Summary

Apple’s trade-secret suit against OpenAI is less about the stolen files than leverage: Apple is “pissed” that 400 employees have defected, and the panel expects it to “twist the knife, depose everyone and see how high it goes.” The tradeable read is that OpenAI’s hardware effort is “probably on the bubble” — Jason’s call is it gets cut like the Apple car, and “if this was a mercy killing, Apple may have done them a favor,” because “LLMs are really amazing at code and the amount of economic value that can be created from that dwarfs everything else.”
Rory’s TAM math is the episode’s bear case: the BLS counts only 1.88M US developers (~$250bn of total wage spend), and with Anthropic exploding from $9bn at the start of the year to “what now might be $50bn” — a huge portion of it code — AI coding revenue is “close to 20% already” of the wage pool it addresses. Either growth “decelerates faster than anything you’ve ever seen,” or “the prize for becoming the fastest growing company in human history is you may hit TAM faster than any other company in human recorded history.”
The next leg is the agentic tax: ~$1.5T of software spend can tolerate roughly 10% of topline going to tokens — “just like Amazon took a tax on software… now everyone’s going to run using LLMs, give me 10%” — worth another ~$140bn, plus an AI co-worker sold at $20 a seat like Office. Salesforce can pay 10% of revenue for tokens; it cannot pay 40 on 22% operating margins.
Meta’s Spark 1.1 — launched with Zuck’s first X post in three years — ends free open-weight distribution and starts charging developers via API at pricing Alexander Wang called “very aggressive.” It’s “a battle for the cheap seats”: every org is moving from “burn all the tokens you want, baby, to budgets,” and “every company with a CIO who’s half awake is going to have a cheap token model to hand to stop this madness.” The right metric, per the Databricks paper: cost per completed task, not cost per token.
SK Hynix’s $26.5bn NASDAQ listing — the largest-ever Nasdaq listing by a foreign company — brings the three-player memory oligopoly public at 5-8 P/Es with net margins that went from negative in ‘23 to ~70%; “Samsung is the most profitable company on the planet right now,” but the bear case is classic capex cyclicality. Memory is already eating budgets: IBM’s 20% crash was blamed on CIOs panic-buying memory instead of servers.
Pre-AI SaaS is decaying faster than expected: Constellation bought TouchBistro, a $70M-ARR Toast competitor, for $70M — 1x — and Jason says Marketo is “without question going to zero” now that “Salesforce did an LLM lift — it took one day to leave, it used to be a year.” “If you don’t want to be worth 1x, like do something before it’s too late, man.” The only public-co metric that matters: net new logos growing 15%+ a year.
Jason Calacanis quitting seed for growth is “in a niche, a huge sign of the times”: a secular new asset class of “private hundred-million checks in companies already worth a billion dollars,” plus the most liquid secondary market Jason has ever seen (Cursor: a $60bn exit in four years). Jason’s caveat stands: “when something looks easy in investing is precisely the time you shouldn’t do it, and late stage looks very easy right now.”
Carta data shows the top 5% of seed rounds hitting $200M valuations — top-decile pricing up 6x vs 10-20% for ordinary seeds — driven by neolabs that need $300M and megafunds buying ownership however they must. Meanwhile late-stage norms flip: “in the early stage, when you’re on the board, you can’t invest in two competitors. In the late stage, structurally you have to.”

Deep dive

1. Apple sues OpenAI: “no gratitude in litigation” — and the hardware project may be the casualty

The facts as the panel read them: 24-year Apple veteran Tang Tang, now running hardware at OpenAI, is named for encouraging hires to bring material over; six-year Apple employee Changlu allegedly took the secrets. Rory (likely Rory O’Driscoll) read the 41-page suit and says the fact-based evidence means the junior man is “screwed”: “there’s no gratitude in litigation” — OpenAI will play Claude Rains in Casablanca, “shocked and appalled to find that there’s gambling going on here,” and he’ll be left high and dry. “He’s already lost his job.”
Tang Tang’s exposure is one discovery away: no smoking gun yet, but multiple people plus email traffic saying “we were asked to do this.” The real game: Apple is furious that 400 employees have moved to OpenAI — obviously for the hardware device — and now “they’re pissed and they’ve got leverage. They’re going to twist the knife, depose everyone and see how high up it goes.”
The irony, per Rory: California made the theft unnecessary. No enforceable non-competes plus the “inevitable disclosure” doctrine means everything in your head is yours — “Anthropic is the biggest beneficiary”: the seven who left OpenAI “probably didn’t bring anything… everything in our brain is owned by us.” Jason’s aside: “I wonder if Anthropic could have been started in Massachusetts. Probably not.” Rory calls it “the rare policy triumph for the great state of California” — so why steal what you don’t need?
The tradeable read — hardware is “probably on the bubble”: it made sense when OpenAI had “an unassailable lead,” but with Sora chopped and “why the hell did we buy TBPN,” Jason (likely Jason Calacanis) says it’s “got to be almost on the cutting board. It’s like the Apple car” — and the suit could push it over the line. Rory, channeling Ben Thompson: consumer success may have been “a distraction,” because “LLMs are really amazing at code and the amount of economic value that can be created from that dwarfs everything else. If this was a mercy killing, Apple may have done them a favor.” Let someone else build the pin that clips to your shirt.

2. Spark 1.1: Zuck pays the clown-car toll to fight for the cheap seats

Zuck broke a three-year X silence to launch it. Jason’s imagined internal algorithm: the comms person suggests X, “he’s like, ‘F* no’… and then finally, ‘all the developers are on X… Okay, I’ll type something on the clown car company.’” Alexander Wang has done 27 posts on X to three on Threads; Threads’ claimed 400M users are “activity,” not engagement — “we know where the action is and it ain’t on Threads.”
The substance: Meta charges developers for the first time — “it wasn’t open source, it was open weight,” and now it’s an API on the same business model as every frontier lab, at pricing Wang called “very aggressive” versus OpenAI and Anthropic. Jason found it decent against Claude as an end user, but “the real test is coding,” and Meta’s past benchmarks flattered. Will the fourth player at a low price be ROI-positive over 5-10 years? “I don’t know. But if you’re OpenAI and Anthropic, you’d have preferred this not to happen.”
Jason’s frame: this is “a battle for the cheap seats.” Everyone will be token maxing (he hit his Claude limits on the max plan for the first time), every org is going from “burn all the tokens you want, baby, to budgets,” so “every company with a CIO who’s half awake is going to have a cheap token model to hand to stop this madness.” Underrated incumbent in that B-tier: Haiku at “a tenth of a cent” — Anthropic could improve it in a day “by adjusting the slider of how much Opus you get in Haiku.” How much margin lives in that bucket is TBD.
Rory’s twist: at these token prices, Meta “might actually make more money renting their capacity out” SpaceX-style than selling inference. Jason: “they’ll probably play it by ear — if this is more successful, they get more capacity; if less, they give it to their competitors.”

3. Cost per completed task, doom coding, and the missing governor

The Databricks paper’s core claim: cost per token is not a useful metric — cost per completed task is, because cheap models can carry expensive reasoning tokens in unknowable quantities. Second, a Pareto curve emerges of which model is most cost-effective per task type — a “cheap” model requiring infinite reasoning can lose to paying up. Third, the harness around the model massively changes efficiency. Rory: “I’m shocked to discover that a company that helps you manage models says managing models is important” — but they produced the goods.
Demand looks limitless: ClickHouse’s Aaron Katz posted that their AI spend is up 60x since February. Jason on AI-pilled developers: “they are coding 24 hours a day now… it’s addictive, it’s like video games… instead of doom scrolling, I’m doom coding” — the best run 10-20 agents 24/7, and “any top-tier developer can consume an order of magnitude more tokens than they are now.”
His own case study is Claude Design: “it consumes more tokens than anything else I do” — five minutes into redesigning one page, “you got to wait till 3 p.m., Jason.” His dream spec: redesign all 100 pages of a site every night and doom-click the two or three winners each morning — “orders of magnitude more tokens,” from every product person and designer, “without even blinking.”
Rory names the management problem: there has never been a product where a worker, on demand and at no cost to themselves, can have the company’s money “do their work for them and make them look amazing.” Sometimes $20 of tokens saves a $500 hour and you’re ecstatic — but “if there’s no governor on that, eventually you’re going to hit the moment where you spend 600 bucks to save 500 bucks.”

4. The TAM ceiling: coding AI is “close to 20% already” of the wage pool it replaces

Rory’s puzzle, worked from the BLS last weekend: Anthropic has exploded “to what now might be 50 billion from 9 billion at the start of the year,” and per the SemiAnalysis breakdown $2-3bn is Claude Code with a huge share of API revenue being code too. But the US has only 1.88M developers — 200k at software companies, 600k at tech companies (HP, IBM), a million at JP Morgan and BofA — at a ~$140k median wage, ~$250bn of total spend. “If a large percentage of that is software, they really are close to 20% already.” Jason: “They already hit it.”
So something has to give. “Newton’s laws of motion are true — things growing at 10x might decline to 8x or 6x or 4x, but they don’t come to a grinding stop” — and at $50-60bn of GAAP revenue, even a 6x next year exceeds total US software wages. Either this “decelerates faster than anything you’ve ever seen decelerate” (Rory: “I don’t think that’s option A”), or “a whole bunch of CFOs are literally going: we used to spend 4 million on tech, all of it in salaries, and now we spend 8 million on tech, 4 million of it on salaries. WTF.” His non-conclusion, kept honest: “the prize for becoming the fastest growing company in human history is you may hit TAM faster than any other company in human recorded history.”
Jason’s extension — three buckets. Coding is one. Two: the agentic tax — software spend is “a trillion and a half, plus or minus,” and companies may tolerate ~10% of topline going to tokens, “just like Amazon took a tax on software 10 years ago… give me 7%. Now everyone’s going to run using LLMs, give me 10%” — roughly $140bn “almost accessible now.” Rory confirms from his own agent experiments: non-coding agentic token costs run ~10% of revenue, not 50 — “Salesforce could comfortably pay 10% of its revenue for tokens; they’re not going to pay 40 because they only have 22% operating.” Three: the co-worker replacing the knowledge worker at $20, like Microsoft Office.
On whether the labs are pivoting to legal and biosciences because they see the ceiling: Rory says there’s no evidence of a slowdown, and reads the life-science push as “one-third TAM expansion and two-thirds” a search for meaning — “I’m putting everyone out of a job. The least I can do is keep them alive.” Comic relief: Jason’s price to join Anthropic is “200 — after tax it’s not worth it otherwise.” Rory: “they filter out for cynicism. You’re just not idealistic enough.”

5. SK Hynix goes to NASDAQ: 70% net margins meet 5-8 P/Es

The $26.5bn listing — the largest-ever Nasdaq listing by a foreign company — popped 13% then gave it back. Memory is a three-player oligopoly (Samsung, SK Hynix, Micron) that has “made out like bandits” off AI capex; Hynix is up 6x, and even post-listing the ADR trades at a 20% premium to the same stock in Korea, purely because Americans can’t easily open Korean brokerage accounts. “It’s just good that these guys are accessing the capital markets.”
Don’t over-read the wobble as the AI-capex bubble bursting: Samsung and Hynix are 60% of the Korean stock market and “roughly 120% of the GDP in terms of market cap,” retail-heavy, “crazy casino-level wealth” volatility — so is it a DRAM correction or just weird trading? The fundamental bet: 5-8 P/Es, “dirt cheap on any screen,” against the bear case that these are capital-cyclical businesses — prices up 6x in a year, margins from negative in ‘23 to ~70%, “not gross margins, net margins,” making “Samsung the most profitable company on the planet right now.” Rory: it lasts longer than a typical cycle, “but in the end it does correct.”
The second-order effect is already visible: IBM missed, crashed 20%, and blamed memory eating the CIO’s budget — a scramble to buy before prices rise crowded out mainframes and servers. Rory buys more than “a germ” of it: capex is finite, “you don’t want to be the last item on the purchase order list on the last day of the quarter” — and the same squeeze hits the modular SaaS vendor: “guys, we have a tech budget of 10 million; last year we spent 100 grand on AI and this year 3 million, so we need 2.9 million in savings… You’ll love Teams when you get to know it.”

6. Calacanis quits seed: secular shift, cyclical top, or craft problem?

Jason’s read on Calacanis tilting from decades of syndicated seed into growth: he expected an annex fund, not a full switch by someone with “multiple billion-dollar winners with material ownership” — “in a niche, it’s a huge sign of the times. Seeds for suckers.” His own confession: a $100M position that took years to build, then a growth investor drops $100M in one check — “why did I bother? All the drama, all the years, all the being the only guy at the board meeting… I don’t love the craft. Who the hell loves the craft?” (David Frankel does: “that man will be doing deals from the old folks’ home — and he’ll be getting good ones.”)
Jason’s three-part frame. Secular: a 5x-bigger late-stage business emerged as companies stopped going public and got big faster — “a new class of financial product: private hundred-million checks in companies already worth a billion dollars. It’s not replacing baby venture, it’s adding a whole new category on top” — nobody was writing those checks in ‘94 or even 2004. Cyclical: “when something looks easy in investing is precisely the time you shouldn’t do it, and late stage looks very easy right now” — the people who wrote cheap rounds in 2022 “are feeling pretty damn smart.” Institutional: if your comparative advantage is being “an awesome first-check-in investor, it’s not clear how you monetize that by putting money in companies at 20 billion pre” — it would be “insane” for YC to abandon early stage.
Jason’s why-now, worth keeping: “I don’t think we’ve ever seen the secondary market be as liquid and mature as it is today… in all of my top names I can get out of them today with ease.” Outcomes are bigger and faster — “your Cursor is a 60 billion exit in four years.” Harry sharpens it: this isn’t stay-private-for-longer — “they’re not public because five years ago they were doing nothing”; companies now go zero to billions in five years. And with 4,000 early-stage companies of which “only 40 matter,” late-stage money is one way to catch what you missed.

7. PG’s machine and Greylock’s “discipline”: own a business, size the fund to the strategy

The Paul Graham tweet — a YC company apologizing for “only” 36% month-on-month growth while fundraising — got plaudits Harry says he’d never get. Jason: “no one’s scared of you, Harry”; PG “might be the goat… who picks up more nine-figure and 10-figure checks off the floor than Paul Graham? Nobody.” (Jason read the tweet itself as a tongue-in-cheek humble-brag — “boasting about your kids, not yourself… people just need to get a life.”)
Jason’s better-than-goat point: PG “put himself in a position where he doesn’t have to be an amazing investor. He owns a business, not an investing thing” — the picking is fully delegated, so he can “drive around England going to bookshops in small country towns, sending fun tweets, while setting yourself up to pick $100 million checks off the floor 10 years later.” An N-of-one business — with carry premiums of “40, 50% in cases. Take it or leave it. You want access.” Harry: you just have to recruit a Sam Altman or Garry Tan every five to seven years.
Greylock 18 at $1.5bn (last fund 2.5 years ago): Harry frames it as discipline; Jason won’t have it — “It’s not disciplined. It’s optimized to make the most money. The guys at Greylock didn’t sit back and say we’d like to make half as much money.” Smaller funds deployed “thoughtfully” in 18 months get you to carry checks “before you’ve got grandchildren” — Rory: “that word thoughtfully is very load-bearing.” His rule: raise what your strategy needs — $5bn for a late-stage multi-platform (a16z, Thrive), $1-2bn for predominantly early stage (Greylock, Menlo) — and note both just passed their 50th anniversaries through three brutal cycles: “there’s no doubt that there’s a survivor bias in that discipline.”

8. “If everybody does it, is it okay?” — Phoebe Gates’s company and the gray-area tax

Phoebe Gates’s product, likely Phia, was criticized for allegedly injecting code to take attribution it probably shouldn’t have received. Jason can’t summon outrage: “the whole area of affiliate marketing is scammy on top of scammy on top of scammy” — Honey did it under PayPal — and his GoDaddy rant lands the point: still paying $9.95/month for Outlook 360 subscriptions stuffed into a domain purchase 11 years ago. “I feel like we got bigger things to be outraged about. They walked it back; we move on.”
His harder question for the industry: “Almost every agentic GTM provider is buying data sources that they probably shouldn’t be buying… you can’t be that good with data if you’re getting it the legitimate way.” If everybody does it — sketchy data, affiliate games, Polymarket — is it okay? “It’s hard to tell a founder to do no when it’s standard in the industry.”
Rory started at no and conceded the round: Uber and taxi regulation, Airbnb, Polymarket, “Anthropic and OpenAI in terms of how they trained” — “When you succeed as Uber, you just get what you did legalized” (his Elizabethan quote: treason that succeeds, no one calls treason). But his nuance survives, looping back to Apple: “there are gray areas, and then there are areas that you step over at your peril.”

9. $200M seeds, blast radius, and the new cross-investing rules

Carta’s data: the top 5% of seed rounds now price at $200M valuations — top-decile pricing up ~6x while ordinary seed pricing rose 10-20%. Rory’s explanation: two consensus-bet categories that didn’t exist five years ago — neolabs (“if it’s going to take you 300 million to get something done, there’s no point raising 20 at 20 pre”; DD’s list counts 60+ neolabs where six years ago there was only OpenAI) and capital-hungry agentic-inference infrastructure with “a wall of money” behind the spend.
Jason’s less romantic take: this is old megafund ownership math — from his one board with John Doerr, “we’ll do the seed and the A at the same time — that’s how we’ll get our 20%.” S-tier team, right space, multibillion fund: “the math just makes sense. It’s just a bet.” Not new, just normalized — “there’s just more potential outliers.” Rory concedes via biotech history: capital-intensive deals used to be tranched 20-on-20 with founders diluted to 7-10% before the big round; now supply and demand hands entrepreneurs the terms.
Rory’s “blast radius” concept: Isomorphic was 8-10x oversubscribed, so shut-out demand spilled into Chai Discovery — $400M at a $3.8bn valuation led by Index, Kleiner, Sequoia and Dimension — and from there into Latent Labs in London. Harry’s advice: often don’t announce at all — it “brings out the daggers, funds your competitors… that press release better be worth it, because it’s like a Dorito or a Pringle. It’s gone the next day.”
On Thrive backing Chai after leading Isomorphic (“Different partners, Harry… paper walls. They’re thin. We can hear the other side”), Rory’s structural rule: “In the early stage, when you’re on the board, you can’t invest in two competitors. In the late stage, structurally you have to” — later softened to “should be able to” — because late-stage privates now replicate the NASDAQ small-cap market, where cross-holding was never prohibited; Fidelity Growth would have owned both OpenAI and Anthropic. Jason adds the enabler: small late-stage checks get “no effing disclosure” anyway — “they’re going to learn literally less than a Google search.”

10. TouchBistro at 1x: venture debt, terminal decay, and the ankle biters moving up the leg

Constellation bought TouchBistro — a $70M-ARR Toast competitor — for $70M. Rory’s anatomy: a 2019 round at $600M pre, then Francisco Partners debt that converted to senior equity on a miss, leaving a cap table where the lender wants their $100M back and everyone below has “nothing to play for.” “The real dangers of adding a big slug of venture debt on top of a slow-growth business — you just end up trapped.” Jason, a convert: “I just hate debt now. I hate it, hate it, hate it… certainly don’t do debt instead of an equity round. That’s the sucker bet.”
Jason’s value of the case is its cleanliness: no egos, no 2021 markups to defend, founder gone since March 2021 — “it’s a clean look at what’s 8%, 6% growth with deteriorating market share at 70 million ARR worth. It’s worth 1x. If you don’t want to be worth 1x, like do something before it’s too late, man.”
The decay is accelerating. Jason, a first-10 Marketo customer: “It is going to zero — one of the slowest decays you can get,” no net new logos, prices just raised 20% while the API gets deprecated — and “Salesforce did an LLM lift. It took one day to leave. It used to be a year… terminal decay is accelerating faster than I would have thought when we started this podcast.” That reframes Constellation’s 3x-revenue (~12-13x cash flow, vs Salesforce below 10x forecast) purchases: “these kind of companies don’t last 10 years is what they’re effectively saying.” Jason: “I don’t think they do.” And versus Bending Spoons paying 12x for consumer assets: Bending Spoons is the only buyer in its town, while “there’s a 100 PE firms that want to buy your shitty B2B company.”
The same force threatens the healthy incumbents from below: Claude Design won’t dent Figma’s million-dollar enterprise deals this quarter, but “those ankle biters move up the leg… they’re the piranhas — then they take over the knees and the thighs.” The next generation “may never graduate to Salesforce because they never started there. They started agentically.” Jason’s single public-market filter: “are they growing net new logos 15% or more a year? If they are, they’ll figure it out” — with the Wix founder’s 2.1bn-revenue, 2.1bn-market-cap company as the cautionary backdrop.

🔬 RL with Verifiable Rewards, but the Verifier is a Lab — Lila Sciences A 4-Hour Interview with 柯丽一鸣 of Physical Intelligence on Robotics