Anthropic Raises $30B from Microsoft & NVIDIA & NVIDIA’s Core Business Faces TPU Threat

Summary

• Anthropic’s $15B from Microsoft and Nvidia at $350B (with $30B in Azure commitments) is read less as validation than as inevitability: OpenAI “wanted an open marriage and Microsoft said well if you want an open marriage I want one too.” Jason’s meta-lesson: “you need infinite capital when there’s no stability,” and roundtrip revenue concerns are shelved — “in a bull market nobody cares about everything and then in a bear market everybody discovers why you were meant to care.”
• The clearest structural threat to Nvidia is customer concentration meeting custom silicon. Four or five customers are 70-80% of revenue at 75%+ gross margins; Rory’s math says Google buying its ~$36B of compute from Nvidia would hand over “north of $20 billion a year of profit” — more than enough to justify a billion a year on TPUs. John’s blunt read: “we’re just ignoring the risk because Nvidia’s numbers are just too good.”
• Even so, nobody on the desk calls Nvidia overvalued: its PE is “lower than the PE on Costco,” and Google’s infrastructure head says Google needs “a thousand times more compute in five years.” The real question is whether 2025-26 capex is steady-state demand or a cyclical peak — and the market is already differentiating, rewarding Google’s spend and punishing Oracle’s (down 40%).
• On Altman’s “war mode” memo, Rory mocks the metaphor (“the US Marines are still taking applicants”) but Jason defends the substance: “nothing happens when you’re not in hyperaggressive mode,” and founders should not go easy — “push them as hard as the business needs to go. And if they leave, it’s great.” Sergey’s return is cited as the proof case at Google.
• Sierra at $100M ARR / $10B valuation only works if it eats labor, not software: Rory’s 5-year compounding math gets to ~$5B revenue against Service Cloud’s $8B — a mere 2.5x — unless it captures the “$200 billion a year services market for customer support agents.” The rate limiter isn’t demand or Brett Taylor’s selling power (“Brett can get multiple $10 million checks”); it’s “the physics of diffusing this technology into the enterprise.”
• The sharpest disagreement: is an installed base an asset or “cement shoes”? Jason says supporting pre-AI customers is “a drag” — his example: $45M of AI revenue growing 100% chained to $50M pre-AI revenue growing zero — and most B2B unicorns will fail the transition. Rory counters that customer data plus distribution wins when there’s an organic path (Intercom, Gong), while Jason allows that “Salesforce could be the next Google… in 18 to 24 months.”
• Lovable at ~$200M ARR / rumored $6.3B vs Wix at $2B revenue / $5.24B cap frames the public-private dislocation: markets pay 5.1x for sub-20% growth, 11.8x for 20-30%, 23.7x above 30% — so if Base44 (0→$50M) can hit ~$250M, “put as much of the 401k into Wix as you can.” Jason takes Wix “on financial engineering” only if the founder stays 24 months; Rory takes Lovable at the margin but is “a bit nervous at six billion.”
• GEO is the episode’s live bet: Jason calls it “snake oil AI… going to die in 26 or 27” because nothing is actionable, while Rory argues advertisers must show up wherever the people are — and ChatGPT advertising is the shoe that drops. Adobe already paid $1.9B for Semrush at ~3x with a ~100% premium; Harry stakes $5,000 that his portfolio company Peak (as spoken; likely Peec AI) is the exception.

Deep dive

1. Anthropic’s $15B raise: infinite capital, open marriages, and nobody caring about roundtripping

• Jason’s meta-learning from a week where Twitter crowned OpenAI, then Gemini 3 Pro, then “Cloud45” (likely Claude 4.5) in three days: “there’s just no stability… I think you need infinite capital when there’s no stability. More power to him.”
• Harry’s structural read — the commentators are less stable than actual market share, and the key fact Harry glossed over is Microsoft’s participation: “Microsoft was in a monogamous relationship with OpenAI and then OpenAI wanted an open marriage and Microsoft said well if you want an open marriage I want one too — this was probably inevitable.” The $15B-for-$30B-of-Azure structure “we’ve seen before,” and buried in the announcement: Anthropic breaking ground on its own physical data center, another model company deciding hyperscaler compute isn’t enough.
• On whether anyone still cares about circular revenue, Harry’s framing is the episode’s most quotable: “In a bull market nobody cares about everything and then in a bear market everybody discovers why you were meant to care. Now we’re in the don’t-care part of the trade.” John’s view is that, of all the roundtripping deals, “a Microsoft-Anthropic-Nvidia deal probably has better principles and upside than most — some of the other roundtripping deals look like they’re already a bit shaky.”

2. The TPU math: why Google building chips is rational and Nvidia’s five customers matter

• Jason’s user-level anecdote first: using Replit two hours a day (“top 1% of users”), he tried Gemini 3 Pro the day it shipped — “maybe 20% better for design” — then rolled back to Claude in his next prompt with zero friction. “TPUs, GPUs, LLMs, all in the space of 60 seconds… the idea that Nvidia is unstoppable because of the software-hardware connection — I know there’s a lot of truth to that, but as an end user I went right back and forth.”
• Rory’s load-bearing arithmetic: 90% of Nvidia customers by count spending $1-10M would be mad to design out CUDA — but four or five customers account for 70-80% of revenue. Google’s ~$90B capex at a ~40% compute rule of thumb is ~$36B of chips; at Nvidia’s 75%+ gross margins that’s “handing them north of $20 billion a year of profit.” A credible chip might cost “a billion, maybe a billion a year for 5 years” — “if you’re giving someone $20 billion of profit a year, you got to look at that. I gotta look at that if I’m Google, if I’m Amazon, I definitely got to look at that if I’m Tesla.”
• The contrast with Intel’s CPU era carries the point: Intel had ~100 million customers paying $200 each for a Pentium — “the more customers you have, the easier it is to charge them a little more and the harder it is for them to take your margin back.” And for OpenAI, “burning even more money,” cutting two-thirds of compute cost “goes from one of the biggest cash-hemorrhaging businesses of all time to a profitable business.”

3. Nvidia’s defense: neoclouds, CUDA inertia, and Google’s game-theory dilemma

• Rory on why Nvidia sponsors CoreWeave and the neoclouds: “a simple humble $20 billion market cap company does not have the capacity to build their own chip” — diversifying the compute buyer base is the only structural defense. For the long tail, CUDA’s “brain debt” and activation energy protect the franchise; “but you’re right, Harry — all you have to do is peel off one or two of those big margin cows and you’re done.”
• The twist: the biggest potential TPU customers are “Google’s sworn enemies” — Microsoft, Amazon, maybe OpenAI. Does Google sell TPUs and take the capital, or keep a structural cost advantage in-house? “Don’t have the answer there, but it’s outside Nvidia’s control.”
• John’s confession, tying back to the opening: “we’re just ignoring the risk because Nvidia’s numbers are just too good… This is systemic risk.” He invokes Twilio losing Uber (12% of revenue gone) as the small-scale precedent — “this would be much bigger” — but “even public market investors have to play the game on the field.”

4. Underwriting four-customer businesses: the data-labeling lesson

• Rory’s change of mind, stated plainly: “I will admit I was scared of the data labeling companies and I was wrong. There was a period of a couple of years where they clearly worked and worked really well.” The open question is whether Scale/Surge/Mercor/Turing keep working “the next three to four years if dollars get a little more scarce” and customers shift from effectiveness to efficiency.
• Harry’s underwriting logic, which Rory endorses: having interviewed essentially all the founders, what made him comfortable was the verticalized, “strange and weird” data requirements — surgical data, bookkeeping data — “so specific that the large customers are never going to churn.” Rory adds the required CEO skill: “extremely good poker — I know your $20 million contract is my biggest contract, but I also know my data is your most important data.” The ASML/TSMC dynamic — one seller, one buyer to a rounding error — proves such businesses can exist, but “you have to have something really unique to avoid them pounding you.”
• The generalization worth keeping: market speed is the get-out-of-jail-free card. “The earlier you are with the more hypergrowth ahead of you, the easier it is to be undifferentiated. And by the time growth slows down, you better either be differentiated or — as Jason was — you better be exited.”

5. Is Nvidia overvalued? Wrong question — is 2025-26 compute demand steady-state or peak?

• Rory reframes: on today’s revenue Nvidia is “far less aggressively priced than Costco or than Cisco in ‘99” — the PE-below-Costco soundbite. The earnings themselves contained “no data of any significance” because every hyperscaler had pre-announced capacity-constrained capex. So “is Nvidia overvalued” translates directly to: is 2025-26 demand steady-state or a cyclical peak where “two-three years from now we’re not going to be spending $90 billion?” The TPU/substitution question is secondary.
• John’s demand anchor: Google’s head of infrastructure said Google needs a thousand times more compute in five years. “It’s difficult to believe Nvidia won’t be a leader in that time… I don’t know many SaaS companies predicting a thousand-x growth at scale.”
• Rory reads the recent 5% wobbles not as correction but as functioning markets: “Who’s doing it well? Google — you can invest more, your stock goes up. Who’s doing it badly? Oracle — your stock goes down… those of you that are a bit out there on the risk continuum probably should be thinking about that.” John’s color: a hedge fund once backed his first fund purely to get real-time Nvidia exposure through portfolio companies with linear Nvidia linkage.

6. Altman’s “war mode” memo: bad metaphor, right instinct

• Rory’s demolition of the rhetoric: “Did you just suddenly discover there was a war? … Let’s be honest, we’re a bunch of pampered West Coast elitey computer people. If you want a war, the US Marines are still taking applicants. Otherwise save the metaphor for someone who cares.” His better question: “forget the metaphor — what exactly are you going to do today that you weren’t doing yesterday?”
• Jason’s counter is categorical and the episode’s strongest management take: across startup, scale-up and Fortune 500, “nothing happens when you’re not in hyperaggressive mode.” Twelve years of technical debt and promised features consume every engineering hour by default — “just fixing the bugs could take all year.” When he evaluates portfolio companies now: “I don’t care about your talk. I don’t want to hear about your pilot. I want to smell that your team is in hyperaggressive mode. If I don’t smell it, you have no chance.”
• The tell of hyperaggressive mode, per Jason: visible velocity increase in every function at the board meeting — “we shipped twice the story points as last quarter and we’re still falling behind and I’m pissed.” His proof case: Sergey returning to Google — “we weren’t even allowed to use our own coding tools, our own chips. I got rid of that in a week.” And a warning: “most of the startups that aren’t founder-run are unable to get back to hyperaggressive mode. It is impossible for them.”

7. How hard can you push? “Just one level below delusional”

• Jason’s advice to founders, flagged as 2026-appropriate rather than 2021-toxic: “I do not think you can push your team too hard. Push them as hard as the business needs to go. And if they leave, it’s great — because they’re not going to get you there… The best ones will always step up. They might cry or lose it for a couple days, but they will step up.” His model portfolio moment: the CRO of his fastest-growing company standing up unprompted to demand more than the stretch plan, with a data-driven case.
• Rory’s calibration — self-aware that “I’m not a good rah type leader, which is why I’d be a mediocre CEO and I’m okay as an investor”: in his best companies the CEO would occasionally “lose his shit” from exasperation, and the board’s job is to “check in on the gear grind and go, okay, I can live with that gear grind.” His limit: “You can push to just one level below delusional… it doesn’t work if you lose your entire team” — pointing at OpenAI’s departures and the risk of creating Anthropic or Thinking Machines. “You don’t want to force two of your best people out and they build your top competitor.”

8. Google vs OpenAI in consumer: not either/or, and Rory owns his miss

• Rory’s explicit correction: six-nine months ago the “Google is dead” narrative — “I got that one wrong.” Search volume is rising as people ask more questions (citing the WSJ interview with Google’s head of search); the decline “isn’t going to be precipitous.” Simultaneously, ChatGPT has carved a genuinely new category — paid consumer AI subscription, ~800M users with ~5% paying — “I don’t see all those users quote-unquote going back to Google.”
• On Altman’s leaked reference to growth falling to 5%: “the therapists call that catastrophizing — if a company growing north of 500% year on year went to 5%, that would be the single most catastrophic growth decline in history.” His base case: both win, ad dollars concentrate further, and “the great American tech 7 suck more of the world’s profit dollars out of the rest of humanity. Go team.”
• Jason’s consumer verdict is simpler: he prefers AI Overviews to ChatGPT “because I don’t have to leave Google… You’re not giving me any reason to leave. They’re truly great software.”

9. Sierra at 100x ARR: the math only works if you eat labor

• Jason’s street-level caveat first: enterprise AI support is “massively overselling what they can deliver today… so many folks I know have bought next-generation AI support tools and have not deployed them at all, or barely, or there’s no AI working” — the same arc as vibe coding, which “bordered on fraud” early in the year before becoming real. “I’m not being cynical. It’s just oversold.”
• Rory grants the category — support alongside coding is one of the two biggest LLM markets, and a reference call put resolution rates at ~60% with LLMs vs ~23% pre-LLM. Then the compounding math: $100M ARR going 10x, 5x, 3x, 2x, then 50% and 20% gets you ~$5B in five years — sanity-checked against Service Cloud, Salesforce’s largest cloud, at $8B today. At a Salesforce-style 5-6x multiple that’s ~$25B, only a 2.5x from a $10B entry. “The only way this math works is if you eat a huge slug of the labor — instead of the $20 billion software market, the $200 billion a year services market for customer support agents.”
• Both agree Brett Taylor is uniquely equipped for the Benioff play — “What’s your number one problem? Give me $10 million and I will get this for you” — CTO of Facebook plus co-CEO of Salesforce means “no one’s got a better package.” Jason’s twist: “the hundred million almost isn’t impressive because he can will it out of the ether with his background. A billion next year with a thousand FTEs — that will be magic.”
• Rory’s rate-limiter thesis, the section’s keeper: “the physics of diffusing this technology into the enterprise will be the rate limiter — not raw demand, not the talent of the CEO, not the product.” Anthropic and Cursor are offered as examples of $100M-to-$1B growth, with self-serve mix left as an open question. And the investing coda: “when something is obviously true, valuation expands to recreate risk.”

10. Installed base: cement shoes or the greatest asset? Jason and Rory genuinely disagree

• Pressed on Intercom (his portfolio), Rory defends the position: Fin’s AI growth rate is “frankly comparable to Sierra’s,” acquired “at a very attractive price” alongside an existing SaaS base — “I feel pretty damn good here, thank you.” John adds that Intercom has done an amazing job transitioning to an AI-first world.
• Harry’s counter with his own specimen: a portfolio company with $45M of AI revenue growing ~100% shackled to $50M of pre-AI revenue growing zero — “they are inexorably linked, but they’re not the same product. Right now it feels like a drag to have a couple thousand pre-AI customers to make happy.” Existing customers “can consume all of your story points” — and AI-native startups sometimes win not on talent but “because they don’t have a thousand complaining customers to support.” His hedge, stated twice: “in a year I may think it’s the greatest thing… Salesforce could be the next Google. It could be, in 18 to 24 months — Agentforce crushed it, we didn’t need all these startups.”
• Rory’s framework: incumbents hold one huge advantage (customer base, data structures, seamless human-agent handoff) and one huge disadvantage (“they can’t get out of their own freaking way”). The precondition is organic linkage — Gong had “an obvious path from here to there”; if your SaaS thing has no AI equivalent, “you could be God’s gift to management and you’re screwed.” Harry’s synthesis: “you got to be as good or better than Owen at Intercom to make this transition. Most of these B2B unicorns are going to fail… your typical PE company can barely get a release out a year.” Supporting detail: Zendesk’s CEO told Harry base-version AI is ~20% automated; Sierra/Fin/Decagon-level requires a month of training with FTEs.

11. Lovable at $6.3B vs Wix at $5.24B: segment the cohorts, and the market won’t pay for a checkbox

• John’s analytical move on Lovable’s $200M ARR (2x in four months): segment the base. Low end churns at “50% or higher” (Suno retains only ~20% at the low tier, per Menlo); mid-pack ~100% with upsells; the high end — he watched Replit close a seven-figure deal live — “probably 140 to 160% NRR.” His accounting treatment: “draw a black line through the bottom and just call that marketing spend.”
• The Wix comparison is the uncomfortable one: $2B revenue growing 14%, $5.24B market cap (~2.6x), with Base44 at $50M ARR and growing very fast — “either Wix is very undervalued or Lovable is very overvalued.” John’s read: “the market’s saying you’re not getting credit for checking the AI box… you need to be Palantir. Wix ain’t Palantir. Show me 500 million.”
• Jason’s tier table makes it tradeable: sub-20% growth = 5.1x ARR; 20-30% = 11.8x; 30%+ (Rubrik, Palantir, Figma) = 23.7x. If Base44 does $250M next year, Wix jumps tiers and “you should put as much of the 401k into Wix as you can.” Rory rejects the “credit” framing on principle: “everyone’s just a set of discounted future cash flows… in the short term markets are a voting machine, in the long term a weighing machine” — though Harry needles him that Palantir, “detached from a DCF model,” proves otherwise; Rory holds: “in the end the truth’s out.”

12. GEO: snake oil or the next Omniture? The $5,000 bet

• Context: Adobe bought Semrush for ~$1.9B (~3x revenue, ~100% premium) — publicly because it was “the number one thing customers were asking for,” as mature CMOs panic about showing up in LLMs. Jason’s verdict is brutal and specific: GEO is “snake oil AI… going to die in 26 or 27.” His evidence: SaaStr gets ~5M blog views a year, traffic up 50% while SEO is down 8%, he’s tried every tool — “I literally cannot find one thing that’s actionable. Put more things on Reddit doesn’t help me.” His heuristic: “if you have to put in your credit card immediately, this is a bad sign… How come I can’t GEO for free if it’s so great?”
• Rory’s counter, drawing on being an Omniture investor pre-Adobe: analytics was “unactionable” once too, but “if your CEO says I typed L’Oreal into ChatGPT and they said 10 mean things about us — what the fuck? — you better have an answer. Just giving that VP of marketing an answer is worth 30 grand.” First principles: “advertisers go where the people are — Yahoo, Google, Facebook — and the people are using answer engines.” The big shoe: when ChatGPT allows advertising, “the wall of money hits here” — though platform risk cuts both ways, since in SEO the platform kept most of the value and Semrush at $2B was the biggest outcome.
• The wager: Harry, an investor in Peak (as spoken; likely Peec AI; the captions specify “15x in 3” without a clear period), stakes $5,000 that Peak is the exception — while Jason maintains the category-level call: “there’s going to be a billion-dollar ARR opportunity in AI marketing in 36 months, but the snake oil isn’t going to do it. Amplitude clones that get you 5% of the way there aren’t going to do it.”

13. Figma’s broken IPO, unearned windows, and stale 2021 marks

• Jason mourns the signal: Figma’s IPO captivated “almost at a consumer level,” and a broken IPO punctures the liquidity-is-back feeling. Rory shrugs: professionals priced it at ~$35 — roughly Adobe’s offer of two years prior — retail madness took it to “10-something,” and now at ~$17-18B “it’s pretty much priced exactly what they thought it was worth. Capitalism works. Lighten up everybody.” Jason’s founder lesson: adjusted for time, dilution and risk, it sits 30-40% below the Adobe deal — “if you turn down a deal, you better want it to go public” (with the Wiz $32B sale, now US-cleared, and iRobot’s post-blocked-deal collapse as the bookends).
• On 2026 IPOs, Jason: “I don’t think we earned it this year… You’re better than Figma? You’re better than Netskope? It’s a high bar.” Rory: “IPOs tend to be Pavlovian — people do things when the last thing they did felt good.” Both want froth back — oversubscribed IPOs to “free up liquidity in our portfolios. I’m not being facetious.”
• The decaying-cohort warning: any unvalidated 2021 mark should be down, and “every year that goes by, the probability that the three-four hundred pre-2022 unicorns not making the AI transition ever get out has to decline.” The growth-quality kicker from Harry: if you’re growing 15%, “probably the majority is from price increases… it’s fake growth.” Rory: “you’re not going to get into the 20s on price increases, probably not even on NDR — you’re going to get into the 20s because new people want your product. If you cut R&D in the downturn, you’re probably screwed.”

14. Would-you-rather: Lovable vs Wix hangs entirely on one founder staying

• Rory takes Lovable at the margin — “the prior is that the AI-first company has the edge” — but only for lack of a Wix thesis: “if I had clarity on how the vibe-code product rolls out across the organization, I would take Wix all day every day, because I think you can make a four or five-x that would be liquid.” And the hedge: “I’m a bit nervous at six billion. A bit.”
• Jason’s condition is the Base44 founder: “if he’s really going to stay 24 months, I’m going Wix just on financial engineering. If he’s going to leave, I’m putting my money in Lovable.” Harry says the founder may stay; Rory says he has a variable package. Jason’s worry is the siren call — “now he can raise a billion pre for his next startup… 150 million to start likely Base44 is a tough one to say no to.” Rory’s comp-committee fix: an accelerated equity grant tied to AI penetration of the installed base — at 20% penetration “you probably jump two buckets” — even if that means “an Elon package… you got to give him 300 million because he’s added three billion to your market cap.”
• The closing through-line, looping back to war mode: “call it what you like — mission clarity is worth a lot. The mission clarity for that company is: take this existing customer base who should be vibe coding, you have the product, you have the founder, make it happen.”

(Note: the episode feed carries a 20Sales/John McMahon title, but the captions are a 20VC roundtable — Harry Stebbings with guests addressed as Jason and Rory.)