Latest / Super Intelligence News and Report / Day 19
Transcript
- Super Intelligence News and Report — Day 19 Welcome to Super Intelligence News and Report. Today is Sunday, November 8th, here are the headlines. Theo: Today's split screen: the capabilities keep jumping, while money and politics keep cutting what investors will pay for AI. Maya, you've got a rumor that moved a whole stock. Maya: I do. On Friday, November 6th, The Information reported that TSMC has started planning 3D DRAM manufacturing at Nvidia's request. That's memory cells stacked vertically. That's reported, not confirmed. According to the report, the effort is run by Hsu Kuo-chin, an advanced-packaging VP who used to run Micron's Taiwan operations. On Saturday, TSMC said it's a pure-play logic foundry and intends to stay one. That's a denial, though it never addressed Nvidia's request directly. Theo: And the stock? Maya: TSMC's US shares rose about six percent on Friday, a day after falling nearly eleven. The memory shortage is now moving stocks on rumor alone. And Tom's Hardware points out that 3D DRAM couldn't ease the shortage for at least five years anyway. Theo: Over to the capabilities side of the ledger. On Friday, a startup called Physical Superintelligence announced a proof of the Aaronson–Ambainis conjecture, an open problem in quantum computing from around 2008. Their research system was built by David Turturean and runs an unnamed Anthropic model that isn't public. It had a complete proof by day three, checked in the Lean proof assistant. The AI wrote the statements, the draft and the Lean code. Turturean reviewed it and takes responsibility for it. Maya: Translate the result for me. Theo: On most random inputs, quantum computers get no exponential speedup on yes-or-no questions unless the problem has special structure. Factoring isn't affected. Stanford's Guy Blanc called it a huge result. It's not peer-reviewed yet, but this is frontier math from a model none of us can use. Maya: Speaking of factoring. Theo: Right on cue. Also on November 6th, Google DeepMind said AI agents guided by human cryptanalysts found two new weaknesses in old 1024-bit RSA, and it took them weeks. One makes forging digital signatures roughly sixteen hundred times more efficient. The other speeds up factoring keys in bulk. DeepMind's Pushmeet Kohli says RSA-2048 isn't affected, and no live system has been named. Maya: So should anyone be worried? Theo: If you're still on legacy keys, yes. The bigger story is that for the second week running, a frontier lab has turned AI on real security assumptions instead of a benchmark. One more from me. On November 5th, Mira Murati's Thinking Machines Lab launched Inkling 2 Small. It's a large mixture-of-experts model with image input, available for now only through their Tinker platform. It scores thirty-one on the Artificial Analysis Intelligence Index. That would make it the top open-weights model from outside China. Maya: Would? Theo: Would. The launch post promised Apache 2.0 weights "next week," after external red-teaming. Their overview page now says "coming weeks," pending more safety testing. That's a quiet softening. If the weights ship, it's the strongest Western open-weight model. Watch whether they land this week. Maya: From weights to concrete. On Friday, Crusoe announced that Google has acquired its data-center site in southeast Wyoming. TechCrunch reported that Google had raised concerns about cost and timing under Crusoe's management. No price was disclosed, and the two still work together on Google's campus near Amarillo, Texas. Theo: How big is the Wyoming site supposed to be? Maya: Google has announced plans to invest seventeen billion dollars there. When a buyer that size takes a project back in-house over timing, it tells you the constraint now is building data centers on schedule, not paying for them. And that brings us to the market's mood. Nvidia closed Friday with its sixth straight weekly decline, even though US tech partly rebounded that day. Reports put it as much as a quarter below its record high. Theo: Six weeks. What's it trading at now? Maya: About twenty-four times trailing earnings, against a five-year median around fifty-five. AI's flagship stock is priced like an ordinary large-cap. And the timing is loaded. Nvidia reports on November 17th, and Jensen Huang has a closed-door House briefing the next day, right in the middle of IPO marketing season. Theo: I'd call that cheap, not broken. Maya: Earnings will tell us which. Next, the first US humanoid-robot stock. Agility Robotics, which makes the Digit warehouse humanoid, started trading on Nasdaq as AGLT on Thursday, November 5th. It came public through a merger with Churchill Capital Corp XI, a SPAC, and that deal brought in four hundred sixty-five million dollars of new capital. Theo: And day one? Maya: It closed about ten percent below where it opened. Public investors finally have a direct way to price the humanoid boom, and their first reading is cautious. Theo: Last one, and it's easy to miss. On Friday, Meta introduced Spaces for its Muse agent. These are persistent areas for each project, like trip plans, gift tracking or events. You describe a goal, Muse builds a tracker or a board, and connected apps keep it updated in the background. Zuckerberg says he's used it for a game, an event tracker and a home-control dashboard. Muse Voice rolled out in the US the day before. Those claims are Meta's own, but the direction is clear: Muse is moving from one-off answers to an agent that keeps working on your behalf. Maya: Which ties nicely to what's next. Up next, we're going deep on Anthropic pressing ahead with its IPO into an AI selloff, and the lower price it seems willing to accept. Now for the Main Episode Maya: Here's the question. Will public investors buy a frontier AI lab at more than a trillion dollars, one week after the AI trade cracked? And whatever price Anthropic ends up accepting, what does it tell us about how the market values AI right now? Theo: And they're not waiting for things to calm down. They're going anyway. Maya: That's how it looks. A quick reset. Anthropic filed publicly for its IPO in late October, and NPR called it potentially record-breaking. Then on November 2nd an amended S-1 came out with the price range left blank. Theo: Which is normal at that stage, right? You fill it in when you launch the roadshow. Maya: Normal, yes. What isn't normal is what the money is for. That filing lists about eight hundred ninety-five billion dollars in commitments. That's roughly eleven times the company's annualized revenue. Theo: So most of what they'd raise is already spoken for. Maya: Already promised to someone, yes. Anthropic needs public-market capital to pay for a very large compute build-out. The IPO isn't a victory lap. It's how they fund the next few years. Theo: That's the part people underrate. The quantum proof from the headlines came from an unreleased Anthropic model, and everything like that runs on this compute. If the raise comes in light, the build-out has to bend somehow. Maya: And the revenue side is genuinely strong. The filing showed third-quarter revenue of almost seventeen billion dollars. A year earlier it was under one and a half billion. That's more than ten times in a year. Theo: And the eighty-one billion figure I keep seeing? Maya: That's the annualized number. The Decoder reported about eighty-one billion, up from about sixty-five at the end of July. But it's a run rate, recent revenue projected over a year, not booked annual revenue. The booked number behind it is that seventeen billion for the quarter. Theo: Still huge. I just think people quote eighty-one like it's a past fact, when it's really a bet that the next twelve months look like the last few. Maya: Hold onto that, because a missing revenue number becomes a big deal in a minute. So that's where things stood. Then came Thursday, November 5th. Theo: The rout. Maya: A tech selloff driven by three things at once: AI spending plans, rising borrowing costs, and the Democratic sweep in the midterms. Democrats had campaigned partly on slowing data-center construction. The Nasdaq 100 fell six point four percent that day and only partly recovered on Friday. Theo: Six percent in one session for the Nasdaq 100 is not a wobble. Maya: No. And over the weekend, one market newsletter called it the AI trade's first real political shock. Its point was that the worry has shifted. It used to be whether AI demand is real. Now it's whether data-center capacity can physically get built. Theo: Which is the Crusoe story from the headlines. Google acquired Crusoe's Wyoming site after reported concerns about cost and timing under Crusoe's management. The bottleneck is concrete and power, not appetite. Maya: And Anthropic's own compute plans run into that same constraint. You're asking investors to fund a build-out right as the market starts doubting build-outs can happen on schedule. Theo: So what did the investors actually say? There's been a lot of noise. Maya: Late on Thursday, Reuters had an exclusive. Four people familiar with the investor talks said potential buyers doubted Anthropic could start its roadshow the week of November 9th and list this month. And a flag that covers all the investor reporting today, from Reuters, the FT, Bloomberg and The Information: it's anonymous investors, so reported, not confirmed. Those buyers wanted a better valuation. And they were waiting on an October revenue figure that the latest S-1 leaves blank. Theo: Ah. That's the run rate thing again. Maya: That's it. If your story is a run rate that climbed from sixty-five to eighty-one in a few months, the obvious question is whether October kept climbing. Without it, investors are pricing the trajectory blind. Theo: That's a fair ask. If I'm building on Claude, I care about the model. If I'm buying the stock, I care whether the curve held through a rough month. Maya: Reuters' sources also said Anthropic could wait until AI markets are strong enough to absorb an offering some expect to raise around a hundred billion dollars. Which is a polite way of saying the market might not take that much right now. Theo: So that's Thursday night. Then what? Maya: Then the Financial Times, citing several investors, reported that Anthropic still plans to list before Thanksgiving and is willing to accept a lower valuation to get it done. The target is to raise eighty to a hundred billion dollars, at a valuation of one and a quarter to one point seven five trillion. Theo: How does that compare? Maya: It's above Anthropic's last private valuation of about nine hundred sixty-five billion. And well below the two trillion the most optimistic investors were floating in October. Theo: So the FT and Reuters flatly contradict each other on timing. Maya: They do. Then on Friday, The Information reported the formal roadshow goes ahead this coming week. A relay of that report said the listing target is still the week of November 16th, but outlets couldn't check that date against The Information's own text, so treat that date in particular as reported, not confirmed. Anthropic declined to comment on the investor doubts. Theo: Help me with the price talk, because the numbers have been all over the place. Maya: The slide is the story. In early October the talk hit two trillion. After Anthropic briefed investors on its finances, Forbes says it settled around one point four to one point five. Theo: So the briefing alone took a chunk off. Before any selloff. Maya: Before any selloff. After the rout, Bloomberg reported some prospective investors were talking one point two to one point three trillion, with some talk as low as one point one. And some told CNBC they now expect one to one and a quarter. Theo: So two trillion has turned into something like one and a quarter in about a month. How bad is that, really? Maya: Jon Markman, writing in Forbes, has the cleanest way to see it. At two trillion, buyers would pay nearly thirty times third-quarter revenue, annualized. At one point three trillion, it's about nineteen times. Theo: So about a third less per dollar of sales. Maya: Roughly a third. That's the real number under all the trillion headlines. Investors haven't decided Anthropic is a worse business. They've decided to pay a third less for each dollar it brings in. Theo: And nineteen times sales for something growing more than tenfold a year doesn't sound crazy to me. But I'd have said the same about Nvidia six weeks ago. Maya: And that's the backdrop. Nvidia's at about twenty-four times trailing earnings after six straight losing weeks. The whole sector got re-rated, and Anthropic is getting re-rated with it. Theo: Now here's the bit I find most interesting. OpenAI. Maya: On Friday, Reuters reported OpenAI closed an initial eleven billion dollars at a one and a quarter trillion valuation, with SoftBank and MGX in. It had first sought up to thirty billion at one and a half trillion. The full round may end up around twenty. Theo: So OpenAI came down too. Smaller raise, lower price. Maya: And look where it landed. One and a quarter trillion is exactly the bottom of Anthropic's reported IPO range. The two leading labs are converging on similar prices. Theo: That's a reference point every investor in the roadshow will have in their head. If you can buy OpenAI privately at one and a quarter, why pay much more for Anthropic publicly? Maya: There's a counterargument. Public shares are liquid. You can sell them tomorrow, and some investors pay a premium for that. Theo: Liquidity cuts both ways, though. You can also sell them tomorrow at a loss, in public, where everyone sees it. Maya: Which brings us to the actual debate. Nobody serious is saying cancel the IPO. I couldn't find a single named analyst arguing that. The fight is about price and timing. Theo: So who's on which side? Maya: Max Kettner, a strategist at HSBC, told Bloomberg Radio, as Axios reported, that a conservative valuation, even one below a trillion dollars, might be the best way for this IPO to lift the whole AI trade. Theo: Below a trillion. That's further than anyone in the investor talk went. Maya: It is. His logic is that a cheap deal that trades up becomes good news for the whole sector. You give away value to buy back confidence. Theo: I get the theory. But the company isn't a charity for the sector. Anthropic has almost nine hundred billion in commitments to pay for. Every hundred billion of valuation it gives up is real money. Maya: That's the strongest objection, and I think it's right, with one wrinkle. If they raise a fixed amount, a lower price means selling more of the company to get it. That's dilution, not necessarily a smaller check. We don't know the share count, so we can't say which. Theo: Fair distinction. Who else is on the go-now side? Maya: Alexandre Baradez, chief market analyst at IG. He says the selloff adds risk. But listing after inflated valuations have already been worked out of the market could actually be better for the deal. Theo: I like that one more. It's less about rescuing the sector and more about not being the company that priced at the top. Maya: Right. If the froth is gone, buyers are closer to a floor. On the other side there's Michael Alfaro, chief investment officer at Gallo Partners. He spoke before Thursday's selloff. He said Anthropic should price at one point three to one point four trillion to build a high-quality order book. Theo: Meaning long-term holders, not people flipping it on day one. Maya: Exactly. His line is that how it trades after the listing matters more than a trillion-dollar IPO headline. He doesn't want a quick pop followed by a sharp drop. Theo: Which is roughly the mood we just saw with Agility, closing ten percent below its open on day one. Much smaller deal, totally different company, but public buyers aren't handing out the benefit of the doubt. Maya: And notice Alfaro's range came before the rout and already sat below where talk settled after the October briefing. He was asking for discipline even when the mood was good. Theo: And then the skeptics. Maya: The four unnamed investors who talked to Reuters. A deeper discount may be needed, the October number is missing, and Anthropic might be better off holding until AI markets recover. Theo: Where do you land? I'm closer to Baradez than Kettner. Price it fairly, don't price it to save Nvidia. Maya: I'm closer to Alfaro, which isn't far from you. I think waiting is the weakest argument. The compute bills don't wait, and nobody knows the market will be kinder in three months, especially with politics now in play. Theo: I'd push on that a bit. Waiting isn't crazy if the issue is one missing month of revenue. Publish October, let people see the curve, then go. That's weeks, not a quarter. Maya: Fair, and that might be the cleanest fix. Put October in the amended filing. It answers the Reuters investors' biggest complaint without waiting for the market to heal itself. Theo: So what's still unknown? Maya: Quite a lot. No price range, no share count, no October revenue. We don't know for sure the roadshow starts this week. And we don't know whether eighty to a hundred billion can actually be fully placed in this market. Theo: That's the biggest one for me. Forget the valuation. Can the market even absorb that much stock? Maya: Agreed. So my read. The answer to our question is probably yes, but at a price set by the market, not by October's hopes. The talk is converging on roughly one and a quarter trillion, the same as OpenAI's private mark. That's what a frontier lab is worth this month. Theo: And the discount isn't really about Anthropic's business. Revenue is growing more than tenfold, and the models keep surprising people. It's about whether the compute behind that growth can get built, with politics and construction schedules in the way. Investors are discounting the build-out, not the model. Maya: And the real test is after the bell. Alfaro's right about that. A trillion-dollar headline followed by a slide would hurt the AI trade more than a modest price that holds. Theo: Then the room it sells into matters. TSMC reports October sales on November 10th, and Nvidia reports on the 17th, right in the middle of the roadshow. Maya: If Nvidia steadies the sector, Anthropic gets a better room. If it doesn't, the bottom of that range starts to look like the ceiling. Theo: And the first real clue could come any day now. If the roadshow starts this week, an amended filing with a price range should land with it, and we'll see whether October revenue is in there. Maya: That's where we pick this up tomorrow. Whatever number Anthropic files becomes the public market's first real price for a frontier AI lab, and everyone else gets measured against it. Thanks for listening to Super Intelligence News and Report. Please subscribe and share it with your friends.