20VC: Valar’s Pro-Rata Fight, Meta’s Muse and NVIDIA’s Bet
Summary
The Valor Atomics–Day One Ventures fight exposes pro rata as a valuable economic option, not a ceremonial investor right. Day One had invested roughly $1 million at seed and $15 million in the Series A, owned about 5%, and resisted being excluded from a $1 billion Series B at a $6 billion valuation. Yet Jason Lemkin and Rory O’Driscoll agreed that, despite a potentially valid grievance, continuing the founder fight was pragmatically irrational: when the existing stake is worth about $300 million, “raise it once and let it go.”
Neither side handled the pro-rata dispute well, but the market’s commentary was even less neutral. Jason faulted Valor’s CEO for dismissing an investor that had previously bridged the company, while calling Day One’s hiring of Quinn Emanuel an escalation too far. His broader point: seed funds that waive pro rata and megafunds that defend it are usually describing their own business models—“You waive, but I ain’t waving mine.”
Meta’s Muse Wonder matters less as a benchmark result whose usability Jason disputed than as evidence that Meta is back near the model frontier. Jason called the evaluation “performative” because, in his view, outside developers could not put the model into their own applications, while Rory credited Meta for returning to the arena with closed models after falling behind. The larger bet is Muse itself: Jason sees “a material chance Meta has just a monster AI product on its hands” if the stronger model makes it 10–50 times more capable.
A rumored $30 billion NVIDIA acquisition of Reflection would be strategic defense for chip margins as much as a model bet. Open models keep more of the AI profit pool in compute, while closed-model providers would prefer multiple chip suppliers; NVIDIA’s ownership of Poolside and Nemotron, and potentially Reflection, strengthens its negotiating position. Rory would “hit the bid,” and Jason warned that declining NVIDIA means facing a better-funded competitor with an integrated hardware, software and model stack within 12–18 months.
Jev’s $870 million Series A at a $7.5 billion valuation is defensible only if a focused team turns a cheap binary classifier into infrastructure used everywhere. Jason estimates it currently resolves perhaps 70%–80% of carefully prompted matches and serves about 10% of AI use cases; hyper-accuracy would be the unlock. Rory’s upside case starts with $200 billion of frontier inference, assigns 20% to classification, compresses pricing by 80%, and gives Jev half the resulting market—roughly $4 billion–$5 billion of revenue and a possible $50 billion valuation.
Private AI valuations are pricing an unprecedented production rate of giant outcomes. The hosts counted 42 new decacorns in 2026 so far versus 34 in 2021, against roughly 20, plus or minus, created per decade historically and roughly 130 public technology companies ever worth at least $10 billion under Rory’s framing. Either AI is “the most economically transformative event since the Industrial Revolution,” or investors are “a wee bit ahead of ourselves.”
The first practical AI deflation may have arrived with Haiku 5.5, Sonnet 5.5 and Jev—but a 90% price cut requires at least 10 times the usage merely to preserve revenue. Jason described Haiku 5.5 as nearly Sonnet-quality for many jobs at one-tenth or one-twentieth the cost, while Sonnet 5.5 is about 10% cheaper and twice as fast. His bet is that customers stop sampling and process everything; Rory’s warning is that this only works if demand is extraordinarily elastic.
The software market has bifurcated between scaled consolidators and a worsening “SaaSpocalypse” for everyone below them. Public indices recovered, but Jason thinks a safe software IPO may now require $1.5 billion–$2 billion of ARR rather than $500 million, while Rory sees hundreds of midsize vendors chasing too few acquisitions. Meanwhile, Bending Spoons-style layoffs and Flock Safety’s security failures raise separate governance questions about employee obligations and product integrity.
Deep dive
1. Pro rata is an economically valuable option, not a courtesy
Harry’s setup: Day One invested about $1 million in Valor Atomics’ seed and $15 million in its Series A, held roughly 5%, then resisted waiving participation in a Sequoia-led $1 billion Series B at a $6 billion valuation.
Rory explained the machinery: pro rata protects an investor’s ownership, a majority can ordinarily waive everyone’s rights, and a spring-back clause restores proportional access when an investor that supported the waiver later receives a special allocation.
Jason corrected the viral framing: Day One had not sued its portfolio company; Valor sought declaratory judgment so it could proceed. His reconstruction was that Sequoia demanded enough ownership to require broad waivers, then Valor allowed an existing “strategic” investor into the round—precisely the favoritism spring-back provisions address.
2. Being right against a founder can still be the losing move
Jason’s verdict was deliberately unsatisfying: “I think everyone’s wrong here.” Valor’s CEO should not have said he wished Day One had never invested after it had supported the company, but Day One’s retention of Quinn Emanuel turned a potentially legitimate grievance into an unnecessarily aggressive confrontation.
Rory reduced the dispute to expected value. If the strategic investor contributed $5 million, Day One’s spring-back allocation would only be proportional to that amount; fighting for perhaps $10 million–$15 million at a $6 billion valuation makes little sense when its existing 5% stake is notionally worth $300 million. “What are you doing here?”
Jason’s personal lesson came from losing an allocation to Harry years earlier despite being the company’s only investor and bridge provider. Challenging the decision “bent” his founder relationship without changing the outcome: “Raise it once and let it go. It’s not your company.”
Keith Rabois supplied the cleanest pricing framework: remove pro rata and early investors must pay materially less, because the right to deploy more capital into proven winners has measurable option value. Rory’s reality check: megafunds asking seed investors to waive are not volunteering to waive their own rights.
3. Meta’s model comeback is real, but Muse is the larger prize
Muse Wonder reportedly landed just shy of Opus 5.5 on an independent index, but Jason scored the announcement “a zero” because, in his view, outside developers could not put the model into their applications. His complaint was commercial, not technical: parity at the same price is irrelevant without usable distribution.
Rory’s counterweight: Meta deserves credit for getting “back in the arena” after falling from the front of open-weight AI to the rear. The shift to closed models suggests that if Meta spends $100 billion, it now intends to create something chargeable; in his shorthand, the earlier $12 billion commitment bought more than expected.
Jason saw Alex Wang’s role as organizationally important: Mark Zuckerberg has had “a lonely stretch,” and a genuinely trusted right-hand partner is scarce. More consequentially, a model that makes Muse 10 or 50 times stronger could give Meta “a monster AI product” with more potential than its AR/VR hardware effort.
4. NVIDIA wants open models strong enough to discipline its customers
Rory framed the rumored $30 billion Reflection deal through industry structure. Chip vendors benefit when models are open and more economics remain in compute; closed-model companies prefer proprietary intelligence and competing chip sources. The parties may lend one another $100 billion, but they are still “watching you closely.”
With Poolside reportedly acquired for $12 billion, Nemotron already in-house and Reflection potentially next, Rory would sell: open-weight monetization remains uncertain, and today’s strategic premium “might be a moment in time that doesn’t come back.”
Jason added the coercive edge of a friendly offer. Saying no means competing against NVIDIA’s capital and engineering stack; within 12–18 months, hardware, software and models could become integrated enough that another open model looks unnecessary. That same compute arms race weakens the case for Thinking Machines at $44 billion: his gut was that independents simply may not secure enough compute or memory.
5. Jev’s bull case is hyper-accurate classification at infrastructure scale
Jason’s mock investment-committee pitch captured late-stage incentives: Jev usage had allegedly quadrupled since Andreessen agreed to lead the $870 million Series A at $7.5 billion, so investors should maximize allocation, block earlier pro rata and mark the position toward $15 billion. Harry’s response: “The terrible thing about that parody is it’s accurate as hell.”
The substantive product thesis is classifier specialization. Jason described Jev’s current system as producing binary answers with perhaps 70%–80% accuracy on carefully prompted matches—useful for roughly 10% of AI workloads because it is “super fast and super cheap.” If the focused former OpenAI team reaches previously impossible accuracy, switching away becomes irrational.
Rory’s market bridge: if frontier inference reaches $200 billion and classification represents 20%, that is $40 billion. Reduce price by 80% and the pool is still $8 billion; half would yield $4 billion–$5 billion of revenue and plausibly support a $50 billion company.
Harry challenged the habitual claim that every richly financed AI company has acquisition downside protection. Rory’s answer was intentionally conditional: with 12 trillion-dollar companies racing to buy time, “it does until it doesn’t”; when the market turns, strategic bids and inflated valuations could disappear together.
6. Forty-two decacorns require either transformation or suspension of history
The headline comparison was 42 startups becoming decacorns in 2026 so far, already above 2021’s 34. Rory contrasted that with roughly 20, plus or minus, decacorns per decade across the prior five decades and only about 130 public technology companies ever reaching $10 billion or more under his tally.
Harry called the implication a bubble; Rory resisted the label but not the anomaly. The production rate is definitively unlike anything in the public-market record: either AI is “the most economically transformative event since the Industrial Revolution,” or investors are “getting a wee bit ahead of ourselves.”
Liquidation preferences partly explain how private marks survive. Investors treat a preferred share as “1X plus option value”; high volatility makes the remote chance of a trillion-dollar winner valuable, even when those companies would not all sustain $10 billion valuations as freely traded common stocks.
Jason cited fund models that require roughly three or four $25 billion-plus exits per fund. Historical supply makes that demanding, but Andreessen’s early insight was to be in every winner despite overpaying: many positions disappoint, while “the price of overpaying for a winner is being in the super winner.”
7. Isomorphic Labs prices AI against discoveries, not displaced labor
After raising around a $10 billion valuation months earlier, Demis Hassabis’ Isomorphic Labs was reportedly discussing $40 billion–$50 billion. Jason’s justification was talent concentration: “Generational talent is generational pricing,” and this may be the last opportunity to assemble an unmatched AI research team.
Rory distinguished discovery economics from automation. Saving $500 million of scientific labor is insufficient at these valuations; the prize is a royalty—perhaps 10%—on a $200 billion drug. Models eventually need to create new value, not merely consume the existing pool of knowledge-worker wages.
Jason said he had only recently changed his mind on AI-enabled longevity because evidence around GLP-1s, including heart-disease benefits, appeared profound to him. Getting a major drug to market five years faster could be enormous; Rory preserved the caveat that molecule identification still precedes phases one, two and three, making value capture harder than discovery rhetoric suggests.
8. Robotics-data revenue is ahead of robot deployment
Maca’s $60 million Series B from Sequoia joined several physical-world data providers reportedly reaching more than $100 million of ARR. Rory recognized the familiar foundation-model pattern: only five or six large competitors serve a relatively small number of enterprise buyers, and those buyers need data immediately and are prepared to spend venture-funded dollars at exceptional speed.
The unresolved issue is downstream deployment. Rory cited a logistics operator’s experience of “42 POCs of humanoids, zero deployed,” alongside doubts about the scale of the BMW–Figure relationship. Robotics-data businesses can reach hundreds of millions now, but billion-dollar durability ultimately requires robots working in production.
Jason’s best demand specimen was a founder who pivoted into one narrow training-data niche, generated $30 million in 60 days and claimed $300 million of demand. The company rejected $270 million because fulfilling it required too many specialized humans across too many countries.
Harry preferred the infrastructure because he believes robotics will win without knowing which form factor or vendor will. Rory’s pushback: if one can identify the winning robotics or foundation-model company, it should capture more value than its data supplier; the picks-and-shovels thesis works best if the model layer remains fragmented.
9. Haiku 5.5 may be the first model that makes inference feel cheap
Jason called Haiku 5.5 and Sonnet 5.5 the first genuinely deflationary model releases he had experienced. Haiku was, in his testing, nearly as good as Sonnet for many jobs at one-tenth or one-twentieth the price; Sonnet 5.5 approached earlier Opus quality while being roughly 10% cheaper and twice as fast.
Speed belongs in the economics. Slow models time out or constrain parallel workloads, while twice the throughput changes which workflows are viable. Jason’s board-level challenge was blunt: if B2B companies can swap models in a week, gain speed and sometimes cut unit cost by 90%, “no more excuses, boys.”
Rory set the elasticity hurdle: after a 90% price cut, usage must rise tenfold merely to hold revenue constant. Jason thinks it will because customers will stop sampling and “just do it all”—including applying classifiers to every row of a database.
Cheaper domestic models could also squeeze Chinese open-weight usage in US enterprises. Jason imagined NVIDIA combining Reflection, Poolside and its stack to capture perhaps 50% of usage but only 15% of revenue; open models could remain strategically important while becoming a difficult standalone business.
10. Workforce cuts expose two different theories of corporate duty
Jason’s objection to Bending Spoons was moral clarity: founders selling Airtable, Miro or similar businesses should know that most employees may disappear. “There’s no way I would sell my company and have 80% of my team laid off,” especially when the company remains growing and well capitalized.
Harry pressed the fiduciary case. Rory said public-company directors must maximize shareholder value during a sale and should not suggest otherwise in the boardroom; the private-company duty was less clear to him.
Rory’s uncomfortable counterargument was that a “forever march” also destroys retention and value. A sale can be better even for employees if their economics are negotiated upfront; otherwise the best people may leave over two years while the remaining business becomes an “empty husk” worth one times revenue.
Bending Spoons’ method is unusually legible: centralize G&A and people functions, then run the asset leaner still. Jason contrasted that with Cloudflare’s 20% cut despite strong results; Cloudflare subsequently reaccelerated, supporting his view that a decisive cut can sharpen a healthy company rather than merely strip it.
11. Bumble’s take-private prices the gap between human value and revenue
Harry reported Ravi Gupta’s Ithaca Holdings taking Bumble private for about $1.3 billion, versus a market capitalization near $500 million before the deal. Jason noted why venture investors historically disliked dating: churn is intrinsic and customer lifetime value difficult.
Rory found the outcome economically strange. Dating platforms address one of humanity’s most basic needs and reportedly account for 50%–60% of people meeting on apps, yet the category captures only a few billion dollars of enterprise value.
Jason’s bull case is that Bumble’s historical data plus AI could replace mediocre visual matching with fewer, better dates—and charge materially more. If a good date costs $200, he joked, spending $150 on Bumble AI still leaves the user “50 bucks ahead.” Harry added that AI avatars could represent users and meet 100 different people without the users spending the time, though Rory insisted the mission eventually requires physical contact.
12. The SaaSpocalypse ended only for companies that already had scale
Public prices recovered—the Nasdaq was up about 22%, while cloud indices cited by Jason rose roughly 25%–28%—but he thinks relief is misplaced. The crisis passed for companies well above $1 billion of revenue; below that threshold, “everyone without massive scale is living a worse SaaSpocalypse than in March or April.”
Rory’s diagnosis was maturation and consolidation. Leaders have cash flow, market capitalization and acquisition leverage; a vendor at $100 million trying to reach $300 million faces a structurally harder market. Jason suspects even $500 million of ARR is no longer IPO-safe and that software companies may need $1.5 billion–$2 billion.
Harry offered TechWolf’s sale to SAP for hundreds of millions and Intercom’s AI-native reinvention as counterexamples. Jason congratulated the companies but said neither supplies the billion-dollar recurring revenue and 50% growth megafunds require; Harry replied that such an outcome could return 24% for his fund.
Rory reconciled the positions: a $500 million exit matters enormously to a normal fund but not a megafund. The graver problem is scarcity—perhaps 200–300 midsize SaaS companies and only around 10 acquisitions a year—leaving founders to become AI-native, engineer strategic relevance or accept a decade-long compounding and dividend business.
13. Flock Safety’s problem reaches beneath privacy into engineering culture
Flock Safety cut 270 people, about 18% of staff, after privacy controversies. Rory still preferred controlled surveillance to crime, but unauthorized personal use had damaged the brand across both red and blue America: “It’s not just a comms problem, it’s a tough business problem too.”
Jason’s larger concern was engineering discipline. He cited unsecured camera streams, unencrypted stored images, credentials sent in clear text, a hardware button that exposed Wi-Fi access, a police interface without authentication and cameras whose security patches had stopped in June 2018.
Startups necessarily cut corners, Jason conceded, but security companies cannot make those choices at the core of their product. Repeated failures suggest a culturally consistent weakness elsewhere in the stack: “It’s always more when you pull on the string.”