Anthropic Raises $45B but Falls Short on Compute & Thoma Bravo Hand Back Medallia Keys to Creditors

Summary

• The model race flipped, and the market is trading last year’s news. OpenAI’s reported miss on users and revenue (CoreWeave down ~5%, Oracle ~7%) is “a late dropping shoe” on a period when OpenAI built weak models and bled share to Anthropic — but reviews of 5.5’s coding are “pretty damn good and arguably better than the current Anthropic model,” and on AI Twitter “Anthropic is the guy that’s getting the slamming right now.”
• Agents will increasingly pick the vendors, and that resets everything. Jason’s core call: “more and more the agent is going to choose what models and just what vendors we use” — his own AI VPs “love OpenAI. They love the API,” so “I got to follow my agents.” His weekend API-grading experiment had Claude, OpenAI and Gemini rank the top 120 APIs: Stripe got the only A+, while Marketo, Outreach and Salesloft were dismissed as “worthless products in the age of agents.”
• Anthropic’s $45bn raise is the price of winning the model war and losing the compute war. Google commits up to $40bn ($10bn cash now at a $350bn valuation, $30bn on milestones), Amazon adds $5bn. Rory’s math: a $10bn run rate compounding 5x then 4x needs roughly $300bn of capex between you and your partners — “it makes running an airline look easy.” And Altman’s “compute equals revenue” is mere correlation: “compute and a shitty model also equals no revenue. See Grok for details.”
• The buy calls: Google risk-adjusted, Nvidia for pure upside. Google “win win win” — it profits whether you use Gemini or Anthropic, has rotatable surplus capacity, and Google and Amazon’s chips attack Nvidia’s 70% gross margin ($14bn of raw profit per gig) — “provided ChatGPT does not erode Google search, which is the mother lode of cash.” Jason’s counter: “just back the truck up to Nvidia. Don’t even think or spell it.”
• Medallia is a mega-zero of the PE-SaaS era — killed by overpaying, not over-levering. Thoma Bravo hands creditors the keys on a $5.1bn equity wipeout from a deal that was ~80% equity: “you can’t service 2 billion plus of debt on a 1 billion low-growth company with a pre-AI story.” Watch list named: Coupa, New Relic, Anaplan, Zendesk, Avalara, Smartsheet — and if AI renders B2B software non-durable, “the whole classic PE model is broken.”
• The exit funnel has collapsed to “fewer but bigger winners.” With PE as buyer-of-last-resort gone and strategics “narrower than you’d ever imagine,” the real IPO bar is a billion in revenue growing 40% — a board-meeting anecdote said exactly that and “the room went silent.” Jason’s predicted micro-trend: founders with no path simply hand the keys to a bigger friend’s company for a third of the equity.
• China blocking Meta’s $2bn Manus deal is deterrence, not clawback. Distributed capital isn’t coming back (“I don’t know if you can hide in Woodside from the CCP”), the leverage is on Meta, and the message — “we’ll show you with Manus like you showed us with Nvidia” — is one early expression of the US-China AI war, alongside coming social dislocation (California’s billionaire tax at mid-40s odds on Polymarket).
• Sell what you can’t defend. Harry said he was selling his Figma and Duolingo positions and specifically sold Figma 40% down; Rory’s rule — “don’t wait for the [expletive] to come up. Just sell it and redeploy” — and Jason’s blessing: “The agents don’t need either of them, Harry.”

Deep dive

DIGEST:

1. OpenAI’s miss is a late-dropping shoe — 5.5 may have turned the tape

• Rory’s framing: a foundation model company has exactly two jobs — “you got to build great models and you got to buy enough compute to be able to run them” — and in the back half of last year OpenAI failed the first job, so its traction and market share “declined markedly” against Anthropic. The reported miss that hit CoreWeave (~5%) and Oracle (~7%) is “a late dropping shoe on facts that were probably knowable three, four months ago if you were paying attention.”
• The present tense looks different: the model OpenAI shipped — “I think it’s 5.5” — draws coding reviews that are “pretty damn good and arguably better than the current Anthropic model,” and in “the super connected Twitter AI universe,” it’s Anthropic getting slammed — “Claude can’t keep up, can’t support the users, and the current Codex model is better.”
• Context on how the share shifted: Elon’s line that Anthropic had “something special in coding” understated how much of overall market growth was coding; OpenAI acknowledged it by going “double coding red” and fixing Codex.

2. “I got to follow my agents” — vendor selection moves from humans to machines

• Jason’s biggest call of the episode: “more and more the agent is going to choose what models and just what vendors we use.” His deployed agents — an AI VP of marketing, an AI VP of customer success — “love OpenAI. They love the API. They love it,” which is “one of the many reasons I’ve come back to team Sam.” His timeline: late 2026 into 2027, most workflows are managed by AI agents — “not crazy open claws blowing up our Mac minis, but running everything.”
• Harry’s paraphrase, accepted: agent choice removes human anchoring bias, making it “every day’s a new day” market. Jason’s caveat from his weekend “996 project” — having Claude, OpenAI and Gemini jointly grade the top 120 APIs — is that agents are “very biased toward momentum” and innovative market leaders. Stripe got the only A+ (“a reason to go long Stripe”); Claude “mocked Marketo, Outreach and Salesloft” because an agent “will just craft and send a better email itself” — “worthless products in the age of agents.” The graders ranked Anthropic just above OpenAI, Gemini below.
• The market read: public markets have “the right idea, but the wrong direction” — the threat isn’t vibe coding, it’s what the agents pick. Hence Atlassian and Monday punished (agents have “no use” for project management tools) while Twilio and Cloudflare outperform (agents still call them). And it’s why the agent wars matter: whoever owns the agentic layer gets lock-in, because “the OpenAI agent will probably pick OpenAI as the API.”

3. Deferred churn is “where the rent-a-CEO and the mediocre hide”

• Harry asks whether big multi-year enterprise deals confer value. Jason: no — “churn that is deferred still exists.” Workday’s standard three-year contract plus five-year renewal means the average customer effectively signs for eight years — which just means “they have eight years to find better agentic solutions.” If terminal value dies, masking churn doesn’t help: “if the customer dies anyway at the end, you never have it.”
• Harry’s ServiceNow proof point: ServiceNow grew 20%+ and got a very negative reaction, with analysts grinding the CEO of a $16-20bn business about half a billion to a billion of agent revenue — “what they’ve recognized is that’s the tell for the future.” Harry said he was “willing to bet in a quarter or two” someone would ask Benioff how many calls to the agent, the headless API, it got.
• The Canva test case: Canva’s agentic suite — “Canva 2, I think it’s called Canva 2.0” — is “actually very very good” and the IPO will be “wildly successful” — but “ask yourself a question, would an AI agent use it? No… It’s just going to create the assets.” Jason thinks enterprise crosses over before consumer — the $18/month prosumer won’t replace themselves: “none of us really want to replace ourselves with agents.”
• Harry’s three buckets: melting iceberg (add leverage and “you’re dead”), system of record (kept forever, bounded terminal value, “there is a price at which you should buy the stock”), and agent-native (increasing returns). Pricing consequence: “SaaS pixie dust credit expired” — Canva and Rippling (~70% growth at $1bn) will trade at fair value, not “that stupid 30 times revenue premium” of 2021. “The dirty little secret of venture is how much of your money you make in that one year in 10 when everybody buys the dream.”

4. Anthropic’s $45bn: the model may have over-succeeded, and the compute math is terrifying

• The round: Google up to $40bn — $10bn cash now at a $350bn valuation, $30bn on performance milestones — plus another $5bn from Amazon. Rory’s symmetry: OpenAI got compute right and the model wrong; “Anthropic did it exactly the opposite way… in fact, they may have over succeeded,” leaving them massively compute-constrained. “No one had a business plan last year when they went from one to nine that said they’re going to go to 30 by the end of Q1.”
• His napkin math on why the second job is brutal: at a $10bn run rate, 5x this year and 4x next implies ~$100bn+ two years out; at $4-5 of capex per $1 of run-rate revenue, “someone between you and your partners has to find plus or minus $300 billion to buy chips, dig holes in the ground, build data centers.” Undershoot demand and “you’re going to look like an idiot”; overshoot and you’re left with $150bn of stranded capacity. “It makes running an airline look easy.”
• On Altman’s “compute equals revenue,” Rory doesn’t hedge: “it’s a f*ing stupid statement… it’s just correlation. No compute equals no revenue. But compute and a shitty model also equals no revenue. See Grok for details.” Jason’s softer version: “it’s possible you look back and see that as the first disconnect from compute equals revenue” — not certain, but “superficially it seemed to have happened.”
• The medium term still argues for leaning in: Rory first cited a number of 50-100x more tokens for agents, then clarified it was probably 10x a Jason and 100x a Rory — Jason’s Salesforce bill went from $12k to $22k while seats fell from 10 to 2-plus-1 — so aggregate demand rises even through “6-month periods where you’re like, I’m an idiot, I don’t have enough demand, and then six months later, I’m an idiot, I don’t have enough compute.” Direct rival-to-rival resale may be awkward, but hyperscalers will re-route it — “just a $10 billion sublet” — as CoreWeave’s real-time data-center reallocation already shows.

5. Google wins either way; hyperscaler chip bundles attack Nvidia’s 70% margin

• Jason’s verdict on the round: “Google win win win.” Anthropic is now deeply tied to Google, so Google profits whether you use Gemini or Anthropic; it has surplus capacity it can rotate between itself, Anthropic and others “when they want,” and the cash flow to fund it.
• Rory’s chip math: GPUs are roughly 50-55% of total build-out capex; a one-gig data center costs $30-40bn, ~$20bn of it compute, and at Nvidia’s 70% gross margin “$14 billion of that per gig is raw profit to Nvidia.” Neither the Google nor Amazon chip is widely available standalone — both hyperscalers bundle chip + capital + equity to keep Anthropic on their silicon and take that margin, “arguably with, as Nvidia would say, a substandard product. But there’s many examples in tech of substandard bundling products succeeding. See Microsoft for details.”
• Quick-fire, Google ($4trn) vs Nvidia ($5trn): Rory buys Google “reluctantly,” risk-adjusted — it “can win if AI adopts fast, it can win if AI adopts slow… provided ChatGPT does not erode Google search, which is the mother lode of cash, they’re golden.” Jason takes the other side: Nvidia is “the best pure play into the AI vector” — “if you want to bet on AI today, cuz we can’t buy Anthropic or OpenAI… just back the truck up to Nvidia. Don’t even think or spell it.”

6. China blocks the Manus deal — deterrence dressed as an unwind

• On China blocking Meta’s $2bn acquisition of Manus (a Singaporean company, distributions already made), Jason’s riff on investor reality: “If I own 20% of Manus and got my $400 million out… I ain’t giving my money back. I’m taking my $80 million in carry for myself, and I’m hiding.” Rory’s dry reply: “I don’t know if you can hide in Woodside from the CCP” — but both agree the chance of returned capital is zero.
• Rory’s read of the real mechanism: China’s leverage runs through Meta, not the VCs — “you have this technology, we’d like it back.” If Meta did serious business in China it would be at the table already, the way Tesla with a massive Chinese plant would be. The deeper purpose is deterrence: “that’s the first, last, and only one of these deals that anyone will do” — future versions mean putting the team on “a 737 in Beijing” and wiring only when they land in Singapore with their families.
• Zoomed out, both see it as an early expression of two decade-defining themes: the US-China AI battle (someone facing prison for selling Nvidia chips into China; from Beijing’s side, “we’ll show you with Manus like you showed us with Nvidia”) and AI-driven social dislocation — Jason expects revolts as layoffs land, California’s billionaire tax passing (Polymarket has it mid-40s), New York’s penthouse tax, “the exodus will continue.” Rory: in 30 years, historians will write about “the revolt against inequality and AI” and the China rivalry — “provided we don’t blow up the world.”

7. Medallia: a mega-zero — overpaying killed it, not leverage

• Thoma Bravo hands the keys to creditors: $5.1bn of equity wiped out against roughly $3bn of debt (Rory thought closer to $2bn; Pluralsight was the smaller precedent, where Vista lost ~$2bn on a heavily levered deal). Rory’s load-bearing insight: the 2021 deal was ~80% equity — “they didn’t way over-lever it. They just way overpaid for it” — so with a couple hundred million of EBITDA, at 8-9x adjusted EBITDA it wasn’t worth putting in another equity dollar.
• Jason’s version of the same epitaph: “you can’t service 2 billion plus of debt on a 1 billion low-growth company with a pre-AI story that has to transform to AI.” Medallia is a customer-survey product, not a system of record — easy to transition away from, with “a whole bunch of way better AI-first products in this space,” and a top target of vendor consolidation (Gartner: 30-50% of AI dollars come from consolidation). “Do we really need that half-million-dollar-a-year dated survey product?”
• The at-risk list Jason names: Coupa, New Relic, Anaplan, “even Zendesk,” Avalara, Smartsheet — plus Proofpoint, Qualtrics, Alteryx, and Cornerstone (multiple term loans reportedly underperforming). Rory’s sequence for how these die: “the horsemen of the apocalypse are first the debt starts trading well below par… then payment-in-kind toggles… and when the refinancing cliff happens, that’s when you face the music.”

8. The exit funnel collapses to “fewer but bigger winners”

• Does it matter if half these PE-owned names go bust? Rory: yes, three ways. LPs take real losses in what was sold as “the safe part” of the private portfolio — “you can seem calm and phlegmatic about the prospect of loss, but when it actually happens it hurts.” And critically: “there goes one of our exit routes.” IPOs must now be huge, strategics are “narrower than you’d ever imagine” (at Adobe, Jason sat in meetings where the ‘perfect fit’ targets had never been heard of), and PE buys only at very low prices.
• The bar, from a board-meeting anecdote: a company just crossed $100m, cash-flow positive, and an investor told them “to achieve your outcome in today’s market, you need to hit a billion in revenue probably growing 40%.” “And the room went silent.” “400 million growing 30% is not good enough” — Navan, Figma, SailPoint, NetSkope are “great companies, but the IPOs are crap.”
• Portfolio construction splits venture in two. Early (anything below ~$100m ARR, “before you can squint and see an IPO”): diversify harder, accept “fewer but bigger winners” — probability per company is lower, the one that goes the distance is bigger. Late (above $400m): massive concentration, Thrive-style — “there’s only 40 names you even have to think about.” For scale: the ’90s had 300 IPOs a year; the IPO “used to be basically the Series C.”
• Jason’s micro-trend for this year: founders giving the keys to their friends. The monologue: “I’m at a hundred million… I’ve never gotten an offer from Google. I haven’t gotten a PE call in three years… so I’m giving the keys to Rory” — merge into the bigger buddy’s company for a third of it. Rory’s industrial version: a mid-career operator rolls up five systems-management companies, runs them “like a hard-ass” to 20% growth and 30% EBITDA — because nobody walks away from the $2-3trn of private FMV sitting outside the few names that can comfortably exit.

9. Is B2B software durable at all? PE’s existential question — and YC polices fake ARR

• Jason’s blow-up rule, delivered on PE: “whenever something looks incredibly easy and it looks like it always works… it’s about to blow up in your f*ing face. And it’s going to happen in venture too.” The tell was uniformity: “20 names all doing the same thing with exactly the same strategy, that was probably a clue.” The current playbook (Orlando Bravo’s, explicitly) is buying AI-enhanced B2B at scale — Jason has seen exactly one soft offer in his portfolio this year, fitting that template.
• Harry’s darker meta-question: if AI renders classic B2B non-durable — “doesn’t matter if you’re Legora or Medallia, none of it’s durable” — then “the whole classic PE model is broken.” Harry’s corollary: if AI-first startups can’t make it on equity alone, “they sure as hell can’t make it with debt on top.” Jason’s own scar tissue cuts the other way: he sold his last startup at ~$1m revenue just before the PE wave — “I never would have sold… if I’d known PE would come to the rescue” — and thinks big exits (Wiz, then Cursor: “I’m going to win the bet of a hundred billion dollar exit in the next year”) eventually cover the LP nut, even as “the other 96 companies wither off on the vine.”
• On Garry Tan’s revenue memo: fake ARR is “rampant” — Jason gets three different ARR definitions each month on one nine-figure-revenue investment, and “how can everybody get to 3 million in revenue by the end of demo day?” His pattern: “if it’s mostly disclosed, it’s been okay. If it’s been hidden, I ain’t going to make any money.” Harry calls the memo good and shrewd — like De Beers policing the diamond market, YC (25% of seed) is protecting trust in the market it makes, so “some version of it’s going to stick” as the “Y Combinator revenue guidelines” seal of approval.

10. Eternal assets, retail venture wrappers, and selling what you can’t defend

• Jason’s correction on Thrive Eternal vs Sequoia’s 2021 evergreen: Sequoia’s hold-forever compounding analysis was right over 10-20 years but launched “in that one year where it blew up in your face.” Thrive is a different bet — assets AI can’t replace (first investment likely the Giants): “a Toyota Corolla can drive faster than people, but we still watch the marathon.” Harry’s bear case, worth keeping: AI-personalized media consumption could “significantly maim digital rights” and thus team revenue; Jason concedes the margin (Messi monetizing Messi is value the owner never gets), while Rory flags relegation risk — Tottenham on the brink, Leicester down twice — and praises it: US leagues are antitrust-exempt “nasty little oligopolies” with “no penalty for failure, which is the definition of socialism.” Counter-example: Ryan Smith’s ~$1bn from Qualtrics is up roughly 4x in the Jazz.
• On Robinhood Ventures and the AngelList fund giving retail access to SpaceX, Anthropic and OpenAI: Rory put in the $500 minimum (“even as we speak, I’m adding the logos to our website”). On the 3.61% fee flap — hypocritical for VCs to scoff when a successful venture fund’s gross-to-net drag runs ~4% with carry; the real question is whether these assets still compound 15%+ from here, and Anthropic (10x from $60bn) is the case that they might. If the big three list at 3 or 4 billion dollars — nearly 5% of the S&P — a ~1% allocation is “kind of logically correct,” though “I’d be angsty about the valuations.”
• The fee-structure kicker: “Who made money from Medallia? Sequoia, baby” (it owned ~40%, basically bootstrapped). “Who makes money from Anthropic with a 17.5% carry and a 1% upfront fee? Goldman. Be Goldman or be Sequoia.” Related signal from Jason: when a top fund doesn’t elbow in for super pro-rata in your growth round, “it’s such a bad sign” — if they have billions to deploy, it goes into their winners.
• Harry’s confession: he said he was selling his Figma and Duolingo positions and specifically sold Figma 40% down. Rory’s lesson: “Don’t wait for the [expletive] to come up. Just sell it and redeploy” — and Jason’s view: “the agents don’t need either of them, Harry.” Rory’s discipline point: the best public investors have “an exit price on every position” (citing Brad at Altimeter) — “if you don’t know why you’re holding it, you shouldn’t be holding it” — which is exactly why venture investors, trained to work it out together with the company, “can be mediocre public investors.”