Anthropic vs The Pentagon: Who Wins? | Cursor Hits $2BN in ARR | Block's 40% Headcount Reduction

Summary

• Anthropic broke with the Pentagon over its $200M Department of War contract, insisting its models not be used for mass surveillance or autonomous weapons while the Pentagon demanded the right to do anything “legal” — the bid-ask never converged and the Pentagon now threatens to designate Anthropic a supply-chain risk. Rory’s verdict: with safety as your organizing principle, “it was naive to try and sell to government at all,” because “the state is more powerful than Anthropic.”
• Jason’s counter: Dario had no choice — labor runs the labs. The “we’ll make weapons safer” justification could no longer be said “with a straight face,” and at a company whose messianic safety culture kept all seven founders while competitors bled theirs, breaking that unity would be fatal. “Whatever it is that’s in the water at Anthropic, it’s working.”
• The shareholder math on the rupture: “same expected return, wider variance” — you should be unhappy. Anthropic just hit #1 in the App Store ahead of ChatGPT and can move whole stock indices with a blog post, but it has traded a popularity boost for “a mild existential risk that probably you can beat in the courts.”
• Sam Altman pounced on the vacated deal despite not being approved at the DoD — and Rory’s irony: he was right on the merits (elected government, not private companies, should decide), “which of course means pleasingly, it’s the one thing that’s going to bite him in the ass,” because OpenAI’s own labor hated it.
• OpenAI’s $110bn round is 4x the largest IPO ever and likely exhausts private capital: Amazon’s $50bn is only $15bn upfront (the rest contingent on IPO or AGI) while Amazon’s own free cash flow has fallen to $11bn a year, and Nvidia already walked $100bn back to $30bn. The call: the next round is a public one — IPO in October at $1.5 trillion.
• The founder-premium test: “Tesla trades at a trillion today. I think if Elon died tomorrow, they’d trade at 200 billion. OpenAI trades at 800 billion today. I think if Sam Altman died tomorrow morning, they’d trade at 600 billion” — Bret Taylor takes over, OpenAI buys Sierra for $80bn, and the company is “possibly stronger afterwards other than the fundraising.”
• Jason rethinks the SaaS-pocalypse: it isn’t vibe coding, it’s that the path to growth has been lost — “it’s all worse than it looks” because retention hides decay, and nearly every public software company will keep guiding down all year. Block’s 40% layoff (10,000 → 6,000, the biggest percentage cut by a public tech company in 20 years) is an abandonment of growth dressed in AI language, and it becomes the default: “Every single CEO I talk to doesn’t think they need 40% of their team.”
• Cursor doubled from $1bn to $2bn in revenue in ~90 days while VC Twitter declared it dead for Claude Code — the portfolio bias trap. 60% of revenue is enterprise (Barclays just made it the first approved agentic coding product), but 60-70% of model calls still flow “back through to mom and dad” at likely near-rack rates to Anthropic, and “the knife fight doesn’t start until the TAM is like 60-70% saturated.”

Deep dive

1. Anthropic picks a fight with the Pentagon — and discovers state power

• The facts, per Rory, with his caveat up front: “Unless you’re in the room, you should assume that everyone who was in the room is spinning a version of this that makes them look good.” Anthropic’s $200M Department of War contract ruptured over two conditions Anthropic wanted — no mass surveillance, no autonomous weapons — against the Pentagon’s counter that it be allowed anything “legal.” On Friday the Pentagon broke off talks and threatened, at minimum, cancellation — at maximum a supply-chain-risk measure barring other government vendors from using Anthropic, “either a mild loss of a contract or something approaching thermonuclear.”
• Rory’s verdict: “it was naive to try and sell to government at all.” The Department of Defense has its own organizing principle — defending the United States — and “the idea that someone’s going to waltz in and say, ‘I want a $200 million contract, but I’m going to tell you the safe way to use this’ is absurd… ‘Go away, little boy.’”
• His supporting analogy, from a book he and Dario both love (Richard Rhodes’ The Making of the Atomic Bomb): General Groves humored the scientists, “but when it come to the crunch, they weren’t even in the room when they decided to use the bomb.” And Tony Benn’s line to civil servants applies: “Who voted for you, guys at Anthropic? Cuz 80 million people voted for us.”
• The bigger lesson in political economy — beyond capital and labor there is a third actor, the state, and “the state is more powerful than Anthropic.” “AI might be scary in the theoretical… The state is scary in the real sense of we have laws, we have men with guns to enforce those laws, and we can take your company if we want to.” Harry’s advice: tiptoe out — “if all it costs you is a $200 million contract in the context of a $14 billion business, you should declare a win and move on.”

2. Dario had no choice — labor runs the labs

• Jason’s defense of Dario: he’d sold the team on “at least we can make weapons safer… at least we’ll minimize the collateral damage” — and events pushed past the point where “I don’t think he could say it with a straight face.” The real constraint: “Not only do I not believe in it, but I’ll lose my team.”
• Harry concedes the premise even while calling the outcome wrong: Anthropic’s “almost messianic belief” in AI’s danger is “not just a likely held belief… it’s a core organizing principle and arguably even the engine that has driven the enthusiasm to build that company” — which is why all seven founders remain while “the other guys across the street have lost almost all but one or two.” “Whatever it is that’s in the water at Anthropic, it’s working and it has created unity.”
• The labor-capital frame: AI labs sit at the extreme of labor power — “people are leaving eight figures of vesting stock at OpenAI… after 13 months” — while non-AI workers at companies like Block have “the least power versus management ever. You’re less than fungible.” His illustrative anecdote: at OpenAI, salespeople’s badges don’t work at the researchers’ building and they can’t attend meetings there. “That’s protecting your talent at all costs… I don’t think the team would tolerate anything else.”

3. The shareholder math: same expected return, wider variance

• Asked whether an Anthropic shareholder should feel better or worse after the lift — #1 in the App Store ahead of ChatGPT for the first time, a real consumer brand — the answer is crisp: “Same expected return, wider variance. By definition, the risk increased and therefore you should be unhappy.” A week ago “you write a blog post and every legal tech company goes down 20%. You had a good thing going.” Now you’re “at the mercy of how the administration wants to handle this.”
• A cynical-realist investor’s rejoinder: “no one really cares as long as the numbers are there.” If an Anthropic investor’s March update said the company just crossed $16 billion, “my email back would probably be ‘great job’ with three exclamation points and maybe an emoji or two” — investors have almost no rights anyway, and only at struggling companies does their opinion matter.
• The structural point on how the deal happened: it was signed last summer when Anthropic was around a $1 billion run rate — $200M direct plus $200M pulled through Palantir, plus blocking a competitor, was worth swallowing principles for, the way most investors would take a legal-but-distasteful contract that turns 60% growth into 110%. “It was a bigger deal at the time than it is today.” Rory adds nobody planned this: they were selling to Palantir, got dragged toward government, “and a bunch of people drifted into a situation that nobody fully thought through” — plus, per the Wall Street Journal, personality clashes.

4. Sam’s pounce: right on the merits — which is exactly why it bites him

• The reaction savors it: “It’s the most entertaining of all possible outcomes.” On the substance — that in a democracy the government, not individual companies, should be the final arbiter — “I actually think he was right on the merits, which of course means pleasingly, it’s the one thing that’s going to bite him in the ass.”
• Because OpenAI’s labor rebelled too: “His team hated that he did that.” Sam’s response — saying he’d unilaterally change ChatGPT’s contract clauses the next day — draws Jason’s scorn: “That was a numbnuts thing to say… oh, we’re just going to change the clauses ourselves… good god.” Rory’s irony deepens: the first display of labor power at OpenAI was 2023, when employees forced Sam’s reinstatement — “labor basically said we want to pick our boss.”
• The competitive backstory Jason supplies: OpenAI wasn’t even approved at the DoD for this application — only xAI was, and they’d lost last summer when Palantir picked Anthropic. Sam did “what most of us in B2B would do. He pounced on the deal” the moment the competitor stumbled — though Rory insists it wasn’t only opportunism: “it actually was roughly trying to do the right thing.” Jason’s hedge on where it all leads: capabilities have accelerated so much since December that “it’s just very hard to predict for a military application what these models will be able to do in 12 months.”

5. The $110bn round exhausts private capital — next stop, public markets

• The round is 4x the size of the largest IPO ever, and Rory’s tape math is staggering: $30bn for Anthropic plus $110bn here is $140bn — “more than US venture for all of last year.” Jason’s dismissal of the perennial hand-wringing: “we’re all worried some B2B company can’t IPO at 400 million growing 40%… It’s so minuscule” next to this.
• The double asterisk: of Amazon’s $50bn, only $15bn is upfront — the rest is contingent on IPO or AGI (“if you needed to pay your workers next week with that 35 billion, it wouldn’t be there”). Amazon’s own free cash flow has fallen to $11bn a year; Nvidia already walked its $100bn pledge back to $30bn. Jason: “the folks that did this round literally cannot do any more.”
• Hence the panel’s call: “Articulate who writes the next check privately and what the investment thesis is. I don’t think there’s more capital there… You’ve got to believe that the next round for this, Anthropic, and SpaceX are all public offerings.” The prediction: OpenAI goes out in October at $1.5 trillion — with the pre-committed $35bn effectively “half your book sold” walking in.
• The valuation caution stands: the S&P is within 2% of all-time highs at a 20-year-high trailing P/E, and these prices demand enormous “leaning into the future.” OpenAI at 40-50x revenues works only if hypergrowth (roughly $3-4bn GAAP revenue last year to $12bn, ~$20bn ARR run rate) continues for two or three more years. SpaceX is harder still: $18bn growing 20% at ~100x trailing, requiring Starship, Starlink 2.0 direct-to-cellular, and data-centers-in-space to all land. “If that was a SaaS company doing 18 billion growing at 20%… Jason would be saying five times.”

6. The founder-premium test: remove the founder, reprice the company

• The method for comparing the Elon and Sam premiums: imagine the founder gone. “Tesla trades at a trillion today. I think if Elon died tomorrow, they’d trade at 200 billion. OpenAI trades at 800 billion today. I think if Sam Altman died tomorrow morning, they’d trade at 600 billion.” Without Elon there’s no robot and no robotaxi — just “a declining car product line” — while OpenAI would simply install Bret Taylor and build the company.
• The argument goes further: if Sam left, “you buy Sierra for 80 billion, 10% of the value… your problem is solved… possibly the company’s stronger afterwards other than the fundraising” — swapping “a very brilliant but ultimately non-technical founder” for one of the great technical CEOs, already on the board. The kicker on relative mispricing: Tesla trades at 10-13x revenue for declining revenue — “the Elon premium is way higher relative to the performance of the underlying asset.”

7. SaaS-pocalypse, rethought: growth is gone — permanently

• Jason’s re-diagnosis: “I think we got the SaaS-pocalypse all wrong… It’s not vibe coding that’s killing us” — vibe coding is “one of those minor threats” — it’s that everyone has lost the path to growth, and because revenue retention makes reported numbers backwards-looking, “almost everyone is worse than it looks.” His call: nearly all public software companies “are just going to do worse and worse for the rest of this year.” The best performer he tracks is DigitalOcean, up 28%.
• Harry’s history of the multiple, worth keeping whole: 15-16 years of the bucket growing ~30% at ~6x revenues; two COVID years at 40% growth and 20x; then deceleration to 15% while multiples stupidly stayed at 6x — the chart “looked like it had gone back to normal, but the growth had gone out the bottom.” Then in January the market’s eyes opened: “It’s not a temporary growth decline… it’s now permanent because of AI. Oh my god, what were we thinking? You should be at four times. That’s the movie.”
• Jason’s own-goal charge: unprecedented spend flowed to cloud and LLMs and public SaaS captured none of it — “it is the biggest fail and own goal in our history.” “These three kids from Stanford figured out how to use Opus, why couldn’t you?… Why can’t you put your 20 best engineers and just freaking clone Harvey or ElevenLabs?… You all deserve to be fired.” The tell is audible on earnings calls: Benioff “believes” (“I was there in the building, the dude believes”), Toby is “dragging his company mercilessly into the agentic era” — everyone else “you can smell the lack of belief… out of the pores,” already falling toward “the event horizon, into a state of terminal decline.”
• Harry holds the boring middle: the straw man that CEOs are ignoring AI “isn’t really true,” Owen at Intercom showed the bloody-mindedness required to gut the business and build Fin (“a year of feeling ridiculous”), and price matters — a stock at 40x next-year EBITDA drops 25% on a modest guidance miss, while names at 7-9x EBITDA growing 8-12% arguably have “more upside than downside.” Most scaled companies “should not be banished” — though they’ll never see 20-30x EBITDA again, “still less, God forgive us, 20x revenue multiples.”

8. Block’s 40%: not an AI story — an abandonment of growth

• Jason called a mega-layoff last week; Block delivered: 40% of headcount, 10,000 → 6,000 — per Rory’s digging, “the biggest percentage change for a publicly held tech company in the last 20 years.” Jason’s read of the press release: everyone got suckered by gross profit per share up 27% — “you only lead with that when you’re not growing.” Top line is growing ~3%; Jack has “completely given up on returning to growth,” so the only button left is profitability — a choice he expects most public software CEOs to copy this year, some throwing in the towel even at 12-13% growth.
• Harry’s decomposition — keep this distinction: there are two uses of “AI,” top line and opex. “This is not an AI top-line story” — Salesforce is (AgentForce either grows or shrinks its revenue); Cash App and Square don’t fundamentally change with AI. Block is just saying “maybe we can use AI to cut some opex” — “I doubt it’s true at the 40% level.” Headcount had ballooned 50% a year, and against best-in-class financial services comparables “there was clearly fat to cut.”
• The contagion mechanism: Harry tweeted and three CEOs of 500-1,000-person companies all said they’re cutting a minimum of 20% — the Overton window expands. Jason: “Every single CEO I talk to doesn’t think they need 40% of their team… it’s a gift.” Rory rejects “acceptable” for “necessity”: “price clears all markets” — grow sub-10% with 10% free cash flow and you’ll trade at 3-5x until “someone will come in and buy you and make those changes, and then a month later they’ll be worth eight times.”
• Jason’s longer arc, with Harry’s example: Block is “a window to the past… like looking out to the galaxies,” because future companies will be permanently smaller to start. Harry’s example is a hot AI company scaling sales from 10 to 250 in a year; Jason estimates how many will be great: “20. Maybe a few more.” Harry’s data agrees directionally: the trend was pre-AI (“Apple has 160,000… Google 190,000” vs. GM’s 3-400,000 in the 1950s), and Cursor’s revenue per employee already matches Apple’s ~$2.6M (Google ~$2.1M). Jason’s coda: we may look back on “20 years of brute-forcing revenue growth with human-led sales and marketing as an anomaly in the era of software.”

9. Cursor at $2bn revenue: reports of its death greatly exaggerated

• The puzzle Harry poses: everyone he talks to has moved from Cursor to Claw Code — his tweet drew thousands of confirming comments — yet Cursor just went from $1bn to $2bn in revenue in three months, with rumors of a secondary round at $50bn. Jason’s mea culpa on portfolio bias: at his fastest-growing companies the joke was “it’s only grandpas at the company that still use it” — “our portfolios are not representative… it’s so funny we got it wrong. Numbers don’t lie.”
• The likely answer: 60% of revenue is enterprise. Barclays just made Cursor the first approved agentic coding product — banks have the biggest budgets and are “super aggressive and super conservative” at once — and Cursor ships what CISOs need: no data retention, full SSO, role-based access controls, audit logs. Jason’s sharpened point on the swarm era: “every agent will leak data if it’s allowed to. That’s how goal-seeking works” — if Cursor can unleash autonomous agent swarms and make them safer, “95% of CISOs want that one.” Enterprise adoption cycles also lock in share: “corporate’s got to approve, purchase orders got to be raised… we’ll do an eval mid-year.”
• Rory’s framing of the market structure — the episode’s most quotable line: “never underestimate big markets and momentum… The knife fight doesn’t start until the TAM is like 60-70% saturated, which by the way happened to SaaS in 2021.” Until then, “we all have a win rate of 60%. We don’t compete against each other.” Both players are exploding — Anthropic says it has added $2bn ARR year to date too — in “a 50 to 100 billion-dollar minimum coding prize.”
• The caveats: it’s the Fortnite effect — “Claude Code will keep doing more and Cursor has to keep ahead… that Fortnite storm’s going to shrink to a pixel” for anyone who doesn’t. And the margin question: 60-70% of Cursor’s model calls still run on Anthropic’s API — “a lot is flowing back through to mom and dad” (a point Michael Cannon-Brookes made) — likely near rack rates, since Cursor lets you bring your own key. The rule for the era: “The prize for winning is to reinvent the company from scratch and the product from scratch every 6 to 9 months. Congratulations, it’s a fun game.” Autocomplete → IDE → agents → swarms.

10. Everything accelerated in December — and demos now carry no signal

• Jason’s case that change will “radically accelerate” this year, against Harry’s skepticism: on Replit and Lovable, apps used to be fake buttons; by December, “especially with Opus 45,” they could finish — his colleague Amelia built an AI VP of marketing running production marketing daily, built by a swarm (architect, security expert, back-end, database). “It’s like Hal. It wakes up every morning and tells us what to do.” “When that 10x’s this year… the rate at which your product has to accelerate will be like nothing we’ve ever seen before.”
• The venture consequence: “There’s no information in the demo anymore” — a great demo just means you kicked off a coding project 24-48 hours ago. Harry’s translation of the viral “software is uninvestable” line (from “what’s-his-face from AngelList,” as Jason puts it): software itself is obviously investable — the real version is “there’s going to be infinitely more software… software alone will not be a competitive advantage,” so moats shift to network, distribution, and vertical knowledge.
• One investor’s lived version: two recent investments did $1M+ in their first week — one of them 45-50 days after he’d told the founders “I put it back in the oven… I don’t think it’s competitive.” “That level of pace we just didn’t see before… how the hell do you pick favorites?” The old credential-plus-demo signal — the David Franco-style “only pre-seed investor in Sonos (likely Suno), 300 million in 2 years” pick — still worked in 2020; the demo half of it is now dead. Rory promises his public-market picks next week: his benchmark, the WisdomTree SaaS index, was up ~6% at one point — “I haven’t looked at it today and it’s been pretty grim.”