20VC · · 73 min
20VC: Will Jensen Return to NVIDIA | Anthropic Set for Largest IPO of All Time | Kalshi, Mercor & Fractile: Mega-Rounds Continue
TL;DR
- A Jensen return would not automatically repair NVIDIA or venture’s acquisition backstop. Jason sees a high chance but says a return should take a couple of weeks; Harry predicts it within three days, while Rory doubts it absent a board change. Jason thinks the personal allegations may fade, but says the China-linked chip-transshipment issue is more consequential and could implicate other executives. Export-control uncertainty could constrain revenue, dealmaking and venture exit backstops even if Jensen returns; Jason’s framing is that “the whole ecosystem’s been maimed.”
- Anthropic’s anticipated $1.4–$1.5 trillion IPO combines extraordinary wealth creation with a radically smaller workforce. Jason expects a $100 billion revenue run rate “by Q1 or whatever” with fewer than 6,000 employees, versus 81,000 at Google and 71,000 at Meta at that revenue milestone. He sees “a 12th” of the workforce as the central economic signal; Rory doubts that efficiency ratio generalizes across the economy.
- Venture is concentrating more capital into fewer companies while accelerating fundraising. Jason cites Q3 2026 Carta data showing deals down 21% and dollars up 2.5x, and expects LPs to grade portfolios on companies worth “20 billion and up.” Both warn that rapid fundraising and markups could compress returns, even when underlying businesses grow exceptionally fast.
- Mercor wins the mock investment committee, but valuation—not just business quality—is central. At $20 billion, its attraction is labs spending heavily on outsourced training data; Jason’s reservation is the durability of that “dirty work.” Kalshi’s $65 billion price carries regulatory risk, while Fractile’s $7.5 billion valuation prompts comparison with revenue-producing Cerebras: Harry calls Fractile “the hope play.”
- Legacy SaaS faces a purchasing slowdown that founder returns alone may not reverse. HubSpot’s quarterly net-new customers have fallen from roughly 10,000 to 5,000, with guidance below that. Jason’s theory is that SMBs increasingly defer conventional software while buying agents that remove urgent labor costs: an $800 receptionist replacement solves a clearer problem than “Kanban crap.”
- Infrastructure reliability and vendor security could become more expensive conditions of doing business. The panel initially treats Microsoft’s roughly 36-hour outage and the FAA disruption as conventional failures, while Jason predicts agents will make overload attacks easier. His concern is the collision between one-person billion-dollar startups and customers needing “S-tier security teams” throughout their dependencies.
- Horizontal agents are building switching costs through accumulated workflows, while subsidies threaten standalone competitors. Harry has roughly 50 Instinct automations; Jason describes 20–30 Muse workflows and calls their moat shallow but increasingly wide. Meta’s ability to subsidize heavy usage matters, and labor substitution may arrive through attrition: “You just won’t backfill them.”
- SpaceX’s technical achievement and its stock valuation require different underwriting. Rory celebrates Starship but says rocket launches represent less than 10% of market capitalization. His investment question is compute demand: “You buy SpaceX stock entirely on the basis of what you think Anthropic will be paying you one year from now for that compute.”
Digest · the substance, structured for research
1. Jensen’s return depends on the board and shared regulatory exposure
- Rory’s first objection is process: “decisions in haste are often wrong.” For a founder who built NVIDIA over 30 years, he would have preferred leave, an interim CEO and time to investigate. He separates the personal allegations from export-control issues and repeatedly stresses that outsiders do not know what the board knew.
- The board’s long history with Jensen makes Rory reluctant to dismiss its judgment. Several directors have had the most consequential business relationship of their lives with him; reversing themselves within a week would be a substantial step. Without replacing the board, his “gut would be no,” though he adds that he could “look like an idiot in two days.”
- Jason sees a high chance of reinstatement because potential export-control responsibility extends beyond Jensen. If Jensen is guilty, the CFO signed the financial statements and other executives helped organize the business. His argument is that removing Jensen alone cannot resolve collective accountability. Rory agrees that a CFO becoming CEO faces awkward questions about reaffirming her own accounting and guidance. Jason also thinks the “spy who loved him” allegation may fade, despite facts he says are getting worse, while the chip-transshipment issue will endure.
- Harry predicts a return within three days, before another news cycle displaces the campaign. Jason thinks it should take a couple of weeks, although it might happen faster. Rory distinguishes desirability from probability: absent the U.S. government trying to imprison his CEO, Jensen is the person he would most want running a wholly owned NVIDIA.
2. NVIDIA’s disruption reaches open models and venture exits
- Harry cites up to $601 billion of market value lost in a morning and thinks Jensen’s return could restore the stock. Rory disagrees that it would return to its previous level. Rory says China represented roughly 15% of NVIDIA’s business ten years ago; Jason says the issue could amount to 20% of revenue if China is counted. Jason argues NVIDIA’s overseas data centers served China-linked users remotely and that chips could be transshipped; he says that arrangement was legal but apparently is no longer legal. Rory also points to Department of Commerce and Justice comments about chips sold to Indonesia that may end up in Singapore and China.
- Rory doubts Jensen could resume deploying capital as aggressively after scrutiny of those arrangements. His strategic winner is Anthropic. Jensen was financing a significant U.S. open-weight alternative, and the reported roughly $10 billion cost of training Anthropic’s latest model illustrates the capital required. If that agenda stalls while Chinese open-weight models face greater resistance, U.S. closed-weight providers gain pricing power. Anthropic also depends less on NVIDIA than OpenAI does, according to Rory.
- Jason argues that NVIDIA’s disruption weakens investors’ confidence in liquidation preferences: those preferences felt safer when NVIDIA or its friends might acquire, backstop or rescue companies. The planned NVIDIA–Hugging Face deal remains conditional on closing. When Harry praises its timing, Jason replies, “Perfect timing would be closed.” Rory regards regulatory entanglement as the larger risk.
- Rory worries that conflicting Washington factions favoring and opposing China sales could trap NVIDIA in bureaucracy for two years. Jason extends the exit concern to 2027 if NVIDIA stops buying and Salesforce pauses to digest acquisitions. Poolside’s disclosed inability to fund the next $10–$20 billion illustrates his concern: without an acquirer, what was its alternative?
3. Anthropic’s IPO exposes the scale of compute-driven labor leverage
- Harry cites an IPO book seven times oversubscribed, with some reports reaching ten times, at up to $1.4–$1.5 trillion. They are recording before the listing. Rory qualifies the “largest IPO” framing as biggest or second-biggest by dollars raised, depending on SpaceX; reaching a trillion-dollar valuation in five years is the more consequential change in expectations.
- Jason’s central comparison is headcount. At $100 billion in revenue, Google had 81,000 employees and Meta 71,000; Anthropic will have fewer than 6,000 at the run rate he expects “by Q1 or whatever.” That is “a 12th,” versus the roughly 30% smaller go-to-market teams he cites elsewhere. Compute costs explain the model, but do not diminish its human implications.
- Rory calls this “compute leverage, human non-leverage”: fewer employees can share a similar compensation pool. He disputes Jason’s broader pessimism, arguing that traditional consulting firms will not necessarily spend $5–$7 on inference for every dollar of labor. Jason pushes back that automation already reaches front offices, back offices and delivery; he regards the staffing difference as a much wider disruption.
4. Venture’s new benchmark is fewer investments with much larger outcomes
- Rory congratulates Spark and Menlo on potentially producing 8–9x funds on substantial capital and “playing with the house money.” He also supplies historical perspective: the median 1996 fund had 6x DPI. Buying at a $20 million pre-money valuation and exiting publicly around $500 million could generate roughly 20x; today’s exceptional distinction is the dollars involved.
- Jason cites Carta’s Q3 2026 figures: deal count down 21%, invested dollars up 2.5x. He expects continued concentration through repeated investment in winners. LP screening, he predicts, will shift from unicorn counts to “How many 20 billion and up” companies a portfolio contains—preferably acquired before they were valued at $20 billion.
- Rory argues that combining early and growth investing fits this environment. If only 10–20 companies become the relevant winners, identifying one at seed becomes harder. Spark’s meaningful Anthropic entry came around a $5 billion valuation, pre-revenue, through its growth fund. Getting exposure at Series C can be rational if smaller outcomes cannot obtain comparable liquidity.
- Harry says mega-platforms that raised in Q1 2026 are already preparing larger funds for Q1 2027. Rory warns that the clever Anthropic checks came in the more fearful early-2023 and early-2024 environment. Jason asks whether deployment pressure helps drive valuations from $10 billion to $20 billion to $40 billion in six months; maybe returns compress from 8x to 4x or 3x.
5. Mercor wins the mock IC while valuation limits enthusiasm
- The mock IC’s reluctant choice is Mercor at $20 billion. Jason likes the founder and labs outsourcing “dirty work,” but cannot predict that work’s durability; he envies investors at $2 billion, $5 billion or $10 billion. He proposes putting in $200 million for 1%, saying the firm will learn, because it is supposedly fundraising again in January.
- Harry imagines $10 billion of lab spending on Mercor’s data, with another $10 billion potentially unlocked if companies own their intelligence, supporting his $100 billion outcome case. Rory notes that roughly 70–75% of Anthropic’s compute goes toward the next model. Five or ten customers is concentration risk, but customers each spending around $1 billion on data create substantial demand. Scale and Surge remain competitors.
- On Kalshi at $65 billion, Jason dislikes lawsuits and regulatory risk even if investors might make money. Harry counters that states may become so dependent on revenue that a new administration cannot ban it, and sees a huge U.S. sports-betting market. Rory cites roughly $5 billion of revenue but wants a sober TAM estimate under full legalization, then an assessment of prediction-market upside: “How many bored 25-year-old dudes are there betting on sports?”
- Harry defends Fractile at a $7.5 billion price with a cited $250 million Anthropic order and chip specialization. Jason objects to another pre-revenue inference-chip startup. Rory’s comparator is Cerebras: roughly $28 billion valuation, just under $1 billion revenue this year and probably tripling to $3 billion next year. He lacks enough information on Fractile’s edge to prefer it; Harry calls Fractile “the hope play.”
6. Returning founders can improve SaaS without restoring its former growth
- Jason says HubSpot reliably added roughly 10,000 net-new customers quarterly, then fell to 6,000, then 5,000, with guidance below 5,000. Its stock has fallen 50% this year, but customer acquisition concerns him more. Monday’s cited net-new-customer measure fell from roughly 1,200 to below 200, a decline he puts at 80%. “Something had to change, and it had to change urgently.”
- Jason calls net-new customers his “lighthouse metric”: price rises and decent net revenue retention can conceal stalled acquisition behind 20% growth for years. Rory adds that CRM markets eventually saturate. Jason expects Brian Halligan and Dharmesh’s renewed founder involvement to help because founders can make aggressive changes; his question is “will it work enough?” Rory cites the pattern of boards bringing founders back when a company is in trouble, saying it was the right move with Dines, Brian, Owen and Dev.
- Intercom is Jason’s deliberately uncomfortable example. Its returning founder corrected pricing problems in 2024 and committed early to agents, producing more than $100 million in agentic revenue. That worked, he says, but the Salesforce acquisition prevents observing whether it could reach $1 billion. “Did it really work?” means the full turnaround thesis remains unproven, rather than that the improvements failed.
- Rory sees no SaaS-era public company reaccelerating above 30%; Palantir’s doubling is a separate case. A founder might restore quarterly additions from 5,000 to 8,000 and achieve 25% growth while Sierra grows 4x. HubSpot’s “customer relationship harness” positioning attempts to address the new product world, but even a successful AI business remains embedded in a much larger, slower-growing company.
7. SMBs may defer software while buying labor replacement
- Rory thinks AI-built applications may suffice for small businesses with simpler workflow and authorization requirements. Harry strongly disputes that ordinary Monday customers are building their own CRM with Claude Code. Jason takes the middle: he sees no evidence of widespread nontechnical SMB coding, and says CRM is not a core Replit vertical even though Replit is at roughly $1 billion in revenue.
- Jason’s explanation is deferral, rather than sufficient migration to another CRM. Ramp data suggests Attio has benefited a little—about another $10 million in revenue—but that is too small to explain the incumbent slowdown. An SMB may keep Google Sheets for another six months, using ChatGPT for analysis, instead of spending $40,000 annually on HubSpot. “I don’t wanna buy this crap. I just don’t wanna buy it.” He says he did not anticipate this behavior.
- Conversely, three of Jason’s portfolio companies sell $500–$1,000 monthly front- or back-office agents and can close qualified prospects that day. His example is concrete: “My receptionist quits every Wednesday.” An $800 agent that works continuously, reschedules appointments and handles billing solves an immediate priority. Buyers might postpone conventional software for a year or two; whether postponement becomes permanent remains unknown.
8. Outages and vendor breaches raise the cost of reliability
- The initial distinction is that neither Microsoft’s roughly 36-hour outage nor the FAA disruption was being attributed to an AI cyberattack. The panel describes an FAA router issue and a Microsoft network or load problem. Jason says the Microsoft bug reportedly existed for four months before an unmodeled load threshold triggered it; explanations blaming engineering cuts remain unverified.
- Jason calls the prospective threat “DDOSing 2.0”: agents make probing networks and generating overwhelming traffic easier, even without stealing credentials. His operating example is cascading failure: a thousand servers lose capacity, the remaining machines absorb more load, and failures accelerate from 900 functioning servers toward 500. Even identifying the fix does not immediately stop the deterioration.
- Harry pushes back that Fireworks’ engineering team is exceptional, making its breach more worrying. Jason says exposure came through a testing corner storing keys, rather than all its data; Notion, Cursor and Harvey publicly named the vendor. His tension is between aspiring to one-person billion-dollar startups and requiring “S-tier security teams” across the products and dependencies customers trust.
- Rory expects substantial spending to make mission-critical systems durable. At DocuSign, a security-experienced director warned that scale made attacks inevitable, and the company funded the necessary work. A week without air-traffic systems would become economically painful; the reported nationwide ground stop with 4,000 planes airborne shows why he thinks operating such infrastructure “on a shoestring” is untenable.
9. Horizontal agents accumulate switching costs through everyday work
- Harry asks whether OpenAI buying travel-assistant company Edesia, founded by Sonder’s Francis Davidson, signals consolidation into horizontal personal assistants. Jason expects about 20 such acquisitions as companies compete for the next phase. His framing moves beyond discrete agentic applications toward autonomous horizontal agents: “It gets good next year,” particularly for people who do not spend their days using Claude Code.
- Jason describes roughly 20–30 Muse workflows spanning outbound, content, analysis and venture-deal review, with full-time email access. “It is a shallow moat, but man, it’s become wide.” His marketing example is selecting 200 CEOs from a CRM and inviting them to dinner without a human assembling the list. He says they have already executed that kind of workflow, potentially displacing expensive marketing-automation software.
- Harry has around 50 Instinct automations and considers rebuilding them burdensome. Jason says another agent must solve an important unmet task to justify switching or adding it. Muse’s browser capabilities provide his example: it handles WordPress interactions for SaaStr’s roughly million monthly content views where other agents he tried did not. Rory asks whether this ostensibly consumer product is already functioning as a small-business agent.
10. Subsidized agents change both vendor economics and hiring
- Jason estimates his $20 Muse plan delivers service worth perhaps $200–$1,000 monthly, and says he is sure his $200 Claude subscription consumes $10,000 monthly. Rory replies that he is not the typical customer, so the business model may survive. Jason’s counter is a possibility rather than a forecast: “Maybe I’m the future, just in the present.”
- Rory argues that heavy users will favor a company able to offer functionality for one-tenth of its delivery cost—probably Meta rather than Instinct. A lighter typical user could leave standalone vendors room to compete through speed. But Meta’s four-to-five-billion-person reach and technically powerful product are formidable; Jason worries that if Instinct cannot let him run agents 24/7, the product’s usefulness is maimed. Rory notes that distribution did not make Threads work.
- Jason expects the staffing model to invert: ask agents first, then put humans into the remaining gaps. SaaStr’s collections contractor previously billed $3,000–$4,000 monthly and now bills one hour. Attrition can implement the change without a discrete firing event: employees leave for another job and “you just won’t backfill them.” Rory agrees that roles can slowly ebb away.
- SaaStr has gone from about 20 people to three humans and 21 agents, but Jason stresses the management burden: someone may spend six, eight or ten hours daily managing agents. His suggested starting point is unfinished work: “Think about something you just didn’t get done and ask Muse.” He thinks it will outperform an available human five times out of ten.
11. SpaceX’s rocket achievement leaves compute as the valuation question
- Rory calls Starship a stunning technical achievement, potentially among the year’s five most significant developments, following roughly half a decade of sustained effort. He separates that appreciation from his frustrations with Musk’s politics and interpersonal behavior. Rockets enable differentiated Starlink service, especially in less crowded areas, yet he says launches now represent less than 10% of SpaceX’s market capitalization.
- Rory initially says “Not at this price,” but ultimately says he might or might not buy after forecasting compute demand. His reading of SpaceX’s S-1: most business value and TAM rests on compute. “You buy SpaceX stock entirely on the basis of what you think Anthropic will be paying you one year from now for that compute.” He likes Sean Maguire’s five-year, three-market framework despite complaints about omitted tax.