Summary

Anthropic’s $10B at $350B looks cheap if growth holds one more year. Roel O’Driscoll’s math: revenue runway went $100M (end-‘23) → $1B (‘24) → ~$9-10B (‘25); assume it “only 3Xs” to $30B and average the ARR endpoints and you get ~$20B GAAP revenue next year — roughly 17x NTM, “a much lower revenue multiple than Palantir… comparable with Cloudflare, for God’s sake.” “It turns out you really, really can pay up for anything that goes 10X year on year.” Roel expects this is the last private round before IPO, and Jason Lemkin reads the small raise as a tell that “the unit economics are probably healthy.”
Claude has won enterprise and is coming for the whole stack — the premium API, coding via Claude Code (capturing 100% of a coder’s spend instead of 50% at 50% gross margin), and now a workspace product for non-coders that aims at the “AI Office Suite,” a “huge-ass idea” that should scare Microsoft. For Cursor, the scorpion-and-frog risk is live: Anthropic already cut off xAI’s access, and could limit, degrade, or simply copy — “I would be nervous if I was a 27 billion pre-Cursor investor,” though Jason “would still invest.”
OpenAI won’t go to zero, but it carries genuine existential risk. Roel’s scorecard: the Anthropic:OpenAI value ratio has converged from roughly 8–10+:1 to 2:1 — “You’re still in the lead, don’t blow it.” Jason’s bear case: LLM shelf life is under 100 days, OpenAI needs ~$100B over 2-3 years, and if a macro shock freezes fundraising while Gemini and Anthropic keep shipping, it becomes “Detroit… AOL and dial-up.” His meta-warning: the whole market “has ascribed the odds of a downturn to less than zero,” and OpenAI “is a bet that the best of times lasts at least a decade.”
a16z’s $15B — over 20% of all 2025 venture dollars — means “they’ve won, and they’ve won really well.” Roel’s equilibrium math: the industry raised ~$75B against ~$300B of exits this year (maybe $500B next), so a16z just has to capture 10% of everything — which a DST partner’s data shows they already do at the Series A. But there are limits: their hit rate is already much lower than Benchmark’s, “index investing is a scale business, stock picking is not,” and the math is brutally top-dependent — the top three private companies could be ~$1.8T of the $3.6T total if SpaceX really were worth $1T.
The “middle is dead” thesis has a wrinkle: a16z is itself four roughly $1B boutiques (American Dynamism, fintech/AI apps, infra), “very Alfred Sloan” — what they add is brand air cover and a $5B late-stage fund whose real function is “clean up on aisle five”: “You can be promiscuous at the A if you have enough late-stage stuff to cover it up.” Harry’s addition: ballooning growth assets buy price elasticity at the early stage — they bid 300 when you bid 150.
Late-stage AI is 100% correlated valuation risk. Roel’s stress test on Databricks — $4.5-5B revenue, 40%+ growth, cash-flow positive, ~25x at $100B: if growth slows to just 20%, history says ~6x, i.e. $30-36B. “Everyone’s leaned in so far that if it dislocates even slightly, the pain impact will be magnified.”
ElevenLabs at $11B: Jason says no, despite calling it the best API he’s ever used ($330M revenue in ~2 years, implemented in ~90 seconds). His own game burned $30 of credits in 48 hours with 20-30 players — “$1,320 a month” at current usage — and he was hunting substitutes in week one. Roel’s underwrite: a 3X needs ~$5B of speech revenue at 6-7x, and survival requires distributed demand, not white-label whales. Jason’s broader call: this is the year substitution risk gets real — “Mark Benioff will actually be right. We will rotate out for cost.”
California’s “entrepreneur’s tax” is a Trojan horse, per Jason: prior versions already targeted annual taxes at $50M then $25M paper-wealth thresholds, and assessment by voting control turns 5% into ~50% for super-voting founders. Brin joins Page in leaving; the rational meme becomes “leave before the Series B.” Behind it, a wealth gap widening toward unrest: 1 in 3 Nvidia employees worth $20M+, startups normalizing at $1-2M revenue per employee — “Now 100 billion doesn’t feel like that much, does it?”

Deep dive

1. Anthropic at $350B: “You really, really can pay up for anything that goes 10X”

Roel’s back-of-envelope, done live: Anthropic “very kindly did their numbers in round units of 10” — $100M runway end-‘23, $1B end-‘24, ~$9-10B end-‘25. Assume next year they “only, quote, ‘only 3X’” to $30B; average opening and closing ARR and you get ~$20B of GAAP revenue next year, so ~17x NTM — “a much lower revenue multiple than Palantir. It’s kind of comparable with Cloudflare, for God’s sake.” “If the growth is there for one more year, it looks cheap.”
Roel expects this is the last round before IPO — “they’ve stated they want to, and it feels like they can.” The investors who paid $170B three months ago “are at 2X in four months. Calculate that IRR, Harry.”
Jason’s tell: only raising $10B signals healthy unit economics — minimal dilution from a company that “owns code creation, application creation… everything we have spent our lives working on.” His hedge-free read: “it’s hard not to believe we’re in the first inning. It’s literally hard not to believe it.”

2. Claude’s three enterprise fronts — and the AI Office Suite prize

Front one is the premium enterprise API, which “has won it” per Jason and birthed Cursor, Lovable, Replit, Harvey — “it’s tough to stop this train.” Front two is Claude Code: having noticed coding was their single largest use case, Anthropic built the app, moving from “maybe 50% of the revenue of a coder, with a gross margin of 50” to 100% of the revenue by selling the product — enterprise coding share now roughly comparable to Cursor and GitHub, per Roel.
Front three, announced the day before recording: Claude for non-coders (“Claude Workspaces, I think it’s called. Don’t quote, I could be wrong”). Roel’s zoom-out: every knowledge worker buys the Office Suite — “What is the AI version of the Office Suite?… The idea that for every knowledge worker there can be some product like this is a huge-ass idea.” His caveat stands as spoken: early reviews say “amazing, but a bit janky,” and “I don’t think this is it yet” — but the direction is scary for Microsoft.

3. Cursor and the scorpion: strategic dependency on your own supplier

Harry’s data point from 20 Products interviews: every CPO now names Claude Code, and Cursor mentions have “gone down dramatically in the last three months.” Roel: nervousness depends on entry price — fine at $200M pre, “I would be nervous if I was a 27 billion pre-Cursor investor.” His consolation frame: graduating leagues — Cursor now plays against Microsoft and its own supplier, “but you’re damn glad to be playing there, ‘cause the other 10 coding agents ain’t even gonna get to play.”
Jason’s scorpion-and-frog: Anthropic cut off xAI’s access this very week, and the sting has many forms — “they just might limit access to the top models. They just might degrade it,” or copy outright: “how hard is it to build an IDE that’s just the same as Cursor?… They can build Replit and Lovable too. These are not the greatest challenges of mankind.” And yet: “I would still invest.”
The posture both land on: “How nervous can you be holding a large position in a product that didn’t work a year ago? You can only be so nervous or quit the game.” Roel’s version — in SaaS you compounded for seven or eight years; “now there’s existential risk every six months, and if you can’t live with that, you probably need to find a different job.”

4. Google takes Siri — and the consumer flank turns promiscuous

Gemini-for-Siri matters because “if you’re in the distribution business, you wanna be on a billion phones” — Google already pays roughly $10B a year just for search placement. Roel flags the odd money flow: Apple may be paying Google for Gemini while being paid far more for search — dynamics that “could flip” if OpenAI ever ran ads. Jason’s blunter read: “Google feels like a far more stable partner than OpenAI. It’s just the reality.”
Harry’s evidence for consumer promiscuity: a 22% drop in ChatGPT usage since the new Gemini models, plus his own confession — “I’m a total slut for a new model.” Jason’s household short signal: his son pays for Cursor out of his own pocket but dropped the $20 ChatGPT sub because Google’s free — “invest in whatever your kids do… if my son’s off ChatGPT, we got a short.”
Roel’s pushback, kept as hedged: he doesn’t believe this is “the kind of user base that churns at 100 days notice” — consumer stickiness is increasing, and he still prefers ChatGPT for his own research.

5. Could OpenAI go to zero? The bear case, fully argued

Roel calls the Sebastian Maltby (name unclear) “OpenAI goes to zero” take “absurd” — 800M users, subscriptions, a real business. His scorecard: the Anthropic:OpenAI value ratio has gone from roughly 8–10+:1 to 2:1 in three years. “You’re still in the lead, don’t blow it… They just need to frankly focus, knuckle down, and make it work.”
Jason’s bear case is structural: LLM shelf life is under 100 days, OpenAI needs ~$100B over the next two to three years — “more than it has spent to date” — and if capital freezes while Gemini and Anthropic keep going, “you die if you don’t have the capital.” Frozen in time, ChatGPT becomes “Detroit… or AOL and dial-up. Grandma’s fine with ChatGPT from a year ago because it helps her with recipes, but the rest of the world’s moved on to broadband.”
Roel’s two-by-two rebuttal: death requires the corner where scaling laws still matter and macro turns — “concatenated probabilities that I think are fairly low,” and airplane crashes are always multifactorial. The mitigation is the old Bill Gates rule: never hold less than two years of opex cash, and OpenAI has “the world’s best fundraiser.”
Jason’s meta-point — the sharpest of the episode: “We have returned to a moment in time, it feels like late 2020, 2021, where we have ascribed the odds of a downturn to less than zero… I think OpenAI has existential risk. It is a bet that the best of times lasts at least a decade.” A 10-year run with no downturn “would be a long one historically.”

6. a16z’s $15B: “They’ve won, and they’ve won really well”

The raise is over 20% of all venture dollars raised in 2025. Jason’s two-by-two: a16z uniquely holds the biggest fund AND the strongest founder brand, plus published top-tier returns — “it’s hard to do both.” His conclusion: “you might as well Hoover up 51% of the capital and then just shut down your competitors.”
Roel reframes the “not enough exits” objection as an equilibrium question: $15B at 20% share implies the industry raised ~$75B, against roughly $300B of exit value this year — “not an amazing year” — and plausibly $500B next once Anthropic goes public. “It’s not like it’s stupid, the amount of money,” especially since venture has been deploying a couple hundred billion a year while raising only $70-80B.
On a sustaining basis it’s ~10% of industry capital, so the job is to get 10% of everything — and a DST partner’s decade-long study showed a16z did roughly 10% of all Series As that became $5B outcomes. “Provided they could execute that all the way up the stack, they make it happen.” The catch: with total private value at $3.6T and the top three companies potentially ~$1.8T of it if SpaceX really were worth $1T, “you simply can’t make this kind of math work without getting those top exits” — you don’t need SpaceX’s A, but you’d better be on the cap table before a trillion.

7. If 10% works, why not 51%? The limits of scaled picking

Jason keeps pushing: seeing every deal is table stakes (“at Insight, I learned this from Teddy — you get fired if we don’t see every deal”), conflicts are solvable the way law firms solve them, and Klaviyo’s Andrew Bilecki asked the same question years ago: why not 40-50% share? “You’re a wimp not to try it if you can access the capital” — especially with no a16z fund below 4X gross or below 3X.
Roel’s counter is the episode’s cleanest structural claim: quality decays with scale. The DST data shows a16z’s Series A market share is higher than Benchmark’s but its hit rate is much lower — more pickers means the marginal picker is worse. “Index investing is a scale business. Stock picking is not… it gets hard to be smart in a room with more than five or seven people.”
He also enjoys the whiplash: “I came in expecting to argue they can’t make the math work at 10%, and clearly I convinced you they can, so now you’re like, ‘Fuck it, if you can do 10, why not do 50?’”

8. “Are you not the middle?” — a16z as four boutiques with air cover

Harry lands Alex Rampell’s line from Monday’s show — “the middle is dead” — directly on Roel: “I mean this in the nicest and most loving way… are you not the middle?” Roel concedes the AUM pressure (“you’d be a fool not to say it”) but rejects the framing: boutique means focused, not small — his firm does 20-30 enterprise deals per fund, no consumer, no crypto, and at Series A/B check sizes “you couldn’t be a $250 million, quote-unquote, ‘boutique.’”
His killer observation: a16z itself has split into four funds roughly $1B each — American Dynamism over a billion, fintech/AI apps ~$1.5B, infra ~$1.5B — “very Alfred Sloan, if you read the founding of GM… each of those funds is a fricking boutique fund at a billion dollars, just like us.” What the $15B adds that a $900M fund lacks: “the air cover of the brand” and the late-stage money to cover misses.
The survival condition for everyone else is the Peter Thiel question — “what do you know that no one else knows?” — because “if you wait till it’s consensus or anything close to consensus, you’re probably gonna lose.”

9. The real power of the late-stage fund: clean up on aisle five

Jason doesn’t buy the standard pitch: “I don’t believe founders highly value the fact that VCs can fund you through every stage… I don’t think an average Series A founder is picking you because you diluted the founders of Navan to 5%.” Founders pick a16z “because I believe Mark and Ben and team are gonna help me build a $100 trillion company.”
Roel’s third argument — the one he flags as the really compelling one: “You can be promiscuous at the A if you have enough late-stage stuff to cover it up.” Get three or four Series As wrong at $20M each, put a billion into the one winner, 2X it, “and you’ve covered your nut… It’s clean up on aisle five.” The deeper point: Benchmark approaches venture as investors trying to pick best; a16z approached it as engineers asking “how do you engineer an overall system such that it works?”
Harry’s addition from the trenches: ballooning growth assets buy price elasticity on early assets — “they can just come in and bid 300 when we’re bidding 150. Doesn’t fricking matter, ‘cause David George is gonna put in a 300 million check at 3 or 4 billion… that’s the real alpha.”

10. The meta-question: can you still find a $10B gem outside the system?

Jason’s framing of why he hasn’t quit: “If you cannot find a $10 billion gem outside the boundaries of this system, then this is all a game of fees… it’s all 25K checks into hot YC companies, and it’s all a lifestyle joke on Twitter.” The counter-evidence he clings to: Anthropic’s co-founder said 22 of 23 VCs passed on their first raise. His remaining niche — the “glitch in the matrix” second seed, when re-acceleration is invisible: Owner, which Redpoint didn’t see, Jason seeded, and Redpoint then funded every round since.
Roel’s numbers on market efficiency: ~20-something percent of unicorns went through Y Combinator, so ~80% didn’t; his “hard to beat” list of ~10 firms appears in 40-50% of Series As and ~80% of cap tables by the Series C.
But picking still matters, because venture “does a stunning job of missing the turn”: Salesforce struggled to get a dime from venture; OpenAI’s first venture round was at $23B pre with Thrive, Anthropic’s first at $4B pre with Spark and Menlo — Khosla the amazing exception. “We’re sitting here saying ‘we have structurally solved all our problems’… in the end, turns out picking matters.”
Harry’s tweet that got him in trouble: “The worst place to be investing is Series A. You either need to be pre-seed or pre-IPO today to make money.” His sharpening of the danger: if AI company growth has genuinely become linear — Lagora, ElevenLabs, Lovable show no faltering — then picking matters less and beauty-contest ranking decides, where the $15B fund and “the guy who invented the product” outrank you.

11. Early stage is uncorrelated business risk; late stage is 100% correlated valuation risk

Roel’s axiom, worth the price of the episode: “In the early stage, you’re taking uncorrelated business risk, and in the late stage, you’re taking 100% correlated valuation risk. When it goes wrong, it’s gonna go wrong for all of them.” When everything becomes obvious, “valuation expands to fill the vacuum” — the best firm wins the round but pays market price; “Lovable is not saying ‘I’ll take six billion when I could get eight billion from someone else.’”
His stress test on “one of the best companies out there”: Databricks at $4.5-5B revenue, 40%+ growth, cash-flow positive, ~25x revenues at ~$100B. If growth slowed to just 20%, two decades of history say such companies trade around six times: $30-36B. “If the growth stays, the valuations stay. If the growth slows even slightly, you have a dislocation to the downside.”
The behavioral kicker: because it’s been “so clear and obvious and linear for three years, everyone’s leaned in so far that if it dislocates even slightly, the pain impact will be magnified.” And his standing rule: “be very nervous when you think everything’s going to work.”

12. ElevenLabs at $11B: the best API Jason’s ever used — and he still wouldn’t buy

Jason’s live experiment: he vibe-coded founderscape.ai (a founder-life simulator), added an ElevenLabs-voiced CTO — “a 99 out of 100 product,” implemented “in less than five minutes, maybe even three” — and burned $30 of credits in 48 hours with 20-30 players; Replit estimates $1,320/month at current usage. Verdict: “At 11 billion? No… because I already wanna substitute out. In less than one week.” But on the founder: “Mati is the kinda guy I would just wanna bet on no matter what. Even if the ship went down.”
Is there a 3-5X from $11B? “If the whole world uses voice the way all the VCs talk about it, of course there is… It’s 330 million in one year. So what’s 100 times that?” — but the same ease of integration that produced $330M in ~two years is what makes switching trivial.
Roel underwrites it on the fly: a 3X means $30B, which at scale trades 6-7x “‘cause that’s the way human life is, dude. Get over it” — so you must believe in ~$5B of annual speech revenue. The escape from “the gravitational pullback of substitution” is a distributed customer base: if revenue comes from Epic Games paying half a billion, “Epic Games is gonna design them out”; if it comes from tens of thousands of adopters spending $20-30K, it’s defensible — and ElevenLabs has already proven the distributed pattern. His counter-anecdote: a speech company whose $2-3M customers all told reference calls they’d swap out — “you fast-forward five or six years, they’re all still on the platform.” At $10M spend, they probably would.
Jason’s broader call, hedged as spoken: “This will be the year, probably the back half, where we have to take substitution risk seriously.” His own SSO example — Clerk at $30/month when Replit itself costs $30/month: “when Replit launches a product as good, I will immediately delete it… Mark Benioff will actually be right. We will rotate out for cost.”

13. California’s “entrepreneur’s tax” is a Trojan horse

The mechanics Roel flags: wealth taxes always underperform projections — Norway and France introduced and unwound them — and this one assesses ownership by voting control, so a founder with 10X super-voting shares is taxed as if 5% were 50% of actual wealth. “Are you gonna sit in California worth $2 billion and give a billion dollars for the privilege of living here? You’re gonna leave.” His optimism, as hedged: the default California referendum voter says no, so “my guess is in the end this loses — but it’ll still have had a cost.” Brin has already joined Page in leaving; Chamath’s claim: ~$700B of $2T already gone.
Jason thinks Roel’s wrong on the vote and that the bill is “much more clever and worse than it looks”: a coalition that has tried versions three times, whose written end-state is an annual 1%+ tax at $50M then $25M paper-wealth thresholds — on illiquid assets marked at the last venture round. “You cannot solve an annual healthcare gap with a one-time tax… This is just stage one of the ultimate plan.”
His prediction if it passes: “leave before the Series B” becomes the rational meme — “come to Dogpatch, do YC, stay a year, build up your team, and then leave… you could have a Detroit in the Silicon Valley.” Winners: Miami and Austin, the places that “almost won in 2020 and 2021.” Would Jason himself go? “I’m on the edge” — an annual 1-2% on top of ~50% California taxes “compounds to 15 to 20% of your net worth… you gotta go to Miami or Austin at some point in your career, right?”

14. The wealth gap behind the bill — “I’m worried”

Jason’s structural point: even B2B is normalizing around $1-2M of revenue per employee — Replit does $300M with 200 people — so “it’s not about AI replacing jobs; we just don’t need that many people, and it’s gonna lead to malaise.”
The dispersion is real but corrosive: one in three Nvidia employees is now worth $20M+, 18,000 worth $25M or more — and in Palo Alto “there are literally zero houses for sale because it’s instantly bought up.” Meanwhile: “If you just got laid off from a previously high-flying public SaaS company growing 4%, who’s gonna hire you? The problem is nobody.”
The closing register, unresolved by design: “When you and I first met, Harry, a billion dollars was a good exit… Now 100 billion doesn’t feel like that much, does it?” Each year, people get “more and more angry at the AI decamillionaires,” which is why he reads this tax as only the first bill. “There is a level of social unrest that will grow over the coming years… I’m worried.”

No Priors Ep. 10 | With Copilot's Chief Architect and founder of Minion.AI Alex Graveley New Year Live 3: The AI Bubble, Chips and Black Swans Through Wall Street’s Lens