Why the venture box keeps shrinking, why CoreWeave is a B-tier asset, and why build-to-prod is the most important category in software
Another week with Harry Stebbings and Rory O’Driscoll, and the theme ran through almost every segment: capital is concentrating into fewer names than ever. Great businesses outside the hype cycle can’t get a meeting. Mega-funds need billion-dollar positions just to matter. Anthropic posted a quarter most public companies would envy and the market read it as slowing. Meanwhile Modal tripled its valuation in five months and Berkshire doubled down on a software holding company. Here’s what we covered.
1. A $20M ARR, Profitable Company Growing 6-7x Can’t Raise
Harry opened with a company he’d just met: $20M in ARR, growing 600-700%, profitable, strong team. The VC response was silence. His tweet asking whether venture is broken drew a big founder response.
Jason’s read: I’m in a similar company. It raised a roughly $3M seed and now plans to never raise again. The investors get the monthly update showing 15-18% growth and nobody wants to meet. Capital is going to fewer names than I’ve ever seen, which is odd for a market as herd-driven as venture.
Rory’s read: It’s not that only 10-20 companies are getting funded. Data centers in space and quantum computing are getting money that wouldn’t have in another era. What’s struggling is vertical software, where there’s an unstated fear that AI makes the whole category go away.
Jason’s tactical advice: If you have top-decile numbers outside the hot categories, you have to meet 100 VCs. It’s the old days again. Benchmark won’t do it. But someone will, at an elevated valuation, because it’s the best deal they’ve seen that quarter. Top decile used to be an automatic term sheet. It isn’t anymore.
2. $5B Funds Need Billion-Dollar Positions in Everything
The debate moved to why the box is so narrow. Harry’s framing: unless you can paint a trillion-dollar market cap, the top firms aren’t interested.
Jason’s read: If you’re raising $5B+ funds every 18 months, a $2B outcome doesn’t work. Unless you own 51%, it’s very hard to build a billion-dollar position. In those IC memos, a billion-dollar gain is the B case. The A case is $25B or higher. It’s just math.
Rory’s pushback: The historical data says 10-20 decacorns per decade across all of US tech. The market is acting like there will be 90 this year. One of those two facts is going to be wrong, and he’s betting on 50 years of data. Maybe five funds can play the billion-dollar-position game. The whole industry can’t.
3. Seed Is Now Harder Than Growth
Harry asked whether the best seed investor should find it easier to return 5x net than a firm writing billion-dollar checks.
Jason’s read: No. It’s much harder. Entry prices are $50-100M at seed, and the multi-stage firms that lead those rounds crush your allocation. You end up with $300-500K in a company that goes to $18B, and it doesn’t move the needle.
Rory’s read: In a rational capital world, early-stage returns should beat growth. Otherwise why do it? But he agreed the last five years will show the inversion. The combined value of SpaceX, Anthropic and OpenAI went from about $400B three years ago to around $5T today. That’s $4T+ created at the latest stage. If you think there will be 10 more of those, growth is all you should do.
4. CoreWeave Drops 25% in a Week as AI Debt Fears Spike
CoreWeave fell 15.5% on October 14 to close at $65.78, its biggest one-day drop since February. Some of its bonds traded at yields above 15%, and default-swap spreads jumped past 1,000 basis points. By October 16 the stock was on pace for a 25% weekly decline.
Rory’s read: Neoclouds will always be the riskiest part of the AI trade. They underwrite high growth with heavy capex and lots of debt. Grow too slowly and customers leave. Grow too fast and any demand slowdown sinks you. There’s no hunkering down to cover your nut like a software company. Every “AI is slowing” week pushes them down, and last week was one of those.
Harry’s read: He’s a holder, and asked whether a 25% drop on a week when demand clearly isn’t cracking is a generational buying opportunity. Rory’s answer: demand going up isn’t enough. With a $20B capex plan, demand has to go up by a lot every year.
Jason’s read: CoreWeave and Oracle are B-tier compared to Nvidia. CoreWeave exists because Nvidia needs it to exist for now. Oracle can’t sell enough enterprise software, so it rents out excess capacity. B-tier assets should be more sensitive to shocks. I’m not putting CoreWeave in my great-grandchildren’s 529s.
5. Anthropic Posts $16.8B Quarter, IPO Talk Drops to $1.5T
Anthropic told prospective investors Q3 revenue hit $16.8B, up 45% from $11.6B in Q2 and more than triple Q1’s $4.7B, with reports putting adjusted gross margin at 85%. Some attendees then discussed a roughly $1.5T valuation, below the ~$2T investors had expected for the November IPO.
Rory’s read: Growth is slowing, but from an unsustainable level. If you add $5B a quarter linearly, you’re at a $20B+ quarterly run rate going into the IPO year and roughly $100-110B of revenue. That’s 12-13x forward revenue for the most exciting company in the world. It doesn’t take much squinting to make that work.
Jason’s read: In absolute terms everything is great. But the stock will be choppy after the IPO. Developers aren’t loyal, and the market share swings between Anthropic, OpenAI, open weights and new entrants have been huge just since this show started. There’s only so much “R” in this ARR. It wouldn’t shock me if Anthropic’s revenue share is down 10% a year from now and the business is still great on an absolute basis.
Rory’s addition: If share keeps shifting, the next shoe to drop is price competition. As long as end-user demand keeps growing, everyone wins. If it slows, pricing becomes the weapon.
6. Replit Flips Its Default Model From Claude to OpenAI
Replit changed its default model from Anthropic’s Opus and Fable to OpenAI. Coverage framed it as Replit dropping a partner.
Jason’s read: That framing was wrong. We use Replit 14-15 hours a day. It has its own router. Replit just changed the default because OpenAI’s models are 30-40% cheaper for roughly the same quality output. That could flip again next week.
The more interesting signal is what Replit didn’t do. Replit, Lovable, Base44 and v0 each run around $1B in revenue, with customers who are extremely cost-sensitive at every tier. They have every incentive to nail the quality-price routing. None of them moved their users to Kimi. If Kimi were as good as people claim, they’d route everything through it. These vibe-coding platforms are a lab we should be watching more closely than the labs spending the most money.
Rory’s read: Agreed. With a billion dollars of revenue, you have the time and the incentive to figure out what you’re leaving on the table by paying top-tier prices for work that doesn’t need top-tier intelligence.
7. Claude Code Review Launches, and Build-to-Prod Becomes the Category
Anthropic launched Claude Code Review this week. Harry asked which startups it threatens.
Jason’s read: It’s scary for every competitor, no matter what they say on Twitter. But independents have a real advantage: one LLM checking another LLM’s work. And code review is arguably more important than code generation now. LLMs get you 95% of the way. The engineer builds the feature on their phone at a barbecue, then spends three days testing and fixing it.
I’d go further. Build-to-prod is the single most important category in software today. Not the largest fortunes necessarily, but the most important. Everyone is building 10-14 hours a day in Claude Code or Cursor, and the time to make that code production-ready is still an order of magnitude more than the time to build it. Generation frameworks are solved enough. Let Sam and Dario duke it out. If I could find the best next-gen build-to-prod teams, I’d make 10 bets.
Rory’s read: Coding is the mother lode of LLM value. The whole software development life cycle changes in an agentic world: code review, memory, testing, sandboxes. The fastest-growing AI apps all sit closest to the foundation models. That’s the Icarus trade. Every six months you have a terror moment. That’s the price of the highest growth rates we’ve ever seen. If you want to avoid it, go back to the dull 3x software companies nobody wants to fund.
8. Modal Raises $750M at $15.75B, Crosses $1B Run Rate
Modal raised a $750M Series D led by Accel at a $15.75B post-money and said it crossed $1B in annualized run rate. The Series C in May was $355M at $4.65B, so the valuation more than tripled in under five months.
Rory’s read: Modal was under $100M mid-to-late last year. That acceleration is stunning. Add up the four or five key inference providers and you get maybe $5-6B of run rate, versus $100B+ for OpenAI and Anthropic combined. So open-weight inference is under 5% of dollars. Think about Replit choosing between a 70% gross margin token from Anthropic and a 30% gross margin token on Modal. If quality is similar, you know where it goes.
The risk: the neoclouds will eventually ask why they take all the capex and debt while the inference layer makes the money. But this growth covers a multitude of sins, and margins are likely creeping up.
Jason’s read: I like it better than CoreWeave. It adds value instead of being outsourced racks of GPUs. The size of the inference market is the one thing I know today that I wish I’d known 12 months ago.
9. Town at $1B and Anduril at $90B: Are Private Multiples Disconnected?
Town, from former Plaid CTO Jean-Denis Greze, raised a $108M Series B at a $1B valuation, co-led by Forerunner and Menlo. Separately, Anduril opened an employee tender at $94.80 a share, implying roughly $90B, five months after a $5B round at $61B. Anduril generated $2.2B in revenue over the last year.
Harry ran it as an investment committee: does the growth fund do Town?
Jason’s read: Pass, but I won’t block it. A billion is the new $100M. Old Series B investors didn’t care about price under $100M because they wanted a $1B+ IPO. Town is a classic early bet with a bow tie on it. Smart team, interesting early revenue, too early to predict.
On Anduril: clearly generational, but is it worth 40x revenue with its margins? Braze was the best exit in public GTM at about 6x. Monday and HubSpot trade around 2x. My worry is AI versions of low-multiple categories. When growth slows to 60-70%, do they trade like Cloudflare or like Braze? Marketing is the worst public category. HubSpot and Klaviyo aren’t dying and still trade under 3x. I’d struggle to get an IC excited.
Rory’s read: The multiples are almost certainly disconnected. A mature $2B defense contractor is worth $8B at best. If you’ve doubled for years and have a line to the Pentagon for five more, maybe you go 2-4-8-20 and justify it. It’s leaning a lot into the future. And don’t count on US multiple expansion to bail anyone out from here.
10. Berkshire Adds $330M to Bending Spoons, Stock Jumps 41%
Berkshire is buying 10 million new Bending Spoons shares at $33, adding to the $600M it invested at the July Nasdaq listing for a 4% stake. The stock closed up 41% at $53.46 on the announcement.
Jason’s read: Bending Spoons is the Berkshire version of software. These aren’t flips. They’re holding companies, like See’s Candy or Coke. And the asset quality is improving. Vimeo, Evernote and Eventbrite were dogs they made work with price increases. Miro and Airtable are better. Berkshire is betting Bending Spoons can make these endure for a decade or more in a way Vista can’t.
Rory’s pushback: A See’s Candy from the 1960s looks like one from today. The Evernote of 2080 won’t. It’s conceptually Berkshire, possibly applied to an industry where the idea doesn’t fit.
11. 37signals’ $370M in Profit Since 2020 and the Case for “None and Done”
37signals disclosed what implies roughly $370M in profit since 2020, based on its 10% profit-sharing payouts.
Rory’s read: Go back to the start of the episode. VCs don’t care about these software companies. Hold it yourself, pull out $370M, and you’ll be fine. Not everything has to be “$25B or die trying.”
Jason’s read: The number may be a little high, but it’s in the zone. The comparison I keep coming back to is Zapier: one venture round, about $300M in revenue, a generational company that probably won’t IPO. What if the three founders had raised nothing and split $200M of the $300M? I used to think one-and-done was nirvana. In the AI era, I think it’s a trap. Dividending profits out with a VC on the cap table is hard. None-and-done is better.
Rory on buybacks: If an investor owns 30% and you buy them out at 6x cash flow, you’ve tied up two full years of cash. Doable, but risky with debt. Buffer and Expensify did it at mediocre outcomes for investors.
12. GA Heads to IPO, Harvard Posts 27.1%, and LP Cash Is About to Recycle
General Atlantic is set to go public. Harvard’s endowment returned 27.1% to $68B, helped by venture, including a SpaceX stake worth about $2.2B.
Rory on venture firm IPOs: You need roughly $800M-1B of fee income to matter, because public investors credit fees and barely credit carry. Only the very largest will do it. And it’s hard to go public without becoming a more risk-averse, corporate institution.
Jason on the LP cycle: If you’re an LP without exposure to Anthropic, OpenAI or SpaceX, you have serious FOMO right now. Cash is coming back from SpaceX, Anthropic is on the horizon, OpenAI maybe in H1. That cash will recycle into venture, and the market is about to get hotter. But 27% is no 2021. Show me the 60% IRRs.
13. Tesla’s Roadster as the Swan Song, and the Elon Premium
Tesla showed off the new Roadster with thruster videos. Elon also continued the push to put a SpaceX share in every American child’s account.
Jason’s read: It’s the swan song before Tesla merges with SpaceX in the next six months. Elon basically said it: China will dominate EVs, and this is the last car a human will drive. On the shares: it’s very hard to hate someone when you’re a shareholder. Giving wealth directly to people beats routing it through a foundation.
Harry’s read: The brilliance of Elon is the Elon premium. Stunts like this justify extreme valuations and buy another 12 months of momentum at no cost.
Rory’s read: The kid accounts are an interesting way for the ultra-wealthy to make meaningful gifts at scale. About $10B in aggregate is real money, and because it’s tied to the share, there’s real marketing in it.
3 Quotable Moments From Each Speaker
Jason Lemkin
- “You’ve got to have billion-dollar positions in everything in big boy venture today. Billion-dollar positions are nothing. There’s no point in investing otherwise with large funds.”
- “Build to prod is the biggest category. It works in dev. It looks great. How do I get that thing into prod? That’s where the big money is going to be made now.”
- “A billion today is the new $100 million.”
Harry Stebbings
- “I have a company that’s $20 million in ARR, growing 600-700%, profitable, brilliant business with a great team. Crickets. VCs do not want it. Is venture broken today?”
- “Would you not say for the first time ever that we’re actually going to see that inverted?”
- “The brilliance of Elon is the Elon premium and his ability to have extremely high valuations because of his presence.”
Rory O’Driscoll
- “Coding is the mother lode of LLM value.”
- “Look, compared to Nvidia, we’re all B-tier assets.”
- “When you have this kind of growth rate, nothing else matters. It covers a multitude of sins.”
