Anthropic Inference Costs Skyrocket |TikTok Deal Closes |The IPO Market:Wealthfront & EquipmentShare

Summary

• Brex sold to Capital One for $5.15B (50% cash, 50% stock) — a heroic zero-to-$5B outcome that still “feels weird” because of the $12B raise in 2021. Jason’s frame: this is the era of “hubristic financings” — you must promise you’ll own all of business finance to win the round, then pay a “weird one-day tax the day you exit.” Harry’s absolution: “the bad feelings last for a day and the 5 billion lasts forever.”
• The 7x revenue print is the real news for Ramp, marked at $32B (~30x). Rory: real money writing a check for the whole asset said “multiply by seven” — apply that to Ramp’s ~$1B run rate and even doubling for growth you get ~$15B, and Capital One now enters card spend with a structural cost advantage via Discover’s closed interchange network. “Not sure that was the best day out there for my stock.”
• Anthropic’s inference costs came in 23% above expectations, and the panel’s advice is to model inference UP, not down — Jason: if your ops team promised 30% inference savings, “I would throw my mouse at the monitor.” But gross margins went from -94% last year to ~+40% this year, so Jason’s boring middle take: there is real leverage, it just converges slower than hoped.
• Jason’s darkest call: unaffordable inference is “the final nail” for mid-tier SaaS — the $50–200M ARR company that dutifully got profitable and shipped a beloved agent now needs $50–100M of inference it can’t fund against rivals who “can raise $200 million on a dream” and treat inference as their sales-and-marketing budget. The only way out is an agent so good it replaces 20 people and earns $10–20K/month with “ROI measured in weeks.”
• TSMC is the bubble tell: skeptical, cynical, and now raising its capex budget for next year saying compute demand is “effectively infinite” — despite ~$600B capex against ~$100B of app revenue, “$500 billion a year in the hole.” Jason: “at some point, Nvidia puts will be a great buy… I ain’t buying them today.” Jason: for 99.9% of participants “there’s no upside in betting this is going to slow in the next 24 months.”
• OpenEvidence raised at $12B from Thrive and DST — a 12x step-up in under a year on ~$150M revenue — and the panel’s question is who catches the “Brex round”: this one, or the $30–40B round next year? Jason’s TAM haircut: direct-to-doctor pharma advertising is only $2–3B, not the headline $22B, so 3x from here needs ~$5B revenue at a 7x multiple. “You’re not going to find the discount here, people.”
• a16z’s claim that its companies generate 2/3 of private AI revenue is mostly one name — Rory: “to a rounding error,” the slide says “we have money in OpenAI and OpenAI is 40–50% of total revenue.” Against Gary Tan’s 10x-the-asset-class push, Rory invokes (likely) Martin Biggs — “no investing business so good that excess capital won’t ruin it” — and likely Druckenmiller: “most of the time we sit around here waiting, reading, and thinking.”
• The IPO bar is ~$3B and it’s time to flush the 2021 unicorns: EquipmentShare’s “effortless” 33% pop at $8B (47% growth, $4B revenue, profitable) vs Wealthfront “barely public” and down ~36%. Rory: winners are those who “got out from under that valuation via a down round, a down IPO or a down M&A” — with ~7,800 unicorns and ~200 IPO-viable, at one a week it takes four years.

Deep dive

1. Brex at $5.15B: heroic outcome, hubristic financing

• Harry sorts the discourse into three questions: is it a great outcome in the abstract (yes — “anyone who builds something from nothing to $5 billion… grow up, kids”), is it great versus the $12B raise in 2021, and is it great versus Ramp. On the second, Jason’s “hubristic financing” thesis: to compete today — Harveys, Loras and OpenEvidence going from $1B to $12B — “you’ve got to keep doing these” raises, promising, as Brex effectively did, that “it would own all of business finance.” On paper everyone but the late-stage investors did fine — roughly a billion-plus of equity and debt raised, no preference stack that ruined the deal.
• Harry refuses the moralizing: in 2021 you raised because you needed the capital, at market price — nobody takes $6B when offered $12B. In return “you pay this weird one-day tax the day you exit… the bad feelings last for a day and the 5 billion lasts forever. You’ll get over it.”
• Jason’s counter-anchor is Ali from Databricks: “I never wanted to raise more than two years ahead of the valuation I was confident I could hit.” Harry calls the 2021 logic a “sin of extrapolation” — at 200–300% growth the sentence “provided I keep the growth rate up for 2 years I’ll have grown into this” is perfectly innocuous, “but buried in it is a whole debt trap”: 2022 arrives, growth fades, capital gets scarce, you converge on profitability, growth fades more.
• The resolution, in Harry’s signature line: “things prove up in the end for what they really are, not what you delusionally think they are at one point in time.” A financial services company was always going to trade at a financial services multiple adjusted for growth — ~$700M revenue, ~7x, and Capital One said done. Even for late-stage funds paying up for hypergrowth, the math works: 30–40% of the book at 1x, no wipeouts, 3–6x on the winners still yields a fine sub-2x fund in a mediocre vintage like 2021.

2. The 7x print is aimed straight at Ramp’s $32B mark

• Rory’s three pieces of news for Ramp: operationally “you’ve won” (started later, ~$1B run rate vs Brex’s ~$700M — even the Ramp CEO’s congratulations post was classy, though “the Founders Fund guy did a straight dance on your grave”); but real money buying the whole asset — not 2% in a thin secondary — said multiply revenue by seven. Ramp at 7x is $7B; grant faster growth, maybe $15B. Against a $32B mark, “if I was the investor who just wrote the $32 billion check, I’d at least pause and go, hm, let me check my assumptions one last time.”
• The structural insight — “we live in this crazy land of VC valuations where they’re made once a year when only one person buys, no one can substantially sell” — plus the third headline: Capital One now enters with a structural cost advantage. These businesses monetize interchange on Visa/Mastercard’s third-party rails; Capital One bought Discover (~$35B at announcement, ~$50B at close, closed less than six months ago), whose closed network keeps the entire interchange. Fold Brex onto those rails and the asset is worth more to Cap One than standalone — “you’re going to be playing against the A team now.”
• Harry’s founder-seat version: Brex is now a comp with basically your revenue. “We could work for three more years to an IPO, suffer lots of dilution, stress, and basically economically be in the same place… so we better want it.” Rory’s summary of Ramp’s day: tally the three data points and “my impulse at the end of that is to tweet and say ‘well done.’ But deep in your soul you kind of go, hm, not sure that was the best day out there for my stock.”

3. TikTok at ~1x revenue: oligopolistic capitalism, and the dogs that didn’t bark

• The deal finally done — US investors own 80%, the algorithm stays with the Chinese owners — can’t be analyzed economically, Rory says; it’s a geopolitical divestiture with “a very directed purchaser program.” On the numbers: ~$15–16B of US revenue bought for roughly one times revenue, “wildly cheap compared to anything else,” caveated by an opex license fee back to the Chinese parent. His verdict on “the new oligopolistic capitalism that we now practice”: the most addictive app in America at 1x — “put me down for some. I wish I had some of that in my 401k.”
• Harry’s analogy is the Skype carve-out from eBay — a structurally weird deal (Silver Lake, per Rory’s memory, with a16z brought in) that 3x’d in 12–14 months. Which sharpens Rory’s one worry: why aren’t a16z, Sequoia and Lightspeed in this deal — the firms “just in the business of minting money now” — while Oracle and UAE sovereign wealth funds are? “It’s free money otherwise, right? We’re missing something.”

4. Anthropic’s inference bill: margins improving, but the break point never comes

• The headline — inference costs 23% higher than expected — lands two ways. Rory’s board-meeting anecdote: a B2B company with a powerful agent costing $100M debating “driving down inference costs in 2026” while facing six mega-funded competitors whose only differentiation is who has the best agent. “It doesn’t make sense.” He channels Amjad’s long-standing point: as soon as we figure out how to do this stuff, “we will actually burn an infinite amount of tokens if we can” — it even happened to Anthropic.
• Jason’s deliberately boring middle take: last year Anthropic ran a negative 94% gross margin; this year it’s ~positive 40%. That’s massive leverage — just slower than hoped. “We thought we’d be at 50, you’re at 40. It may take two years to get to 70, or you may asymptote out at 60.” He doesn’t doubt a free-cash-flow-positive business emerges; the open question is whether it converges at a 10% or 30% operating margin.
• Jason’s practical warning to founders, from building his own memory-everything Claude variant, REN (renchat.ai, as spoken), over the weekend — “if you want to save every chat you’ve ever had… it’s going to burn a lot of tokens”: model inference costs going up this year, not down. “If you walked into your board meeting and said inference costs are going down 30% because our IT team’s really good at managing costs, I would throw my mouse at the monitor.”
• Rory’s zoom-out: infinite demand for an expensive digital good is “a wonderful problem” versus no demand at all — the business challenge is metering it, viz. Anthropic’s $200 plan users burning $1,000 of tokens. And the cost structure has inverted: AWS at 10% of revenue used to cause board meltdowns; now inference runs 10–15% for a high-priced app and 50–70% of revenue for a coding app — “if you don’t manage that correctly you don’t have a business.”

5. The mid-tier squeeze: “I think it’s the final nail”

• Jason’s nightmare category: the 50–200M ARR B2B company that did everything right — got to breakeven because investors demanded it, finally shipped an agent its 10,000 customers love — and now needs $20M+ more inference this year at $2–2.50 per interaction across 50 million interactions, against DecaGon or Lora funded to burn. “You told me to get break even. I did that. Now I built it. How am I going to fund the 50 million of inference?” Rory concedes the point is exactly right and unresolved: “rather than telling you I have the answer, it’s an issue I’m wrestling with — I can think of two boardrooms in the next two months.”
• Rory’s gut-level test is pure capitalism: if customers get value they can’t get elsewhere, you can charge enough to pay for your tokens; “if you’re locked in a war with someone who has infinite capital and is willing to give it away longer than you, you’re probably going to lose and you should figure out how to exit now” — the mid-year coding wars, where likely Windsurf looked to exit, had some of that dynamic. Jason’s way out is narrower: an agent so epically good it replaces 20 people and commands $10–20K a month — “not pretend that good… literally so good that the ROI is measured in weeks.”
• The asymmetry that makes it brutal, per Jason: for the new entrants, “inference is your sales and marketing team in essence” — Harvey at ~$200M last year, OpenEvidence ~$100M, some AI leaders with four salespeople or hiring their first marketer at $200M revenue — while incumbents already carry massive traditional S&M. Even Salesforce, with its resources: “talk to folks at Salesforce today, they’ll tell you this is the most stressful time they’ve ever worked there.” Hence Jason’s escalation on the is-SaaS-dead debate: “I worry this is the next final act… you did all the right things, and the final nail in the coffin is we just can’t afford the inference.” He recalls Canva’s Cliff making the same point — “I could build Gamma, but I can’t burn the way Gamma burns those tokens.”

6. TSMC is the tell — and the Nvidia-puts discipline

• Rory discounts what model-company CEOs say about demand (“they’re talking their book”) and even pities Oracle chasing it, but trusts TSMC — sharp, cynical, the people who mocked Altman’s trillion-dollar talk a year ago (“go away, AI boy”). Their earnings call just said compute demand is “effectively infinite right now” and they’re raising their capex budget for next year (after a prior peak of $40B and a $22B trough two years ago). The logic: “you can cut employees, you can turn off your GPU, but if you dig a big deep hole in the ground in Phoenix… and no one uses it, you’re out 20 billion bucks. They’re leaning in.”
• On the bubble question Jason splits grandiosity from the marginal call: you don’t need Dwarkesh-podcast trillion-dollar data centers or mass unemployment to believe investment continues on trajectory for 12 months — even with capex at ~$600B against ~$100B of app revenue, “you’re still $500 billion a year in the hole.” His tangible discipline: “at some point, Nvidia puts will be a great buy, cuz every semiconductor cycle for the last 40 years has ended up in a massive downswing. I ain’t buying them today… if you knew when it was going to happen to the day, you’d be trading Nvidia puts and you wouldn’t be talking to us.”
• Jason agrees the pop comes “someday” — even if it’s after SpaceX IPOs and Elon’s thousand data centers in space — but recalls December 2021, when 90% of tech thought it would last longer “and then bam… HashiCorp went public and then it just stopped for 2 years.” Still: “for 99.9% of us there’s no upside in betting this is going to slow in the next 24 months.” Rory’s only actionable advice: scenario-plan, and raise while money is cheap — “going back to Brex, you’re really glad you raised that money in 2021.”

7. OpenEvidence at $12B: who does the Brex round?

• The raise — led by Thrive and DST, a 12x step-up from the ~$1B likely Sequoia round at the start of the year, on revenue Jason believes is ~$150M — draws no quality quibbles: commanding share of doctors’ technical-question mindshare (vs Doximity, the pre-AI-generation comp), Journal of New England Medicine relationships, medical-professional-only access, HIPPA compliance, and a founder who sold likely Kensho to S&P. Harry: “you’re not going to find the discount here, people.”
• The TAM is where Jason cuts: headline pharma media spend of $22–30B is roughly half consumer TV ads; actual doctor-directed advertising is a $2–3B market, shrinking slightly, with much of the rest spent on reps bringing donuts and sample packs. To 3x from $12B you need ~$5B revenue at the 7x multiple Brex just validated — so OpenEvidence must either pull rep budgets online (“totally credible”) or expand into adjacent doctor products, the way Doximity added scheduling with a non-personal phone number. “It’s credible they do it, but they got to do it.”
• Jason closes the loop to hubris: “this is the back to hubristic fundraising… who does the $12 billion Brex round here — is this that round, or is it the round in March at 30?” His guess: Thrive has done the math and this is the right insertion point, and someone else does 30–40 next year. Jason notes the multiple hasn’t actually expanded — they 10x’d revenue as the valuation 10x’d, and every prior round (3, then 6) looked like “the one going too far.” Harry’s confession makes it human: he had Henrique and Pedro on in 2020–21 talking Amex fragility and a $100B business — “12 billion did not seem that crazy.” Rory: “Hubris is like that… the good stuff is so good, it’s so easy to walk into the partners meeting and advocate.”

8. a16z’s 2/3 claim, deflated — and is venture finally an asset class?

• Harry’s astonishing stat from a16z’s report — $8B invested in 2025, and 2/3 of private AI revenue generated by a16z-backed companies (OpenAI, Databricks, Cursor, Harvey, Replit…) — gets Rory’s cold water: add up AI revenue and “you’re going to get 13 billion for OpenAI, 4 billion for Anthropic, and everything else is in the noise… 200 million for Harvey, whoop-de-doo.” So “to a rounding error, another way of saying the same slide is: we have money in OpenAI, and OpenAI is 40–50% of total revenue.”
• Harry takes the report seriously anyway, alongside Gary Tan’s claim that venture should be 10x bigger: the classic knock is that venture isn’t an asset class because the bottom 75% “isn’t worth getting out of bed for” — but if a16z’s penetration is repeatable like clockwork and YC industrializes the low end, maybe it finally is one. The a16z aha, in his math: at ~18% of last year’s funding, they must be in ~10% of the good deals, and the early-stage bundle lets them deploy 3–4x more late-stage — “they’ve structurally figured out a way to make that happen.”
• Rory grants “two asset classes” — traditional early-stage, plus late-stage venture “that used to be called small cap growth and is now privately held” — but rejects the more-capital conclusion, citing Martin Biggs: “there’s no investing business so good that excess capital won’t ruin it.” His confessional on activity: 2025 was the most active year since 2021, “in retrospect, I wish I’d just gone home” — everything from 2021 was either priced wrong at 1x or just wrong at less. The likely Druckenmiller line stuck with him all week: “most of the time we sit around here waiting, reading, and thinking.” “That’s a real investor — activity is not everything.”

9. Succession: the graveyards are full of indispensable men

• Jason’s glum scenario — what if Mark or Ben steps down? “Even health scares happen… Dustin Moskovitz quits Asana out of the blue. You just don’t know.” Rory asserts vigorously it survives: “the graveyards are full of those indispensable men” — then recites “Kleiner, Perkins, Caufield, and Byers,” firms that outlived every name on the door.
• His deeper point inverts the intuition: succession is harder for a Benchmark, “small and brilliant,” where the asset is four or five individual brains. a16z’s fundamental bet is that venture goes the way of investment banking — from individuals and small partnerships to institutions (“Mr. Goldman, Mr. Sachs, Mr. Salomon Brothers”) — and the cynical incentive check: Kravis and Roberts retire comfortably on KKR’s management fees “provided those fine forty-something Ivy League graduates keep it on the straight and narrow.” Whoever builds the public-market equivalent has every incentive to manage the transition.
• The Khosla coda, kept for the exchange: Jason isn’t sure Vinod even wants a generational fund; Rory — “I thought Vinod’s plan was not to die. So none of this matters.” Jason: “It’s a good plan… either for real or in the GPUs, one way or another.”

10. The IPO ledger: EquipmentShare effortless, Wealthfront broken, Ethos as capitulation — and SaaS gets an army

• The bar is now legible. EquipmentShare — tech-enabled construction equipment rental, a 2015 YC company (“YC and Lead Edge both made a lot of money on that one”) — popped 33% to an $8B cap, growing 47% at $4B revenue and profitable: per Jason, “an effortless IPO… oversubscribed, you trade up, no drama. This is what an IPO is supposed to be.” Wealthfront is the anti-case — a genuinely admired company but “a deeply broken IPO,” down ~36% to ~$1.3B and “barely public”: no liquidity, no analysts, “a long haul for everybody to get their money out.” Rory’s rule: ~$3B plus or minus is where going public gets easy; slip below and “you’re down into who-cares land, which sucks” — though he likes Wealthfront enough to note “a mental note to go check on it and maybe buy some.”
• The talent disagreement is worth keeping. Jason, from recent calls with such companies: “people are just blinking at the camera. They joined these companies to not work” — one resisted shipping a big release this year — so you need a driven, charismatic CEO on a mission who can recruit two or three needle-moving leaders (“failed founders — the hottest recruiting category in tech right now”). Rory’s pushback: outcomes follow a power law but humans follow a bell curve — “the idea that all the great people are in a great company and everyone in the okay companies is mediocre… I think that’s an overgeneralization.”
• Ethos (life insurance) prices at $1.3B against a $2.7B private peak; Harry asks if that’s “being fed to the dogs.” Rory turns it around — “what’s your alternative plan for this company if it does not go public?… price clears all markets” — venture cost of capital runs ~30% versus ~11% public, so there comes a point you’re better off out. Jason predicts a drop but reads it as healthy capitulation, and asks whether it’s “time to flush our 2021 unicorns out the door in 2026 and 2027.” Rory, catching him sneering at the very outcomes he defended on Brex: the winners “got a crazy valuation in 2021 and then got out from under that valuation via a down round, a down IPO or a down M&A” — and with ~7,800 unicorns, of which SVB’s analysis says ~30% (~200) have IPO-viable growth and scale, “at one a week it’ll take four years.” Jason’s kicker from an M&A review at a unicorn he advises: “everyone’s for sale — I was shocked at folks that have been on 20VC that I did not know were in market, aggressively looking for an exit.”
• The Benioff hat-tip closes it: the Army awarded Salesforce a $5.6B contract over 10 years — “SaaS is not dead. And now SaaS has an army.” Rory says such deals bury the claim that “you’re going to vibe-code your way to a product that can replace a $500 million Army order,” but Jamin’s inventory of daggers tempers Harry’s bull case on Salesforce: seat contractions are “existential” (Workday: seats “perpetually under pressure”; Shopify flat headcount for 3 years while growing 40%+), and SaaS price increases of ~40% over 3–4 years “crowd everything out” — a clever pricing model can’t force customers to pay more than they want to. “Slack never raised prices and still grew at 140% NRR. I don’t know those days are ever coming back, no matter how good our agents are.” Rory’s landing spot: “you end up in that boring quadrant of it ain’t going away, but it ain’t exploding.”