Anthropic Files to Go Public | Cognition Raises $1BN at $26BN Valuation | The 996 Work Ethic

Summary

• Anthropic filing to IPO the same week it raised $65B resets the bar for the whole ecosystem — “five years to a trillion,” with Cursor’s $60B exit in four years alongside it. Jason Lemkin’s response is a new hurdle: “I’m not interested if it can’t be a billion dollar position anymore,” while Rory O’Driscoll counters with base rates — realistically “not going to be more than four or five $10 billion plus outcomes a year” — and the adult’s move after missing the one-in-a-decade deal: mourn it, then “go on and do perfectly good deals.”
• The stay-private era is officially dead: “We are done with the ooh I don’t want to do the public markets… we are f*ing done with that.” Google announced an $80B equity raise, SpaceX formalized $1.75T for early June, Anthropic filed while OpenAI may have filed — call it $300-400B of AI-related equity issuance as everyone rushes the door. The structural read: these have gone “from capex-light cash flow machines to capex-heavy cash consumptive machines,” historically bad for medium-term stock returns.
• The SaaS apocalypse panic is over but the disease isn’t: WCLD round-tripped -30% to flat on the year while semis “murdered SaaS.” The clean learning — every leader that agents consume more of is up (Twilio +57%, Okta +56%, Datadog +100%) while “the classic human per-seat software really is dying… no one wants to buy this crap. In fact they’re cutting it because they got to come up with money for all these goddamn tokens.”
• Token budgeting panic hit corporate America simultaneously: everyone cranked on Claude in Q1, CFOs found the accrual when bills came in, and Uber capped spend at $1,500/month. Rory’s zoom-out is bullish for model providers — nobody said stop, so “you’ve just established a category that probably has a market size of half a trillion to a trillion dollars.”
• Jason’s boldest prediction: “by the end of the year we’re going to choose tokens over humans” — he predicts QA and CS departments will be gutted as budget-holders trade heads for tokens, with December emails reading “It’s not you, but we needed the tokens.” The discussion centers on the single most important number in every AI model: token spend as a % of engineering comp — Uber implies ~10%, Jason’s thought experiment ~33%, EDA (the most automated precedent) runs 13%. If it’s 33%, “buy at any price in the IPO.”
• Cognition’s $1B at $26B with Devin at $492M ARR validates the autonomous-engineer thesis Jason finds “more compelling than all this Claude Code crap” — but developers won’t tolerate model downgrades: “I would quit as a developer if you told me I could not use the model of my choice. I would quit.”
• Apollo’s warning on PE software returns is “probably correct”: if private credit — senior in the stack — is struggling, “the equity is dead.” Buy a mature SaaS asset at 10x that’s now marked 5-6x and “there’s no accelerant… you just own a mature SaaS company” — best case a decade of bolt-ons grinding out 1.2-1.3x, with Harvard now 41% in privates.
• 996 is neither new nor toxic — but it demands a quid pro quo: early employees need “a shot at eight figures,” and Harry nails the valley’s core contradiction — “we’re all here with a plan to automate white collar work… yet you talk to every single person in this valley and they’re like, I’ve never worked this hard.”

Deep dive

1. Anthropic files to go public — the trillion-in-five-years reset

• Jason’s setup, week 58 of “this week in Anthropic”: ARR up 28% since the last show, “the fastest growing enterprise software startup of all time,” and now on track to be the fastest to IPO at anything near its scale — “five years to a trillion,” with Cursor acquired for $60B in four years (assuming the deal closes). His provocation: “Why would you rationally do anything else? Why would you even try to have a $400 million exit, $2 billion exit? Isn’t it just a waste of our time?” The bar has reset not just on valuations but on time.
• Rory’s pushback — worth keeping in full: after missing the best startup of the decade you have three options — fool yourself that another arrives next year (“foolish… it is by definition one in ten”), quit psychologically damaged, or “grow up and be a f*ing adult and say, I wish I’d done that deal, I’d give my left arm… but I didn’t. Now I got to go on and do perfectly good deals.” Business is psychologically healthier than politics: the 700th most successful politician isn’t a backbench MP; the 700th businessperson is worth ~$1B — “an okay consolation prize.”
• Rory’s confession: he lies awake reconstructing his calendar from early Feb/March of the relevant raise — “I can tell you what I’m doing every day… and I’ll tell you what I wasn’t doing. Meeting with Anthropic.” His consolation: going public strips the mystery, as with SpaceX’s S-1 — “there wasn’t any magic pixie dust.” Harry’s precision on which round to mourn: the Series B was the Sam Bankman-Fried round, “a poison chalice on every dimension”; the Series C, which Spark led and Menlo did, is the one you kick yourself over.

2. Jason’s new bar: billion-dollar position or pass

• The changed mind, stated flatly: “I’m not interested if it can’t be a billion dollar position anymore” — which given dilution means the company must be worth north of $10B. His method isn’t forecasting, it’s black-marker blockers: the “pretty good but not jaw-dropping” CTO, complainers (“I just haven’t seen a lot of great outcomes from complainers”), small TAM with no drive to grow it. You can still luck into a billion-dollar outcome with those flaws — but not a billion-dollar position.
• Harry’s challenge: a thousand GP interviews say the biggest winners were underestimated at entry — nobody called Twilio a $2B company at seed. Rory’s base-rate math: his old mental model was ~10-20 $1B+ outcomes a year, two to three $10B+, one to three $100B+ per decade; scale up 10-20% for AI and “there’s not going to be more than four or five $10 billion plus outcomes a year.” His resolution: underwrite a realistic base case but “never do a deal with just capped return” — you need a “credible upside story,” which is how he lands where Jason does despite disagreeing on the math. The key word difference: can be, not will be.
• Rory’s scar tissue on the risk appetite everyone suddenly has: he started investing in 2001 and watched the NASDAQ fall 90%. “Losing money is like sex. You can talk about it all you like, but until you feel it, you don’t know what it’s like.”

3. The founder aspiration bar just went up too

• Jason’s practical point: “There’s meetings I won’t take now that I would have taken in 2024 or 2023… it’s not because they’re not great human beings building real companies. The bar has gone up so much. I’m not sure all founders get this.”
• Rory agrees on the anchoring: founders now ask where to sit on “the grandiosity versus boring stakes,” and the level below which you’re perceived as boring “probably has increased significantly because people have seen what quantifiably amazing looks like — 10x growth for three years.” Jason’s darker read: the Anthropic IPO isn’t all net positive — it will reinforce “a general sense of not being good enough” across the ecosystem.

4. Everyone is rushing the public markets at once

• Rory’s framing of the week: “We are done with the ooh I don’t want to do the public markets. Stay private is cool. We are f*ing done with that.” Elon had 20 years private, now it’s go-go-go at $1.7T; Anthropic filed while OpenAI may have filed after going from “maybe 2027” to filing now; Google announced an $80B equity raise the same day. Across four names that’s $300-400B of equity issuance, all effectively AI-related — “like one of those airline flights in countries where they just don’t queue… a mad rush to get on the plane.”
• Why Google issues equity when it could borrow: stock’s at an all-time premium, “equity is cheap,” and it insulates against the debt market’s “micro panics.” Jason’s answer is that they can do both; Harry notes that payback on AI datacenters is normally 2-3 years (Elon’s storage deal with Anthropic, paying back in “a year and a bit,” is the outlier), so a fortress balance sheet makes sense if you might spend $300B a year before it comes back.
• The through-line, and it’s not bullish: “All these businesses have gone from capex-light cash flow machines to capex-heavy cash consumptive machines… across history, things that eat money tend to be bad investments.”

5. SaaS apocalypse: the panic is over, the problem isn’t

• Rory’s round-trip: WCLD (likely — the cloud ETF he calls “worldcloud”), which he bought when it was down 30% a month ago, is up 25-30% since — “but we’re just back flat on the year.” Fun fact: SaaS outperformed semis over the last month, but on the year “semis have murdered SaaS to the upside.” The lesson: “the narrative got way overdone, and then people looked up and said, these things aren’t going to zero. If they’re not going to zero, they have cash flow value.”
• Jason’s numbers: his cloud basket is up 5% vs NASDAQ +21%, semis near triple digits. The fundamentals haven’t changed — seat contraction, and Gartner’s AI software spend +60% this year “has to be cut somewhere else.” Multiples “all bounced off the hard deck and went back to crappy from worthless”: Atlassian 4x ARR, HubSpot 3.8x, Salesforce 4x. “I just don’t think investing really works without 10x or higher outcomes.”
• Both agree on the harder second act: a month ago you could buy the whole discounted sector; now you have to name which stocks have a genuine reacceleration or AI-attach story — “that’s a harder message.”

6. Agent-attach wins; human per-seat software dies

• The “captain obvious learning” of the year, per Jason: Jeff Lawson said on this show that agents would need more Twilio voice — took a quarter or two, and the stock is +57% with growth from 4-5% to 20%. Okta +56%, Datadog +100%. “Every software leader where agentic products and agents need more of it is up… the ones that only humans use is kind of down.”
• Salesforce split its reporting into Agentforce and the rest for the first time, guiding the classic software business to “perpetually single-digit growth” with the rest at 12-13%. Jason’s obituary: “the classic human per-seat software really is dying… no one wants to buy this crap. In fact they’re cutting it because they got to come up with money for all these goddamn tokens.”
• Rory’s single test for the survivors: “Are you growing quicker? If you’re growing quicker, you’ve gotten lift” — Replit as the pre-AI company that brilliantly attached, and the Postgres database he cites (likely Supabase) “brilliantly co-attached to everything.” It requires deft product management; otherwise you’re not dying quickly, just facing a slow question of how to create and realize value.

7. Cognition at $26B: the autonomous engineer, not the copilot

• Cognition raised $1B at $26B with Devin at $492M ARR. Jason’s history: the two highest-IQ CTOs in his portfolio ran Devin early — one drove it entirely from Slack, letting it make commits autonomously “when things were still pretty crappy.” His thesis: “The vision to me is actually more compelling than all this Claude Code crap… Why do we want to empower mediocre engineers? Let’s get rid of them, man. Devin doesn’t argue. Devin doesn’t only want to work on interesting problems like most of your best engineers.”
• His own hedge: “maybe it’s not that impressive because there’s just so much money in the space” — possibly just repackaging models for one workflow. And Rory’s market caution: “this is very much a market where leads change hands at a furious pace” — every trillion-dollar company wants to eat their lunch. Harry’s tradeable footnote: if Cognition is $26B here, was Cursor at $3B ARR cheap at $60B?

8. The token panic: CFOs found the bill, and it validated the category

• Rory’s reconstruction of why the ROI panic hit everywhere at once: Claude Code’s “magical version” shipped around November/December, pricing moved to pay-as-you-go, “literally everyone cranked in Q1,” and the accrual accounting caught up when bills came in — “we told these guys to crank in Q1 and f* me, they cranked… it looks like we spent our entire budget.” Uber’s response: cap everyone at ~$1,500/month, slightly above observed averages.
• The pricing-curve consolation: frontier models are getting slightly more expensive, but today’s frontier will be 5-10x cheaper in a year once it’s off-frontier — so adoption “might slow markedly” but the value stays accessible. Rory’s zoom-out is the bull case: 3-5% of tech spend got noticed and nobody said stop — “you’ve just established a category that probably has a market size of half a trillion to a trillion dollars. Corporate America has said… we’re spending too much, but we’re going to have to have a plan to spend it. It’s enormously validating.”
• Jason’s tell: two of his fastest-growing portfolio companies blew through their token budgets; none of his slowest did. His advice on where to see the future of cost engineering: watch Replit and Lovable, under massive customer cost pressure — Replit builds complex features in Sonnet to save money and brings in Codex as the architect agent to check the work: “every single time it finds issues. It’s so powerful.”
• But he draws a hard line on developer choice: apps will optimize across models (OpenRouter blowing up, dollar-per-chat economics), but developers won’t be downgraded — “I would quit as a developer if you told me I could not use the model of my choice. I would quit. It’s not worth my time.” As an ecosystem, “we are all in on Opus” — paying into the inflationary side, not harvesting the deflation.

9. Tokens over humans: the year-end budget choice

• Jason’s central prediction: “I really do think by the end of the year we’re going to choose tokens over humans for engineering and product.” At Adobe his EPD budget was pure headcount at a flat $300K/head; 2027 budgets become dollar pools where leaders ask, “do I want another 20 mediocre engineers or do I want to give my best guys unlimited tokens?” QA departments “will get destroyed,” CS gutted to the head of CS — “the ones on the bubble that weren’t already cut in the first wave will be cut for tokens.” The December emails: “It’s not you, but we needed the tokens.”
• The discussion forces quantification, and Jason calls it the most important number in every AI model: his 400→300-heads thought experiment implies 33% of engineer comp in tokens ($66K on a $200K engineer); Uber’s cap implies ~10%; EDA software — the most automated engineering precedent — runs ~13%. “This is the number that will determine: is $1 trillion a fully priced company that could slow down for a year, or is it going to eat one-third of engineering salaries and get you to 4 trillion two years from now.” If it’s 33%: “buy at any price in the IPO.”
• The live data point from Harry’s own show: Brendan of Mercor (likely — captions garble the name) says they now spend more on tokens than engineering salaries. Harry’s caveat — that’s ~80 engineers, not 1,200 — but also the concession: “if the new companies starting with a clean slate really can do 50% tokens, 50% people, and they are able to ship, then that is the future… and we are underestimating the size of these markets even now.”
• Where Harry disagrees: outside engineering. For customer support, “the cost of tokens is so low relative to the total cost, it’s in the noise” — the interesting fight is engineering, and there’ll be “VPs of engineering on both sides of that trade who lose their job.” Jason’s counter on capping: the Uber-style limit is transitory — “for 2027, 2028, you should give department leaders a choice, and they’re going to choose tokens in good companies.”

10. The weakest link caps the upside — and orgs will rebloat

• Jason’s countertrend to his own layoff thesis: as agents let you ship far more products (not features), you still need humans to manage them — one portfolio company crossing $100M will end the year with 3x the products, and even at half the humans per product, “everyone will be a Rippling with 22 products and you got to have 22 PMs.” Startups “will achieve the same historic level of bloat, maybe half the size.”
• Harry’s macro version, via an academic paper cited that day: when one part of the org speeds up, “it doesn’t matter if you can make a gazillion pieces of software if you can’t package it, price it, sell it, train it.” “The weakest link in the chain is what determines the speed of the convoy.” His standing bet against the 10%-GDP crowd: 200 years of ~2% real growth — “it’ll stay at 2%.”

11. Kirkland builds its own Harvey — the crown-jewels question

• Jason deflates the headline: Kirkland & Ellis committing $100M a year for 5 years is ~$100M/year against $11B of revenue growing 20% — under 1%, “coming out of their Windows NT box or some other crappy budget.” It doesn’t preclude buying Harvey or Legora too, and the pressure is healthy: “Everyone should spool up and try to build their own CRM and see if it’s worth it… It’s great for everybody that we’re under AI pressure.”
• Rory’s sharper reads: first, K&E “already won because they said it first” — AI-forward publicity without having done anything, classic hard-headed transaction-business move. Second, the real buy-vs-build logic: you buy horizontal commodity (case management, Westlaw — “everyone has the same s*, that’s not how they compete”) but pause if AI encapsulates your secret sauce — “if we pay Harvey 10 million, but in return they know the K&E way… hmm, maybe not,” especially with chatter about vendors going full-stack law firm, which “nothing could piss your clients off more” than.
• And the escalation beyond Harvey: “if you thought Harvey and Legora were fast and loose with your IP, Mr. K&E, wait till you see what Claude does with your IP” — relevant given Harry’s note that Ironclad’s Jason (likely Boehmig) just joined OpenAI ahead of foundation-model legal entries expected within 2-8 weeks. Rory’s historical rhyme: mid-90s everyone said Microsoft would be a bank; lines blur, then it becomes obvious what goes where.

12. AI legal grows the bottom of the market; humans keep the top

• Jason’s market call: AI is expansionary at the individual and SMB level — cheap divorce, cheap wills, decent legal advice for $100 when a lawyer starts at $2K. But full-stack AI firms won’t displace K&E: “when you’re doing a $20 billion transaction, at some level you want a human to hold your hand and tell you these are the last ten of these I did.” Rory: $10,000/hour “is nothing” on a hundred-billion-dollar deal — commodity work goes to Claude, mission-critical judgment keeps its premium.
• The anecdote that carries it, from Rory circa 2018-19: a smart Stanford-trained big-law associate stress-tested an NLP tool over a weekend, reported it was 98% accurate — and said “I wouldn’t touch it with a 10-foot pole. My boss will sack me if I’m not 100% accurate.”
• The joke that nearly cost a sponsor: Rory doubts “the Anthropic safety committee will allow them to build a model quite as mean as your average K&E bankruptcy attorney — it will literally fail the safety test.” Jason’s rejoinder: K&E paid every partner an $11M bonus this year, “and they didn’t do that by being pussies.”

13. Robinhood’s AI agents: planning is solvable, trading isn’t

• Jason’s enthusiasm comes from pain: wealth management humans are “the lowest quality of any professional I’ve ever worked with — they all put you in the same crappy models and the same 11 proprietary products.” An agent that ingests your full picture — risk tolerance, house purchase in 3 years, carry coming — and gives the right answer means “so many folks will not get ripped off… Fidelity doesn’t do it. Vanguard doesn’t do it.”
• Rory’s crucial split: correct financial advice and asset allocation are “pretty well understood,” knowable, and should be automated — LLMs have a meaningful role (his firm invested in Range for exactly this). But trading to outperform is different: “that task cannot be accomplished by that agent… if there was an edge, Jane Street would be doing it” quietly. And the demographic mismatch: “I watch my son trade his Robinhood account. I don’t think he’s focused on where he’ll be at 65.”
• Jason’s generalization, via Klaviyo’s Andrew Bialecki (likely) at SaaStr Annual: the most important agents are the ones that make every customer a true expert in your domain — he cites YouTube’s agent, “better than any human could be” at explaining video performance. “All applications should make you an expert… you log in and the first day I’m a f*ing expert.”

14. Apollo’s PE warning, Anthropic distributions, and who keeps playing

• Rory on Apollo: “talking their non-book, but they’re probably correct.” Private credit is the senior lender at ~5x EBITDA leverage; if the debt’s worried, “the equity is dead” — the SaaS assets bounced 30% but still trade 3-6x, and if you bought at 10x, “it’s kind of like overpaying for a real estate transaction. There’s no accelerant.” His worked example: buy Salesforce at 14x revenue in 2021, half debt half equity — the public market now says 5-6x. Best case: own it 10 years, bolt-ons, “grind out a miserable 1.2, 1.3x.” Context: Harvard is now 41% privates.
• Jason’s cynical-realist take: Anthropic distributions will let LPs “give you a pass on all these” bad funds. Rory’s rebuttal on incentives: LPs can move on, but PE managers shouldn’t get to — the difference between giving up at 0.5x and grinding to 1.5x is skin in the game, which is why “capital commitment really matters.”
• On what $10B carry pools (Menlo, Spark, Founders Fund on SpaceX) do to firms: Jason shrugs — OpenView and most of Emergence rationally retired rich, “I don’t think every VC firm has to last until the 23rd century.” His own math: he only entered venture demanding 10x his founder outcome — “if I made a couple billion in carry and my next fund might make 20 million, I would quit… give the young kids the keys.” Jason’s fix is Thiel’s: be a third of your own fund as LP — “as always when Peter Thiel does something, assume it’s the entirely rational cold-blooded correct solution.” And yes, Rory would show up tomorrow: “large amounts of money and large amounts of free time tends to be pretty destructive.”

15. 996: performative theatre, or just the old deal restated

• Jason’s demystification: he worked six and a half days a week in services before tech and rolled into his first startup on a Saturday at 9am; the only real question is “how deep does working Saturday and Sunday go” in the org. He liked how Cognition’s Scott handled the Windsurf cuts — “we work seven days a week,” said without being douchey — because at $26B those first 50-100 people make $40-50M. The quid pro quo is the whole point: “the $150 million exits don’t justify that… just do it, man. Pay up. Give them four times the equity, and if you don’t like it, go work somewhere else.”
• Harry’s agreement with a warning label: it’s not new — the Apple/China history “talks heart attacks,” big law bills 2,100-2,200 hours, and “sometimes to do really hard things you need small numbers of people to concentrate 24/7 and will it.” But intensity erodes judgment: “instead of rage baiting, rage working — you’re just performatively working and not achieving… make sure your psychological health and judgment is good.” Jason: implicit promise required — “you better give them a shot at eight figures.”
• Harry’s closing contradiction, the best line of the episode: “we’re all here in the valley with a plan to automate white collar work such that there’s going to be mass unemployment in three years… yet you talk to every single person in this valley and they’re like, I’ve never worked this hard” — and Jason’s number one problem is hiring. On the fate of the B players they disagree: Jason — “we absorbed so many B’s in tech… I’m not convinced” there are jobs for them; Rory, benign — they’ll get jobs, “maybe you won’t get another job that pays 400 grand and allows you to work from home three days a week. Life will go on.”