Plus: why LPs will now grade VCs on $20B+ outcomes, why SMBs stopped buying software, and what a two-day Microsoft outage says about how fragile everything has become
There was more news this week than in almost any week since Harry, Rory and I started doing this together. Nvidia lost its founder. Anthropic got set to price the largest IPO ever. HubSpot went back to its founder. Microsoft 365 went down for most of two days. We recorded Monday, before Anthropic priced and before anyone knew whether Jensen would come back. Here are the 11 things that mattered most, followed by the best lines from each of us.
1. Jensen Huang Resigned After 33 Years, and Rory and I Disagree on Whether He Comes Back
What happened: Jensen Huang resigned as Nvidia’s president, chief executive and a member of its board, effective immediately, with CFO Colette Kress serving as interim CEO while the company searches for a permanent successor. The trigger: Huang admitted to an affair with an alleged Chinese spy and Megaspeed executive, and after the board opened an investigation, a second alleged affair was reported involving a direct report. Separately, Commerce and Justice are looking at chips sold into Southeast Asian data centers that may have served Chinese customers. A petition to bring him back drew around 200,000 signatures, and by Monday Jensen said he’d return if asked.
Rory’s read: Decisions made in haste are usually wrong. He would have pushed for a leave of absence and an interim CEO to buy time. Four or five of these directors have been on the board for 30 years, and Jensen is the most important business relationship of their lives. For them to vote against him, something serious had to be going on. His gut says no return without a board change.
My read: I think there’s a high chance he comes back. The spy story will fade. The export issue is bigger, possibly 20% of revenue counting China, but if Jensen is guilty, so is the CFO who signed the financials. Colette Kress now has to stand behind the guidance she signed as CFO. Her job, more than his, was to know where every GPU went.
Rory’s twist: If he owned 100% of the stock, Jensen is the person he’d most want running Nvidia. But even if he returns, Rory doesn’t think he gets back the freedom to finance the AI ecosystem. The minute he left, everyone started trashing the circular deals they’d cheered for two years.
2. Anthropic Is the Biggest Winner of the Nvidia Mess
What happened: Jensen’s biggest strategic push heading into 2027 was a credible U.S. open-weight alternative, including a pending acquisition of Hugging Face. Reports say some Anthropic people were pleased with how the week went.
Rory’s read: The S-1 says roughly $10B went into training Anthropic’s last model. Nobody but Nvidia had the money and motive to fund a serious U.S. open-weight effort. If that stalls, and China pressure pushes buyers away from Chinese open models, the winner is the closed-weight U.S. provider with pricing power. Anthropic isn’t a big Nvidia customer. OpenAI is.
My read: Nvidia’s remote data centers near China were legal at the time. They no longer are. But if you’re Anthropic watching your models get distilled a few hundred miles from your biggest competitors, that loophole looked like theft. I don’t know that Dario is gloating. It’s still a real boost.
3. Anthropic Will Hit $100B in Revenue With 1/12th the Employees Google Had at the Same Size
What happened: The 825 million-share offering could raise up to $64.4 billion at a valuation of up to $1.4 trillion, with pricing expected November 18. The IPO was reported six times oversubscribed, and could raise as much as $90.4 billion including a $26 billion private placement. Trading on Nasdaq as ANT was set to start November 19.
My read: The oversubscribed book is the least interesting part. At $100B in revenue, Google had 81,000 employees and Meta had 71,000. Anthropic will get there, likely by Q1 on run-rate, with fewer than 6,000. Yes, $10B goes to compute instead of humans. But that’s 1/12th the headcount, versus the roughly 30% smaller go-to-market teams at AI-native companies like Harvey and Sierra. For jobs, that’s worse than anything we’ve discussed.
Rory’s pushback: This is the extreme case of compute leverage, which is why the people there get paid so much. An old-school consulting firm won’t end up with $5 to $7 of inference per $1 of labor. He thinks it’ll be less grim for jobs than I do.
Harry’s addition: The rumored 40% layoffs at ServiceNow. Unconfirmed, but I think it’s real.
4. LPs Will Now Grade VCs on How Many $20B+ Companies They Own
What happened: Anthropic went from founding to a trillion-dollar listing in about five years. SpaceX took more than two decades. Spark and Menlo led the key early rounds. Sound Ventures turned roughly $435M into about $7B.
Rory’s read: Spark and Menlo will play with house money for years, and everyone else will work just as hard for a tenth or a twentieth of the return. That’s an extreme power law, and it can get in your head. No meaningful venture money got into Anthropic below $5B, and Spark came in from its growth fund while the company was pre-revenue. If only 10 or 20 companies a cycle reach $20B+, the odds of landing one at seed or A go down. That favors the early-plus-growth fund structure.
My read: When I started, one unicorn was enough for LPs to come find you. Now it’s how many $20B+ companies you have, ideally bought below $20B. Carta’s Q3 2026 data shows deal count down 21% and dollars up 2.5x. Fewer rounds, much bigger checks.
Harry’s addition: LPs getting Anthropic distributions will redeploy, and early Anthropic employees are already angel investing in U.K. companies.
5. Every Mega Fund Is Raising Again in Q1 2027
What happened: The mega platforms all raised in Q1 2026. Every one of them is now lining up to raise again starting January 1, roughly 12 months later.
Rory’s read: When you raise money quickly, you usually don’t make as much. The smartest Anthropic checks were written in early 2023 and 2024, when people were scared.
My read: Deployment pace is feeding the markups. Do the best AI companies justify $10B to $20B to $40B in six months? On a spreadsheet, sure. But some of it is the need to deploy justifying itself. If 8x funds compress to 3x or 4x, maybe that’s fine.
6. IC: $200M Into Mercor at $20B, a Pass on Kalshi and Fractile
What happened: Kalshi is in talks to raise fresh capital at a valuation of $65 billion, roughly 62.5% above the $40 billion valuation it locked in only weeks earlier. Mercor, a leader in AI training data, is raising at $20B. Fractile, an inference chip startup competing with Groq and Cerebras, is raising $720M at $7.5B.
My read: A weak lineup after last week’s Zipline, Fal and Revolut. Kalshi at $65B carries too much regulatory and litigation risk. Fractile has no revenue and is the 11,000th inference chip startup. Mercor is a great space because the labs don’t want this work, but I can’t predict the future of the dirty work. I’d pick it because we have to put money to work.
Harry’s read: Bull on all three. Kalshi leads a monster U.S. sports betting market, and states will get too addicted to the revenue to ban it. Mercor could see $10B of spend from the labs. Fractile has a reported $250M order from Anthropic, and chips are specializing.
Rory’s read: At this stage, every company is good. The only question is value relative to TAM. Cerebras is public at roughly $1B of revenue this year, valued in the high $20Bs. Would he buy Fractile over that? He doesn’t know. Mercor has five or ten customers, but each may spend a billion on training data. For Kalshi at around $5B in revenue, the question is the sober TAM: how many bored 25-year-old guys bet on sports?
The decision: $200M into Mercor at $20B.
7. Without Jensen, the Liquidation Preference Got Weaker
What happened: Nvidia’s run of acqui-hires and backstops, including talks around Reflection and the pending Hugging Face deal, is now in doubt. Poolside sold because it couldn’t raise the capital it needed.
My read: For two years, a lot of AI deals carried an unspoken safety net: Nvidia or a friend would buy the team or backstop the round. I don’t think the CFO-turned-CEO is going to buy all these kids for $15B or $17B. If Salesforce also sits out 2027 digesting this year’s deals, it’s a very different year from the “you can’t lose with a liq pref” year of 2026. Will Thinking Machines survive? I don’t know.
Rory’s read: Washington is split, with Commerce and Lutnick in the sell-to-China camp and Rubio in the don’t-sell camp. You’re dealing with a government that’s mad at itself, which could mean two years of regulators kicking back any acquisition. He hopes Hugging Face got a good breakup fee.
Harry’s read: Nearly perfect timing on Hugging Face, assuming it closes.
8. HubSpot Went From 10,000 Net New Customers a Quarter to Guiding Below 5,000
What happened: Co-founder Brian Halligan returned as CEO, replacing Yamini Rangan in an abrupt leadership reset as the company races to reposition around AI agents. Its shares have halved this year, and Adobe’s Karen Burgess will become finance chief. Halligan called it a refounding, with HubSpot as a “customer relationship harness.”
My read: HubSpot added about 10,000 net new customers a quarter like clockwork, even through the 2024 downturn. Then 6,000, then 5,000, and now guiding below that. Monday.com went from roughly 1,200 net new 10-plus-seat customers to under 200. Net new logos were always my lighthouse metric, and price increases can mask the decline for years. Founder energy will help. The question is whether it’s enough, fast enough. And Intercom doesn’t prove what VCs say it does. Eoghan came back and went all-in on agents, and it worked, but it got to roughly $100M of agentic revenue and we never saw the next act.
Rory’s read: No B2B company from the pre-AI era has reaccelerated above 30%. At a billion in revenue, you’re playing to push growth back to 25%, while a Sierra growing 4x gets all the sizzle. Scale backed HubSpot’s Series C in 2010, and he isn’t sure a 100-person HubSpot customer has an opinion on “harnesses.” But when boards’ backs are against the wall, they bring the founder back, and in each recent case it was right.
Harry’s read: He asked Daniel Dines why he’d go back to UiPath. Man loves to struggle, and founders who’ve already made their money come back with an aggression outside CEOs can’t match.
9. SMBs Stopped Buying Software, but They Buy an $800 Receptionist Agent the Same Day
What happened: Net new customer counts at HubSpot and Monday.com have fallen sharply over three quarters, and Ramp data shows newer CRMs like Attio picking up only a small amount of that spend.
My read: This isn’t vibe coding. Your average SMB isn’t building its own CRM. They’re just not buying, running Google Sheets plus ChatGPT another six months rather than spend $40K on HubSpot. Meanwhile, three of my portfolio companies selling $500 to $1,000 a month front-office agents close prospects the same day. The SMB’s P0 is “my receptionist quits every Wednesday,” not a Kanban board.
Rory’s read: The truth is in the middle. Every associate at his firm has built a Claude front end on Salesforce, and nobody logs into Salesforce anymore. Granted, they went to HBS and work in San Francisco.
Harry’s read: Monday’s users are not vibe coding their own CRM.
10. Microsoft 365 Was Down for Two Days, and Fireworks Got Called Out by Its Own Customers
What happened: Microsoft says its Microsoft 365 services are fully operational again after a global outage that ran from November 12 to November 14, 2026. An issue linked to an earlier network-capacity upgrade surfaced after a more recent network change. The FAA also had a nationwide ground stop tied to a contractor’s router problem. Notion, Cursor and Harvey all publicly named Fireworks after a breach, and an Optimizely compromise hit sites including CNN, the BBC and Zoom.
Harry’s read: Air traffic control, Microsoft, the NSA’s internal comms. He’s never seen systems this fragile. The FAA outage was straight out of Die Hard 2.
Rory’s read: These were old-school IT failures, not AI attacks. When DocuSign got big, he brought in a board member with a cyber background who told the team: you’re going to get hacked, here’s what you spend. Mission-critical companies need that seriousness now.
My read: It’s DDoS 2.0. Microsoft’s bug sat in production for four months until load hit a new level. Now anyone can spin up 10,000 agents and probe every network with traffic, load and hacking at once. We want one-person billion-dollar startups, and those products will have close to zero security. I’m not sure you should trust any web analytics plugin, or any new payments startup besides Stripe or Adyen.
Harry’s addition: Fireworks has a world-class engineering team. If it still happened to them, everyone should be worried.
11. OpenAI Bought Adessia, and Horizontal Agents Are Already Locking In
What happened: OpenAI acquired Adessia, a travel-focused personal agent founded by Sonder founder Francis Davidson. Harry’s question: are vertical personal assistants over?
My read: I’m all-in on Muse, Meta’s horizontal agent. It runs our outbound, content editing, clips and attendee analysis, and reranks every inbound venture deal each night. Mine and Amelia’s coordinate with each other without telling us. With 20 or 30 workflows running, a shallow moat has gotten wide. And this is how SMB software gets displaced: you get a $30K HubSpot quote, realize Muse can do more of it than you thought, and push the purchase out.
Harry’s read: He has about 50 automations on Instinct. Switching would be a real pain.
Rory’s read: Power users will gravitate to whoever can give them the most functionality at a tenth of the cost, and that’s Meta. If most users don’t need that much compute, Instinct has a chance. Still scary when a company with 4 to 5 billion users ships something more powerful than yours. It didn’t work for Threads, though.
On the workforce: Harry’s biggest challenge is getting teams to ask “could an agent do this?” first. My view: our agents now do all our collections, replacing a contractor who billed $3,000 to $4,000 a month. You won’t fire people. They’ll leave, and you won’t backfill. We went from 20 people to 3 humans and 21 agents. Rory’s counter: someone then spends 6 to 10 hours a day managing agents.
On SpaceX, briefly: Rory called Starship one of the year’s top five achievements. But launch is under 10% of SpaceX’s market cap, and most of the TAM is compute. You buy SpaceX based on whether Anthropic is buying more compute a year from now.
Quotable Moments
Rory O’Driscoll
- “If I owned 100% of the stock of this company, the person on the planet I would most want to run this company is Jensen.”
- “You’re not just dealing with a government that might be mad at you. You’re dealing with a government that might be mad at itself.”
- “You buy SpaceX the stock entirely on the basis of what you think Anthropic will be paying you one year from now for that compute.”
Jason Lemkin
- “At $100 billion in revenue, Google had 81,000 employees. Meta had 71,000. Anthropic will have less than 6,000.”
- “I think the liquidation preference has been weakened by the Nvidia drama.”
- “You know what my P0 is? My receptionist quits every Wednesday. That’s what people want to buy. Not Kanban crap.”
Harry Stebbings
- “Every single fund is queuing up out the door for Jan 1. We have to get the bag.”
- “Man loves to struggle.”
- “Right now the clips that you make, we can’t do with AI. In a year we will be able to. You need to get better or you will be replaced.”
