Mega-funds need $1 billion positions, so profitable startups go unfunded

1. Profitable, yet ignored — ▶ 1:15

   A company at $20 million ARR, growing very fast, profitable, with a great team, gets silence from VCs. A similar agentic company one speaker backed is "literally planning to never raise again" because no one wants to meet it. The workaround is old-school: meet 100 VCs. Top-decile numbers once meant an automatic term sheet. Now the marquee firms will pass, but a growth investor outside the famous names will take the bet at an elevated yet reasonable valuation, because it is the best deal they have seen that quarter.

2. Power law on steroids — ▶ 2:35

   Capital is concentrating in fewer names "more so than I've seen ever in venture before". It is not risk-off: data centers in space and quantum computing still get funded because the upside is ultra-high. What starves is software for specific markets, where an "inarticulate fear" that AI will make it go away dominates. The test is a "power law on steroids": "unless you can paint me the trillion dollar market cap, I'm not interested." For Sequoia, Benchmark or a top Andreessen partner, the box is narrow: "Cursor better", OpenRouter.

3. Billion-dollar position math — ▶ 5:47

   A fund of $5 billion or more must hold $1 billion positions in everything, or "there's no point in investing". A company like Listen Labs is not enough because, short of 51% ownership, which founders resist, a $1 billion position is hard to build. 1% of Anthropic gets there even after dilution. Memos assume a 10% probability of a $25 billion outcome or higher. Funds of $5-10 billion are raised every 18 months, so many billion-dollar positions are needed to even play. "It's just the maths."

4. Decacorn stats don't add up — ▶ 6:36

   The pushback: US tech has produced 10-20 decacorns per decade across 50 years of data, yet the current thesis expects 90 this year. "One of those two facts is going to be at variance with the other." Rory O'Driscoll bets on the history, lifted for growth. He says investors should want some $25 billion outcomes but cannot confine themselves to deals where a billion dollars can work. The whole industry cannot chase them: "maybe five" funds is enough.

5. Seed returns inverted — ▶ 8:08

   Seed is harder, not easier. Entry prices are $50-100 million, ownership is lower, and multistage firms lead rounds and crush allocations to $300-500K. In a company heading to $18 billion, that still doesn't move the needle. In a rational market early-stage multiples should beat growth, but the last five years are "clearly inverted". SpaceX, Anthropic and OpenAI together went from about $400 billion three years ago to about $5 trillion. "Nothing else matters" if 10 more arrive next year; if not, make different bets.

6. CoreWeave's no-slack model — ▶ 10:51

   CoreWeave fell 15-20% in a week. Neoclouds are always the riskiest part of the AI trade because they underwrite high growth with heavy capex and debt: by O'Driscoll's recollection, about a $7 billion run rate, $10-20 billion of capex and $35-40 billion of debt. Grow too slowly and customers are unhappy; too fast and a demand wobble leaves you "screwed". Demand must rise "by a lot every year". The gauge is OpenAI plus Anthropic revenue combined against CoreWeave's capex. Lemkin calls it B-tier: it exists because Nvidia and friends need it to.

7. Anthropic's slowing growth — ▶ 17:05

   Q3 revenue was $16.8 billion, up 45% on Q2, at an 85% gross margin. Growth of 10x a year for three years cannot repeat. But if revenue added $5 billion a quarter (5, 11, 16 so far), a $20 billion run rate entering next year puts a $1.5 trillion IPO at 12-13x forward revenue. IPO talk has slipped from $2 trillion. The risk is share: developers are not loyal, Anthropic's share could be 10 points lower by end of next year, and share defense usually means price cuts.

8. Replit's routing signal — ▶ 21:53

   Replit made OpenAI models its default over Anthropic's. Lemkin says the press story was wrong: it flipped a default inside its own router because quality is peer-level and cost is 30-40% better. Replit, Lovable and Base44 are strongly incentivized to nail routing since they can only partly pass savings to users who want everything for $25 a month. Enterprises fear variable bills. If Kimi were as good as claimed, they would route to it entirely. They have not, so watch what these builders do.

9. Build-to-prod is the category — ▶ 25:37

   Lemkin calls code review, better framed as "build to prod", "the single most important category today in software". LLMs get you 95% of the way, then you spend three days testing and fixing. Review takes an order of magnitude more time than building. It may not mint the largest fortunes, with no "600 billion code review cursor", but he would make 10 bets even if Anthropic kills them. Apps close to the foundation labs grow fastest but face a terror moment every six months. O'Driscoll: coding is the "mother lode of LLM value"; a $10 billion outcome earns his check.

10. Inference is under-sized — ▶ 31:37

    Modal closed a $750 million Series D at $15.75 billion, with run rate above $1 billion, up from under $100 million late last year. The four or five main inference providers total roughly $5-6 billion, while OpenAI and Anthropic exceed $100 billion, so inference is under 5% of lab revenue. A Replit can pay a lab a 70% gross margin on a token or pay Modal 30% for the same token. The risk is that neoclouds do it themselves. Lemkin likes it better than CoreWeave and regrets passing.

11. Multiples have detached — ▶ 38:50

    "We are just taking revenue multiples too far today." Braze, the best public GTM exit at 6x revenue, compares with HubSpot and Monday near 2x and Klaviyo under 3x. A mature $2 billion-revenue defense contractor is worth about $8 billion best case, so a $90 billion defense company at 40x revenue needs years of doubling. O'Driscoll agrees multiples are almost certainly disconnected, justified only by growth. Fund size explains the rest: a billion dollars is now "the new 100 million" for a Series B, and a $200 billion upside case justifies paying up.

Fund size decides what gets funded before any founder walks into the room. A $5 billion fund, raised again every 18 months, cannot be moved by small wins. Every memo has to answer one question: can this return $1 billion to us? That implies a 10% chance of a $25 billion outcome, and a position large enough to matter.

The arithmetic is brutal at seed. Entry prices run to $50-100 million. Ownership is low. Multistage firms lead the rounds and crush seed allocations to $300-500K. Jason Lemkin holds one of those checks in a company heading to $18 billion, and it still does not move the needle. Further up, there are two routes to a $1 billion position. One is 51% ownership, which founders refuse. The other is something like 1% of Anthropic.

The filter this creates is severe. Capital is concentrating in fewer names than at any point the hosts have seen. Data centers in space and quantum computing still get funded, because the upside is ultra-high. Vertical software does not, because of an inarticulate fear that AI will erase it. The test is a power law on steroids: "unless you can paint me the trillion dollar market cap," the answer is no. So a profitable company at $20 million ARR, growing fast, with a great team, gets silence. The workaround is the old one. Meet 100 VCs. Someone outside the marquee firms will take the bet at an elevated but reasonable price, because it is the best deal they have seen that quarter.

Is the filter irrational? Not if you look at where the returns went. In a rational market, early-stage returns should beat growth-stage returns, or nobody would take the risk. The last five years are clearly inverted. The combined value of SpaceX, Anthropic and OpenAI went from about $400 billion three years ago to about $5 trillion. That is $4 trillion of value created at the late stage. Rory O'Driscoll's conclusion is conditional. If 10 more companies like that arrive next year, that is all you do. If they do not, you make different bets.

The same logic of betting on the top runs through the AI trade, and so does its fragility. CoreWeave fell 15-20% in a week on a whiff of slowing growth. O'Driscoll's explanation is structural. Neoclouds underwrite high growth with heavy capex and debt. By his recollection that means a $7 billion run rate, $10-20 billion of capex and $35-40 billion of debt. Grow too slowly and customers are unhappy. Grow too fast and a small demand wobble is fatal. It is not enough for demand to rise. It has to rise by a lot every year, measured against the capex. The cleanest gauge is OpenAI and Anthropic revenue added together, which strips out share shifts. Lemkin calls CoreWeave and Oracle B-tier assets. They exist because Nvidia and friends need them to, for now.

The labs look stronger but less steady than their growth suggests. Anthropic's Q3 revenue was $16.8 billion, up 45% on Q2, at an 85% gross margin. Growth of 10x a year for three years cannot repeat. Yet O'Driscoll shows how little squinting the IPO price needs. Add $5 billion a quarter, and a $20 billion run rate becomes about 12-13x forward revenue at $1.5 trillion. He does not think that looks crazy. The fragile assumption is share. Developers are not loyal. Anthropic's share could end next year 10 points lower and still be a great business in absolute terms. But then the next move is price. Replit shows how fast buyers move. It made OpenAI models its default because quality was peer-level and cost 30-40% less. Replit, Lovable and Base44 can pass only part of the savings to users, so they are strongly incentivized to get routing right. When share fights become price fights, the comfortable idea that everybody can win ends.

Value outside the labs gathers where model output has to become something usable. Lemkin argues that code review, better called build to prod, is the single most important software category. LLMs get you 95% of the way. Then come three days of testing and fixing, and review takes an order of magnitude more time than building. It may not mint the biggest fortunes. He would still make 10 bets, even if Anthropic kills them. Inference is the other pocket. Modal closed a $750 million Series D at $15.75 billion, with a run rate above $1 billion, up from under $100 million late last year. The four or five main inference providers total $5-6 billion against more than $100 billion at the two labs. That is under 5%. A Replit can pay a lab a 70% gross margin on a token or pay Modal 30% for the same token. The risk is that the neoclouds notice and do it themselves.

Which is where it gets harder. Prices. "We are just taking revenue multiples too far today." Braze, the best public GTM exit, went at 6x revenue. HubSpot and Monday trade near 2x. A mature defense contractor with $2 billion of revenue is worth about $8 billion best case. A $90 billion defense company at 40x revenue needs years of doubling and a line to the Pentagon. O'Driscoll agrees the multiples are almost certainly disconnected from public comps, and that only the growth rate makes them plausible. Then the loop closes. A fund that must move $1-3 billion can pay up for a company with a chance of being worth $200 billion, and be happy at 2.2-2.5x. Fund size sets the price, and the price confirms the fund size. A billion dollars is now the new 100 million for a Series B.

O'Driscoll's decacorn statistic sits under all of this. US tech has produced 10-20 decacorns per decade across 50 years. The thesis driving the mega-funds expects 90 this year. One of those facts is wrong, and he bets on the history, lifted for growth. If he is right, the billion-dollar-position model is a business requirement presented as a forecast, and the funds that stayed small were not the fools. If he is wrong, they missed everything. Nobody in the conversation settles it. Lemkin admits he passed on Modal and would now pay $20 billion to get back in.