20VC: Trillion or Bust, CoreWeave’s Sell-Off and the $90B Anduril Test

Summary

• Mega-fund return targets are narrowing what gets funded while making seed ownership harder to monetize. Harry cites a profitable $20M ARR company growing 600–700% that attracts little interest; Jason’s explanation is “billion-dollar positions or nothing” for large funds. Meanwhile, $50–100M seed valuations and $300–500K allocations can leave even an $18B winner immaterial to a fund. Rory’s objection: an industry historically producing 10–20 decacorns per decade cannot support everyone pursuing only enormous outcomes.
• CoreWeave’s sell-off tests whether AI demand can outrun infrastructure spending, rather than whether demand is growing at all. Rory describes a leveraged business that must keep filling capacity while financing the next buildout; Harry’s proposed “generational buying opportunity” therefore requires more than confidence in AI adoption. Jason prefers Nvidia’s differentiated assets, but Rory challenges that hierarchy with the returns he cites: Nvidia up 18% versus Micron up 350%.
• Anthropic’s reported slowdown leaves room for a huge valuation, but its revenue deserves less predictability than conventional SaaS. Harry cites $16.8B of Q3 revenue, 45% sequential growth, an 85% gross margin and adjusted profitability, alongside a rumored IPO valuation falling from $2T to $1.5T. Rory can construct a roughly 12–13x forward-revenue case; Jason warns developers switch providers and Anthropic could lose market share while still growing substantially. “There’s only so much ARR in the ARR here.”
• The panel’s strongest software call is that production readiness has become more important than generating another feature. Jason argues models get developers most of the way there, while testing and fixing can consume the next three days. He broadens “code review” into “build to prod,” encompassing verification, testing and deployment. Rory agrees these are compelling markets, while warning that their exceptional growth comes with repeated foundation-model competition.
• Modal’s $750M raise at $15.75B exposes an inference opportunity the panel admits it underestimated. Its run rate exceeds $1B, against Rory’s recollection of less than $100M around mid-to-late last year. He contrasts roughly $5–6B across leading independent inference providers with more than $100–110B at OpenAI and Anthropic combined. The opportunity is cheaper, sufficiently capable tokens; the unresolved question is whether competition eventually overwhelms today’s growth and improving margins.
• Town at $1B and Anduril at $90B force a distinction between exceptional companies and attractive entry prices. Jason calls Town an early venture bet with limited data, dressed in a billion-dollar valuation; Anduril’s roughly 40x revenue demands substantial execution. Rory agrees growth might justify the prices, but expects little help from broader public-market multiple expansion. His warning is straightforward: “amazing companies just priced wrong.”
• Cash-generating software can reward founders even when it fails the venture industry’s outcome requirements. Jason revises his former enthusiasm for raising one round and stopping: “none and done” may preserve more freedom to distribute profits. The panel questions the exact $370M profit claim for 37signals but accepts the broader wealth-creation point. Bending Spoons offers another ownership model, although Rory challenges whether software assets can endure like Berkshire’s See’s Candy.
• A venture firm IPO creates an obligation to keep growing assets, while returning LP liquidity could fuel another funding cycle. Rory estimates roughly $800M–$1B of annual fee income is needed for a meaningful public listing and warns that the structure can change investment culture. Harry expects SpaceX distributions and prospective Anthropic/OpenAI IPOs to recycle money into venture. Harvard’s reported 27.1% return illustrates the rewards—and resulting FOMO—of exposure to the few dominant private assets.

Deep dive

1. Billion-dollar positions are narrowing the venture market

• Harry’s opening puzzle is a profitable company at $20M ARR, growing 600–700%, that encounters “crickets” from VCs. Jason describes a comparable agentic company planning to stop after a $3M seed. Rory accepts that strong businesses struggle, but challenges the diagnosis: investors also fund previously implausible projects, including data centers in space and quantum computing. Capital is willing to take risk when the perceived upside is enormous.
• Jason’s explanation is fund arithmetic: firms raising $5–10B every 18 months need billion-dollar positions, sometimes much larger ones. A $2B exit cannot deliver that without unusually high ownership. “It’s billion-dollar positions or nothing.” Rory objects that the industry historically produced 10–20 decacorns per decade, whereas current expectations appear to assume 90 this year; there cannot be enough giant outcomes for everyone.
• For merely excellent companies, Jason’s prescription is to meet 100 investors. A less fashionable growth fund may fund the best opportunity it has seen that quarter, even when the marquee partnerships decline. What changed is that top-decile metrics once practically guaranteed a term sheet; now founders must search much more widely.
• Seed reverses Rory’s initial intuition: Jason says $50–100M entry valuations, lower ownership and lost pro rata make 5x net harder. Harry’s $300–500K allocation in a company reaching $18B still “doesn’t move the fucking needle.” Rory agrees the next five years look inverted, citing SpaceX, Anthropic and OpenAI rising from roughly $400B combined to $4–5T over three years.

2. CoreWeave needs demand to outrun its next construction bill

• Rory reports a roughly 15–20% decline the previous week; bond prices rose, implying lower default risk. Neoclouds amplify changes in sentiment because they underwrite rapid growth with large capital expenditures and debt. Recalling figures from memory, he sketches something like a $7B runway, $10–20B of CapEx and $35–40B of debt. Grow too slowly and customers are unhappy; build too quickly before demand softens and there is no software-style option to hunker down.
• Harry, a shareholder, asks whether a roughly 25% decline creates a “generational buying opportunity.” Rory’s qualification matters: filling existing capacity is insufficient when another $20B of spending requires demand to rise substantially every year. He would track OpenAI and Anthropic’s combined growth to remove market-share transfers, then compare that demand trajectory with CoreWeave’s investment plans.
• Jason calls CoreWeave and Oracle “B-tier” relative to Nvidia. CoreWeave, he says, depends on Nvidia and lacks unique silicon and software; Oracle cannot sell enough enterprise software and is renting excess capacity. He calls them more sensitive to shocks, while saying B-tier assets “should trade at a premium.” Rory rejects quality ranking as a complete investment rule, citing Nvidia up 18%, Nebius up 50% and Micron up 350% over the year he references. Yet when Jason asks about putting 20% of personal assets into those riskier names, Rory admits: “I wouldn’t dare.”

3. Anthropic can slow sharply without becoming a weak business

• Harry reports Q3 revenue of $16.8B, up 45% from Q2, an 85% gross margin and adjusted profitability; the rumored IPO valuation shifts from $2T to $1.5T. Rory sees deceleration from an unsustainable starting point after three years of 10x growth. His illustrative continuation of roughly $5B in quarterly revenue additions produces around $105–110B the following year and a 12–13x forward-revenue valuation. “I’m not saying they will.”
• Jason’s concern is post-IPO volatility: “developers are not loyal,” and provider shares have moved dramatically. Anthropic’s revenue market share could be down 10% from where it is today by the end of next year while its absolute business remains excellent. “There’s only so much ARR in the ARR here.” Rory adds the next risk: companies defending share through price cuts, especially if aggregate end-user demand slows.
• On OpenAI’s supposed $1.4T financing, Rory declines Harry’s strategic-anchoring interpretation. “When you need money, you take money.” His simpler explanation is compute spending and an unfinished IPO: raise above the last round, but leave enough upside for new investors. Jason relays Chamath’s suggestion to raise what Anthropic needs at $1.5T and move on, noting that a $2T price could stress its capital needs.

4. Application routers reveal what customers will pay for intelligence

• Jason dismisses the significance attached to Replit switching its default to OpenAI. Replit already routes among models; his explanation is comparable output at a 30–40% better price. Unlike users manually configuring routing, Replit, Lovable, Base44 and v0 have a strong incentive to optimize quality against cost, because they cannot simply pass every expense through to customers.
• Those applications are useful laboratories, Jason argues, because SMB users seeking a roughly $25-a-month product and enterprise customers alike are cost-sensitive. He says the default could change again, mentioning Sonnet 5.5 and Haiku 5.5. His narrower inference about Kimi is that Replit is not seeing sufficient benefits to route everything through it. Rory agrees: at roughly billion-dollar revenue scale, these applications can run an economically meaningful comparison of intelligence against price.

5. “Build to prod” is becoming the critical software bottleneck

• Claude Code Review threatens specialists because coding sits at Anthropic’s core, Jason says, but independent products retain benefits—including having one model check another’s work. He mentions Greptile’s $75M raise and CodeRabbit as examples in the category. His example captures the remaining workload: an engineer builds Harry’s feature from a phone during dinner, then spends “the whole next three fucking days testing it and fixing it.” Generating the feature does not finish the job; Jason says review and processing can take an order of magnitude longer than building.
• Rory’s “Icarus moment” framing preserves the trade-off: the fastest-growing applications sit closest to foundation models precisely because those models unlock their capabilities. Coding, code review and legal startups therefore face a fresh competitive scare every six months. Investors cannot demand extraordinary growth while also expecting complete insulation from the technology producing it. “You can’t have it both ways.”
• Jason initially calls code review software’s most important category, explicitly separating importance from the largest eventual fortunes, then broadens it: “build to prod.” Code may work in development without being production-ready. Rory agrees that testing, memory management, sandboxes and deployment create related opportunities. “Coding is the mother lode of LLM value”; the unresolved work is getting that output into reliable production.

6. Modal reveals an inference opportunity the panel underestimated

• Harry describes Modal’s $750M Series D, led by Accel at $15.75B, with run rate exceeding $1B. Rory recalls revenue around $60M or below $100M in mid-to-late last year: the acceleration is “stunning.” Jason prefers its developer-facing value creation to simply supplying GPU racks, although he acknowledges seemingly limitless competition.
• Rory estimates four or five leading inference providers together generate roughly $5–6B of run-rate revenue, versus more than $100–110B for OpenAI and Anthropic. His economic example returns to Replit: why buy a comparable token while supporting a model provider’s 70% gross margin if an inference supplier can deliver it at a 30% margin? Harry and Jason acknowledge underestimating how much revenue could accrue outside the leading labs.
• The valuation debate remains open. Neoclouds might ask why they finance infrastructure while another layer captures developer economics, then enter that business themselves. Against that threat, Rory says extraordinary growth “covers a multitude of sins.” Harry puts Fireworks’ margins in the mid-30s and says they will be hitting 50% “pretty soon.” Jason says even mid-30s margins can work publicly, with consequences for the eventual multiple.

7. Town and Anduril test how far future growth can justify today’s price

• In the mock investment committee, Jason passes on Town’s $108M Series B at $1B. He sees a smart team from Plaid but insufficient data: essentially a conventional early venture bet at a radically larger price. “A billion’s the new 100 million.” If the partnership wants this as one of ten bets over the next year, he would let it proceed.
• Jason’s comparison of Town needing $500M of inference with Instinct needing $5B turns toward future financing obligations. Rory wants the clip revisited in two years: a billion dollars remains real money, however casually investors discuss it. Jason instead keeps asking whether they can get back into Modal.
• Anduril at $90B is the sharper valuation test. Jason calls it generational, with contracts and depth, but questions roughly 40x revenue given its margins. Rory contrasts a mature $2B-revenue defense contractor worth perhaps $8B with one doubling toward $4B, $8B and eventually $20B. The latter might justify the price, but investors are “leaning in a lot into the future.”
• Jason’s broader concern is whether AI companies inherit their categories’ depressed public multiples once growth slows—even to 60–70%. He cites Braze at around 6x revenue in one comparison, versus Monday.com and HubSpot at 2x and Klaviyo at 2.2x; he separately notes Braze had been trading at roughly 3x revenue. Rory agrees investors can own “amazing companies just priced wrong,” and doubts broad US market multiple expansion will rescue them.

8. Bending Spoons tests the Berkshire model in software

• Harry introduces Berkshire’s reported $330M purchase of additional Bending Spoons shares. Jason sees “the Berkshire version of software”: select a handful of enduring assets from thousands of opportunities, improve them and hold them. He contrasts earlier, challenged acquisitions such as Evernote and Vimeo with what he considers better-quality assets such as Miro and Airtable.
• Rory’s objection is durability. See’s Candy from the 1960s resembles today’s product; software cannot be assumed to remain similarly stable. He calls it “conceptually the same idea as Berkshire Hathaway, but perhaps in an industry where the idea doesn’t apply.” Jason accepts the distinction but argues this may be as close as software gets to that ownership model.

9. Bootstrapping gives founders more freedom to keep the cash

• Harry cites $370M in 37signals profits since 2020. Jason questions the precision: the figure appears inferred from $37M of profit sharing at 10%. His own rough guess is about $60M a year, with perhaps $40M reaching the bottom line; he calculates that as $240M over six years. Nevertheless, both accept the central point: unfashionable software can make its owners extraordinarily wealthy.
• Jason changes his earlier view that raising one small round and stopping is an ideal outcome. Using Zapier’s approximately $300M revenue as an illustration, he asks what founders might keep if they had raised nothing. “None and done” may be preferable because dividends become harder with venture investors. He says 37signals took money from Bezos or others, while allowing that they presumably worked out their arrangements.
• Harry proposes buying investors out. Rory explains the constraint: repurchasing a 30% stake at six times cash flow absorbs roughly two entire years of company cash, creating risk if financed with debt. Jason says such deals happen but can be expensive and deliver mediocre investor outcomes. Rory’s broader conclusion: “not everything has to be through the lens of venture.”

10. Public venture firms must keep growing assets as LP liquidity returns

• On General Atlantic’s prospective IPO, Rory estimates a meaningful listing needs roughly $800M–$1B of fee income. Public investors value management fees much more than carry, with firms trading around 4–5x revenue in his framing. Only the largest venture and growth platforms seem plausible candidates. He separately warns that public ownership can make a partnership more institutional and risk-averse.
• Harry questions whether already enormous firms would change that much. Rory’s clearer distinction is the obligation to expand assets and launch products. A publicly listed version of old Benchmark, raising $450M every three years, would face pressure for growth funds and crossover strategies. Jason agrees that managers could no longer simply enjoy substantial personal income without growing.
• Harvard’s reported 27.1% return brings its endowment to $68B, helped by a SpaceX stake worth roughly $2.2B. Harry expects SpaceX distributions and hoped-for Anthropic/OpenAI IPOs to recycle cash into venture, with OpenAI potentially in Q1/Q2. Jason resists declaring a peak boom: he cites 2021 university venture/PE returns averaging 36.6%, 41% for the top quartile and 62% at Cambridge. “This is no 2021 yet.”

11. Jason predicts Tesla’s latest spectacle precedes consolidation

• Jason calls the new Roadster Tesla’s last major standalone celebration before an eventual merger with SpaceX over the next “X months.” He categorically predicts Chinese dominance in electric vehicles and autonomy, and an end to people driving themselves within a few years. Harry focuses on marketing: spectacular demonstrations sustain the “Elon premium” and can give the valuation another 12 months of life.
• The discussed idea of placing a SpaceX share in every child’s Trump account combines distribution with shareholder affinity. Rory cites roughly $10B in aggregate value. Jason argues direct ownership could soften hostility toward extraordinary wealth creation: “It’s that much harder to hate someone when you’re a shareholder.” Harry’s counterexample is his own unhappiness holding CoreWeave; Rory notes he did not receive those shares free.