Summary

OpenAI’s restructuring turns a capital-constrained hybrid into an investable PBC and moves an IPO “one enormous step closer.” Microsoft’s agreement and approvals from the Delaware and California attorneys general remove the corporate obstacle to raising another $100–200 billion. With a possible $2 trillion retail-magnet IPO now structurally feasible, the remaining risk shifts from governance plumbing to whether AI economics support roughly 40 times GAAP revenue.
Microsoft, OpenAI’s charitable foundation, employees, and future investors all emerged as winners, while Sam Altman still owns no shares. Microsoft turned roughly $13 billion into a stated 10X while retaining some extended IP rights, revenue share, and prospective Azure business; the foundation ended up with about $135 billion of value. Jason’s striking contrast: Elon Musk was arguing for a trillion-dollar pay package while Altman retains enormous influence with “no shares in the combined entity.”
Andreessen Horowitz’s $10 billion raise makes scale itself a competitive weapon, even if its individual sleeves are less intimidating than the headline. The $6 billion growth fund and three specialist funds let a16z pay above intrinsic value for early “options,” finance talent and events, and maintain a relentless media presence. Rory’s image captured the strategy: “Andreessen Horowitz is the Red Army of the venture industry,” because “quantity has a quality all its own.”
Mercor’s $350 million round at $10 billion is fundamentally a leveraged bet on three to five more years of AI CapEx hypergrowth. Its reported rise to $500 million of revenue in 17 months reflects foundation-model companies’ urgent need for specialist human training, data-quality measurement, and implementation. The catch is pass-through-heavy margins and extraordinary customer concentration: two buyers reportedly account for more than 50% of revenue.
Ramp’s move from $23 billion to a reported $30 billion exposes a new asset class: “public stocks hiding in private.” Jason called the step-up too small to matter after dilution, but Rory argued that a billion-dollar-revenue company should deliver public-market-like returns regardless of its legal wrapper. Ultra-late investors may happily turn $1 billion into $2 billion even when an early-stage fund still needs 3X outcomes.
Carta’s 547-company Series B dataset says venture remains a picking business unless a fund can afford to buy follow-on options at scale. Roughly two-thirds of 2018 Series B deals returned less than 2X, while about 10% exceeded 10X and Figma reached 100X. Rory separated three models—picking, spraying, and optioning—and argued that large multi-stage platforms can relax selection discipline through options, while YC is the clearest structurally advantaged mass-production case.
Synthesia’s rejected $3 billion Adobe offer split the panel on the real sell decision: expected value is only one variable; founder ambition, duration, and IPO fitness matter too. Jason would have advised selling unless the team truly wanted to build and run a $10–20 billion public company; Rory said reported growth from $100 million to $150 million ARR in under six months would make the answer easy: “There’s no way they should sell.” Their shared test was whether the founder is genuinely “an IPO guy,” not whether the VCs want another turn.
Amazon’s difficult fortnight illustrates how abruptly AI can challenge an incumbent’s lead. The host cited layoffs affecting 10% of white-collar staff, cloud share falling from 50% in 2018 to 38%, and a forecast of only 7% AI-cloud share; Jason wondered whether Bezos exited just before the strategic regime changed. Rory’s prescription was narrower: retail distribution remains formidable, but AWS must become AI-relevant without accepting Oracle-like economics he fears are subpar.

Deep dive

1. OpenAI escaped its corporate straitjacket

Rory’s framing: OpenAI completed agreements with Microsoft and the attorneys general of Delaware and California, enabling the long-delayed restructuring and reopening access to capital. Elon Musk may still litigate, but “possession is nine-tenths of the law”; unwinding a completed conversion after regulatory approval becomes much harder.
The resulting structure places a heavily capitalized charitable foundation above an investable public-benefit corporation. That PBC is a genuine for-profit entity required to consider more than shareholder maximization—not the old structure whose disclosure effectively said, “You should regard this as a donation. It can all go to zero.”
An IPO is neither promised nor required, but Rory called it “one enormous step closer.” OpenAI can now approach public markets as a recognizable American corporation rather than asking investors to tolerate an idiosyncratic governance trap.

2. The restructuring paid each constituency according to its leverage

Jason highlighted the rough ownership map: Microsoft at 27%, employees and the nonprofit each in the twenty-something-percent range—and Sam Altman still holding no shares. After the revolt that restored Altman as CEO, Jason found it unprecedented that the leader could remain economically empty-handed yet potentially gain moral and political power from that fact.
Microsoft put in roughly $13 billion and, by Rory’s account, earned about 10X while preserving some extended IP rights, revenue share, and meaningful prospective Azure business. It pushed hard without breaking the company; Rory awarded its corporate-development team and lawyers a “gold star” and noted the stock’s positive reaction.
The charitable foundation ended up with approximately $135 billion, making the original public-benefit ambition economically real even if its future impact remains TBD. Employees gained liquidity, SoftBank could put in its roughly $22 billion and own about 10%, and other investors could finally exhale.
Rory gave Brett Taylor his second “best board chairman” award in five years: first for enforcing Twitter’s $44 billion sale to Elon Musk, now for unravelling OpenAI’s structure. The practical losers were Musk and nonprofit purists; Rory’s verdict was that nobody received much more than the risk and work justified.

3. A retail blockbuster could finance OpenAI’s next capital wall

Jason’s financing thesis: a public valuation potentially approaching $2 trillion could unlock perhaps four times more equity capital than private markets, including another $200 billion if necessary. That makes suppliers’ huge forward commitments less fantastical because OpenAI can plausibly finance the purchases.
Rory remained skeptical that Oracle would collect the final dollar of its cloud contract, but the failure mode is no longer a nonsensical corporate structure. If OpenAI needs another $100 billion, he said, it becomes “just banking and math,” subject to the underlying returns on AI investment.
Both expected extraordinary retail appetite. Investors may ignore profitability caveats and roughly $250 billion of third-party cloud commitments in favor of “Let me get some of that OpenAI”; Jason could not imagine a more popular retail IPO.
SoftBank was reportedly closing a direct investment near a $300 billion valuation while buying employee shares at $500 billion. Rory joked that wiring at the lower price and marking to the secondary the next day creates “40% IRR in an hour.” At $500 billion and roughly $12 billion of GAAP revenue, he described the company as trading at about 40 times GAAP revenue; assuming roughly $20 billion of revenue and a 25-times run-rate multiple, he guessed $1 trillion would normally take two years—unless euphoria accelerates it.

4. Andreessen’s $10 billion makes scale a strategy, not just a fund size

The raise divides into $6 billion for growth, $1.5 billion for AI applications, $1.5 billion for AI infrastructure, and $1 billion for defense. Jason initially found $10 billion unprecedented, then surprisingly modest when decomposed: even $1.5 billion does not stretch far across rounds for ElevenLabs-, Replit-, or YC-scale companies.
Rory saw two durable advantages. A mega-fund can treat a seed check as an option on the B and therefore pay above intrinsic value; it also creates a “wall of news” through constant portfolio activity and media reach. Rory also invoked Marc Andreessen’s stated view that investing is not a media business, while crediting a16z’s media strategy with tilting the table.
Harry added what AUM buys: management fees for exceptional staff, larger carry pools, impressive offices, and events that create founder serendipity. Rory’s memorable summary—“Andreessen Horowitz is the Red Army of the venture industry”—came from the maxim that “quantity has a quality all its own.”
The structure also aids retention by giving senior investors specialist “fiefdoms.” LPs may have to commit $2 elsewhere for every $1 they want in a coveted early-stage sleeve—what Rory called bundling and rent extraction. The unresolved question is whether late-stage returns eventually disappoint LPs enough to revive smaller funds; until then, “having $10 billion is better than not having $10 billion.”

5. Mercor monetizes the human labor hidden inside model progress

Mercor reportedly raised $350 million at a $10 billion valuation, led by Felicis only eight months after its $2 billion round. Harry said it had reached $500 million of revenue—“I believe” faster than anyone else—and estimated that it took roughly 17 months.
Rory rejected the claim that this is merely GMV: Mercor supplies specialist work, gets paid, and books GAAP revenue. Foundation-model companies need mathematicians, physicists, doctors, and other experts to pose questions, evaluate answers, and provide feedback that can “pound the model into submission.”
The labor market has moved from “labeling cats” to probing the outer edges of human knowledge. Mercor’s achievement was assembling high-end talent as the models became able to handle simpler tasks, while its customers went from spending zero on this category five years earlier to spending billions.
Harry described three expanding layers: talent acquisition, delivery and quality measurement of the resulting data, then implementation inside customers’ model-development processes. As complexity rises, pricing rises too—but two customers reportedly generate more than 50% of each major provider’s revenue.

6. Mercor’s $10 billion price assumes the AI CapEx train keeps running

The bull case is an urgent customer with abundant funding and too little time: “Money I have in spades, time I don’t got.” The bear case is pass-through-heavy gross margins, stressful eight-figure renewals, and buyers eventually repricing the work. Jason also noted that OpenAI is building its own chips; Rory focused on the risk that the model companies eventually use their leverage against vendors.
Rory would not speculate on the exact payout ratio, but said the gross-margin profile is not amazing because a substantial percentage goes to the doctors and mathematicians doing the work. He would rather own a business with 800 million customers than one with two customers capable of renegotiating a $200 million contract.
At a required 3X, a $10 billion entry needs a $30 billion outcome. Assuming a five-times revenue terminal multiple implies $6 billion of training-data revenue—“a lot of training data,” and enough to make an investor pause.
The alternative underwriting method is momentum: if revenue grows 5X and the multiple holds, value can rise 5X within a year. The prior $2 billion round near $100 million of revenue now looks cheap beside $10 billion and $500 million. Rory called the entire trade “one big-ass bet on AI CapEx hypergrowth.” Jason Calacanis then invoked Chuck Prince’s 2007 Citigroup line about staying on the dance floor while the band plays.

7. Ramp is a public-market stock wearing a private wrapper

Ramp was reportedly discussing another round at $30 billion. Jason said this company genuinely consumes capital: advancing customer spending might require about $5 of capital for every dollar of added revenue, so a $1 billion revenue base could support roughly $5 billion of assets and need a substantial equity cushion beneath its debt.
Jason dismissed the move from $23 billion to $30 billion as barely an up round. In his example, a diluted 2% seed position might move only from roughly $400 million to $480 million; with annual dilution potentially around 10% in some AI companies, the headline valuation can obscure a flat per-share result.
Rory’s pushback: Ramp is already “public stock’s hiding in private.” A billion-dollar-revenue, near-profitable company should resemble a high-growth mid-cap stock, where 30–40% annual growth is excellent and financing valuations move incrementally. Stripe’s change from $91 billion to $110 billion was another example.
The pronoun mattered: early-stage investors may underwrite to 3X, but ultra-late funds investing at $30 billion probably do not. Someone can be “very happy indeed” turning $1 billion into $2 billion—and earn more dollars than a small fund generating a higher multiple.

8. Carta’s data rewards picking—and capital-backed optioning

Carta examined all 547 Series B investments from 2018. About 35% returned below 1X, roughly two-thirds returned under 2X, 18% exceeded 5X, about 10% exceeded 10X, and Figma delivered the lone 100X.
Rory found the distribution reassuring because it approximated his fund model: 30% below 1X, 50% between 1X and 5X, and 20% above 5X. Filling those buckets correctly produces approximately 3.7X gross and 3X net—not through diversification alone, but through disciplined selection.
The panel separated three strategies: picking, spraying, and optioning. Harry cited a separate graph showing that a16z had made 72 seed investments versus Sequoia’s 27; its growth capital lets those checks function as options, because it can concentrate enormous follow-on dollars into whichever companies emerge.
Even a 50-company seed portfolio remains selective against an estimated 1,600-company universe. Rory calculated that one hypothetical 300X winner contributes only about 0.2X if capital is spread equally across all 1,600. YC is the exceptional “mass production seed business” because its structured economics give it a distinctive ability to spray and retain option value.

9. Venture’s losses remain personal, operational, and expensive

Harry described receiving acquisition stock valued at what he considered an absurd price as his bad outcome. Rory offered a worse one: sit on the board, fail to close a sale, then wire another $300,000–400,000 for shutdown and severance before immediately writing it off.
Jason joked that SAFEs provide “no rights, no visibility, no financial statements,” letting him ghost failures. His serious point was narrower: a post-money cap simplifies dilution and option-pool questions, while the instrument’s limited governance commitment can spare an investor a decade on the board of a struggling company.
Rory said he never becomes numb to failed investments: “I get all sad.” The exchange punctured the tendency to remember only 10X and 20X winners when Carta’s data shows how much of the business consists of disappointing outcomes.

10. Synthesia’s sell decision depends on growth and founder temperament

Synthesia reportedly rejected Adobe’s $3 billion offer and was preparing to raise above that level at around $150 million ARR. Jason would have advised selling unless the founders were certain they wanted to build a $10–20 billion public company, even while conceding that rejection might be correct.
Rory made the financial condition explicit: if ARR truly rose from $100 million to $150 million in under six months, “There’s no way they should sell.” Even at roughly doubling growth, he liked the category and its runway as a new human interface to computing.
Jason saw divergent incentives. A founder owning 10% may live essentially the same life after a $3 billion or $10 billion outcome; a VC owning 15% can turn a roughly 1X fund return into 2–3X. Harry raised DPI pressure, but Jason said even conservative LPs with a hot manager usually prefer “another card” over a trivial 20% fund distribution.
Rory’s board-room test begins with undisclosed risk: “If there’s something about the business that’s really worrying you and you haven’t told us, now would be a good time to share.” Jason added the human question—“Are you really an IPO guy?”—because beyond $3 billion, the realistic paths narrow to compounding privately, going public, or changing CEOs.

11. Paying up for the category leader beats funding its imitation

Harry’s lesson was blunt: if the company you want is the category leader, pay its higher price rather than searching for “the next” one. He challenged Rory on why the same logic did not demand owning Synthesia itself.
Rory distinguished his investment in Tavus: real-time avatar interaction rather than Synthesia’s asynchronous, pre-generated avatar product. It expresses the same broad interface trend without being a “Synthesia wannabe” competing head-to-head from behind.
Rory’s firm learned the distinction painfully in 2013: it chose a normally priced company B over a late, expensive round in market leader A. B returned 2X; A would have returned 15X. “You’re not in the business of doing modest number twos.”

12. Antitrust delay can turn a premium acquisition into a bad trade

Jason used iRobot as the warning case: Amazon offered roughly $1.7 billion, antitrust intervention killed the deal, iRobot borrowed about $200 million to bridge the gap, and bankruptcy became a possibility. Rory blamed the FTC and Lina Khan for what he called an outdated and foolish assessment of vacuum-cleaner competition.
Jason’s acquisition lesson was that strategic buyers sometimes pay a multiple unavailable in public markets, so founders should take the offer seriously. Harry’s pushback: Synthesia could spend 18 months waiting for approval, reach $400 million of revenue, then discover that $3 billion no longer looks attractive.
Rory agreed duration erodes the headline. A nominal 30-times-revenue offer can become 10 times by closing; similarly, an apparent 2.5X in six months may actually be 2.5X over two years. Prolonged review puts “sand in the gears” and pushes companies toward independence or talent-only acqui-hires.
Harry used that duration to argue for early secondary sales and called IRR king. Rory disagreed with the single-metric framing: optimize “the maximization of multiple subject to a constraint on a minimum IRR.” A 25% IRR over four years beats 30% for one year, but holding until 25% slips toward 17–19% violates the constraint.

13. Amazon’s retail engine is intact, but AWS missed the AI reset

Harry opened with a severe scorecard: layoffs affecting 10% of white-collar employees, cloud share falling from 50% in 2018 to 38%, Raymond James projecting 7% AI-cloud share, and an outage causing billions in damage. Jason contrasted Bezos’s departure with Sergey Brin’s return to Google.
Jason argued Bezos left when Andrew Jassy took over on July 5, 2021, when products appeared frozen and the old regime felt permanent, then failed to “punch back in” as AI changed the market. Rory countered that Bezos may rationally be maximizing psychic joy rather than money—but Jason insisted Bezos would lay off half the company “in a fortnight” if necessary.
Rory separated the businesses: retail remains dominant through distribution, speed, robotics, and cost control; AWS is simply not relevant enough to new AI compute. Google had its own model, Microsoft “rented” OpenAI’s, and Amazon mainly had Anthropic. It must regain relevance without accepting the subpar economics Rory fears Oracle has taken on.

14. Relative value still turns on growth, access, and mission

Asked to choose Brex at $13 billion or Ramp at $30 billion, Rory refused without Ramp’s growth rate. Brex was cited at roughly $700 million of revenue and 50% growth versus Ramp around $1 billion; the core venture question is the equilibrium: how much additional multiple should an investor pay for each increment of growth?
Harry nominated a16z as the best-performing mega-platform of the prior 12 months, reversing his earlier skepticism after working with multiple partners. Rory agreed its operational delivery and returns had been excellent; Jason’s test was founder preference, and a16z increasingly appears among the top two choices across many stages and founder types.
Rory preferred Anduril at roughly $50 billion for its mission and its achievement in becoming a defense prime, while questioning roughly 50-times-revenue economics against OpenAI or Anthropic near 20 times with millions of customers. Jason acknowledged the party-bragging value of Anduril but declined: weapons were not his interest, and he no longer attends enough of those San Francisco parties for that psychic benefit to matter.

"Is there an AI bubble?” Gavin Baker and David George 118. Li Xiang’s Second 3-Hour Interview: CEO Models, VLA, Wisdom