Summary

DPI ultimately settles whether a venture fund made money, but TVPI remains a noisy early signal rather than “nothing.” Sam Lessin backs Chamath’s “you can only eat net DPI” line and separates investors from fee-driven asset gatherers; Rory counters that a three-to-four-year 2.0x TVPI fund probably contains more information than one at 0.8x. Sam’s practical compromise: early on, show him “five names that matter”; later, return cash.
Venture is splitting between small funds with entry-price leverage and mega-funds able to overwhelm late rounds, leaving roughly $200 million-to-$1 billion vehicles under pressure. Sam calls $1 billion “the death zone,” while Rory argues a $900 million fund can still work with about 30 investments and $20 million-$30 million checks. The genuine danger is ownership: when a normal Series A consumes $20 million-$30 million and giant funds can deploy at almost any price, mid-sized managers need exceptional selection and pricing discipline.
Selling winners is a separate skill from finding them, and the right answer depends on whether an investment still has “infinity” potential. Sam’s first rule is “you cannot sell the things that matter,” but once the thesis breaks—as it did for him in Allbirds and Astra—available liquidity should be taken. Chime crystallizes the dispute: selling at its prior $25 billion valuation looks smart against an approximately $14 billion market cap, yet a later speaker notes that selling Revolut at $25 billion could miss a path to $100 billion-$150 billion.
The panel’s sharpest disagreement is over whether merely good companies matter when only a handful become generational. Sam says OpenAI “might matter” but that even its long-run importance remains unsettled; Harry and Jason push the logic through Box, asking whether Aaron Levie therefore wasted 16 years building a roughly $5.5 billion company. Rory accepts he may never back a $100 billion outcome, yet rejects the resulting nihilism: $5 billion-$10 billion businesses and an $80 million carry check remain real value.
AI demand is historic, but expectations and infrastructure commitments leave almost no tolerance for an ordinary outcome. ChatGPT reached 800 million users in 17 months, while the Big Six spent $212 billion on CapEx and Sam frames a roughly $600 billion CapEx business. Rory’s warning is that if OpenAI proves merely “as good as Google” on the comparable growth curve, it could miss next year’s projection by about 40%—a fundamentally great result that markets could still treat as failure.
Falling inference costs and Chinese competition undermine any thesis built on scarce, expensive intelligence. Mary Meeker’s report puts token-cost compression at 99.7% in two years; DeepSeek reportedly reached 93% of o3-mini’s performance for a fraction of the cost, while Baidu ERNIE was cited at 0.2% of GPT-4.5’s cost. Sam’s call is categorical: “The price of a unit of intelligence is plummeting every month,” so teams should build for capabilities and economics likely to exist by launch.
The immediate SaaS risk is organizational inertia, followed by agents and MCP erasing the application interface. Jason calls cautious pilots the “AI slow roll” and says startups need “existential dread”; Sam expects systems of work to capture value from systems of record that increasingly become invisible databases. Mangomint’s CEO supplied the concrete warning: if ChatGPT can choose and book a spa without exposing its scheduling software, the SaaS vendor can “become a pipe overnight.”
Liquidity is returning, but price and ownership determine who benefits. Thoma Bravo raised a record fund of roughly $34 billion-$35 billion after distributing $30 billion, normal IPOs are resuming, and Snowflake and Databricks paid roughly $250 million for Crunchy Data and $1 billion for Neon as AI blurs data-platform boundaries. Meanwhile, YC AI companies are being pushed toward $50 million-$60 million post-money valuations; that is excellent financing for founders, but a venture fund ending with 3% ownership must be dramatically better at picking.

Deep dive

1. DPI is the verdict, while TVPI is contested evidence

Harry opens with Chamath’s provocation: “TVPI is bullshit vanity metric. You can’t eat IRR. You can only eat net DPI.” Sam reluctantly agrees despite his “default instinct” to oppose Chamath.
Sam sees two businesses sharing one label. Investing means finding companies early, paying correctly, selling, and returning cash; asset gathering means accumulating fee-bearing capital. The latter is “actually a better business,” but he considers it “a stupid game.”
Rory calls Chamath’s claim trite for 2013 and 2015 vintages: after ten years, promises are inadequate. During venture’s five-to-seven-year illiquid period, however, TVPI is a “loose proxy”; funds at 2.0x and 0.8x after three-to-four years are unlikely to have identical eventual distributions.
Sam distrusts marks inflated by financing rounds or inconsistent valuation policies, sometimes viewing them as negative signals. He concedes that “zero is an extreme statement,” but wants only “five names that matter” early and, once sufficient time has elapsed, proof that the manager made money.

2. Institutional incentives make paper marks useful marketing

Sam’s LP explanation is less about portfolio truth than careers: the junior institutional allocator who selected a fund cannot wait seven years for promotion. TVPI gives that person “some marketing thing they can then use for their own internal purposes.”
This is why statements can be laughable yet organizationally useful. Funds preserve the highest marks they can come up with, LPs receive evidence of progress, and everyone satisfies near-term incentives even when sophisticated participants know “this is not real.”
Thoma Bravo supplies the cash counterexample: it raised a record fund of roughly $34 billion-$35 billion after approximately $30 billion of distributions in the prior year. In a quarter when reportedly nobody raised even a $5 billion PE fund, liquidity itself unlocked more capital.

3. Venture’s hollow middle faces incompatible check-size economics

Sam is confident a roughly $200 million early-stage fund can produce DPI, but calls $1 billion “the death zone.” At $10 billion, the objective changes into asset gathering and deployment; consistently generating venture multiples is no longer the same game.
Rory, managing about $900 million, defends stage-appropriate construction: roughly 30 investments, with $20 million-$30 million checks into $30 million-$40 million rounds. His concession is that $10 billion conglomerates make disciplined deployment harder for managers with $500 million-$900 million.
Jason reframes the SVB data through founders: the old $8 million Series A is now “three SAFE notes,” while $20 million-$30 million has become normal. If middle-sized funds disappear, founders may need relationships with $5 billion-$10 billion firms before demo day.
Sam resolves the apparent contradiction between abundant mega-fund capital and companies struggling to raise: only a few businesses in each generation matter, while much Series A money is misallocated. Large funds can overwhelm the perceived winners and let the remainder go unfunded.

4. Seed survives through entry multiples and private liquidity

Sam sees durable room for investors who explore ignored parts of the economy and become “an N of 1 or an N of a few.” At near-zero starting values, a broad seed portfolio can still make the DPI math work.
Early investors also possess secondary optionality unavailable to a fund holding a $30 million-$40 million Series A position. Sam believes “private to private is an important future”: small holders can sell into later private demand without being too large to exit.
Mega-funds have the opposite advantage. They can put “a gajillion dollars at almost any price” into the tiny set of presumed winners and make meaningful dollars from scale, even if the multiple would disappoint a smaller return-maximizing fund.

5. Selling requires fewer repetitions—and more honesty—than buying

One investor describes an offer that could return his entire fund at once, but selling would eliminate all further upside. The tension is not whether 1.0x DPI looks good online; it is whether the remaining position is the portfolio’s irreplaceable outcome.
Sam’s first rule is “you cannot sell the things that matter.” His second is to recognize when the thesis has broken, as it did in former holdings including Allbirds and Astra, and sell to buyers whose cost basis, ownership, or allocation objectives differ.
Early investors receive hundreds of opportunities to practice buying and very few to practice selling, so exit judgment matures more slowly. The decisive question is whether the company remains “an infinity shot”; if not, a sensible cash price can serve both sides.
Rory adds the time value of life itself: a founder who sold after ten years used the proceeds to marry, buy a house in Spain, and start again. A later comparable company became worth more, but “that was my life choice.”

6. Chime proves both the value of liquidity and the middle-stage model

Sam argues an early Chime investor should have sold at the prior approximately $25 billion valuation. Jason notes that seed holders did not appear among principal shareholders in the filing, suggesting at least some did exactly that: “That means they did their job.”
Rory separates that conclusion from Menlo Ventures’ Series B decision. Buying around a $200 million-$300 million valuation and potentially earning 10x-15x validates savvy A-to-C stock picking by a mid-sized fund, even if Chime is not generational.
The same defense applies to Hinge Health: a position returning $400 million to a $6 billion fund may look modest proportionally, but 20% carry is $80 million. Rory’s rebuke to dismissive colleagues: “Every single one of you will cash the fucking check.”
A later speaker’s Revolut pushback preserves the upside risk: rounds in the billions repeatedly looked expensive, yet the company could reach $100 billion and perhaps $150 billion. Sam says he would “probably sell” both Revolut and Chime now because each has roughly 15% of its true TAM, where acquisition gets harder and “CAC only goes up.”

7. “Mattering” is not the same as building a valuable company

Sam calls Chime good but questions whether it is important. His candidate list for paradigm-level impact includes Microsoft, Facebook, Google, Bitcoin, possibly Solana and Venmo, and companies such as Anduril or OpenAI that “can matter” but have not necessarily secured that status.
Harry presses Sam on the uncertainty around OpenAI. When Sam says, “Will we look back in 20 years and say OpenAI was a fundamentally important company? Maybe,” Harry calls the posture arrogant; Sam replies simply, “The game is young.”
Rory offers a distribution: many companies fail, some reach $1 billion, a few reach $5 billion, perhaps one annually reaches $10 billion, and one or two per decade reach $100 billion. He treats market capitalization as a rough proxy for importance but is content to create real $5 billion-$10 billion value.
His objection is the nihilism embedded in “only the things that matter.” Sam embraces the aspiration—being called “a market participant” feels insulting to him—while Rory answers, “I am a cog in the system,” without believing that makes sound investing or company building worthless.

8. Box exposes the gap between a fund manager’s duty and a founder’s life

Sam says Box itself does not matter in his narrow paradigm-shifting sense, though he deeply respects Aaron Levie for grinding a difficult, unglamorous company into existence. AI might now give Box “an opportunity…to matter.”
Harry and Jason drive the framing to its uncomfortable conclusion: if a roughly $5.5 billion company is irrelevant, Levie should have sold to Citrix around 2008, taken approximately $100 million plus retention, and avoided “wasting” the next 16 years.
Sam distinguishes the roles. A fund manager likely should have taken that liquidity; Levie is a person who wanted to keep running the company.
The broader AI excitement partly comes from mature companies sensing a new route out of “meh.” Some reinventions will prove real and others will not, but Sam finds “the race to matter” more compelling than merely participating efficiently in an existing market.

9. AI adoption is historic, but infrastructure is far ahead of applications

Jason’s headline from Mary Meeker’s report is ChatGPT’s climb from zero to 800 million users in 17 months—the fastest adoption curve cited. Netflix took 15 times longer and TikTok five times longer, making three-month-old intuitions about AI “super dated.”
The Big Six spent $212 billion on CapEx, yet Jason cannot locate the complete application demand: coding, support, custom analysis, or subsidized $20-per-month ChatGPT usage. He believes the foundation is transformational while remaining “so far ahead of the application level.”
Sam says hyperscalers converted excellent cash-efficient businesses into “CapEx hogs” that would break Buffett’s heart. Free cash flow itself grew enough that the hit was only about 10%, allowing investors to tolerate spending later discussed around $260 billion despite limited AI revenue.
Timing is the unresolved variable. If business adoption fills capacity within two years, the economics may work smoothly; if it takes four or five, companies could endure an extended gap between costs and revenue.

10. OpenAI’s greatest risk is being merely excellent

Jason warns that public investors are “mean VCs on steroids”: enthusiasm can reverse from “spend more” to “why are you spending so much?” Core-growth deterioration or a weaker economy could bring pressure resembling Meta’s partial retreat from its maximal VR posture.
Sam Altman’s Stargate figures—$500 billion to begin and potentially $5 trillion—are, in Harry’s reading, market socialization. The bet requires not one Amazon-like spender but every major participant accepting an order-of-magnitude expansion in infrastructure.
A scary moment could be two weak OpenAI quarters, 30%-40% growth, or unexpected saturation. OpenAI and Google’s revenue trajectories were described as almost exactly 20 years apart through roughly $1.3 billion-$1.4 billion and then $3 billion, before OpenAI’s projections accelerate sharply.
Rory’s key calculation: “If all OpenAI is is just as good as Google, then it’s gonna miss its number next year by about 40%.” That miss could create panic—and an investment opportunity—even though matching Google’s historical achievement would hardly constitute business failure.

11. Cheap Chinese models and collapsing token prices keep AI competitive

Meeker’s China data showed DeepSeek reaching 93% of o3-mini’s performance for a fraction of the cost; Alibaba reportedly outperformed both, while Baidu ERNIE cost 0.2% as much as GPT-4.5. Harry worries the US conversation has moved past both China and safety too quickly.
Rory expects ChatGPT may retain Apple-level quality, surrounded by Android-like alternatives that enforce price discipline. Corporate America may avoid Chinese-built models, but DeepSeek still demonstrated that competitors can get “quite close quite cheaply”: “We ain’t a monopoly anymore.”
Token costs fell 99.7% in two years. Usage and training expense may rise, but Jason has no patience for non-coding engineering leaders claiming AI is inherently too expensive for B2B products.
Sam’s operating advice is to build against the direction of travel: if a capability fails today, it may work by launch or six months later. “The price of a unit of intelligence is plummeting every month,” making delay more dangerous than current model limitations.

12. The “AI slow roll” is an organizational failure mode

Jason calls phased experimentation—limited Q4 release, followed by slightly broader use next year—the “AI slow roll” and “the number one thing killing B2B companies.” His forecast for these teams is blunt: “They’re just gonna be slaughtered.”
Sam points to Facebook’s post-IPO mobile pivot as the necessary top-down model: leadership committed the organization rather than treating incremental effort as credit earned. Incumbent inertia makes that kind of turn unusually difficult.
Windsurf founder Varun’s formulation is “startups beat incumbents because of existential dread.” A startup whose product does not convert loses; a strong incumbent engineer whose project fails is usually reassigned, muting the personal and organizational consequence.
Jason sees the right posture in Aaron Levie and HubSpot’s Yamini: “We’re so excited and we’re scared.” He wants board meetings shaped by urgency—recent hackathons and AI voice agents shipping Wednesday—not reassuring plans for another controlled pilot.

13. Agents will capture workflow value while records fade backstage

Meeker’s next 32% of humanity coming online will be AI-first, using voice, agents, and natural-language interaction. Jason expects them to skip familiar constructs such as leads, contacts, opportunities, and even files: “Kids do not know what a file is.”
Sam expects a system of work to sit in front of every system of record. The AI closest to the employee becomes the daily product, while the underlying Salesforce database may persist for decades without retaining the user relationship.
A sales representative directing automated agents will regard that control layer as the CRM, regardless of whether Salesforce, HubSpot, another product, and internal databases sit behind it. Value accrues to software that helps perform the job, not merely “keeping score.”
Legacy vendors can keep compounding through installed bases while losing the next dollar of new functionality. Platform shifts put every company’s market share “up for grabs”; giants usually slow rather than vanish, but unprepared mid-tier vendors can hit a wall.

14. MCP could turn vertical SaaS into an invisible pipe

Sam initially found Harry’s MCP thesis “too nerdy” and “too Microsofty.” Seeing early integrations changed his view: once API-key friction disappears, users may conduct every HubSpot, Notion, or calendar action through Claude or ChatGPT without learning the source application.
Mangomint, approaching $25 million and serving spas and medical offices, made the threat concrete. An AI could find and book a Palo Alto spa without revealing whether Mangomint or another scheduler fulfilled the request, leaving the vendor to “become a pipe overnight.”
Sam translates “agentic” jargon into the consumer question: why can’t ChatGPT find and book the best haircut? The founder’s response was to rethink the value proposition immediately: if customers no longer recognize the application, maintaining pricing power becomes difficult.
Rory’s hierarchy is decisive: “I wanna be the decider. I don’t wanna be the tracker.” Software closest to the user’s desired outcome gains the right to swallow the tracking products behind it; even HubSpot must choose how freely agents can abstract its structured data.

15. AI is redrawing data infrastructure, venture entry prices, and exits

Circle interests Rory as “the boring version of crypto”: a roughly $43 billion-$44 billion pool earning around 4%-5%, producing just under $2 billion of revenue before distribution sharing, operating costs, and a couple hundred million dollars of profit. Its transparent, bank-like economics also bound the upside.
Snowflake’s approximately $250 million purchase of Crunchy Data followed Databricks’ $1 billion Neon acquisition within roughly 60 days. Both are becoming broader database companies as agents require more data structures; ClickHouse’s cited raise near a $6 billion valuation reinforces that “AI plumbing” is attractive territory.
YC is pushing many pre-revenue AI companies toward $50 million-$60 million post-money valuations, with AI around 70% of the batch. Jason’s rule of thumb requires the next seed round at 3x and an A or B at 2x; otherwise waiting may offer better risk-adjusted entry.
Founders receive perhaps $6 million instead of $2 million for similar dilution, while investors may finish with only 3% or less than half their historical ownership. SAFEs soften visible down rounds through many conversion prices, but only five-to-eight-year fund returns will reveal whether paying 60 instead of 20 worked.
The healthiest market signal is that ordinary IPOs are happening again, alongside Chime, Groww, Omada Health, Circle, and other transactions. Rory’s governing maxim is “price clears all markets”: realism about valuation restores liquidity, while venture investors remain far better at decade-long themes than predicting tomorrow’s stock reaction.
In the closing bets, Jason expects the first OpenAI-Jony Ive device may lack a screen but the eventual family will include screens, audio, wearables, and a phone app; Rory takes under five million units in its first full year, Harry takes over. Rory says Meta stays open absent government intervention, while Jason puts a closed release above 35%.
On Elon Musk leaving Tesla before 2027, Rory says it is not the base case but the odds allow for a “rage quit.” Jason moves to roughly 50%: Musk may seek a Gwynne Shotwell-like operator, while legal compensation complexity, Tesla’s brand, and the demands of SpaceX and xAI make delegation increasingly plausible.

No Priors Ep. 117 | With Co-Director of Stanford's HAI & Founder of World Labs Dr. Fei-Fei Li E194｜From Fenty to Rhode: How Western Celebrity Beauty Brands Conquered Gen Z