Summary

AI’s revenue evidence now makes outright pessimism harder than arguing about the slope of growth. Harry cited Altman saying OpenAI would hit $20 billion of ARR this year, Anthropic projecting $70 billion by 2028, and Gamma reaching $100 million of revenue with 50 people. Rory’s verdict: “Right here, right now, the revenue has shown up”; the investable uncertainty is whether next year’s CapEx is $80 billion or $40 billion, not whether demand exists.
Michael Burry’s $1.1 billion NVIDIA-and-Palantir short can be directionally right and still lose because options force the clock. With NVIDIA at $188, Rory used a 47-day $180 put and a roughly $9 stake: the stock must reach $160 to double the stake, while failing to fall below the strike wipes it out. A two-year put costs $50–55, loses money unless NVIDIA falls below roughly $150, and needs to fall below $100 to produce the same 2X. Moving from an “arm-wavy bullshit podcast statement” about overinvestment to a profitable timed trade is “damn hard.”
The next software TAM opens when AI stops being a copilot and becomes part of the team. Jason’s Replit v3 remembers prior work, discusses mistakes, and shipped a new production page in 15 minutes; Gamma already turns company data into bespoke sales collateral in roughly 10 minutes instead of weeks. The threshold is being “sufficiently autonomous, knowledgeable, and powerful” to complete material work with human check-ins — and then “the amount of revenue that’s accessible is so high.”
AI has not killed moats; it has moved the point at which they become credible further to the right. Incumbents that once took a year and a half or more to decide whether to clone a product can now produce credible copies in 30–90 days, so an early revenue explosion is “not as defensible” as it used to be. Jason still sees a later “plane of stability,” while Rory argues defensibility emerges only after distribution, engineering depth, and market anointment narrow the field.
Later-stage evidence reduces operational risk, but valuations are expanding fast enough to consume that benefit. Rory thinks the probability can narrow from one-in-ten at seed or early A to perhaps one-in-three by the B; Harry argues horizontal AI markets may still be one-in-seven to one-in-ten once Codex, Claude Code, Vercel, Figma Make, Salesforce, Atlassian, and other adjacent competitors are counted. Harry’s cited company moved from roughly $4–5 million of revenue at a $200 million valuation to $80–100 million at $2 billion, showing both genuine de-risking and the price of consensus.
More outcome variance logically calls for more diversification, but the ownership math can push a seed fund toward institutional scale. Jason’s example requires 40 initial $5 million checks, another $200 million of reserves, and roughly $100 million for fees and flexibility — a $500 million “little seed fund.” Harry raised the possibility that larger outcomes permit smaller ownership; Fabrice cited a 100–150-position LP fund using $100,000–$150,000 checks and argued that $1 million on a $50 million valuation can work if the company becomes worth $100 billion.
Fundraising has become brutally binary, and the best process is cultivated before the company formally raises. Jason sees a “Captain Obvious era” in which YC, Neo, or South Park Commons pedigree—or hot AI-native growth—gets funded while conventional SaaS struggles; Harry’s example grew from $400,000 to $3 million, took 120 meetings, and received one $10 million term sheet at a $40 million post-money valuation. Their shared conclusion: “The best-run processes don’t feel like a process” because several investors are already primed before diligence begins.
Public-market AI credit goes to companies capturing new budgets, not merely adding AI features. Datadog’s 23% jump fit the rule “sell shit to the people who are making AI,” while Duolingo’s 25% fall showed that “sprinkling AI dust” earns no premium without new economics. Jason’s hierarchy is explicit: attach to compute spending, replace human labor, or use AI to displace an incumbent and steal its revenue.

Deep dive

1. Sequoia’s handoff exposes the strain on the old venture playbook

Harry framed Roelof Botha’s departure as steward after three years, and Pat Grady and Alfred Lin’s elevation, as unusually consequential. Rory’s outside-in interpretation was blunt: “Whenever you have a CEO change happen, it’s ’cause something is wrong.”
Rory carefully stopped short of declaring Sequoia behind in AI. His claim was that an internal perception likely existed that the firm “could do better” after missing rounds, passing on strong companies, and competing against a newly aggressive market.
Jason widened the issue beyond Sequoia: many investors, executives, and founders from the last 10–15 years may not suit the next decade. “The old playbook doesn’t work”; for those unwilling to relearn, his prescription was to take their NVIDIA shares, buy a beach house, and check out.
Harry’s structural framing was “Walmart versus Chanel”: mega-platforms and walls of capital on one side, focused boutiques such as Benchmark and USV on the other. Splitting early and growth leadership may help specialization, but Rory warned that the person above them becomes a precarious “manager of managers,” increasingly removed from “eat what you kill.”

2. Sequoia’s ruthlessness may be healthier than partnership stability

Rory admired Sequoia’s willingness to reject the idea that leadership belonged to someone because “it’s so-and-so’s turn.” If the partnership believed change was needed, it acted rather than protecting an untouchable figure for another five or six years.
Jason’s less romantic view was that partnerships are intrinsically dysfunctional because performance rarely maps cleanly to economics. Equal carry feels collegial until one partner produces the winners and another makes the large losing bet: “The daggers are always out.”
Rory’s pushback — worth keeping — was not that the daggers disappear, but that economic performance should drive professional change. Sequoia’s willingness to say “It’s not working. Let’s make a change” was, in his view, both healthy and characteristically Sequoia.

3. Michael Burry’s short is a timing trade disguised as a macro thesis

Harry introduced Burry’s reported $1.1 billion short exposure to NVIDIA and Palantir. Rory agreed with the broad premise that AI CapEx will eventually overshoot and correct, but tested whether that belief could actually be turned into a profitable trade.
With NVIDIA at $188, Rory used a 47-day December $180 put as an example. For every roughly $9 staked, NVIDIA reaching $160 would produce about a 2X, reaching $100 would produce roughly an 8X, and failing to fall below the strike would mean “you lose it all.”
Burry’s unusually early filing mattered to Rory because a tightly timed position benefits from publicity. Having reportedly filed on the final permissible day previously, releasing this position early looked like an attempt to “pile on the bad news” and shit-talk the stock without saying so directly.
Extending the horizon does not remove the difficulty: a two-year $180 put might cost $50–55, lose money unless NVIDIA falls below roughly $150, and require a drop below $100 to double. Options are zero-sum, and being vaguely correct about an eventual correction is insufficient.

4. Present revenue makes fighting the AI trend “dumb as rocks”

Harry challenged the show’s repeated worry about whether AI revenue would arrive: Altman was said to expect OpenAI to hit $20 billion of ARR this year, Anthropic projected $70 billion by 2028, and Gamma had reached $100 million of revenue with only 50 people.
Rory put more weight on estimates being raised during the current year than on distant forecasts. Anthropic’s upward revisions were therefore meaningful, while CoreWeave’s problem — getting data centers operational, not finding customers — reinforced the evidence of compute demand.
His conclusion was unusually categorical: “The cynics sound smart and optimists get rich.” AI is perhaps the largest megatrend since the early internet, so “leaning into it is the only sensible thing to do, and playing against it is dumb as rocks.”
The legitimate second-order debate is sizing. A company spending $80 billion next year rather than $40 billion creates a consequential delta, even though both figures describe explosive demand; questioning extrapolation is therefore risk management, not denial of the trend.

5. Gamma and Replit show how tools become revenue-bearing teammates

Gamma raised $100 million at a $2.1 billion valuation after reaching $100 million of revenue. Jason’s SaaStr use case goes beyond “AI PowerPoint”: Gamma pulls Salesforce and marketing-automation data, calculates a sponsor’s prior leads and ROI, identifies competitors and similar companies, and builds bespoke collateral in about 10 minutes.
SaaStr pays roughly $100 a month, or $1,200 annually, for work that previously used free Google Slides or Microsoft Office and might take a marketing-operations team three weeks to complete badly. Jason called that stealth TAM expansion. He also argued that if Gamma continued toward $1 billion of ARR, a 20× revenue valuation would not sound expensive; Harry added that the business was profitable.
Jason had launched 10 applications in 125 days without an engineer. Replit v3 was the first agent he considered literally part of his team: it retained roughly a month of context, remembered earlier implementation choices, discussed mistakes, and put a new idea into production in 15 minutes.
His dividing line is autonomy, knowledge, and the ability to complete material high-value work with periodic oversight. The copilot was the 2024 story and failed as a paid-on tool; the current story is that AI finally works; the 2026 opportunity is AI embedded inside the team rather than merely helping individual employees.
Jason framed the surge as a recent capability break: Gamma was founded in 2020 but had no revenue before this year, while he said Replit and Vercel likewise exploded only once the models became good enough.

6. Faster clones move defensibility later, not necessarily to zero

Harry described the investor’s new burden as continually relearning what became technically possible during the prior 12 months and what will become possible in the next 12. “What I knew six months ago” can become useless quickly, making stale technological judgment equivalent to betting blind.
Jason contrasted the old response cycle, in which a company might take a year and a half to decide whether a clone was worth building and roughly two years before staffing it seriously, with an investment that attracted five clones — including one from a cloud leader — in 30 days. Canva’s presentation product had also become “borderline competitive” with Gamma in the short period since an earlier interview.
The counterweight is a later, still-fragile “plane of stability.” Jason argued that sophisticated products such as Replit can pull away because competitors cannot reproduce the underlying agent; Bolt, once an early leader, had fallen to third and outsourced its agent layer to Claude. He also warned that companies now need to work at a “996” intensity because a three-year product cycle can compress to 90 days.
Harry proposed vertical data accumulation through Solve Intelligence’s patent workflow, while Jason noted patents are public and ingestible by rivals. Rory’s broader rule was that seed defensibility is mostly imaginary: teams must run fast, stay technically superlative, win distribution, and become one of the market’s “anointed” winners.

7. The valuation question is whether investors are paid for wider variance

Jason accepted the high-risk team-and-speed model at a $3 million or $5 million post-money valuation, but questioned it at $50 million with $5 million checks. If innovation can be cloned immediately, investors need to ask whether the entry price compensates them.
Rory distinguishes stages probabilistically: at seed, investors mostly know the people; around an early A, five or six companies near $1 million of revenue may still imply one-in-ten odds; by the B, rate of change and competitive ordering can sometimes narrow the plausible leaders to roughly one-in-three.
Harry disputed that confidence using code generation and vibe coding. Cursor, Cognition, Replit, Lovable, Vercel, Codex, Claude Code, Salesforce, Atlassian, and Figma Make leave the field “one in many”; being a top venture-backed startup means nothing if an adjacent platform ultimately wins.
Rory conceded the remaining platform risk but argued the information gain is still real. Harry’s cited company moved from about $4–5 million of revenue at a $200 million valuation to $80–100 million at $2 billion, removing category and execution risk; valuation then expands to fill the space, leaving the final question: is the TAM large enough?

8. More uncertainty pushes fund construction toward diversification

Jason’s arithmetic started with 40 seed investments at $5 million each: $200 million deployed initially, another $200 million in reserves, and perhaps $100 million for fees and flexibility. Maintaining both ownership and diversification could therefore require a $500 million seed vehicle.
Harry argued that expanding outcomes permit lower initial ownership. Fabrice then cited an LP investment holding 100–150 positions through $100,000–$150,000 checks that had produced a 7× fund; he also argued that a $1 million investment at a $50 million valuation can work if the company becomes worth $100 billion.
Rory observed that Harry’s answer implicitly accepted greater diversification rather than denying the higher risk. Longer exit times and wider operating variance logically imply somewhat more positions, achieved through smaller checks, larger funds, or consciously lower ownership.

9. Investor meeting cadence must fit the strategy and the person

Harry said the partnership does 20 new-company meetings per partner weekly; with his four investing partners, that is 80 net-new companies met in person per week and more than 3,500 company meetings annually. Jason’s response was that he would “give you all my carry back” rather than adopt that schedule.
Rory called himself a “meeting junkie” because even an average opportunity can reveal a market insight unavailable in a deck. His interrupt-driven style skips most slides to extract the operator’s “crucial kernel of knowledge.”
Jason prefers a strong email, deck, financials, and five prior investor updates, now augmented by Claude. Only a truly great founder adds enough beyond that work to justify the meeting; founders with a 0% chance of receiving his investment gain nothing from forcing a coffee.

10. The best fundraising process is built before it officially begins

Harry found it abrasive when a founder rejected his ready-to-sign term sheet to launch a process on November 19. If he had already met the requested price, he reasoned, the founder was implicitly optimizing for a different partner and should simply say so.
Rory separated a committed offer from vague interest. A real term sheet deserves serious consideration, but sharing data serially with one or two uncommitted firms creates an “accidental process”; if they decline, the company has effectively failed a financing before formally starting one.
Jason’s preferred founder cultivates several investors through updates and relationships, then announces a round only when those investors are primed. The optimal version needs no traditional data room: once investors already want to invest, the remaining material can be a diligence file, such as the example Jason labeled “Box-diligence investment 12/21/25.”
Rory’s synthesis sharpened the point: “The best-run processes don’t feel like a process, but they are.” Timing pre-existing interest creates a competitive outcome without overt games, although only a sufficiently attractive and high-performing company can execute it.

11. Fundraising is binary, while AI rewards new budget rather than features

Jason called the market the “Captain Obvious era”: affiliation with YC, Neo, or South Park Commons can unlock pre-seed attention, while later companies need hot AI-native positioning and top-quartile venture growth. The formerly fundable middle has largely disappeared.
Harry’s concrete specimen was a conventional enterprise SaaS company growing from $400,000 to $3 million. It held 120 meetings and received one term sheet: a $10 million round on a $40 million post-money valuation, about 12× revenue for a business that had grown 10×.
Datadog represented the opposite outcome: its shares rose 23%, and Harry cited “$15 million plus AI-native customers,” without clarifying the metric. Rory summarized the playbook as “Sell shit to the people who are making AI”; observability, switches, routers, and interconnects all scale with unprecedented compute consumption.
Duolingo fell 25% after slightly softer guidance despite a good quarter. Jason called it the “wrong kind of AI”: improving an existing product earns no special credit today. A company must attach to compute, replace human labor, or use AI to seize an incumbent’s revenue.

12. Education and Hummingbird show two versions of capital efficiency

Rory resisted reducing every application to layoffs. In education, AI can give students individualized instruction closer to one-on-one tutoring than a class of 20; for language learning, the credible budget may come from adults already paying human coaches, not cash-constrained public schools.
Jason accepted the product benefit but kept asking, “Where does the budget come from?” If AI neither replaces labor nor captures compute spending, it must displace a legacy vendor — difficult for incumbents such as Duolingo that may need to cannibalize their own existing revenue.
Harry closed with Hummingbird’s first biotech investment, BillionToOne, producing an approximately $800 million position at IPO. He also cited BillionToOne at $5 billion and Nirvana at $4.5 billion, both as examples of the power of capital-efficient companies and concentrated venture ownership.
Jason called maintaining ownership from an eight- or nine-figure fund “God tier,” while Rory defended accepting dilution: going from 20% to 12% after investing $4 million can still create a spectacular fund. The small, high-MOIC vehicle is often the more compelling home for the marginal LP dollar.

No Priors Ep. 140 | With Benchling Co-Founder and CEO Sajith Wickramasekara Satya Nadella – How Microsoft thinks about AGI