Summary

Venture is in a “low-low” regime: AI entry prices remain extreme while IPO and M&A liquidity remains scarce. Reggie calls venture unloved after no distributions in 2022–25, yet Jason sees an even stronger gold rush than 2021, extending “all the way down to high school kids.” The tension is that almost every GP will still deploy through the boom, even as the margin for underwriting error disappears.
Fast AI revenue is not automatically durable revenue. Lovable’s reported move from zero to roughly $18 million ARR in three months demonstrates the speed, but Reggie invokes a profile-photo app that surged from about $250,000 a month to $30 million before 99% churn pulled it back toward $500,000. His alternative megatrend is sub-1% digital penetration across trillion-dollar B2B categories: slower and less vulnerable to an OpenAI stack extension.
The right fund objective is to maximize multiple subject to an acceptable IRR, not maximize IRR at any cost. Jason’s 2017 fund sits at 4.31x and 32.56% IRR, yet could reach 5x, 6x or even 8x without lifting that IRR; Jason argues that continued compounding above roughly a 20% hurdle remains valuable. Fabrice’s “cheat code” is selective secondaries when a valuation no longer supports a 10x, while concentrated managers counter that “1x is not good enough”—a true winner should return 3x the fund.
Private marks remain unverified until public buyers or acquirers grade the work. The host says that, with few IPOs and M&A exits, VCs are “grading each other’s exams,” awarding themselves A’s while “the teacher appears to be on strike.” Cheap private capital lets companies postpone scrutiny; earlier IPOs return only when founders must accept public capital, not because investors tell them they should.
Longer private holding periods may now exceed the useful life of a software architecture. One panelist’s central fear is that every privately held software company will face at least one existential reinvention before IPO, while the panel predicts faster “terminal decay” for products such as website builders. This increases the value of founders who can survive a 10–15-year marathon, ship a second product and reaccelerate growth rather than settle into the “20-percenter club.”
AI may strengthen deep enterprise systems and vertical applications while compressing horizontal SMB products. ServiceNow’s roughly 20% growth at $12 billion ARR, its $3 billion Moveworks acquisition and Box’s opportunity to make trillions of documents conversational suggest incumbents can use data and distribution. But better products do not guarantee value extraction: falling coding-tool prices and eight well-funded competitors can transfer enormous customer value into capital and pricing wars.
At today’s prices, distinguishing genuine product-market fit from social-circle revenue is the venture job. The panel challenges whether $400,000 ARR from 20 friendly YC contracts proves anything; ten unaffiliated customers might, but even Windsurf and Cursor were radically different businesses when first funded. The old warning has shifted forward: investors now risk “paying Series A prices for seed risk,” with too little valuation cushion to bury mistakes.
Geopolitics and talent geography are becoming underwriting variables, not side considerations. The panel treats Benchmark’s reported $75 million Manus round at a $500 million valuation as carrying deal, firm and political risk; one panelist instead backs low-cost Ukrainian defense manufacturing and tracks “cost per kill.” On company building, the compromise is European R&D where talent can cost half as much, US go-to-market where buyers concentrate—and an unresolved bet over whether AI makes elite engineers 100x performers while eliminating half of mediocre sales and support roles.

Deep dive

1. The AI boom is colliding with a venture liquidity drought

Reggie’s macro framing: AI absorbed roughly $100 billion of venture investment, doubling from Q1 to Q4, while other categories received little attention. After no distributions in 2022, 2023, 2024 and so far 2025, the IPO and M&A “spigots” remain closed; that makes venture contrarian, but he would avoid funds pursuing “AI all the time.”
Rory calls the current market the difficult “low-low quadrant”: new deals are expensive and old investments cannot exit. In 2021, capital was easy to return but hard to deploy cheaply; today both sides are hard. His reality check is equally blunt: outsiders hear VCs complain about paying founders more and respond, “Nobody cares about our problems.”
Jason sees less drought than gold rush. A roughly 15% Nasdaq fall around April 7 frightened growth investors for three days, yet a company that had just closed a unicorn round soon received a higher-priced top-up offer. “The shock doesn’t even last until the next 20VC comes out”—and the rush now includes 17-year-olds dropping out to build.

2. Explosive AI revenue can reverse as quickly as it arrives

The host’s challenge to Reggie: avoiding the consensus means watching AI companies such as Mercor and Lovable scale in days while conventional vertical software takes years to move from $1 million to $4 million. A company still needs some megatrend to travel from $1 million of run-rate revenue toward the roughly $300 million associated with an IPO.
Reggie says roughly 9% of his investing remains in AI, focused on differentiated data, defensible business models and less-crowded categories. His warning specimen is the forgotten profile-photo app that moved from about $250,000 monthly to $30 million MRR, raised at the peak, suffered 99% churn and fell toward $500,000: “These things are riskier than people think.”
Platform extension compounds that churn risk. Reggie built Fabric AI with LangChain, Pinecone and separate voice tooling, then ripped out his backend stack when OpenAI released GPT-4o and became dramatically better. His concern is “the orthogonal risk of disruption”: an application can reach $100 million ARR quickly and still become a zero when a model provider expands.
His slower megatrend is B2B digitization. Petrochemicals, gravel and other trillion-dollar markets still lack catalogs, factory-capacity visibility, ordering, payments, tracking and financing; penetration can be below 1%. He prefers entering around $4 million pre at pre-seed, $10 million pre at seed and roughly $20 million pre later, accepting slower adoption for lower disruption risk.

3. Multiple, IRR and time are different dimensions of fund performance

Reggie breaks returns into three inputs. Picking is most controllable: in a 20-company fund, perhaps four must become great, and “if you can’t get that right, you should lose your job.” Entry and exit valuations are partly controllable; exit timing is least controllable, leaving managers exposed to “IRR erosion” when a 2024 outcome merely slips into 2026.
He credits one full turn of an excellent 2014 fund not to selection but to multiple expansion before its 2021 exits. The honest distinction matters: managers sometimes receive an undeserved valuation windfall and sometimes suffer the reverse, while delayed exits only help if operating compounding offsets the extra years.
Jason’s 2017 fund illustrates the measurement trap: 4.31x with a 32.56% IRR, yet scenario analysis suggested even a 5x, 6x or 8x result might not increase IRR. His LPs said they did not care about the distinction, prompting the question of why a manager should sacrifice another good turn merely to protect an already-high headline rate.
Jason’s answer is “maximize multiple subject to a constraint on IRR.” If LPs require roughly 700–800 basis points above small caps and an asset can keep compounding at or above about 20%, selling to preserve a 32% fund IRR can destroy value. Reggie’s supporting point is that replacing a six-year holding you understand with a higher-multiple company you barely know is “such a risk escalation.”

4. Secondaries help liquidity, but portfolio construction determines when to sell

Fabrice sells portions of winners when the next valuation supports only a 3x rather than the 10x he seeks. Competitive rounds can combine 30% primary with 15% secondary, and marketplaces such as Forge or SharesPost make that practical. Many of his distributions over the past three years came from this “anti-VC strategy of selling my winners.”
Price can override conviction. Fabrice sold numerous companies near 100x ARR in 2021 because merely justifying those marks required “every star in the multiverse aligning.” The panel also notes that institutional liquidity needs and two years of forward revenue credit can justify trimming, while buyers only want the obvious winners—not the portfolio’s dogs.
Jason’s concentrated-fund objection is that “1x is not good enough for me anymore”; he wants a winner to return 3x the fund. Rory also values smaller wins and late additions that approach a full fund return. The apparent paradox resolves through role: core winners create the fund, while secondary wins become highly valuable once carry is active.
Fabrice’s 500-deal construction works differently. Roughly 2% of companies return the first 1x at an average of about 4–6x; 8% return another 1x at about 8x; the remaining 90%, including profitable outcomes across somewhat under half, produce the third 1x. That diversified “1-1-1” model has produced roughly 3x and 30% IRR over 28 years.

5. Venture marks are an exam that public markets have not graded

The host questions whether reported IRRs are reliable at all. Unicorn financings and internal markups create apparent value, but without IPOs or M&A there is no external feedback loop: “We’re grading each other’s exams and we’re all saying we’re getting A’s.” The S-1 is the moment a company “takes its clothes off” and outsiders finally see the numbers.
Companies stay private because they can, not because remaining private is inherently superior. The panel’s maxim is, “People don’t do what they should; people do what they must.” When abundant cheap private capital retreats, founders will accept public-market discipline earlier and liquidity cycles will shorten toward equilibrium.
Public capital should normally cost less because it is liquid; private investors are locked up for five years. The current reversal is a “point-in-time absurdity,” although public-company bureaucracy, activists, quarterly budgeting and full disclosure make the founder experience materially worse.
The panel distinguishes viable and miserable IPOs. A business at $150 million revenue growing 50–70% should be able to list; one near $200 million growing 30% may receive a thin, illiquid $600–800 million market cap and little analyst coverage. The dot-com peak’s roughly 350 IPOs at median trailing revenue of $18 million was clearly too early; today’s near-zero bar is too late.

6. A 15-year private-company marathon selects for obsessive founders

Roughly 90% of B2B companies still have a founder CEO at IPO, but the journey keeps lengthening. Founders must reinvent themselves and sign up for another “tour of duty” every four or five years; what began as a three-year sprint became a 10-year race and is now a marathon followed by several sprints.
The Discord and Ironclad leadership departures prompted the host to ask whether founders are concluding that public-company leadership “sucks.” One panelist views leaving roughly 12 months before an IPO as rational for a founder who enjoys scaling teams but not Wall Street. Reggie says replacing a founder is an extraordinarily high bar: a founder CEO with managerial limitations usually outperforms an excellent manager with no founder DNA.
The longer horizon can make selection easier because the successful founder must be visibly abnormal—obsessed enough to treat the problem as a life’s work. The host’s negative filter is the “20-percenter club”: nine-figure-revenue CEOs reassuring one another that 20% growth is good enough, rather than trying to reaccelerate.

7. Software may now become obsolete before it can reach an IPO

One panelist’s largest fear is that “the technology life cycle to obsolescence is now shorter than the holding period” of private software. If so, every scaled company will face at least one existential architecture, reinvention or second-product crisis before public liquidity. Squarespace, Wix and Pinecone are offered as signs that technical cycles can produce much faster decay.
Plain-vanilla SaaS is particularly exposed after 10–15 private years. One panelist describes a portfolio company that spent 18 months near 20% growth, kept doubling down on engineering during the downturn, made acquisitions and returned above 40%. That path—accepting a temporary stall to build the next engine—separates renewed compounders from companies fading toward 15% and 10%.
Reggie’s pushback is category-specific: scaled marketplaces can gain from AI because liquidity, transaction data and operating efficiency already create a moat. Instead of being displaced, a vertical marketplace can use AI to improve pricing, matching and funnels; the risk is greater for slow, horizontal incumbents than for focused startup incumbents.
eBay illustrates the distinction. Its challenge is not simply poor teams but a horizontal stack that cannot deliver the best workflow in every category; a Pokémon or handbag marketplace can specialize. Reggie cites Rebag’s photo-based AI assessing model, authenticity, condition and price—an end-to-end vertical solution that eBay’s data alone does not reproduce.

8. AI could favor horizontal models but deeply vertical applications

Reggie expects the application layer to reward hypervertical products that solve an entire insurance, retail or manufacturing workflow. At the foundational LLM layer, however, GPT or ChatGPT might win broadly as Google did in search: DALL-E became good enough inside his existing subscription that he stopped using Midjourney, even if coding remains open to Cursor-like specialists.
ServiceNow’s 24% stock pop followed growth around 20% on roughly $12 billion ARR, but one panelist notes that this rate has been consistent for five or six years. Another cautions that enterprise CEOs overstate immediate AI impact: Salesforce’s claimed 500,000 Agentforce transactions sound large until compared with more than 100,000 on SaaStr AI.
One panelist’s competitive map has three cohorts: the pre-AI behemoth, an “AI teenager” building since roughly 2018 and a post-LLM startup. ServiceNow’s approximately $3 billion purchase of Moveworks—around 1% of its market capitalization—was a shrewd way to absorb the teenager and strengthen its claim to remain a post-AI winner.
Box is the harder test. AI can turn trillions of previously unsearchable documents into conversational data, and the panel respects Aaron Levie’s 20-year record of persistence through activists and competition from Microsoft and Google products offered free. Yet the open question is whether an “order of magnitude better application” reaccelerates revenue to 20–30%, or merely preserves the franchise.

9. AI pricing wars can create user value while destroying investor value

A panelist’s concern is the gap between value creation and extraction: products can generate a billion lines of code daily while prices fall from roughly $30 toward $15–20. Another panelist sees Windsurf’s tiering as rational rather than charitable—similar to OpenAI’s free, $200 and $2,000 levels—because free or cheap adoption feeds higher-value paid tiers.
Jason calls the structure a barbell. Easy entry produces marketing and a long tail, while enterprise customers can pay roughly $60–100 per seat through six- and seven-figure contracts supported by 200–400 sellers. That revenue base may absorb substantial OpenAI or Anthropic costs even when the visible starting price looks uneconomic.
Reggie still expects capital destruction before a winner emerges. Eight well-funded teams may overspend on acquisition and underprice products because a winner-takes-most market rewards survival first; GitHub and Microsoft have existential reasons to attack Cursor and Windsurf through Copilot. Durable pricing power may arrive only after investors have funded the war.
Dilution makes headline marks deceptive. A friend’s model-company investment rose from a $4 billion to $60 billion valuation yet produced only a 3.1x multiple after roughly 9% annual employee stock compensation and repeated multibillion-dollar raises. The panel nevertheless rejects abandoning the megatrend: extraordinary technology with excessive capital is still preferable to no technology trend at all.

10. Product-market fit is hardest to read precisely when prices demand certainty

The seed framework has three inputs: “awesome freaking founders,” a directionally correct market and economics that make sense. At an early-revenue Series A, the investor needs evidence that the company has “locked into something that can hunt”; paying for product-market fit and then funding a complete reset means the original price was wrong.
The disputed evidence is $400,000 ARR. The panel argues that 20 friendly YC companies signing $20,000 contracts can reflect social proximity, not market pull. Ten customers outside the founder’s batch might indicate genuine early fit—the investor is paid to decide which revenue is real signal.
Earlier entry does not eliminate that problem. Cursor and Windsurf were radically different businesses when seeded; Windsurf’s predecessor Codeium had its engineers repurposed only about three months before the discussion. Those bets were principally on exceptional founders, and the panel doubts they work most of the time even when the founder looks S-tier.
The pricing hazard remains: Series B investors once feared “paying Series B prices for Series A risk.” Today the same mismatch appears at Series A, where investors pay Series A prices for seed risk. With less valuation room to absorb mistakes, selection skill and win rate—not a forgiving capital market—must do the work.

11. China and defense investing put firm-level risk beside deal returns

The reported Benchmark-led $75 million Manus financing at a $500 million valuation raises three questions. A China-based AI company may offer compensation for idiosyncratic political risk at the deal level, but the panel also sees firm risk—congressional scrutiny or reputational spillover—and a moral question one panelist declines to resolve as a geopolitical amateur.
One panelist’s position hardened through experience. He invested early in Alibaba and Ant Financial, then withdrew after Jack Ma disappeared from public view and Xi Jinping stopped Ant’s IPO. He similarly exited Russia after the 2014 invasion of Crimea stranded companies without Western capital, and moved away from Turkey as Erdoğan departed from Atatürk’s principles.
Fabrice offers the portfolio simplification: the US represents roughly 25% of world GDP but about 50% of software and enterprise technology spending. “If I can’t make money on 50% with the rule of law,” adding another 10% where the rules can change arbitrarily does not improve his odds.
One panelist draws a different moral line around defense, backing Ukrainian startups and investing in Anduril. His preferred metric is the grim but explicit “cost per kill”: Ukraine’s low-cost manufacturing could provide deterrence that expensive Western systems cannot. He concedes that many new defense investors are trend-following lemmings, though momentum into the eventual winner might still pay.

12. Build where talent is efficient, but sell where the money is

One panelist tells European founders to build in the US because roughly 300 million wealthy early adopters make it “the game of life on easy mode.” He argues that once a business reaches $100 million in US revenue, moving to $200 million domestically is easier than building from zero to $100 million elsewhere—and that may remain true even at $1 billion.
The European case is that salaries are lower, retention is higher, capital is available and companies such as Pigment show that founders can sell into America from Europe. Parisian elite engineers may cost half as much and stay beyond their vesting cliff. The synthesis is clean: R&D can be anywhere; enterprise go-to-market must largely face the US.
One panelist disputes Europe’s operating cadence more than its technical competence, citing founders who say their teams cannot match San Francisco’s pace. Another counters with Revolut and “Project Europe”; the panel rejects judging an entire continent but agrees that globally distributed outliers differ from the average workforce. The founder sets the cadence, not the passport.
AI sharpens the disagreement. One panelist says ordinary engineers have already become roughly 2x and elite “10x engineers” could become 100x; he cites about 50% productivity across Windsurf and 20% of Salesforce code committed through AI. His two-year forecast is categorical: half of mediocre sales and customer-success roles disappear, while larger, more intense engineering teams fight an accelerating arms race.

Trump's First 100 Days, Tariffs Impact Trade, AI Agents, Amazon Backs Down Clashing Over Commerce (Fintwit Book Club April 2025)