OpenAI & SpaceX S1 Drops | Layoffs at Cloudflare & ClickUp | OpenRouter & Polsia Raise Mega Rounds

Summary

• OpenAI confidentially files its S1 ($852B–$1T, Q4 listing) because it’s been lapped. Q1 GAAP revenue: Anthropic $5B — more than its entire prior year of $4.5B — versus OpenAI’s $5.4B, only ~35% of last year’s $13B. Rory’s read: within a couple of quarters Anthropic is “visibly and obviously ahead — profitable, growing more quickly and bigger… pareto dominant on all three vectors,” so number two must go first rather than later emerge as “a pale version” of the leader.
• Anthropic turned profitable ahead of its own 2027 guidance — gross margins expanded from 38% to 70%, with a projected $559M Q2 operating profit and a headline $44B ARR. The bear case, per Jason: after Opus 4.6 it’s a premium product at twice its competitor’s price, and if even half the world says the ROI isn’t there, the response is token budgets — “a combination of GPT-4, it’s older, and some Sonnet and some DeepSeek” — which may degrade Anthropic’s pricing power.
• Nvidia printed $81.6B in revenue and $56B in profit — “the most profitable company on the planet” (Google does $100–120B a year; this annualizes to ~$200B) — and the flat stock is the market working: mid-20s PE, the print telegraphed by hyperscaler capex to ±2%, and “beat, raise, and accelerate” already the bar. With Nvidia ~7% of the index, “we’re all AI investors whether we realize it or not.”
• The question of the year is ROI on the next $2 trillion of capex. Jensen’s $3–4T-by-2030 call implies ~$1T of Nvidia revenue, but constraints are economic, not technical: Uber’s COO burned a year of Anthropic credits in four months with gains “probably there but not measurable,” and Rory argues that once AI spend eats wages and forces layoffs, ROI must become provable — “when you’re spending $300 million, someone probably needs to know.”
• SpaceX’s S1 — the largest IPO in history — doesn’t tie to its numbers. Sum-of-parts on launch (10–20% growth), Starlink (~$14B revenue, EBITDA-positive), and an xAI that is “a more efficient CoreWeave” lands far below ~$2T; the residual is the Elon premium. Jason, self-declared team Elon: “It could be the GeoCities deal of the AI era… It’s Solar City on steroids and we’re all here for it because we love AI, but it makes no sense.”
• The Colossus–Anthropic rental is the get-out-of-jail-free card: $1.25B/month ($15B/year) with ~90-day cancellation against roughly $12–19B of xAI capex — “public storage but for compute,” where the tenant thinks it’s four months and stays five years. And don’t count Elon out of owning the datacenter buildout outright: CoreWeave and Nebius “have nothing, no special technology,” and a $2T pre-money “is a pretty damn low cost of capital.”
• “COVID overhiring” as the layoff explanation is “the dumbest take” — natural attrition of 15–25% a year since 2020 already exceeds the cuts at Intuit (16,000), Coinbase, and LinkedIn. The real story is comp compression upward: ClickUp cut 22% to pay high performers $1M, $2M revenue-per-employee becomes the new normal, and the token-ROI debate and the layoffs “are in fact the same discussion” — winner Anthropic, loser employees.
• Pick-and-shovels is the trade: Exa raised $250M at $2.2B for agent search (“agents don’t hop on Google”), OpenRouter $150M at $1.3B, and five or six agent primitives that “don’t come in the box” will support standalone companies. The venture meta-lesson: “seeds are for suckers” — with PMF verification compressed from years to weeks or months, “you want to invest the hour it blows up.”

Deep dive

1. Nvidia’s $81.6B quarter — the flat stock is the market working

• Jason’s correction of the headline: the compelling number isn’t $81.6B in revenue or the $91B Q2 guide — it’s $56B in quarterly profit, making Nvidia “the most profitable company on the planet.” Google clocks $100–120B a year; this annualizes (with some non-recurring items) to ~$200B. The stock moves on delta, not totals: after the 2023 gap-up it’s grown ~20% over six-to-nine months at a mid-20s PE — the market saying “this is amazing, but we’re not sure it’s going to 5x from here.”
• Jason’s frame: in this tape, not falling after your print is a strong sign — “beat, raise, and accelerate, that’s the mantra” — and the stakes are systemic, with Nvidia ~7% of the S&P 500: “when Nvidia falls, we all fall… we’re all AI investors whether we realize it or not.”
• Why there was no surprise available: GPUs run ~50% of total AI capex — call it $400B — and Nvidia’s commanding share implies a ~$300B run rate, “and there it is at 320 billion.” Last week’s hyperscaler announcements telegraph Nvidia to ±2%; “at that point, market reaction is in the noise.”

2. Jensen’s $3–4 trillion: extrapolation gets you there, ROI decides if you arrive

• Jason runs the math on likely Jensen Huang’s $3–4T infrastructure call for 2030: 50% GPUs → ~$1.5T semiconductor capex, and at maybe 70% share (Trainium and others taking the rest) that’s roughly $1T in Nvidia revenue from $300B now — “I don’t think the market’s quite anticipating that, to say the least.” The constraint isn’t technical: “is the ROI there economically on the next $2 trillion?”
• Even asking is heresy: “when I say it in San Francisco, it’s literally like wandering around the Vatican saying, is this Pope guy the right guy?” Yet the willingness to spend is historic — “I can’t remember, except maybe for a year in the mid-’90s internet, where corporate America was so convinced of the ROI of something” — with Anthropic’s jump from $5B to $10B GAAP revenue as proof the sluice gates are open.
• The arxiv paper Rory’s data scientist sent (“The Price of Progress”): benchmark pricing together, not benchmarks in the abstract. Price-per-token falls every year, but agentic reasoning multiplies consumption — the task one level out may cost 100x, so net cost to serve rises even as unit costs drop. “When you’re spending $3 million, you can be laissez-faire… when you’re spending $300 million, someone probably needs to know.”

3. Uber’s COO versus the token maxers — AI spend bifurcates

• The trigger: Uber’s COO said they spent the whole year’s Anthropic credits in four months and the gains were “probably there but not measurable” — alongside Microsoft allegedly moving off Anthropic because Opus is too expensive. Jason discounts the messenger: “a COO is often a bean counter with a different name,” usually without deep product-engineering background, living in “a pre-AI world where products are static” — which, he argues, Uber largely is.
• Jason’s bifurcation call: companies already at $1–2M revenue per employee “will token max until there’s no tomorrow,” while larger, traditional orgs grow skeptical as prices rise (Anthropic and Nebius both raising). Margin structure sets posture — Uber’s 39.75% gross margins make job one protecting them: “we can’t all spend like a drunken sailor with $2 million of free credits out of YC.”
• His falsification test, as told: DoorDash — still founder-led, “one of the most aggressive” — says engineering savings alone justify the token spend. “If Tony says the same, then I’ll consider myself chastised — we got ahead of ourselves for real companies.”
• Rory’s pushback, worth keeping: a developer at Uber has the same software problems as one at Meta — if AI lifts software development at all, the lift should be similar across industries. What differs is the burden of proof: “there’s no one at Facebook yet saying, are we getting value from our billions in AI — we just assume we are.” And once AI spend eats wages: “before we terminate a thousand people, shouldn’t we check perhaps, take another week here, guys?”

4. Anthropic goes profitable — and the bear case is the premium

• The numbers: gross margins 38% → 70%, a projected $559M Q2 operating profit, $5B GAAP revenue in Q1 versus $4.5B for all of last year. Rory saw it coming from his internal model — margins went from negative 60% the prior year to positive 34% last year, “that’s a trend,” and “high growth and improving margins meant profits were inevitable” even though Anthropic had guided to profitability in ‘27. One asterisk: early months may be flattered by a SpaceX discount.
• Jason’s bear case: after Opus 4.6, Anthropic charges enterprises directly per token and is a premium product at twice its competitor’s price — and it “lucked out by not going into video,” OpenAI’s biggest cash sink. If even half the world lands where Uber did on ROI, the choice is cut prices or hold the premium like Apple and cede share. The likely path isn’t back to hand-coding — it’s token budgets: “a combination of GPT-4, it’s older, and some Sonnet and some DeepSeek.”
• Rory’s counter: the Q1-to-Q2 traction says the default switches in corporate America are set, and only two things flip them — the green eyeshades showing up on ROI, or competitive dynamics changing. Right now Anthropic “stole the show in enterprise” while OpenAI was distracted by consumer and Google “despite having amazing raw tech just didn’t do a great job productizing.” “Absent one of those two things, the trend is your friend.”

5. OpenAI files first because it’s been lapped

• Rory on the confidential S1 ($852B–$1T, Q4 listing): “they have to and they should and they must.” The leaked comparison: Anthropic did $5B GAAP in Q1 — more than its whole prior year — while OpenAI’s $5.4B is ~35% of last year’s $13B. Play that out and one company grows ~10x year-on-year, the other 2–3x; within quarters Anthropic is “visibly and obviously ahead — profitable, growing more quickly and bigger, pareto dominant on all three vectors” while OpenAI is “smaller, growing less quickly, still unprofitable — a horrible strategic place to be.”
• The logic of going now: sell “a chance to buy the first foundation model” rather than wait and arrive as “hey, we’re like Anthropic but smaller and not as good.” The rule: “number one gets to choose what happens and number two has to respond” — and waiting to 2029 to grow out of it is “craziness.”
• Harry’s mechanical addendum: IPO capital pools are thinner than assumed — Cerebras was maximally oversubscribed and priced at the exact ceiling short of refiling with the SEC — so going first avoids exhausting the pool. Reception would be “overwhelmingly positive” per Rory: risk-on, no pure-play AI besides CoreWeave and Cerebras, and “you go outside the valley and utter the word Claude and people look at you. ChatGPT is the thing.”

6. Does Anthropic even need to IPO? The capex math says yes

• Should Anthropic react? Rory reaches for Jackie Fisher on why Britain had a navy: “to go anywhere you damn well please with whatever force you want.” That’s the point of being number one — operational excellence buys strategic independence, so at most they nudge timing.
• Rory’s contrarian: if you’re profitable with infinite demand for secondaries, and Dario has pledged 90% of his shares to charity, maybe you never float — “stay like Stripe… why would I want to float a small amount of my shares and deal with the headaches?”
• Harry’s rebuttal is the load-bearing math: Stripe’s capital needs are light; Anthropic’s are not. Every $1 of revenue requires $4–5 of upfront capex someone must fund — $100B of revenue implies $500B invested — and “maybe two years from now Microsoft doesn’t want to be your hyperscaler provider.” SpaceX, famously capital-efficient, raised only $23B privately in 20 years; above some dollar sum “only the public markets can do it. The capital needs here dwarf anything we’ve seen.”
• Harry’s scorecard on the trio, none “pitching a perfect game”: Anthropic nailed the model but botched capex forecasting by underestimating its own success; OpenAI nailed compute but lags on enterprise features; xAI builds datacenters faster than anyone but “weren’t able to fill it with their own stuff.” Still — three roughly trillion-dollar companies from a standing start in five years. “What did you do on your summer vacation?”

7. SpaceX’s S1: the gap between sum-of-parts and $2 trillion is the Elon premium

• Rory’s cynical read: the numbers don’t matter. Launch is good and boring (capped at 10–20% growth), Starlink is ~$14B of EBITDA-positive revenue growing 30–40%, and xAI is now “a more efficient CoreWeave.” Sum the parts and “I’m so far from $2 trillion in DCF that my head hurts” — the residual is the Elon premium (“Tesla trades six times fundamental value, so maybe the premium is 6x, I don’t know”).
• The tell in the filing: a claimed ~$28T TAM, “the largest TAM in history” — with 90% of it in AI, the one story where they’re not unique or differentiated. Grok “is a foundation model that’s gone nowhere,” and Twitter’s revenue has fallen 50% since the acquisition — “Elon’s the magician of revenue compression.”
• Jason, team Elon (five Teslas, three Starlink subscriptions): “It could be the GeoCities deal of the AI era… 100 times trailing sales, my god.” His verdict: “It’s Solar City on steroids and we’re all here for it because we love AI, but it makes no sense” — financial engineering that bails out the failed Twitter acquisition and the buy-chips-to-compete-with-OpenAI idea in one confab.
• Rory’s own close: “I love the S1… I think it’s all madness. I wouldn’t buy a share. I just love the optimism.” The narrative coheres only if the launch business enables Starlink and data centers in space, leading to “100 gigawatts of capacity a year, five years from now” — a claim he assigns low probability. “And the tweets don’t fit in at all.”

8. The Colossus trade: public storage for compute — and don’t count Elon out

• The deal that rescued the math: Colossus rented to Anthropic at $1.25B/month — $15B a year — with a ~90-day cancellation clause either way, against roughly $12–19B of total xAI capex. “One year and a bit of Anthropic revenue and you covered your nut” — an unusually high cash-on-cash return in a low-return-on-equity business, on a product built a year and a half ago.
• Rory’s analogy, as told: this is public storage for compute. You send furniture to storage thinking you’ll be out in six months; five years later you’re paying $400 a month. Anthropic “went in thinking we’re only going to pay 1.25 billion for four months,” but it’s relying on other people building datacenters, “and life is tricky — not everyone will bulldoze it through like Elon did.”
• Jason’s steelman: the man’s record is doing the radical thing for real — a Model 3 leases for $299/month, self-drives, needs no fuel, “and the only reason we don’t all drive them is 60% of the country hates them.” CoreWeave and likely Nebius “have nothing — no special technology”; Elon “could accelerate past everybody… we could turn around in 5 years and say he owns every data center in the galaxy.”
• Jason’s kicker: the one input a capex-intense datacenter business needs is cheap capital, “and a $2 trillion pre-money is a pretty damn low cost of capital” — sell $75B, put $25B into another Colossus. The compromise Jason sees: give up Mars for the moon, build a $50B+ CoreWeave to make the math tie. “If the one human who’s earned the right to say ‘give me a trillion dollars’ is Elon — if you want to scratch that itch, go scratch the itch.”

9. Layoffs aren’t COVID overhiring — they’re a comp revolution

• Jason on the consensus explanation (likely Marc Andreessen included, “who’s got 40 IQ points on me”): “the dumbest take I’ve seen since our last show.” Natural attrition runs 15–25% a year; compounded over six years it exceeds the cuts at likely Intuit (16,000), Coinbase (thousands), and LinkedIn (800), so blaming overhiring “is clickbait because it’s not mathematically true.” His own tweet: “It’s 5 years since COVID, people — haven’t you heard of performance reviews?” Harry raises one exception — monopoly-profit orgs anchored on headcount rehiring churn out of myopia — “but it’s weird to see it across the board.”
• What Jason liked in ClickUp’s Zeb (22% cut, hazed worse than Cloudflare’s public 20–21%): laying off to pay current high performers $1M — if a 10x engineer is now 100x and a $1M-bookings rep now books $2.5M, “I got to pay them.” He’s living it: “we’re doing with two and a half people what we did with 20, and everyone should make a million bucks.”
• The new paradigm: $2M revenue per employee becomes the new normal at scale — see Anthropic hiring a King’s Cross social media manager at $450K, which “all ties” at their revenue per head — and top performers should make 5x midpack, not 40–50% more. Rory’s synthesis: the token-ROI debate and the layoffs “are in fact the same discussion — in one case we’re focusing on the winner, which is Anthropic, and in one case the loser, which is employees.”
• The dark residue: Jason worries the agentic experts “get another 10x better and everyone else falls further and further behind and are unemployable… I’m the most tired I’ve ever been, but my productivity rate is off the charts.” Rory reaches for Max Planck: “science advances funeral by funeral… it must be pretty dystopian to have to do that or die.”

10. Pick-and-shovels rounds, “seeds are for suckers,” and the idle-agent problem

• The tape: likely Polsia raised $30–40M at $250M for a single founder; Exa raised $250M at $2.2B for agent search (Benchmark again — “this fund is going to be one of the best performing funds in history”; someone bid higher on the prior round and won — “in retrospect, you should have bid 750”), OpenRouter $150M at $1.3B led by CapitalG, and the likely Manus founders attempting a contorted buyback. Jason on Polsia: a 10/10 pre-paywall journey, better than Replit or Lovable, “then it immediately asked for a credit card before I got any value… I’m out.”
• Why Exa is the more interesting bet: these are products with no use without agents — “agents don’t hop on Google… they don’t need Zoom, they don’t need a traditional CRM.” Jason’s market structure: not winner-take-all (developer tool, no network effect, two comfortable players), and the thesis is that five or six primitives — a database, a search engine, observability — “don’t come in the box from Anthropic or OpenAI.” Jason’s caveat: at Databricks’ Neon, over 90% of databases are built by agents, “but they don’t pay as much” — these plays need vast volume to work.
• The venture meta-lesson: “seeds are for suckers… you want to invest the hour it blows up, the minute it blows up you want to get the DM and just wire the money” — PMF verification has compressed from a year-plus to weeks or months, so you must pay more on less information and abstract away why-didn’t-Google-do-it risk. “Corporate America flipped the switch a year ago and said thou shalt do AI in 2026.”
• The rage-bait finale: a CEO vibe-coded away a $600K Salesforce contract in three weeks. Jason: “a dated 2025 take — of all the threats to Salesforce, HubSpot, Monday, and Atlassian, this ain’t in the top 10.” On the same CEO shrugging off a 2x Anthropic price hike, Rory flips it: then you’re underspending — “tell your people: use twice as much.” Jason’s closing constraint: “our agents are idle… humans just can’t process all the output” — his $257/month AI VP of marketing nags three ideas daily at 7:13am. Rory’s fix: hire another human. Jason: “a $125K laid-off ClickUpper is worth nothing to me. I need a million-dollar person — and I will hire them today for a million bucks, no joke.”