20VC: NVIDIA Crushes Quarter and Buys Hugging Face | OpenAI Cuts Off Cursor | Instinct Hits $2.5BN Valuation and The Race for AI Assistants | Cognition Raises at $46BN, Linear $2.5BN and Clay $7BN

Summary

Harry Stebbings, Jason Lemkin, and Rory O’Driscoll argue that AI agents are now the defining venture theme, but the category comes with major technical and security constraints. They read Nvidia’s blowout quarter and forward guidance as evidence that compute demand is still expanding faster than supply, and frame the rumored Hugging Face deal as a way to make compute more efficient so Nvidia can sell even more of it. The OpenAI-Cursor cutoff and the Hugging Face/OpenAI agent hack are treated as signals that distribution fights, IP leakage, and reward-hacking risks are becoming central to the industry. In the software section, they argue that AI is creating “compound startups” that ship far more functionality, making tools like Cognition, Clay, Linear, and Salesforce increasingly agent-friendly and potentially far more valuable than legacy software categories.

Notes

• Nvidia’s quarter is interpreted as evidence of persistent, supply-constrained compute demand rather than a near-term peak.
• Rory says 2027 guidance implies roughly 70% revenue growth versus analyst expectations around 40-50%.
• He argues the dominant semiconductor market can keep growing at a 70% rate for another year, which he calls a major statement.
• The discussion links Nvidia’s demand to hyperscaler capex and the wider circular AI economy.
• The panel frames the AI supply chain as circular: chips go to hyperscalers, hyperscalers sell compute to model providers, and model providers sell intelligence to end customers.
• Rory connects the capex cycle to pressure on Google and Oracle free cash flow.
• He says Nvidia’s quarter shows the largest semiconductor market can grow at 70% rather than a normal 10% for another year.
• The rumored Hugging Face transaction is discussed as Nvidia buying efficiency to sell more compute.
• Jason’s read is that Nvidia is trying to win open weights and every market segment, even if that means overpaying.
• OpenAI cutting off Cursor is treated as both theater and a real competitive move.
• Rory says Cursor and OpenAI were already on a collision course because coding is the main node for LLM competition.
• He adds that OpenAI is worried about distillation and IP leakage through misuse of model terms.
• Jason points out Anthropic previously cut off Windsurf, so the move has precedent.
• Both hosts say personalizing a business dispute makes it much harder to unwind.
• The Hugging Face/OpenAI hack is described as involving roughly 500 to 1,000 agents swarming together.
• Jason warns against anthropomorphizing agents or treating them as human-like collaborators.
• He says OpenAI released hundreds of long-running super agents and let them run long enough to goal-seek.
• Jason argues current LLM behavior is fundamentally goal-seeking and that reward hacking is just another name for it.
• Rory says persistent agents plus enough compute can manage complexity, probe weaknesses, and exploit systems at scale.
• He says CISOs have months, not years, to adapt because open-source models are only about six months behind.
• Harry says Instinct let him book dinner but then quickly ran into credit-card access concerns.
• Jason says a different agent product misread instructions and tried to buy Patek Philippe watches for his team.
• He says that kind of scenario is plausible today if an agent has credit-card access.
• Jason notes that storing emails and credit cards is itself a security issue, separate from whether the agent behaves correctly.
• Rory describes Instinct as a life-management assistant that acts on calendars, email, and credit cards.
• He says Instinct has raised at roughly a $2.5 billion valuation, showing strong venture enthusiasm.
• Jason says he does not believe current agents can solve reward hacking, so he would not trust them with spending authority.
• His near-term control recommendation is hard caps on spend accounts.
• Rory still thinks agents will be used for life transactions and to-do lists as models learn user preferences.
• He compares that behavior to Tesla FSD, which can suggest a destination based on routine.
• Rory says this is basically what Siri was meant to become, and Apple may deliver it in the next two or three years.
• Harry says Instinct has already removed low-level booking work from his EA.
• Jason says every application will add more autonomous scheduling and email functionality.
• He also argues richer autonomous products can become moats because they are hard to replicate.
• Rory says agents are the big venture idea for 2025-2026, and investors are buying into momentum as much as fundamentals.
• Harry says Cognition is raising around a $46 billion valuation and could end the year at $1.6 billion ARR.
• Jason’s point is that Cognition can be a major business even if it is not one of the top two or three players.
• Rory estimates about $500 billion a year of U.S. labor spend sits in software-related work.
• Jason says the core mistake was underestimating how much more software people can build with AI.
• They argue AI does not shrink ambition; it pushes companies to ship far more code and features.
• Rory says the winner is the company that compounds fastest and grinds out the most software 7-by-24.
• Jason says even newer AI-native companies like Owner are being pushed into the compound-startup model.
• Rory cautions that more AI features can create product slop and too many buttons, so strong product leadership still matters.
• Harry says European founders may need larger rounds to support the token and speed requirements of compound startups.
• Jason cites Iconiq data suggesting below-hypergrowth companies are not hiring, while 100%+ growers are adding 133% headcount on average.
• Rory explains the paradox: AI should reduce engineer count, but growth capital and sales intensity create more total spending.
• Jason says the ‘do more with less’ thesis was the wrong assumption; in practice, teams are doing more with more.
• They say capital is being pushed toward what is already working, and Andreessen’s expanded growth fund is read that way.
• Harry lists a burst of private-market pricing: Clay at $7B, Linear at $2.5B, and Shein at $26B.
• Jason says Salesforce’s AI strategy is more substantive than it looks, especially multi-service access and outcome-based pricing.
• Rory estimates Salesforce could plausibly move from about $212B to $220-230B in a year.
• Rory says Salesforce is spending about $300M this year on Anthropic, roughly 5% of its $6B engineering budget.
• Rory cites a rumored Stripe-PayPal price gap, with Stripe in the 60s and PayPal wanting in the 70s.
• Jason explains the failed deal as a moving stock-price negotiation in which PayPal’s valuation shifted as the process unfolded.
• Jason says he started as a Clay skeptic but is now convinced agents will use Clay exclusively.
• He argues agentic GTM can generate 10x to 100x more usage than human GTM and sees a path to a very large outcome for Clay.
• Rory says Clay and Linear are examples of founders turning AI disruption into advantage at the application layer.
• Jason says Linear is especially valuable because it is highly agent-friendly and pairs output quality with agent usability.
• Harry says Linear is around $100M ARR, growing over 100%, and valued near $2.5B.
• Jason says project management is a dying category in a human-first world, and Linear is the clear winner in the agent era.
• He says agent usage has exploded 50x in 90 days and that adjacent companies like Replit and Lovable make Linear look cheap.
• Rory clarifies that ‘agent-friendly’ means third-party agents will choose the product, not just that the product has its own agents.
• Jason says agents will test APIs directly and quickly spot blocked or broken integrations.
• Rory says agent-driven workload growth can overwhelm human-first systems like GitHub because volume scales 10x to 100x.
• Rory says Flock backlash stems from identity errors, police misuse, and abuse of private lookup data.
• He says products that expose large private databases need real controls because abuse creates backlash, citing cops tracking exes and other improper lookups.