Alphabet loses about $500B in market cap, DoorDash buys Instacart for $17B, Kalshi hits $40B a day after criminal charges, and we each had to pick one deal to take to IC: Cognition, Baseten or Serval.
This was one of the busiest weeks we’ve had on the pod. The AI trade wobbled hard. OpenAI took a lower price to fill its round, Alphabet had its worst day in about a decade, and Anthropic’s latest S-1 amendment showed quarter-over-quarter growth cut roughly in half. At the same time, money kept pouring into whatever is working: Kalshi, Cohere, Baseten, Cognition, Serval and Physical Intelligence all raised or are raising at numbers that would have sounded crazy a year ago. Harry Stebbings, Rory O’Driscoll and I went through all of it. Below are the top takeaways, followed by the best lines from each of us.
1. OpenAI Cut Its Round From $30B at $1.4T to $20B at $1.25T
What happened: OpenAI has secured about $11 billion of the $30 billion it sought and now aims for up to $20 billion at a $1.25 trillion valuation, per the WSJ, FT and CNBC. The FT says OpenAI pitched $1.4 trillion and cut it to $1.25 trillion a week before the deadline, as tech stocks were falling after the results from Google, Amazon and Intel.
Rory’s read: Drop the “AI wave is cracking” story. Rory thinks people are projecting their own biases onto what is really a simple pricing question. Two securities are about to be sold over the next couple of months. One belongs to a bigger, faster-growing, profitable company that’s about to list, and you’ll be able to buy it liquid at $1.5 to $2 trillion. The other belongs to a smaller company that’s growing more slowly, still losing money and illiquid. Of course the second one prices below $1.5 trillion. The only open question is how far below.
Jason’s read: I agree with the logic, but I’d still be a little nervous. AI stocks lost something like $5 trillion in a week. CoreWeave debt is paying 15%. Broadcom is guaranteeing Anthropic’s debt. IPOs are always thinner going in than the “20x oversubscribed” headlines suggest. None of that has to matter in 12 to 18 months. Right now, though, there’s real wobble. Maybe the second tranche prices below the first.
Rory’s counter: All of that is price discovery on the supply side. The demand side is what decides it. If AI demand holds, every GPU and data center gets used and the CoreWeave bonds are money good, because someone will buy the company. If demand cracks, Google and Amazon won’t get hurt first. They have huge balance sheets. The CoreWeaves will.
2. Alphabet Fell 12% Even Though Google Cloud Grew 103%
What happened: Google Cloud posted its fastest growth on record, with revenue jumping 103% to $30.7 billion, but the gains were overshadowed by weakness in the core search business. Revenue growth at Google Search rose 12% to $63.4 billion, marking its slowest growth rate since 2023. Free cash flow fell to negative $9.07 billion, the second straight negative quarter, as capex of $51.36 billion outran cash generation.
Rory’s read: Search cannibalization has been the obvious risk for two or three years. This quarter it showed up in the numbers, and it landed in the middle of a broader capex scare. Rory also pointed out how quickly the narrative keeps flipping: first “Google is dumb as rocks and Satya is a genius,” then “Google is the best company ever and poor Satya missed the boat,” and now back to “maybe the Google worries were right.”
Jason’s read: A cloud business at a roughly $120B run rate growing 100% is about as good as it gets. Google says it can sell compute as fast as it can build it. But the market is nervous AF right now. Shopify dropped 15% while we were recording because it came in a smidge short, even after raising guidance. Any metric that misses gets punished.
3. Anthropic Went From 12x Growth to a Projected 2x, and the Likely Cause Is Pricing Pressure
What happened: Anthropic’s quarterly S-1 amendment showed GAAP revenue of roughly $5B in Q1, $11B in Q2 and $17B in Q3. That means net new revenue dropped slightly from Q2 to Q3 even though absolute bookings went up.
Rory’s read: Rory went straight to GAAP, because GAAP numbers are audited. On an absolute basis, Anthropic added about $6B from Q1 to Q2 and a bit less than that from Q2 to Q3. On a percentage basis, quarter-over-quarter growth went from about 100% to about 50%. The bull case is that 50% QoQ still compounds to roughly 5x year over year. The bear case is that growth halved in a single quarter, so does it go to 25% next? Zoom out further: about $4.5B of GAAP revenue last year, probably ~$56B this year (12x), and something like $120B next year at the same trajectory. That’s 2x. Rory called it a sustained slowdown to numbers that are still magnificent. Everyone is betting on which way you squint.
Harry’s read: If a startup did $6M in net new one quarter and $5.4M the next, Harry would not invest. He also pointed out that Anthropic is still around 0.1% of Fortune 500 spend, so the deceleration isn’t a TAM ceiling.
Jason’s read: That’s why I think this is pricing pressure, plain and simple. Demand isn’t down, and OpenRouter data doesn’t show a decline. But OpenAI is going around selling roughly equivalent products at 30% off, and open-weight models put pressure on prices too. I had my own data point this week. When I moved SaaStr Connect, our AI matching app, onto Haiku 5.5 and a cheaper model mix, total costs fell 72%. Same B2B workflow, running constantly, faster, and higher quality. Since early 2025 nothing I’ve run agentically had actually gotten cheaper per use, no matter what the eval charts and token prices said. This was the first time. If mature apps everywhere are seeing that kind of deflation, Anthropic’s ability to charge a premium may have hit a local maximum.
Harry’s read: This is exactly why venture outcomes are so binary right now and public markets are so volatile. For the first time in 30 or 40 years, nobody knows what five years out looks like.
4. I Turned Off Canva After 12 Years. Squarespace and Wix Are Feeling the Same Shift.
What happened: Reports this week varied on Canva’s growth. One said it was flat. Others said it had slowed meaningfully and was preparing layoffs. Squarespace announced 20% layoffs.
Jason’s read: I turned off Canva after 12 years. Our 21 agents run on Higgsfield, Opus and similar tools, and agents don’t log into Canva. What surprised me is that the real world is pulling back too. I assumed Canva, Squarespace, Wix and Shopify would be fine because most companies aren’t running 21 agents. That isn’t what’s happening. ChatGPT, Grok and others are simply that good for images. The worst case for Canva is that it missed its IPO window and becomes a private company growing more slowly, like Automattic, which I think stays private forever.
Rory’s read: Software companies are like sharks. If they stop moving forward, they sink. Rory remembers mainframe companies in the ’90s and 2000s living off 20-year-old contracts. You can stay private, keep cutting costs and run your own little Bending Spoons playbook, but the asset turns into a terminal-value story and it’s very hard for investors to get value out of it. If you don’t climb onto the AI train, even late and even for 10 to 20% growth, you’re heading toward quiet obsolescence.
5. Consumer Agents May Cost More Than $20 Per User Per Month to Serve
What happened: Persistent consumer agents like Muse and Instinct are growing fast. Per Jason, Elon also had to cap free usage of xAI’s agent product this week because it was too expensive. Replit’s Amjad Masad said the era of free agent subsidies is ending.
Jason’s read: I got the cost of persistent agents wrong. Replit and Lovable have large teams tuning their VMs to spin up and down, and they’ve gotten that down to about $4 to $5 per user per month. I assumed always-on agents would land near that. They don’t. Instinct could cost $20+ a month per user to work the way it’s supposed to, and at $40 or $50 a month that’s a brutal business model. Cursor’s V2, which was just reselling Anthropic tokens at a loss, would have died. If someone with Elon’s resources had to cap usage, that tells you something.
Rory’s read: Every digital product takes off when the economics don’t quite work. Remember people panicking about Facebook’s early server bill? The bet you keep making is that digital costs fall 50 to 60% a year. Venture markets are aggressive enough that you can’t wait for clean unit economics, because someone with VC money will launch 12 months earlier and grab the mind share. So almost everything new and digital looks crazy on economics for a while. The only question is whether the economics improve before the money runs out. As for Cursor: “I’d take the $60 billion and call it a win.”
Harry’s read: Power users like Harry are the worst possible economic profile. Blended across people who use it once or twice a day with light automations, the cost is probably far lower.
6. DoorDash Bought Instacart for $17B. Agents Placed Only a Few Hundred DoorDash Orders.
What happened: DoorDash will pay $65 per Instacart share, a 26 percent premium to the stock’s closing price on October 22, for an equity value of about $17B, which is less than half Instacart’s 2021 private price. Separately, DoorDash revealed that AI agents like Muse and Instinct have generated only a few hundred orders, a tiny fraction compared to its 3 million daily deliveries.
Jason’s read on the deal: I was surprised Uber Eats buying Postmates turned out accretive. I assumed you’d get a brand and a customer list and then shut it down. But DoorDash’s share grew from 44% to 59%, and Instacart brings most of the rest. If antitrust lets them through, and arguably it shouldn’t, that’s how you dominate a network. It’s the same pattern as Salesforce buying Braze. Companies circle each other for years, and the acquirer buys once the asset gets cheap.
Jason’s read on agents: The “only a few hundred agent orders” comment is self-serving. It reminds me of Brian Chesky saying housing didn’t need AI, or Parker Conrad saying B2B didn’t. When I had Uber’s president on, he was very clearly worried about agentic usage disintermediating Uber. DoorDash has two big competitors. One is Uber Eats. The other is the restaurant itself. Most restaurants offer pickup or their own delivery to avoid the fees, and an agent can figure that out for you. “Hey Harry, Lemkin’s Curry down the street will deliver for $2. You just saved $14.” Agents are an existential risk to DoorDash because their number three competitor is nobody.
Harry’s read: Harry doesn’t buy it. People love food. They love the photos, the scrolling and the discovery. Agents only make sense for automation and scheduling, like “drop lunch at the office every day.” He’s with Bill Ackman that Uber is a buy at these prices. Rory agrees that Uber’s bigger risk is Waymo disruption, not agentic disruption. Telling WhatsApp “I’d like an Uber, I’m at this spot on the map” isn’t a better experience.
7. Kalshi Raised at $40B a Day After Felony Charges, and Canada Put Up to $1.14B Into Cohere
What happened: Kalshi raised $1 billion at a $40 billion valuation, led by Sequoia Capital and Wellington Management, and just one day earlier, the State of Utah filed criminal charges against Kalshi and its two co-founders. Meanwhile, Canada’s government agreed to invest up to $1.14 billion in Cohere, and Cohere says ARR grew 165% year over year in Q3 2026 and customers rose 20x to 300.
Rory’s read on Kalshi: The business is growing like a weed. Deep down, everyone knows it started as prediction markets with some social utility and has turned into a national sports betting market. That’s why Utah is suing: sports betting is regulated state by state, and Kalshi is pushing through nationally. The takeaway is that VCs will put a lot of money into whatever is working. It raised at $22B a few months ago, so this is a classic momentum round.
Rory’s read on Cohere: This is close to the opposite case. It isn’t a momentum play, it’s a sovereign AI play. Credit to Cohere for being one of the survivors from the original group of LLM companies. In Kalshi’s case, a government wants to stop the company and put the founders in prison. In Cohere’s case, a government wants to make the company work even if the economics don’t. They’re the yin and yang of this week’s rounds. Governments can keep a company alive for a long time, but they can’t force the private sector to use the product.
Harry’s read: In 2026, not disclosing revenue is hardly a flex. “ARR grew 165%” doesn’t tell you anything without the base.
Jason’s read: Sequoia led a $1B round right after felony charges against the founders. And Binance just had $6.6B taken, reportedly by North Korea, in a matter of hours. If that number had been $13B, I don’t think Binance could have covered it, and they’d have been one step from an FTX situation. It made me wonder whether we’ve all just changed how much risk we’re willing to take. Rory’s addition: FTX’s $500M Anthropic position would be worth about $39B at a $1.5T valuation. Timing matters.
8. A Chinese Open-Source Pen-Testing Tool Called ARTEX Was Turned Into a Cyber Weapon
What happened: State-linked Chinese hackers used the open-source ARTEX AI penetration testing tool to compromise a series of Microsoft 365 environments linked to ConnectWise’s Cloud Backup offerings, Microsoft said in an advisory. Jason also flagged a WorkOS breach that Michael Grinich posted about this week.
Jason’s read: This sounds minor, but it isn’t. ARTEX is only a couple of months old. It’s free and open source, you plug it into DeepSeek or another LLM, and it tries to hack you, which is what you’re supposed to do to yourself. Turning it around and attacking other people at scale was easy. Attackers could always do this, but now they can do it with LLMs and aim it at anyone. Black October was much blacker than we made it look.
Rory’s read: It’s hard to overstate how wild the cyber dynamics will get over the next 12 months. At some point something truly weird will happen, like Fedwire going down for a day, and everyone will say we should have seen it coming. Harry pointed to the recent European hack where someone could remotely raise refrigerator temperatures. Now imagine that with traffic lights or heating systems.
9. IC Pick: Serval at $3B, Cognition at $80B or Baseten at $30B?
What happened: Serval has raised $150 million in a Series C round at a $3 billion valuation. Sequoia led the round, up from $1B in December. Just weeks after announcing a funding round at a $48 billion valuation, Cognition has received proposals from investors suggesting a valuation between $70 billion and $80 billion. And Baseten’s valuation stood at $13bn in June, after a $1.5bn round; it’s now set to raise at up to $30B.
Jason’s pick: Cognition. The one thing in Anthropic’s Q3 that never stops working is coding agents. Cognition made a lot of its own luck. I’m not sure the original Devin worked when I brought it to committee, but the Windsurf acquisition, the IP and open weights all broke their way. We’re already up 5x on paper. Close out the fund with the last $2B in Cognition, and add any recycling we get before the round closes. My only worry is that getting a 10x from here means nearly a trillion-dollar outcome, for a company that owns only part of its own IP stack.
Harry’s pick: Serval. Bill McDermott is great, but like Slootman he’s a sales-led CEO, and we’re in a product-led generation. Harry says Serval’s enterprise adoption is the best he has ever seen, and a 10x from $3B looks very likely. The catch: Serval only raised $150M. If you’re a16z and need to put $500M to $1B to work, you do Cognition because you have to move money. On multiple alone, Serval is the easiest call. Between the two inference names, Harry would pick Modal or Fireworks over Baseten.
Rory’s read: It’s a great question because it forces a choice between three different bets. With Serval, you can point to an incumbent worth something like $110B, about 30x Serval’s price, and say you’re going after that value. With Cognition, there’s no standalone coding agent worth $80B today, so you’re betting on a new market where Cognition is fourth behind Codex, Claude Code and Cursor. Rory is also an investor in Factory. And ITSM is bigger than people assume, because ServiceNow has become the to-do list for legal, finance and every other back-office function.
Jason’s challenge: Does Serval actually expand the TAM, or is it limited to a slice of ServiceNow’s? I think 2027 will be the year a lot of vertical AI plays turn out to be nifty startups that never expanded TAM. They’re competing for slices of ServiceNow, Salesforce and Atlassian against five to eight times as many competitors, while the incumbent still has a pulse. Rory’s test: look at ACV on a like-for-like basis. If a mid-market company spends $200K on ServiceNow, will it pay $300K to $400K for Serval because Serval resolves tickets ServiceNow can only track? That’s what TAM expansion would look like.
10. Weil Gotshal Is Dissolving, and AI Law Firm Crosby Bid for Its Brand
What happened: Crosby, a two-year-old AI-first law firm, has bid for some of the assets of Weil, Gotshal & Manges, including the 95-year-old firm’s trademark, brand and unspecified internal data. Crosby is also working to hire some of Weil’s roughly 1,200 lawyers.
Rory’s read: Weil collapsing is a law story, not a tech story. A big firm is a collection of connected rainmakers. When the person running a practice like private equity, which might be 20 to 30% of the firm, walks out with his book of business, things can unravel fast. Rory also thinks Crosby’s model is the counterpoint to Weil, not the same story. Some legal work depends on one trusted partner at $5,000 an hour. A lot of it is NDAs and Series B stock purchase agreements, where you don’t need to know who did the work.
Jason’s read: I loved the chutzpah of bidding. But I’ve personally lost seven figures to incompetent junior lawyers on “routine” deals. More to the point, Crosby still runs on people. Why would a good lawyer take $170K and 0.01% of a company that may or may not IPO when they could make $250K starting at a big firm and $10M as a partner? Crosby can build a business. I don’t see a $10B one, and $10B is the minimum exit to take anything to IC in 2026.
Harry’s counter: For every lawyer who wants to be a 24/7 rainmaker, plenty want 40 hours of good legal work and to go home to their kids or an aging parent. Crosby will be able to hire.
11. Physical Intelligence Is Raising at $20B With No Revenue
What happened: Physical Intelligence is negotiating at least $1 billion in financing, with a valuation expected to nearly double to about $20 billion. Existing investors Avenir and Lux Capital are in talks to lead the investment. In the same week, Dropbox cut 9% of staff and its CTO left.
Jason’s read: Physical Intelligence, with little or no revenue, is worth $20B. Dropbox, with billions in revenue, isn’t worth much. Revenue multiples don’t get you anywhere. It’s all about the future.
Rory’s read: Rory has done robotics deals for 10+ years, and he says the dexterity of the new models is a real breakthrough. But robots take far longer to deploy than LLMs, and there are only about 300,000 robots in the US. If you charge something like $3 an hour per robot for model access, you need far more robots to justify $20B. Rory expects the foundation model companies to buy their way in. It makes for an easy prospectus line: 81 million knowledge workers plus roughly 87 million physical workers, and you’ve doubled your TAM. If you try to reason step by step to deployment, you get lost in the details. The investment thesis is that the technology and the teams are amazing and it will change the world somehow.
12. Venture Needs M&A Like Clockwork: Nvidia–Hugging Face Under DOJ Review, Salesforce Buying Cresta
What happened: The DOJ is asking how Nvidia’s $12.93 billion Hugging Face purchase would affect AI competition. Meanwhile, Salesforce today announced it has signed a definitive agreement to acquire Cresta, the voice-first operating system for customer experience.
Jason’s read: Politics aside, the past year felt like normal M&A was back: normal deals, normal tax treatment, no convoluted team-and-IP structures. If Hugging Face doesn’t sail through, that’s a bummer for all of us. We need Salesforce, ServiceNow and Microsoft buying companies for $1B+ every couple of days. Otherwise all our liquidity rides on IPOs. Salesforce’s deal machine also isn’t Marc betting the company. I’ve been in those rooms. You go around the table and ask every SVP for their top three ideas. Cresta was the service team’s best idea. These are SVP chips.
Harry’s read: Compare Cresta with Listen Labs. Listen Labs was a ~$20M ARR company bought for ~$2B. Cresta is well over $200M ARR. The price tags are similar, so the revenue multiples are about 10x apart, and Sequoia is an investor in both. Rory’s read: That’s the buyer’s real decision. Do you pay a scary multiple for the ultra-modern LLM-first company, or about 10x for a more mature company with real enterprise scale? When you trade at 7x, spending heavily at 40x is hard. Cresta looks like a smart deal for Service Cloud.
Jason’s last point for founders: If you’re confused about why acquirers aren’t interested yet, most of them want at least $100M of revenue before they’ll put their go-to-market machine behind a deal. Below that, either buy a small team cheaply or wait. Fin’s agentic business had to hit $100M before it became material to a $50B company. Until then, the Borg can’t even assimilate you. You’re too small.
Quotable Moments
Jason Lemkin
- “Agents are an existential risk to DoorDash because their number three competitor is nobody.”
- “Nothing got cheaper until Haiku 5.5. I saw a 72% decline for a B2B workflow running constantly, for the same task, and higher quality.”
- “2027 will be a year where we find out a lot of vertical AI plays are nifty little startups but didn’t in the end expand TAM enough.”
Rory O’Driscoll
- “I’m using GAAP numbers here, not run rate, because GAAP numbers are audited and the CFO will go to prison if they’re wrong.”
- “Almost everything that’s new and digital will look crazy for a while in terms of economics. The only question is can the economics get better before the money runs out.”
- “If you’re going to regard Cursor as a failure, then yeah, you can win the argument. I’d take the $60 billion and call it a win myself.”
Harry Stebbings
- “Bill McDermott’s amazing, but he’s like Slootman. He’s a sales-led CEO. We’re in a product-led generation.”
- “If I’m Andreessen, I want to put $500 million, a billion out the door. I need to do Cognition because I need to move money. But for the multiple, it’s the easiest decision.”
- “The whole world does not know what five years out looks like in a way that we generally have for the last 30, 40 years.”
