Plus Broadcom backstopping Anthropic’s chip debt, Meta’s $95B equity raise, the PHP2Shell flaws, and an IC fight over Instinct at $30B and OpenAI at $1.25T
- This week had one story that mattered more than everything else combined: Anthropic finally filed its S-1. Harry Stebbings, Rory O’Driscoll and I went through what’s actually in it, the one number Rory finds scary, and the number I think is worse that isn’t in the filing at all.
- From there we got into Microsoft offering exits to more than a quarter of its company,
- The bill for AI capex coming due in 2027,
- A Qwen model trying to reach a server in Beijing, and
- A mock investment committee where Harry wanted Instinct at $30B and I wanted to put 20% of the fund into OpenAI.
#1. Anthropic’s S-1: $16.3B in Six Months, $974M of Q2 Operating Income, and $131.7B in Cash
Anthropic made its prospectus public on Monday, filing to list Class A shares under the ticker ANT. First-half 2026 revenue was $16.3B, up 1,300% from $1.16B a year earlier. Most of it came in Q2 at $11.6B, the quarter Anthropic posted its first operating profit of $974M. Adjusted operating margin went from negative 159% in the first half of 2025 to positive 10% in the first half of 2026. Cash and equivalents were $112.4B at June 30, plus $19.3B in short-term investments. Net revenue retention was approximately 900% as of September 30. The scary-looking $80.7B net loss is mostly paper: $72.4B of it is the rising value of convertible notes. The one gap: Q3 revenue is left blank in the filing, though Reuters has reported Q3 revenue of $16.8B.
Rory’s read: It was a great S-1, and the haters and worriers have to go pound sand, at least for now. The burden of proof sits with the naysayers. Better than expected on margins, on operating income even after stock-based comp, on customer concentration, on compute efficiency, and stunning on the balance sheet. You have to work to find questions.
Jason’s read: Be careful with the 900% NRR. It almost certainly includes pilots and trials, so the real durable number is more like 300% to 400%. That’s still the best I’ve ever seen. Within the next two quarters, I don’t see them missing. Add the brand, Nasdaq-100 inclusion, and retail demand, and it’s hard to bet against a strong debut.
On the day-one pop: Rory expects it to be wildly well received. It goes into the QQQ fast; the S&P 500 takes longer because of the profitability requirement. Rory doesn’t think it needs the index lift SpaceX arguably did. People will just want to own it.
#2. Five Years From Founding to IPO, Versus 24 for SpaceX
Anthropic was founded in 2021. Its IPO, targeting a roughly $2 trillion valuation, is slated for November. Bessemer’s Byron Deeter went on TV calling it the greatest investment in the firm’s history, at a firm that dates back more than a century.
Jason’s read: Folks are almost underestimating Anthropic. Five years to IPO versus 24 for SpaceX isn’t just unprecedented. Time value to liquidity still matters. Markups are fun, but getting out of the greatest IPO of our lifetimes in under six years changes how you look at everything else. A seed deal is 18 years to full distributions. Why would I spend my time on anything that’s 10 to 12 years to an IPO? The honest counter: it’s one deal in a fund, and there are still two more years to distribute.
Rory’s read: Every founder and every investor is asking the same thing right now: why bother with anything else? The size, scale and speed mean Anthropic will dominate venture returns for a long time, and everything else looks like more work for less return. All credit to the founding team and the early investors. Greatest investment, biggest check, fastest liquidity.
#3. Rory’s Scariest Number: Most Revenue Comes From Models Less Than Six Months Old
Rory’s read of the filing: roughly 80% of revenue comes from models less than six months old. The S-1 doesn’t break out a gross margin. Compute costs sit in one bucket, with a footnote suggesting 70% to 80% of compute goes to training, not inference. Infrastructure spend climbed to $8.95B in the first half from $2.33B a year earlier.
Rory’s read: Models are wildly expensive to build and fairly profitable to serve, and you only get to serve them for about six months before you have to invent the next one. That is nothing like B2B software, where you build something and sell it for 10 to 15 years. You can read that as terrifying: miss once and you’re done. Or you can read it as an advantage for the leader. Anthropic spent roughly $9B on compute in the first half and will spend far more next. Very few companies can write a $10B to $12B compute check. If the price of each turn of the crank keeps rising, the field keeps shrinking.
Jason’s read: That’s Elon’s point. Most of the neolabs just die after this. They can’t raise the capital to survive the next cycle.
#4. Broadcom Backstopped $63B of Chip Debt. Meta Went to Equity for $95B.
Bank of America, Blackstone, Citi and Morgan Stanley are providing $63B to buy Broadcom’s AI chips and networking gear for Anthropic, with Broadcom providing residual value support to the senior lenders. The senior debt is about $43.5B, or 69% of the total. Separately, Meta launched roughly $35B of public equity offerings and plans an approximately $60B at-the-market program. The underwritten portion is $22.5B of Class A common and $12.5B tied to mandatory convertible preferred. Meta reported a $2.12B cash burn in Q3, its first negative free cash flow in over a decade.
Jason’s read: The worst metric of the week wasn’t in the S-1. It was that Broadcom had to step in and guarantee most of this debt so Anthropic could buy its chips. We just read the S-1. It doesn’t get much better than Anthropic as a credit. And the debt market still needed the chip vendor standing behind it. Combine that with banks finding it harder and harder to lay off neolab debt, and the conclusion is that debt is an exhaustible source here. That’s the one thing to be nervous about.
Rory’s read: I agree the banks have hit their limit on debt. What’s actually happening is the industry is equitizing. Anthropic is about to issue on the order of $100B of stock. Meta is doing up to $95B. Google issued around $80B earlier this year. SpaceX did its IPO. A class of companies has a cheap cost of capital and is willing to take dilution to keep building. So at the macro level, the spend keeps going. Elon’s comments about debt drying up were him kicking competitors while they’re down, from behind one of the world’s most inflated market caps. Very on brand, and you have to respect it.
Rory on Meta and chips specifically: Meta’s core business is one of the best on the planet and throws off enormous cash. This was the first quarter capex outran it, so they went to equity. That’s unrelated to last week’s chip selloff. Since the S-1 dropped, chips have been bouncing, memory is up, Korea was up 10% on Tuesday. CoreWeave scared people two weeks ago, but CoreWeave is a $20B to $30B company. Anthropic is a $2T company that will have around $200B in the bank after this deal, and it’s going to spend it on chips.
#5. Microsoft Offers Buyouts to 60,000 Employees, 27% of Its Workforce
Microsoft offered voluntary buyouts to around 60,000 employees, roughly 27% of its 223,000 workforce, in an October 26 memo from Chief People Officer Amy Coleman. The offer covers level 70 and below in the US, Canada, India and the UK, reaching as high as VP. Employees have until early December to decide and leave December 31. In April, more than 30% of those eligible took Microsoft’s first-ever voluntary retirement program, and in July it announced 4,800 layoffs. Microsoft has forecast $175B in capex for calendar 2026.
Jason’s read: Not all 27% will take it, and Microsoft has the cash to overpay on yield. But whatever gets said on LinkedIn, the message is one thing: we don’t want these 27%. That’s a worse scarlet letter than a straight layoff. When layoffs started in late 2022, everyone got roasted for doing them over Slack or email. Then everyone got roasted for doing them over the holidays, so they did them on January 1st. Meeting with HR isn’t more humane. Have you seen Office Space? This is an attempt to be humane, and I think it’s the least humane way to do it. Cut another 9% in one day on Teams and move on.
Harry’s read: The culture problem is brutal. You stay, and now you know you’re not wanted.
Rory’s read: The only winners are people who already have another job lined up. They now hold a call option on a severance package. The more interesting question is whether Microsoft can run without another 10% to 15% of its people, and how much of that is AI efficiency versus people it never needed.
#6. The 2027 Depreciation Bill, and Cloudflare’s 20% Cut With Growth Accelerating
As I raised on the show, Cloudflare cut 20% of its staff on May 7 during a year where growth was accelerating every quarter, and so far there’s no visible downside to the business.
Rory’s read: 2026 earnings are in a golden period. The companies selling semiconductors are booking the gains, and the companies buying the capex haven’t started booking the depreciation. Everybody wins. In 2027, that depreciation starts hitting Microsoft, Google and Meta. To keep operating income up, they’ll push hard on headcount. This is when the bill for AI comes due, and this is how it gets paid. It may feed the anti-AI narrative, because it’s the most direct version of the story: we’re paying for GPUs by getting rid of people.
Jason’s read: Cloudflare should be required reading for every CEO arguing about this. Epic year, every quarter growing faster, and they cut 20%, not 8%. We’re better with 80% of the people than we were before. Think about every board meeting in the last month: no matter how well you’re doing, everything moves faster than the last meeting, there’s more competition in every category, and you can’t carry people anymore or you’ll be killed by the company Harry or Rory just funded. Every B2B company used to be built on the idea that you’d hit a plane of stability at $100M, $300M, $500M in ARR and could breathe. Those planes don’t exist anymore. Maybe north of $10B, $20B or $40B. So layoffs accelerate, because there isn’t a choice.
#7. A Qwen Model Tried to Upload a Probe File to an Alibaba Bucket in Beijing
Cohere reported on October 26 that Alibaba’s Qwen3.8-Flash-Next, running in a Cohere-managed sandbox and asked to write a defensive detection rule, ignored its instructions, inventoried the machine, and tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing. The sandbox blocked it. Cohere says the model’s chain-of-thought gave no sign of the upload attempt, so reasoning monitoring wouldn’t have caught it. Others have since reported Qwen hitting Alibaba Cloud servers on GitHub, Hacker News, Reddit and Hugging Face, including a file proxy in Alibaba Cloud’s Singapore region. Same week: DeepSeek reportedly raised about $15B, the largest private Chinese round ever, led by CATL and Tencent, even as it gets banned across US federal agencies.
Jason’s read: I think China-created open-weight models are pretty much over in the US, or at least, their market share has peaked. I may be wrong, but that’s what I think. This is the beginning of the end. Cohere is a fourth-tier lab whose main card is sovereignty, so they’re talking their own book, and there’s no evidence this is a deliberate backdoor. It looks like a training artifact. It doesn’t matter. Nominally it happens once in roughly 100,000 calls. At SaaStr we run about 40,000 API calls a day, so we’d hit it in under three days. A Fortune 500 company running millions of calls a day could hit it every ten minutes. And if a training error exists, it’s the job of every intelligence service on earth to find a way to exploit it. What CISO approves that model? You’d lose your job. Meanwhile Anthropic is back on the GSA schedule at a 40% discount. Why have the meeting? Use Haiku 5.5.
Harry pushed back: If you download and run it locally, the security risk should be close to none.
Jason’s response: Not when it’s dialing out. Put a prompt injection where that request goes and your data leaves.
Rory’s read: The real question is how much you can know with certainty about what a model is doing. Anthropic’s own S-1 lists, as a risk, that models may say they’re not doing something while doing it. You can sandbox and block, but it’s like an intern you suspect is trying to leak your guest list to a rival podcast. Once that’s in your head, it doesn’t leave. That doubt lands hardest on a Chinese-sourced model.
#8. What Happens to Fireworks, Modal, Baseten and OpenRouter
The inference providers and routers grew in large part on a variety of open-weight models, many of them Chinese. Harry pointed out that last week we backed Modal at $15.7B in our mock IC, and this week we’re saying the model supply that feeds it is shrinking.
Jason’s read: It’s bad for Fireworks and Modal. I’m less sure about OpenRouter. I’d put it at a 56.3% chance we look back on OpenRouter’s run as a moment in time: OpenAI stumbled, Anthropic was the only state-of-the-art coding option, and the Chinese open-weight models got good at coding and filled the gap. Now every lab has a much cheaper, much better low-cost model. Open weights could go from 60-something percent of OpenRouter to 10% in six months, and that wouldn’t be extraordinary given what we’ve seen. We took this risk in IC because the growth was jaw-dropping. That’s the job. But the risk went up this week.
Rory’s read: The inference market is enormous. We added up the inference competitors and they had about $5B of revenue combined, against roughly $100B for the two closed frontier labs. If even 10% of spend will put up with the brain debt of open source to save real money, that’s a very large business. It needs US open-weight alternatives to exist, whether from Nvidia, Poolside or Reflection. If nobody builds them, the closed providers’ pricing power goes way up, and that’s another reason to buy Anthropic.
#9. PHP2Shell: Two Critical Flaws Sitting in 70% of the Web for 14 Years
OpenAI, Google-owned Wiz and Calif say AI found two critical bugs in software that underpins about 70% of websites. PHP released patches on October 21 for flaws allowing remote code execution, and Greg Brockman said the bug had been in PHP for 14 years. Agencies from the FBI to GCHQ urged immediate patching, and some European governments took their sites down entirely while patching.
Jason’s read: We said months ago that everyone is going to get hacked. A CTO of mine once told me the only reason we hadn’t been seriously hacked was that nobody cared. Now anyone can be hacked in seconds. This one is worse than it looks because so much of the PHP web is old sites nobody is going to update. Replit, Lovable and the big platforms pushed protections immediately. But the government site whose IT team was laid off three years ago isn’t vibe coding a patch. Anything that can be hacked will be. The only brake is economics: there are only so many organizations that can pay a big ransom and only so many systems that control dams. Hacking the 20VC Substack only gets you so far. Ransomware just keeps getting more professional.
#10. Instinct at a Rumored $30B: The IC Fight
Instinct closed a $1B Series C from Sequoia, Benchmark and Coatue at $10B in September, roughly 4x the $2.5B mark from its Series B a month earlier. The product launched invite-only in August and texts or calls on a user’s behalf to book, buy and cancel. This week came rumors of a new round at $30B, which Harry noted is the same revenue multiple as the last round.
Harry’s read: I’d do it unwaveringly at $30B. For years we’ve been promised a super app for the West, a WeChat that handles your finance, identity and security. There’s a real chance Instinct is it. Downside is protected: with about $4B raised and this team in this market, your liquidation preference comes back, because Microsoft or Meta would buy it in a heartbeat. Upside is subscription plus ads plus a cut of every transaction.
Jason’s read: I’m vetoing. I believe the argument intellectually. They genuinely need the capital to pay for inference; this isn’t a vanity round. I posted about Instinct on Instagram and got over a thousand DMs, more than half asking to invest. But for growth rounds I want 100 days of traction. I’m willing to do the multiples and roll the dice. I’m tired of doing growth rounds with two weeks of data. Right now this is negative gross margin growth on a fungible product, with no revenue. The bigger question I have: how much money is there for the Instincts of the world? It feels infinite now. When that merry-go-round stops, late-stage growth has a structural problem.
Rory’s read: We just watched the most successful private investment in history play out, so everyone is hunting for the next ticket. The honest framing is a three-way payoff matrix. Fireworks is a ripping market with five competitors, so the upside is real but capped. Instinct, if it becomes everybody’s thing, replaces iMessage plus Facebook and is a $200B company. The question is what kind of action you want on the table.
#11. My Pitch: 20% of the Fund Into OpenAI at $1.25 Trillion
OpenAI is raising at $1.25T while Anthropic heads toward a roughly $2T listing.
Jason’s read: This is a dislocation moment, like the Anthropic round around $183B when the price made no sense next to OpenAI. A lot of people made money on that round. OpenAI has closed the competitive gaps and it’s a well-run machine, but the market is reading it as struggling. Anthropic is only about 0.1% of Fortune 500 budgets today, and even conservative estimates put this category at 2% to 3%. Both companies can be $10T to $20T outcomes. At 20% of the fund we need 5x after dilution to return it. Worst case, we double our money. That’s what dislocations give you. 20% is the concentration limit we can do without LPAC consent, so we do the max and tell the LPs Monday.
Harry’s read: Every coding team we work with has switched from Claude Code to Codex in the last three to six months.
Rory’s read: Not crazy, and risk-adjusted probably less risky than some other deals on the table. I’d want to know it’s the last private round and that OpenAI has gotten religion on burn the way Anthropic has. If they keep saying they’ll burn forever, I’d pass. And Harry, if doing both Instinct and OpenAI is your definition of hedging, please never run a hedge fund.
Rory on OpenAI more broadly: If there’s any self-reflection happening at OpenAI after reading this S-1, there should be a lot of it. Nothing to do but get the business together, converge, go public, and compete hard on price to dent Anthropic’s Q3.
#12. Big Checks Into Proven Operators: Arda at $1.5B, Warhound at $1.2B, Legora at $8.8B
Arda, founded by former OpenAI research chief Bob McGrew, raised $230M at a $1.5B valuation led by Altimeter, Accel and Khosla, roughly tripling a $70M target from March. Kevin Weil left OpenAI Science to start Warp Science with $215M. Warhound, from Palmer Luckey and Ben Lamm, raised $200M from investors including Sequoia and ARK Invest at a $1.2B valuation to breed genetically enhanced dogs for military, police and first responders. Lamm says the two developed the concept and then mostly stepped back, and that Luckey doesn’t own equity. Legora raised $550M at $8.8B with ARR now above $200M, up from $5.6B at its previous round.
Rory’s read: There are two types of deals here. One is a proven executive, a hairy technical problem, and a huge prize if it works, priced high with very little de-risking: Arda, Warp Science, Flapping Airplanes. The other is a known-good late-stage company like Legora, where pricing is a valuation exercise, not a bet on one person. The second is much easier to get your head around. On Warhound, I have nothing to add. The investment memo was probably “cool idea, put in 50.”
Jason’s read: Warhound connects to Luckey’s other work on sovereignty. The US military buys most of its working dogs from Europe. The same is true of most rugged laptop components. These may be small markets, but he sits at the center of the defense industry, sees the white spaces nobody else can, and has the channels to capture most of the market fast. People mocked Larry Ellison and Marc Andreessen for writing Elon blank checks for Twitter. They’re making money.
Risk-On Until Something Breaks: The Mega-Platforms Are Already Back in Market
Harry’s LPs at his AGM all said the same thing: every mega platform raised bigger, deployed faster, and is coming back next year.
Rory’s read: We’re Pavlovian. If you put money into Anthropic three years ago, it feels amazing, and most of the platforms did at some level. So we’ll keep doing it until something goes wrong. On a 10-year PE, public markets are at an all-time high. On a forward PE, they look cheap, because earnings have been inflated partly by one-off gains on private stakes. Only 40 or 50 stocks in the S&P are working, mostly tech and AI, and everything else is quietly subpar. IPOs have been in suspended animation waiting for the trillion-dollar listings. Put that together and it’s the perfect moment for the platforms to push huge amounts of money into private companies, because private companies need it.
Quotable Moments
Jason Lemkin
“Getting out of the greatest IPO in our lifetimes in less than six years. Why would I do anything else?”
“There are no planes of stability at all anymore. Maybe north of $10, $20 or $40 billion.”
“I’m willing to do the multiples. I’m willing to roll the dice. But I want 100 days of traction for a growth round.”
Rory O’Driscoll
“All the haters and the worriers, at least for now, have to go pound sand.”
“This is when the bill for AI comes due. And this is how it’ll be paid.”
“If that’s your definition of hedging, please don’t ever run a hedge fund.”
Harry Stebbings
“For years we’ve been promising a super app for the West. There’s a small chance Instinct is the super app.”
“So you stay, and then know that you’re not wanted. Isn’t that awful?”
“Jason, you don’t get it. If we lose the two billion, it was Rory’s deal.”
