Latest / Super Intelligence News and Report / Day 22: Anthropic's $58B IPO Range Tests What Public Investors Will Pay
Transcript
- Super Intelligence News and Report — Day 22: Anthropic's $58B IPO Range Tests What Public Investors Will Pay Welcome to Super Intelligence News and Report. Today is Wednesday, November 11th, here are the headlines. Maya: Today's thread: public markets are starting to price the AI build-out, the price is coming in lower than the builders hoped, and the money behind it is getting more expensive. Start us off. Theo: Google opened Gemini 4 Carbon to every paying developer on November 10th, through the API in AI Studio and Gemini for business. It's also rolling out to Ultra subscribers, AI Mode in Search and the Antigravity coding tool. When Google announced it on October 28th, it was locked inside a gated cybersecurity program. Intro pricing is two dollars per million input tokens, and that doubles on December 1st. Google also previewed slash swarm in Antigravity, where one prompt spins up a team of coordinated coding agents. Google's top model is finally broadly usable, and the cheap window is short. Maya: So run your evals before the bill doubles. Okay, money. The Information reports OpenAI told investors its third-quarter revenue was about ten billion dollars, up roughly half from the second quarter. That's reported, not confirmed, since OpenAI doesn't publish financials. Enterprise sales reportedly doubled, paid subscriptions passed a hundred million, and the operating loss widened to almost eleven billion. It's the clearest look yet at the race with Anthropic, whose filed third-quarter revenue is about 1.7 times that. Theo: So roughly a dollar lost for every dollar coming in. Maya: Roughly, yeah. Then CoreWeave, after the bell on November 10th. Revenue beat, up a hundred sixty-eight percent. Backlog climbed to about a hundred thirty-two billion, from a hundred four in June, and guidance went up. But the line that mattered came from CFO Nitin Agrawal: at current yields, no new unsecured debt at the parent level through 2027. Some of its bonds recently yielded above sixteen percent. Demand is real, but the debt market is now setting the pace. Theo: And the stock liked hearing that? Maya: Up five to eight percent after hours, after sliding more than forty percent in a month. A bounce out of a deep hole. Same day, SoftBank reported a quarterly loss of about seven billion dollars, its first loss in almost two years. The biggest piece was a three point seven billion dollar write-down on advance payments for power equipment for its Ohio project and other US data centers. Shares fell as much as eleven percent in Tokyo. CFO Yoshimitsu Goto says Ohio isn't being scaled back, but how they buy equipment is being reassessed. Power is the bottleneck, and even its biggest buyers are getting the bet wrong. Theo: On equipment they won't even identify. Maya: From suppliers they won't name. Then the macro hit. October CPI came in hot, up four tenths of a percent for the month against three expected, and core beat too. Odds of a December rate hike went up, the 2-year yield jumped about seventeen basis points, and the Nasdaq slipped. Nvidia fell about five and a half percent, its second straight drop with the board's investigation of Jensen Huang hanging over it. Its relative strength index hit 24, the most oversold since 2019. Higher rates make debt-funded data centers cost more, and Nvidia reports November 17th. Theo: While Broadcom and AMD went the other way. Maya: Broadcom up almost four percent. And one more for the suppliers-funding-customers file. A TSMC filing from its November 10th board meeting approves up to five hundred million dollars each for Anthropic shares at the IPO price, for OpenAI preferred stock, and for a new vehicle for chip-industry deals. TSMC is also one of seven partners in Anthropic's twenty-six billion dollar side deal. So the world's main AI chipmaker has approved buying a piece of its two biggest end customers. That's support, sure. It's also circular. Theo: Over to research. Google DeepMind says a system called AlphaProof Code formally verified key security properties of the core of the Signal protocol in about four weeks, proving the compiled ARM machine code behaves as it should. It found no new vulnerabilities. As a check, researchers put back two old, already-fixed bugs, and it proved those versions insecure. That's about six million lines of machine-checked proof in Lean, with only around one percent flagged for human review. The caveat: selected properties, not the whole app. Still, this used to be impractical at this scale, and it follows DeepMind's RSA work from November 6th. Maya: Somebody's reviewing that one percent. What's Anthropic shipping while the bankers do the roadshow? Theo: Claude Meetings, in beta for Team and Enterprise, starting with Google Meet. Claude joins as a visible participant and drafts plans and answers in the meeting chat, and without admin permission it bows out when outside guests are on the call. Whiteboards and Sheets are in beta too. That's a chat window turning into workplace software. Maya: And the less flattering Anthropic news? Theo: The ClaudeDevs account on X gave every Pro, Max, Ultra and Team user one free full reset of usage limits, good until December 10th. Anthropic says bugs ate more of people's limits than intended, and earlier resets missed people already at their weekly caps. That's an admission of over-counting, while a class action over Claude Max usage claims moves ahead in court. Maya: I'd guess the plaintiffs have already read that post. Anything out of China? Theo: Alibaba's Qwen team put Qwen4-Max into preview, built for coding, front-end work and agents, and only available through Alibaba's Token Plan subscription and early access in its own apps. No benchmarks, no model card, no word on open weights. A million-token context window is reported, not confirmed. China's most-downloaded open model family has a new generation, and whether it ships open is the whole question. Maya: Meanwhile OpenAI is eyeing Google's oldest business. Theo: Pretty much. ChatGPT got a beta Business Hub, where companies claim verified profiles and see limited insights on how users interact with them. TechCrunch notes OpenAI didn't explain how verification works. This is the brand-listing layer that turned Google Search into a business. Maya: And ads usually follow the listings. Last one's xAI? Theo: Last one. In posts on X, reported by TeslaNorth, Elon Musk said Grok 4.8 finished its first training phase and moved to reinforcement learning, after xAI rewrote its training and serving software in C and C++, which he says caused the delay. That's reported, not confirmed. He wants a limited rollout to Grok Bot users later this week, concedes 4.8 trails Anthropic's Opus 5.5, and predicts Grok 5 beats it in two or three months. It's xAI's first concrete release signal in weeks, on Musk's clock. Maya: Next, we go deep on Anthropic's IPO range, seventy to seventy-eight dollars a share, and what it says about what public investors will actually pay. Now for the Main Episode Maya: Here's the question for the next fifteen minutes. Anthropic just set its price below what everyone was whispering about all summer. Is that a smart setup for a big first day that steadies the whole AI trade? Or is it the first hard sign that public investors won't pay private prices for a company that's promised to spend about nine hundred billion dollars on compute? Theo: And you think we can answer that before the stock even trades? Maya: Partly. Start with who's selling. Anthropic was founded in 2021 by Dario Amodei and others who left OpenAI. Its private round in May valued it at about nine hundred sixty-five billion dollars. Over the summer and into October, the talk was about two trillion. Then the talk started sliding. First to about one and a half trillion, then to one point two or one point three, after the AI and chip rout that followed the midterms. Theo: Meaning the Democratic sweep. Maya: The sweep, and data centers losing at the local level almost everywhere. Votes went against them in forty of forty-one ballot measures. So the market was marking down AI infrastructure at the exact moment Anthropic was getting ready to sell itself as the best AI infrastructure customer there is. Then, late on November 9th, a document emailed to banks showed a range of seventy to eighty dollars on about eight hundred sixty-seven million shares. Anthropic said that pricing was inaccurate. Theo: And the real one came the next day. Maya: Tuesday, on its own newsroom. Roadshow launched. Eight hundred twenty-five million shares at seventy to seventy-eight dollars, on the Nasdaq under the ticker ANT. So it's a bit lower and a bit smaller than the leak. Nearly all of those shares are new stock from the company. A sliver comes from existing holders, and Anthropic gets nothing from those sales. Underwriters can take up to about fifteen percent more on top. Theo: How much money does that raise? Maya: Roughly fifty-eight to sixty-four billion before fees. That's enormous, but it's short of SpaceX in June, which CNBC calls a record at about eighty-six billion. CNBC's sources also say pricing is planned for November 18th and trading for the 19th. Anthropic's own announcement doesn't give a first trading date. Theo: Okay, the valuation. I've seen three different headline numbers, and they can't all be right. Maya: They're all from the same filing, and Forbes points out the filing itself doesn't state a valuation. Count the shares outstanding after the deal, about thirteen billion, and the top of the range works out to roughly one trillion. Count every option and convertible security too, so fully diluted, and the top is about one point four trillion. The midpoint is around one point three. Theo: So 'Anthropic is a trillion-dollar company' and 'Anthropic is a one point four trillion dollar company' are the same sentence with different share counts. Maya: Exactly. I'd treat fully diluted as the honest number for what you're actually buying into. On that basis it's still well above May. It's just way below two trillion. Theo: So the May investors are up, and the summer chatter got cut by about a third. Maya: Roughly. Then there are two details on how they're selling it, both from CNBC's sources. First, Anthropic is targeting about twenty percent of shares for individual investors. The typical share is five to ten. Theo: That's a lot of retail. From where I sit, Claude has a very loyal paying user base. Some of the people buying the stock will be people who use the product every day. Maya: Could be, though we don't know who's actually in that bucket. Second, those same sources say the order book is already covered across the whole seventy-to-seventy-eight range. They also say institutional demand is stronger than SpaceX's was at the same stage. Theo: 'Covered across the range' is a slippery phrase. Covered at seventy isn't covered at seventy-eight. Maya: That's one of the real unknowns. It tells you there's enough demand to fill the deal. It doesn't tell you how much of that demand stops at the bottom of the range. Theo: Then there's the side deal, which is where my eyebrows went up. Maya: Twenty-six billion dollars of a separate share class, called Class N, sold at the IPO price to seven partners: AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC. The Motley Fool puts Micron's piece at a billion and AMD's at two and a half. According to CNBC and the updated filing, those shares aren't registered in the offering and aren't covered by the underwriters' lock-up agreements. Theo: So the usual 'insiders can't sell for six months' rule doesn't bind them. Maya: Not those lock-ups, no. How investors treat that block is an open question nobody can answer yet. And on top of it, TSMC's board approved up to five hundred million dollars more for Anthropic shares at the IPO price. Theo: Here's my question from the build side. These are the companies Anthropic buys its chips and cloud from. If I'm Nvidia and I'm buying the stock, am I investing, or am I making sure my biggest customer can keep paying me? Maya: Both, and that's the circularity critique. It's real support. It's also money that comes back around. The way I'd put it: it reassures you the deal clears. It doesn't reassure you about the economics underneath, because the people vouching for it are the counterparties to those contracts. Theo: Okay, then the business itself. What are buyers actually getting? Maya: Third-quarter revenue of seventeen billion dollars, up more than eleven hundred percent from a year earlier. Theo: So more than twelve times bigger in a year. Maya: Right. But here's the wrinkle the skeptics grabbed. Growth over the second quarter slowed to forty-seven percent. The quarter before that, it was a hundred forty-four. Theo: Forty-seven percent in one quarter would be a miracle almost anywhere else. Maya: It would, and Jim Cramer made basically that argument on CNBC. He said forty-seven percent still supports more than a trillion. But a valuation like this is a bet on future growth, and the growth rate just bent. Then there's the run rate at the end of October, about eighty-one billion. That's up ten billion from September. Theo: And a run rate is one recent month multiplied out to a year. Maya: One recent period annualized, not revenue already booked. Booked revenue for the first nine months was about thirty-three billion. So eighty-one is where they're heading, and thirty-three is what's actually in the bank. Theo: And the loss? I saw a hundred billion and nearly fell out of my chair. Maya: A hundred four billion for nine months, against about twenty-five billion in the same stretch last year. That one misleads on its face. CNBC says it comes mostly from changes in the fair value of convertible notes and derivative liabilities. Those are accounting remeasurements, not cash spent running the business. Theo: So part of why the reported loss got bigger is that the company got more valuable. Maya: That's the mechanism. When the company's worth more, the liabilities tied to its value get marked up, and that shows up as a loss. What we don't have is a clean figure for cash burned on operations, so I can't tell you what that is. Theo: Which makes the OpenAI comparison a little apples to oranges. Maya: A little. The Information reports OpenAI told investors its third quarter brought in about ten billion, with an operating loss of nearly eleven. Reported, not confirmed. Anthropic's seventeen is about one point seven times that revenue. Whatever else you think, that goes a long way toward explaining why Anthropic is the one testing public markets first. Theo: The number that actually gets me is the nine hundred. Maya: Nine hundred billion of committed compute over roughly the next decade. That's about eleven times the current run rate. Theo: From my beat, that's the product roadmap written as purchase orders. They shipped Fable 5.2 and Mythos 5.2 on November 9th. Those are their first frontier models since Amodei called in September for slowing AI development, and every model like that runs on this compute. And users feel compute directly in their limits. Just this week they handed every paid user a free reset because bugs were over-counting usage. Maya: That's a trust problem landing in the middle of a roadshow, with a class action already moving. Theo: Right. So for developers the IPO question is weirdly concrete. Does public money make that compute easier to pay for? Or does it put a quarterly scorecard on how generously it gets handed out? Maya: Good question, and I don't think anyone knows yet. So let's weigh the people arguing it. On the bull side, the loudest voice is Patrick Moorhead of Moor Insights and Strategy. He thinks Anthropic may be pricing the stock to soar on day one, and he puts the fully diluted value at about one point two to one point three trillion. Lior Prosor at the venture firm Deep33 told Fortune the steep discount is a smart move. Cautious AI investors get some upside, and the compute still gets funded. Theo: And Darian Shirazi had the line about buzz. Maya: Shirazi, a general partner at Gradient. He says a perfectly priced IPO creates little buzz, and a deal that falls after listing gets months of bad press. So you leave room to rise. Theo: That one I buy, especially with twenty percent retail. If a fifth of your holders are individuals watching the stock sit in the red for three months, that's a brand problem for a consumer product. Maya: And Bastiaan Janmaat of Lithos Capital says the lower valuation doesn't signal any weakness in the business. Theo: That's the part I'd push back on. The range didn't come down because Anthropic decided to be generous. It came down over weeks, as the market fell. Calling that strategy is a little flattering. Maya: Fair, and that's basically Aman Verjee's point. He's the founder of Practical Venture Capital, and he told Fortune that heavy compute spending, slowing third-quarter growth and the political sell-off in data-center states have made buyers hesitant. Dan Ives of Yorkville Ives says the valuation has to be 'digestible' and the market will decide. And shareholders told CNBC on November 6th that even one and a half trillion looked like a stretch. Theo: And the far end of the bear case? Maya: Kyle Guske the second, at New Constructs. Back in October, when two trillion was the talk, he called the deal an unprecedented test of investor gullibility. His bearish model values Anthropic at about a hundred forty-four billion. Theo: That's about a tenth of the range. Does anyone take that seriously? Maya: As a number, it's an outlier. As a framing, it's useful, because what you're paying for is a decade of growth that hasn't happened yet. TradingKey's read of the amended filing says that at the seventy-four dollar midpoint, new buyers take an immediate hit of about sixty-six dollars a share. Net tangible book value is only about eight. Theo: So you pay seventy-four dollars for about eight dollars of hard assets. Maya: Every growth stock looks something like that. This is just a lot of it. And a dual-class structure gives the founders fifty point one percent of the vote. So you're paying for the growth, and you get very little say over how the nine hundred billion gets spent. Theo: Moorhead had the question I keep coming back to. Can eighty-one billion of annualized revenue carry nine hundred billion of commitments, now that the suppliers on the other side of those contracts are also buying the stock? Maya: And the cost side just moved against them. Inflation came in hot, the odds of a rate hike went up, and CoreWeave says it won't issue new unsecured debt at today's yields. SoftBank took a write-down on power equipment. The compute Anthropic signed up for gets more expensive to build when money costs more. After those ballot measures, it's also harder to find places to build it. Theo: Does that hit Anthropic directly, or does it hit the suppliers first? Maya: The builders and cloud providers take it first. Whether it flows back as higher prices or delays depends on the terms of those contracts, and we haven't seen them. So I can't tell you how much of that risk sits with Anthropic. Theo: Then here's my read. The discount is real, but it isn't a verdict on the business. Seventeen billion in a quarter is a real business. The verdict is on the nine hundred billion. Maya: I'd go a step further. I think it's both halves of the opening question at once. It's a smart choice, and it's a sign of the ceiling. Those aren't opposites. A smart issuer reads the room, and the room said about one point three trillion fully diluted, not two. Theo: Where I part ways with you is the first day. I don't think a pop proves anything. With twenty percent retail and a book that's already covered, a first-day jump is close to built in. If it trades up on November 19th, people will say the AI trade is healed. One good day doesn't tell you whether growth picks back up. Maya: Agreed, the pop isn't the proof. The proof is the next couple of quarters. If forty-seven percent was a pause and the run rate keeps climbing ten billion a month, this range looks cheap a year from now. If growth slows again, Guske stops being an outlier and starts being the comparison people reach for. Theo: And the calendar before then is tight. Nvidia reports on November 17th, the last big AI read before pricing. Maya: And Nvidia is one of the placement partners. Jensen Huang is doing a closed-door House briefing on November 18th, the same day the deal is expected to price. So watch whether it prices at seventy-eight or above, and whether that twenty percent retail allocation survives into the final deal. Theo: Pricing above the range would be the market answering your question for them. Maya: And pricing at the bottom answers it too, just less happily. Either way, for the first time, the price of a frontier lab isn't being set by the lab or its suppliers. It's being set by whoever shows up to buy on November 19th. Thanks for listening to Super Intelligence News and Report. Please subscribe and share it with your friends.