barclays data shows flat-to-down usage at Replit/Bolt/Base44; Jason reads it as looky-loo churn shaking out while sticky users like him ($300-3,000/month, “impossible to leave”) compound. Rory’s kicker: “there’s deceleration in ChatGPT too… how much deceleration can you do at 12 billion and still hit 100 billion?”

Deep dive

1. OpenAI got warrants for up to 10% of AMD at a penny — “Sam Altman understands power”

The deal: OpenAI commits to buy up to six gigawatts of AMD’s Instinct chips and receives warrants for up to 10% of AMD at a penny — effectively free warrants, contingent on the chips being bought and the stock going up. Rory’s contrast with the Nvidia deal a week earlier is the whole story: “Nvidia is strong, when they get to get equity in OpenAI in return for giving OpenAI chips… AMD, ‘cause it’s weaker, has to give their own equity to OpenAI for the privilege of having OpenAI buy their chips.”
His reconstruction of the negotiation is worth keeping verbatim: OpenAI to AMD — “I bet you your stock will go up just ‘cause you’re doing business with us, ‘cause you’re kind of a no-hoper, and now we’re saving you, and we wanna get some of that upside.” As of the announcement, AMD’s stock price went up thirty-something percent and $60B, against warrants worth $30-40B — “so you’re up.”
Jason’s initial read ran the other way: AMD is doing more for OpenAI short-term than vice versa, and it reminded him of Shopify/Stripe — Toby so angry about putting Stripe on the board that he went to Klaviyo demanding 10%. “It felt like not getting egg on your face.” Rory’s caveat on the whole structure: the warrants only land when chips ship and get purchased — “two years from now, OpenAI still has to need whatever vast sum of chips that they said.”

2. The Wintel replay — and Microsoft as IBM

Rory’s zoom-out: “what you’re seeing here is the Windows-Intel game beginning again.” OpenAI is the Microsoft of today — “they have the consumers, they have the eyeballs, they’re building this new monopoly” — Nvidia occupies Intel’s seat, and AMD shows up “thirty years later with exactly the same shtick: we’re not as good as Intel, we’re not as good as Nvidia, but we’re here. We’re the second source.”
The role of IBM is being played by Microsoft, which “set this viper in motion.” Harry’s pushback — didn’t IBM at least not own the monster? — draws the concession: “history doesn’t repeat, it rhymes.” Owning 10% or 30% would be a far better corp-dev outcome than IBM’s nothing. But DevDay’s pitch — run your apps inside ChatGPT — attacks Microsoft’s own franchise: “if you’re Microsoft, you’re like, WTF? We’re the place where you should run your other apps. That’s what we do. Who the hell are you?”

3. Jensen’s elaborate dance — and why users are leverage even at a loss

On how Nvidia feels about being two-timed a week after investing $100B: Jason thinks it’s all choreographed. “Nvidia is making so much money it’s almost incomprehensible. So I think Jensen knows he’s got to give up some of it” — an elaborate dance of conceding a little market share politely so he can minimize price erosion and keep ~90% share. The sequencing (Nvidia announced first, homage paid to Jensen) was deliberate; the counterexample is Elon — “being impolite has consequences in this space… I don’t know if xAI would exist if it wasn’t for his bone to pick with Sam Altman.”
Rory’s deeper point is where OpenAI’s leverage actually sits: the users, not the P&L. “You can be sitting there losing money like hand over fist, and still you get credit for committing hundreds of billions of dollars you don’t have. You can bestow market cap on your vendors, for God’s sake” — because the whole world believes, rightly or wrongly, that $12B of revenue is going to $200B and will take $100B a year in chips to get there.

4. Components economics, inverted — and the memory-market warning

Jason sold components in his first startup, and the rule was universal: everyone brings you coffee and says “cost plus twenty percent.” Against that, Nvidia’s fifty percent operating margins are an anomaly: “I’m fine if Nvidia makes twenty cents on a dollar… but fifty cents, I mean, F me.” Normally the vendor has lower margins than the software layer; “here it’s highly inverted.”
Rory’s resolution: the only thing that defeats cost-plus is architectural lock-in — a monopoly selling to an oligopoly. The buyer can rage that Nvidia pays TSMC $50 a chip and charges $300, “and Nvidia sits there and goes, ‘Well, we won’t,’ and you’ve got no other choices.” Jason’s interjection, in Jensen’s voice: “We will, but we’re sold out… Maybe in 2031 we can provide you with some of those chips.”
The history: venture walked away from semis around 2003-04 (Rory’s firm had one of the last exits, Monolithic Power, ~2007-08), while public semis consolidated into wildly profitable near-monopolies. The tail risk, exactly as hedged: “if the GPU market ever turned into the memory market — which I’m not saying it will — then that business looks very sad. Just take a look at how Samsung and Micron trade versus how Nvidia trades.”

5. DevDay underwhelmed: apps-in-ChatGPT is Slack 2.0

Jason wanted exactly this — he told Benioff on this show he wants to talk to his apps (“I’ve literally been a Salesforce customer for twenty years and haven’t logged in in a decade”) — and still came away flat: “I didn’t see an aha moment. I didn’t see magic. I didn’t see something so great my jaw dropped.”
His killer analogy: this is Slack 2.0. Slack was the OS Benioff paid $27B for, every app has a Slack integration — “but how often in Slack are you creating a Spotify playlist or creating a Canva image or even pulling up a CRM record? I bet never.” Nor is he sure the world needs “the tenth app marketplace.” His charitable out: “maybe that’s ‘cause they built it in eight weeks.”
On AgentKit killing the n8ns of the world, Rory hedged: enterprise agents “probably are going to require a lot of orchestration, a lot of management” — real product surface area — and over the next two years, will OpenAI grind on that “because they got bigger fish to fry? Or will they just make it easy to connect and move on?”

6. Naveen Rao’s $1B at $5B pre — does this break venture math?

Harry’s framing: for a 10x with dilution, this has to be a $100B company. Rory’s answer is the star effect in hard infrastructure — the number of people who can credibly say “I can solve this level of technical problem” is tiny (see Thinking Machines, Safe Superintelligence), and Naveen has built two deep-tech winners (a hardware company sold to Intel, then the one sold to Databricks). Lux backed him twice; the third time “he gets a nice coffee and a nice seat, and you’re like, ‘What do you need?’” The compressed logic: “Proven person, hard problem. Hard problems are getting rewarded.”
Jason’s gloss: it’s a confidence game running through the founder too. A Databricks alum just watched $100B happen — “five to 100 seems plausible.” His own generation couldn’t see past a billion, which is why he sold: “I own 30%, but if I get to IPO… I can’t even make more money.”
Harry’s counter, via his Mike Cannon-Brookes interview: yes, “there will still be some Amazons in this AI wave” — but Amazon was priced as Amazon was. “These are not priced in any universe of Amazon-level returns. Even if you have Amazon-level plus-plus-plus outcomes, they’re still not venture-star returns.” Rory concedes the destruction test: if only one in three does something wildly amazing and you paid for all of them as if they would, “you don’t have much of a return.”

7. Comps, entry price, and Pavlovian deployment

Rory’s comps rant, verbatim: “whenever anyone uses comps to discuss what we should pay for a deal, I wanna bludgeon them to death.” Comps rank assets against each other today; investing asks what they’re worth in seven years — using comps in 2021 meant buying at 50x revenue because the other stuff was at 80x. Harry’s live example: Alex Wang’s Scale at $14.8B resetting everyone’s priors — “Jesus, Ilya’s worth 30, and Mira” — with the required belief being that companies at $500-700M revenue “can trade at 20 times revenues over an extended period that includes the next decade. That’s a lot harder story to believe.”
The saving grace, and its limit: venture “is the most forgiving equity business of getting the price wrong” because maximum variance and exponential growth bail you out — unlike PE, where overpaying 50% on a 3x asset is fatal. “But maximally forgiving on overpaying doesn’t mean entirely forgiving.”
On deployment pressure — a16z’s fund at $7.5B implies finding 15-20 such deals in roughly two years — Rory is fatalistic: “we’re all Pavlovian. We do the things that feel good… we’ll just keep doing this until Mr. Market delivers a sad lesson” that says you’ve overreached. “And so far, that hasn’t happened.” Harry doubts it will: a16z is “about to get the payday of paydays in venture land with a minimum of forty billion back from Databricks.”

8. LP liquidity: endowments selling, and Jason’s Evercore fight

Brown and Northwestern selling VC stakes (after Yale and Harvard) is, per Rory, a readjustment of the Yale endowment model under political pressure, not a new normal — “if you’re selling a lot of venture assets, you’re probably not gonna buy a whole load more. So it will tail off.”
Jason just lived his first LP secondary, managed by Evercore, and discovered his own docs give LPs literally no right to sell — “no exceptions for if you’re struggling.” The process was “pretty broken, frictionful, weird”: the seller claimed rights they didn’t have, then retreated to “we have the moral authority to sell your position. I’m like, ‘What? What?’” His conclusion anyway: “everything would be better if there was more liquidity down the stack. Everything. The market can decide the discount.”
Jason notes that even good funds can take 20 years to wind down; Rory agrees that funds dragging on make secondaries structurally necessary — and plants a flag to remember: “Liquidity doesn’t evaporate ‘cause people run out of money. Liquidity evaporates ‘cause people get scared and wanna keep their money, and at some point when that happens, you’ll go, ‘Oh, that’s what the public markets were for.’”

9. Snyk and the stranded cohort: high-class companies, no bid

Snyk: ~$300M ARR growing 26%, down from 150% in 2022, last raised at $7.2B, PE rumored circling. Rory’s data: 15 IPOs year-to-date with a median revenue run rate of “a stunning $931 million,” but a couple got out at $200-300M growing ~30% — “it’s not like they’re miles away from it.” Jason’s mark-to-market, using Netskope ($8B at $700M growing 33%) as “the roughest VC math”: Snyk is worth “in the mid twos” — and the triage is brutal: “it’s great, but it’s not Netskope, which isn’t Rubrik.” Even then: do you want to be “one of these invisible public companies” ignored by Wall Street?
The scarier signal is on the private side. Jason has three portfolio companies in the same bucket — right rule-of-40, right NRR — and zero PE offers: “Crickets. Crickets from the PEs. Is your phone ringing off the hook? Are they banging the door to get in this morning?” In 2021 or even early 2023, these should have gotten PE offers. Rory confirms: PE has capital but is “not in any rush to buy subscale assets that aren’t defendable market niches” — which is why “whenever the liquidity window opens, you should pay attention.”
Rory’s board playbook for taking control of destiny: first, EFG — “equity for growth” regrants so a fully-vested founder on what turned into a 15-year journey says “fuck it, I’ve got another seven years of equity ahead of me”; second, profitability; third, a second act “almost certainly related to what’s going on in AI and how workflows become agents.” The payoff line: “I have a plan that doesn’t rely on the kindness of Thoma Bravo.”
The honest flip side of the regrant conversation: “if we’re going to be at 20, going to 15, going to 10, then with dilution you’re way below my cost of capital. Let’s admit that, and let’s go for a sale, and you get what you get, and you don’t make a fuss.”

10. Replacement CEOs and the stupid $500M question

Harry asked what percent of the time growth re-accelerates when a founder is replaced. Rory’s split: with product-market fit and an entrepreneurial-but-bad-manager founder, a competent operator can give it a lift; without PMF, hiring a manager to find it is delusion — “You are what’s called wrong, and you should sell for what you get and move on… the answer is zero in that case.” Jason’s version: “I never want the CEO to ever leave. Good God” — but if you’ve done the regrant, begged them to stay, and they still raise their hand, “most humans should take the millions and relax.”
On the classic VC question “would you sell for $500 million today?” — Rory never asks it: “there is zero information content to it. I’ve seen ‘I’ll never sell’ people take the money, and mercenaries keep going.” Worrying a good founder might sell in a good-not-great outcome is “so low down the list” versus backing B-level people who never make it happen.
Harry’s pattern: you can tell your best and worst within the first 60 days — the messy middle is TBD. Rory agrees “at the 70% level”: on all his best deals there was a first-year board moment of “Oh, Rory, you clever boy. You’re gonna make money here.”

11. Vercel at $9.3B and Supabase: Captain Obvious, not suicide

Jason, 200 hours into vibe coding, sees both raises as structural: Supabase is “the default choice to how to host and manage Postgres in this world,” Vercel is where the exploding wave of apps gets hosted. “Valuation aside, these are actually Captain Obvious bets. These are the leaders… follow where the developers are going.” Rory’s echo is the episode’s most quotable investing principle: “the more you do this, the more you just say to yourself, you just need to do big, exciting deals in trends that are absolutely obvious. And every time you try and make it harder than that, you lose money.”
Both companies, he notes, predate OpenAI and vibe coding — they “inserted themselves into relevance” and are riding the train. The price may have doubled in six months, but so may the company: the revenue multiple is flat as scale 10x’s, which inverts the normal model where multiples compress as rounds go up. “The B was at the same multiple as the A. Maybe we need to update your priors.”
The stated risk, not smoothed over: if the market is finite — “even if you have a hundred percent share of vibe coding” — growth hits a wall and “you’re wildly wrong at scale.” But a suicide round requires two failures: rapid deceleration plus burning back into the market for a down round. Supabase still has the $300M already sized for it. “I don’t think that will be the problem that brings the whole thing to its knees.”

12. Kingmaking: Harvey vs LawGora settles the argument

Harry’s thesis: kingmaking exists but works inversely to TAM — “the smaller the TAM, the more prominent the kingmaker ability.” Rory initially disagrees, arguing OpenAI’s capital strategy is kingmaking in the biggest market of all — a duopoly where “Anthropic clearly can” raise the required capital but the rest of the field can’t — then audibly changes his mind on the word itself: “I don’t like the word king-making, I’ve decided… the entrepreneur is the king,” with capital as the accelerant on a lead the company already earned. Harry’s rebuttal: it’s happening earlier and earlier — sometimes literally pre-execution, $50-200M successive rounds at $3-5M revenue off a tier-one lead — “and everyone goes, ‘Oh fuck, we don’t wanna go in after Sequoia and Iconiq.’”
The case study that cuts against fatalism: “Harvey looked like they’d been king-made… and then LawGora came in from Sweden, for God’s sake, and killed it. They shipped a good product, a very good product. They got Benchmark in… So there was room for a second — and by definition, if you have two people being kings, there can’t be a king.” Would Rory fund a third in corporate-law software? “I think it’s hard… at some point it gets kind of cooked.”
Jason’s structural point on why AI is different from SaaS: Cannon-Brookes said he couldn’t build Atlassian today — “I had five years to be left alone.” In SaaS you’d catch up to funded competitors around $10-20M ARR; in AI, capital advantage compounds, so “king-making becomes a prophecy because you just bow out — if I believe I need a hundred million to compete with Replit and Lovable, I sell to Wix, because where would this poor guy in Israel with eight developers get fifty million of tokens?” Rory adds that in enterprise deals “the balance sheet becomes a criteria for qualification.”
Should a portfolio company take kingmaker money? Rory: game theory forces aggression — “you can say you’re gonna be careful and slow and rational, but if they’re not, you just end up outclassed.” Jason: it’s founder DNA now — one of his best companies “will consume an infinite amount of capital, the other has sixty years of runway… it no longer matters what I think.”

13. SPACs are back and Polymarket is being legitimized — the sign of the times

Rory, unprompted: “Chamath’s terms are almost legit.” The 2021 vintage failed because sponsors got paid just for closing — investors in at $10 could lose half while penny-stock sponsors dumped at $5. New structure: the sponsor gets nothing until the stock hits $15, then a 30% promote — “not cheap, but a little better than before.” The still-broken part: SPACs are legally exempt from IPO forward-statement rules, so “you’re allowed to articulate any future-leading story you like… ‘It’s gonna be freaking amazing, I’m tweeting like crazy.’” Net: “marginally more attractive competitor to the IPO, but a well-run IPO beats it by a head.”
Jason’s nomination for the week’s real story: Intercontinental Exchange, owner of the NYSE, investing $2B at a $9B valuation into Polymarket — “which was essentially illegal last year,” with the Biden administration moving to shut it down as offshore gambling, and now Trump’s son on the board and investing in the company. “Forget about AI… they legitimize self-dealing in it. It’s just a different world — I don’t know if it’s better or worse, no politics, but it sure is different.”
Rory takes a side anyway: “I think the deregulation’s great, to be clear” — Biden choosing that hill to die on was “both probably wrong and definitely stupid,” and explains the polling collapse among 20-30-year-old men. The open commercial question, exactly as hedged: a market Rory thinks he checked is still 70-80% sports betting, and whether a legitimate non-sports prediction business exists “is TBD.”
The $2B size is not described identically by both speakers: Jason reads it as a bonding/exclusivity investment, while Rory says ICE is a strategic investor that “likes to own some of” any electronic market in interesting financial stuff, not a 2x-return trade.

14. Quickfire: the vibe-coding plateau — and ChatGPT’s too

Replit and Lovable over/under $250M ARR by year-end (both ~$160-170M now): Jason takes over, “but barely.” A Barclays web-traffic report — credible because Base44’s numbers tied to Wix’s public disclosure — shows category traffic “flattened to down.” His read is bullish: the “looky-loos” churn (“Abigail that wanted to build her own CRM and was told you could do it in 60 seconds and it didn’t work — she’s gonna churn”), while sticky users compound: “I’ll never churn off Replit. I’m 200 hours in, I’ve got eight apps into production… somewhere between $300 and $3,000 a month for a long, long time.” And the platforms “are so much better than 100 days ago — this is not SaaS of 2016.”
Rory’s needle: “For the record, there’s deceleration in ChatGPT too… how much deacceleration can you do at 12 billion to make sure you still hit 100 billion is my question.” He’s thinking about it; no answer offered.
On the Tim-Cook-leaves-Apple-this-year prediction-market question, both call the odds right to be long “no”: a leaked succession rumor concerns the SVP of engineering, 50, with Cook turning 65, but “unless it’s a health issue, there’s no way it’s gonna be this year.” Rory remains an Apple holder — bought in 2009, trimmed a little when Warren did — “my basis is so low I just can’t bring myself to pay the 37% tax.”

Media M&A - [Business Breakdowns, EP.230]116. Wu Minghui on 19 Years, Enterprise Agents, and the IPO