Databricks at $100BN, CoreWeave’s $11B Debt Bet & Nubank’s $2.5B Profit Shocker - Ep.19
Summary
Databricks at $100 billion looks less like private-market fantasy than a concentrated bet on growth persistence. Databricks and Snowflake are both around a $4 billion run rate, but the panel put their growth at 50% and 26%, respectively, making Databricks look relatively inexpensive even at 25x revenue. The catch is duration: “If you get two more years at 50, 60%, you’re golden”; if growth falls abruptly rather than stepping through 50%, 40%, 35%, then the 2021-style valuation trap reappears.
The next IPO wave could release enormous venture returns without preserving today’s scarcity premiums. Figma was cited at roughly $1 billion of revenue, 48% growth and 40x ARR; a hypothetical Canva at $4 billion growing 40%, plus Databricks and Stripe, would bring even stronger companies into public markets. Rory expects appropriate high-growth-adjusted valuations but fewer 3x pops once supply expands; the larger consequence is potentially moving $1 trillion of equity from private balance sheets into a $40-50 trillion public market.
Chamath Palihapitiya’s SPAC revival is a bubble indicator because the structure rewards completing a transaction, not necessarily selecting a good one. Rory would not declare a peak before prices turn, but argued that a promoter receiving substantial economics when a deal closes creates predictable adverse selection: “If you pay people to do deals, deals get done regardless of the quality.” The prospectus line that “there can be no crying in the casino” drew sharper pushback because Wall Street’s purpose is allocating savings into productive investment, not operating a negative-sum casino.
OpenAI’s $6 billion staff secondary is becoming necessary compensation infrastructure, not merely an extravagant payout. When Meta can confront an employee with a real $10 million or $100 million offer, mission language cannot substitute indefinitely for liquidity: “At some point you got to match the dollars.” These private companies now resemble public companies in scale, while their employees face a 15-year marathon rather than a four-year sprint; regular sales of vested shares would weaken old retention handcuffs but better match that reality.
Nubank’s quarter shows what happens when a fintech replaces the incumbent bank rather than merely skimming one profitable product. The company reported $2.5 billion of net income, up 42%, with 123 million customers and an approximately $60-63 billion valuation; Rory’s punch line was that “Nubank is old bank,” spanning deposits and lending across Brazil, Colombia and Mexico. A further 2x looked conceivable if it avoids regulation, bad credit and “dumb bank stuff,” while Harry argued its limited current product set leaves room for a much larger $300-400 billion outcome.
CoreWeave’s $11.2 billion debt is manageable only to the extent that its financing and customer commitments are genuinely matched. Against a roughly $22 billion capex plan, large borrowings are inherent to its role as the financing vehicle for AI data centers; seven-year debt funding an ironclad seven-year Microsoft take-or-pay contract is structurally sound. The warning signal would be customers paying rather than taking capacity, new facilities no longer receiving contracts, or weaker imitators losing liquidity: CoreWeave is “the canary in the coal mine.”
Sam Altman’s “trillions” of infrastructure spending is directionally plausible but may be a metaphor until AI captures labor budgets. Jason expects AI usage, tokens and compute to rise roughly 10x, while four major companies were said to be spending about $365 billion on infrastructure; Rory countered that this is colliding with the limits of corporate finance, including Microsoft capex near 25% of revenue. Trillions become economically rational only if spending shifts from technology budgets into human-labor budgets, potentially unlocking $10 trillion of value.
AI-agent adoption can support the boom, but fragmented six-figure products are already setting up an unusually fast consolidation cycle. Jason’s team uses 10 production agents that replaced five people, yet their combined list price is already about $500,000 and could reach $1 million; “everyone thinks their thing is going to save labor, but you can’t all get credit for the same labor.” Buyers will favor suites spanning multiple workflows, while investors are simultaneously overlooking classic platform risks because “the only thing between us and Armageddon is AI adoption.”
Deep dive
1. Databricks earns its valuation only if extraordinary growth persists
The first surprise was emotional rather than financial: a $100 billion private company barely registers beside Anthropic at $170 billion, SpaceX at $360 billion and OpenAI at $500 billion. Five years ago, the number would have been inconceivable; today the response is “yeah, like whatever.”
The decisive comparison is Snowflake: both companies are around a $4 billion run rate, but Snowflake’s cited growth is 26% against Databricks’ 50%. Harry therefore called 25x revenue “very reasonably priced,” particularly for an infrastructure provider sitting directly beneath AI demand.
Rory’s underwriting rule converts the startling multiple into a conventional terminal-value question. If Databricks grows for several years and reaches $10-12 billion of ARR or GAAP revenue, normalized multiples can make the investment “money good”; if growth suddenly collapses, private investors cannot use a public-market stop loss and will own it throughout the decline.
Databricks CRO Ron Grarisco told Harry the company was technologically “five years” ahead of Snowflake. The panel treated that as striking confidence, while keeping the valuation argument anchored to observable growth rather than the executive’s claim.
2. An IPO flood could enrich venture funds while ending scarcity pricing
Jason’s case for more froth: the year’s IPOs—including CoreWeave, Circle, Hinge Health, Rubrik, Klaviyo and Figma—are not yet the “epic ones.” Databricks, Canva and perhaps an accelerating Stripe could create “a bubble on top of a bubble” through public gains and LP distributions.
Rory’s counterargument is scarcity value. Figma can receive an extraordinary pop when few high-growth public companies exist; once Canva and Databricks join it, each may trade at an excellent growth-adjusted multiple without tripling immediately.
The transfer remains historic even without speculative premiums: perhaps $1 trillion could migrate from private portfolios into public markets that can digest it. Some companies may be worth $50 billion, several more than $100 billion and perhaps one $1 trillion—bizarre only about seven years after Apple first reached that level publicly.
Fund economics become extreme under the power law. The panel modeled a hypothetical 15% Databricks stake at a $200 billion IPO as $30 billion of proceeds; even an accumulated $500 million-$1.5 billion investment would produce a remarkable fund-level result, while an $80 billion IPO would perversely feel disappointing.
3. SPAC economics make the promoter’s transaction the product
Harry called Chamath’s return to SPACs “peak bubble,” but Rory refused the timing claim: “You won’t know it’s peaked till it goes down the other side.” If the anticipated IPO boom continues, Chamath may simply be selling into strength again.
The structural objection was categorical. A SPAC sponsor receives meaningful economics once a merger closes, so incentives favor finding any deal; the rough example was Harry’s: a $5 million upfront risk could produce perhaps $50 million when the transaction completes.
Jason wondered why someone already extremely wealthy would accept the reputational drama for that payoff: “The juice doesn’t seem worth the squeeze.” He placed SPACs “at the edge of grift,” even while acknowledging the unusually candid disclosure.
Rory’s response to “no crying in the casino” was moral rather than cosmetic: capital markets direct ordinary people’s savings toward long-term enterprises. “I personally would never let anyone manage my money who thought it was a casino.”
4. Retail access to private assets creates a disclosure-versus-protection dilemma
The panel’s concrete warning involved a private-equity vehicle buying a secondary portfolio of 50 positions at a 20% discount, then marking it up the next day. A retail holder could perceive a genuine 20% investment gain while the explanation sat behind tiny footnotes, asterisks and daggers.
Rory framed the regulatory problem as two imperfect philosophies: disclose everything and impose “buyer beware,” or accept that individuals cannot spend professional-level time decoding every instrument and protect them. Harry’s rule of thumb was: “Whenever a bank is explaining something complex, you’re losing money.”
Yet blanket protection also excluded ordinary investors from rounds like Databricks. Long-term private-equity and venture returns can be good, unlike a casino’s negative median expectation, so the challenge is allowing informed access without sending the weakest assets and most confusing fee structures to retail.
5. OpenAI’s secondary turns a private giant into a costly pseudo-public market
Rory’s reaction to the $6 billion employee sale was “Good for them.” In the face of Meta’s alleged $10 million-$100 million cash offers, that pool represents only 600 employees receiving $10 million or 60 receiving $100 million—astonishingly small relative to the competitive stakes.
His behavioral rule came from acquisitions: “What people say they’ll do when faced with a large amount of money is meaningless.” An abstract commitment to mission changes when a life-altering offer reaches the kitchen table and must be discussed with a spouse.
Jason proposed making late-stage liquidity nearly automatic: after the cliff, employees might sell each newly vested fraction monthly, approximating public-market compensation. Rory agreed that recreating a public market privately is expensive and cumbersome, but liquidity must be one of its features.
Wealth will not uniformly destroy productivity. Jason expects money to reveal preferences: people who love the work can remove daily distractions and continue harder, while someone whose real dream is teaching high school can leave.
6. Fifteen-year companies cannot retain people with four-year handcuffs
Traditional venture equity used illiquidity as a retention weapon: departing employees might be unable to exercise options, returning stock to the pool. That logic fit a four-year sprint toward an IPO at $50 million of revenue, not a company remaining private for 12-15 years.
Harry cited Hopin’s founder leaving after the company had raised roughly $130 million. Rory dismissed the episode as an ordinary venture outcome rather than a scandal: “It’s only a marathon if you finish the marathon.”
Jason expects 10-20 nine-figure founder outcomes resembling Hopin across an exuberant cycle. Rory supplied the base rates: perhaps 30% of Series A/B deals fail, more than 50% at seed, and Harry recalled the loss rate spiking to 60% in 2001-02; a heavily funded company unraveling therefore “didn’t even rise to the level of interesting.”
7. Nubank won by building the bank incumbents failed to build
Nubank’s reported $2.5 billion net income, 42% annual growth and 123 million customers prompted Rory to compare three geographically distinct attacks on banking: Nubank in Latin America, Revolut in Europe and Chime in the United States.
Latin American incumbents offered the weakest service and richest opening. Nubank went beyond deposits into lending and the “full monte” of financial services, becoming the largest or second-largest regional bank by market capitalization despite holding perhaps one-fifth the assets.
Chime found a narrower US opportunity because incumbents were better run. The Durbin Amendment supported debit-interchange economics, allowing Chime to become profitable without lending, but its roughly $11 billion value remained far below Nubank’s despite the larger US economy.
Rory could envision Nubank doubling if it expands beyond three countries while avoiding “dumb bank stuff.” Harry pushed much further, arguing that matching Revolut’s product breadth could support $300-400 billion; the unresolved risk is that a mature Nubank eventually trades on book value like a traditional bank.
8. Revolut turns a profitable wedge into a primary financial relationship
Revolut began where banks were “robbing you blind”: cross-border transfers. Its product worked immediately, produced attractive unit economics and consumer affection, then expanded into everyday banking, FX and crypto until the transfer account became “my account for money.”
Revolut’s snack analogy was to entice users with the FX “snack,” let the snack become the meal, then the entire meal. Harry cited roughly 35% of Revolut users now treating it as their primary account, versus about 50% for Nubank.
Internet distribution changed banking segmentation from geography to demographics or product propensity. A physical UK or German bank had to serve young people, grandparents and local businesses together; Revolut could isolate people moving money across borders and compound inside that cohort.
Harry highlighted 26 new products being run in an internal venture lab with weekly metrics. Jason saw the ability to recruit double-digit numbers of founders and CEOs as product-line GMs as a genuine scaling “superpower,” also visible at companies such as Rippling.
9. On is exceptional precisely because its success is hard to systematize
On was cited at $4 billion of sales, 38% annual growth, 61.5% margins and a $15 billion valuation after founding in 2010 and listing in 2021. Rory readily called it amazing but rejected the inference that venture firms should build a repeatable consumer-goods strategy around it.
Fintech had a knowable enabling trend—digitization plus weak banking infrastructure—allowing investors to anticipate that some Chime or Revolut would emerge. High-end running shoes lack the same broad causal tailwind; On is therefore a “category of one,” not proof of a scalable portfolio thesis.
10. CoreWeave is a leveraged real-estate company financing the AI buildout
CoreWeave’s cited presentation paired approximately $4-5 billion of annual revenue with an $800-900 million loss and a $22 billion capex plan. Rory’s answer to $11.2 billion of debt was “duh”: once the company commits to that construction budget, borrowing is intrinsic to the model.
His preferred description was a sophisticated real-estate company—borrow money, build data centers and sign long leases—serving as a financing vehicle for the GPU appetite of major AI companies. “CoreWeave borrowed $5 billion in debt. It won’t be the last time.”
Matching determines survival. Seven-year financing against a binding seven-year take-or-pay contract can work; short-lived inference demand against a decade-long lease recreates the classic banking mismatch. Customers may also fight obligations when capacity is unwanted, regardless of how secure the disclosure appears.
Jason watches CoreWeave as the earliest public signal of stress: a customer paying rather than taking capacity, or a new center failing to secure another contract, would matter before weakness reached Microsoft. Rory assigned at least a one-third probability that demand slows enough within 12-24 months to compress extrapolative valuations.
11. The stock market is already a concentrated AI-capex wager
CoreWeave had fallen roughly 50% from its June 20 peak of $183 even while the AI narrative strengthened. That volatility complicates the easy instruction to “bet on CoreWeave”; broad QQQ or VTI exposure already owns Microsoft, Nvidia, Alphabet and Meta.
Nvidia and Microsoft together were estimated at 15.8% of QQQ, with the top five or seven companies around 35-36%, though the speakers were not sure whether the latter figure referred to the top five or top seven. Jason’s conclusion: “We’re all living in the bubble”—not only venture investors, but anyone whose retirement account tracks the dominant indices.
The long tail of US equities is not carrying those gains. That makes the market’s aggregate performance unusually dependent on the same infrastructure thesis underwriting Databricks, CoreWeave and the application layer.
12. “Trillions” becomes financeable only if AI captures labor spending
Sam Altman simultaneously said OpenAI would spend trillions on infrastructure and that AI contains a bubble in which many people will lose money. Rory admired the rhetorical preemption: acknowledge economists, losses and excess, then effectively say, “just give me a trillion.”
Jason treated the figure as thoughtful. He said AI use could rise 10x, with tokens and compute also rising roughly 10x; in a later thought experiment, 20x daily usage and 10x tokens would imply 200x demand under simplifying assumptions. Amazon, Alphabet, Microsoft and Meta were said to be spending roughly $365 billion on infrastructure.
Rory called the direction credible but “trillion as metaphor for lots.” Total US corporate investment was framed at roughly $2 trillion annually, half in buildings and half in equipment; Microsoft already directs about 25% of revenue to capex, while Meta has begun borrowing.
Jason supplied the condition that reconciles the arithmetic: AI must move from technology budgets into human-labor budgets, potentially unlocking $10 trillion of value. Without measurable labor replacement, trillions of investment do not pencil; with it, they may become rational over a long horizon.
13. Labor replacement is visible, but adoption pace determines the return
Jason had been skeptical but changed his mind after deploying 10 production AI agents that replaced five people on his small team. A recent stand-up involved only one person alongside the agents: “It’s early, but this wave may accelerate.”
Within six months, he thinks CIOs and CEOs might demand an AI-first answer before approving human headcount. Rory also moved from skepticism after seeing Shopify’s numbers and Jason’s automation, but distinguished fast-moving edge adopters from large enterprises with long implementation cycles.
The investment question is therefore not merely whether the labor transition happens, but when. At roughly $400 billion of annual infrastructure spending, “the time value of money is going to eat your ass at 4%” if enterprise deployment trails capacity by years.
14. AI software will consolidate before platform risk disappears
Jason’s 10 agents carry about $500,000 of aggregate list pricing, potentially $1 million by year-end. Many vendors begin around $60,000-$100,000 annually plus onboarding costs, creating a budget collision even though AI remains the one category CIOs will fund.
Rory’s synthesis: “You can’t all get credit for the same labor.” Rippling-like suites offering five agents plus orchestration can absorb point solutions rapidly; the mythical one-person billion-dollar company still faces competitors eager to claim its excess economics.
Vertical software may benefit when a dominant solution uses an AI wedge to broaden. A voice bot must progress from answering calls to bookings and refunds; the robotics analogy was Locus, whose warehouse picking automation mattered because it attacked a major labor pool, unlike automating two stations among 300 workers.
Abridge versus Epic captures the platform danger. Rory said anyone investing in Abridge without expecting Epic to launch its own product would be naive; yet “there’s only one thing worse than partnering with Epic and getting smacked around,” he said, “and that’s not partnering with Epic,” because partnership provides access to roughly 40% of the market. The bet is that product quality and clinician preference can overcome bundling. Growth is now so compelling that investors routinely ignore risks that would have killed deals in 2019.
15. The quickfire bets favored economics over headline certainty
On whether Anthropic releases Claude 5 this year, Rory liked the payout—$100 returning $322 for yes—despite viewing the event itself as perhaps below 50%. Jason said “no chance”: Anthropic can prioritize an excellent next Sonnet, such as Sonnet 4.5, because coding revenue matters more than a consumer-app spectacle.
On a Mistral acquisition, yes returned $420 versus $113 for no. Jason argued that any sub-$100 million-revenue lab offered $20 billion should take it while such transactions remain possible; Rory agreed on acceptance but questioned whether any buyer would actually make the offer.
On Deel versus Rippling listing first, the quoted odds favored Rippling, but the panel leaned toward Deel because it says it is profitable and can access liquidity now. Rippling remains in ambitious investment mode and may need up to two more years—unless Parker Conrad chooses to de-risk early with public cash and a tradable stock.