Latest / Super Intelligence News and Report / Day 17
Transcript
- Super Intelligence News and Report — Day 17 Welcome to Super Intelligence News and Report. Today is Friday, November 6th, here are the headlines. Maya: The midterms just changed the price of building AI, and nearly everything today runs through money, coding or cost control. Theo, start us off. Theo: OpenAI's president, Greg Brockman, got unusually candid at WSJ Tech on November 4th. He said the first half of this year was, his words, pretty rough, while Anthropic pulled ahead in enterprise coding. Then Codex took off. Business customers now bring in more than half of OpenAI's revenue, up from about forty percent earlier in the year. He also disclosed a JPMorgan Chase deal to use Codex, and said he thinks OpenAI can be ready to list in 2027. Maya: Ready to list. That's a carefully chosen phrase. Theo: It is. And there's more around it. Cat Wu, who ran product for Claude Code and Cowork at Anthropic, posted on X that she's joining OpenAI to lead product for Core Products. AI Weekly counts her as the third Claude Code departure in about two weeks. Axios says OpenAI is buying Pierre Computer, a Git storage startup, to fold its tech into Codex. That one's reported, not confirmed by OpenAI. Put it together, and the coding war is now about people and plumbing, just weeks before Anthropic's IPO. Maya: From one company scrambling to catch up to another. Qualcomm's guidance for the December quarter came in soft on Wednesday, with the profit midpoint below estimates. It's also blaming rising costs for wafers, packaging and memory, and raising its own prices. But in the same breath it agreed to buy the AI software startup Modular in an all-stock deal worth nearly four billion dollars. It also signed Meta for its future Dragonfly server chips and struck a multi-generation chip deal with Amazon. Theo: So the phone-chip company is buying a software stack. Maya: Exactly. It's buying its way into data-center AI. And when even Qualcomm is passing along higher input costs, you know the supply squeeze has spread well past Nvidia. Next, some actual spending data. Economists looked at about fourteen million UK bank accounts. They found Claude's share of directly billed AI subscription spending roughly quadrupled between last December and this July, to almost thirty percent. ChatGPT still has most of the subscribers. Claude's gains come from pricier plans. Theo: So fewer people, paying a lot more. Maya: Right. It's rare hard evidence that Claude is winning on revenue per user, while ChatGPT is still winning on reach. Then defense. Shield AI agreed on Thursday to buy Echodyne, a counter-drone radar maker. It's funding the deal with a new Series H, structured equity from Blackstone funds, and JPMorgan debt. Bloomberg puts Shield's new valuation at nearly twenty-nine billion dollars, more than double what it was in March. Defense AI is now raising money and doing hardware deals at frontier-lab scale. Theo: Back to Anthropic, and something more personal. Claude can now read your bank accounts. Personal Finance launched Thursday in beta for US Pro, Max and Ultra users. You link accounts through Plaid, and it's read-only. It can see up to two years of transactions, and it can't move money. Anthropic says chats in the finance hub aren't used for training. The same day, Claude Enterprise got a security center and inference hooks, which route prompts and tool calls through a company's own security checks. Maya: Two years of my transactions is more than I want to read myself. Theo: Fair. But chatbots moving into data this sensitive raises the stakes on privacy for everyone. OpenAI, meanwhile, shipped ChatGPT Images 2.6 late Wednesday, for every user, free tier included. It makes animated GIFs, saves reference images to reuse, and runs faster. Developers get two new API models: Flare for speed and Sunburst for quality and SVG export. Volume is now over four billion images a week, up from three billion in September. So image generation is one of ChatGPT's fastest-growing uses, and developers now pick their tier. Maya: Does Google have an answer on video? Theo: It does. Gemini Omni 1.2 Flash is now generally available. Clips run up to fifteen seconds, there's conversational editing and better lip-sync, and every clip carries a SynthID watermark. It's priced at seven cents a second for 720p, thirty percent cheaper than the last version. And it's built into Google Ads, which tells you exactly who it's for. AI video is turning into a cheap, production-ready tool for ads and product shots. Then Perplexity launched Teammates, persistent agents that take on ongoing jobs without being prompted. They can message coworkers in Slack and escalate decisions for approval. It's a research preview for Max subscribers at two hundred dollars a month. Aravind Srinivas expects organizations to eventually run hundreds of them. Maya: Hundreds of agents messaging me in Slack. Who signs off on their permissions? Theo: That's the real question once agents become standing jobs instead of one-off tasks. Maya: Speaking of oversight, people on X are talking about Rep. Yassamin Ansari, the Arizona Democrat. On November 4th she posted that the new majority's first priority should be pausing advanced AI and banning artificial superintelligence. The post trended. She co-sponsors the Ban Artificial Superintelligence Act. A related bill from Bernie Sanders and Greg Casar carries prison terms. Semafor reports most Democrats are unlikely to back a pause, and they'd need veto-proof majorities anyway. Still, a superintelligence ban has moved from the fringe onto the new majority's agenda. Theo: Meanwhile, here's how far from superintelligence the agents actually are. FrontierPhysics, posted Thursday on Hugging Face Papers, has fifty-six full research tasks written by physicists. The agent has to do everything: review the literature, plan, run the experiments, and write the paper. The best agent, Claude Opus 5.5, passed about twenty-seven percent. The authors' own published papers scored ninety-six on the same rubric. The authors also dug into GPT-6 Astra's failures. Most came from sloppy reasoning or missing expert intuition. Hardly any came from gaps in knowledge. Maya: So it knows the physics. It just doesn't think like a physicist. Theo: Right. In science, the wall is rigor and judgment. Maya: Dropbox made its biggest acquisition ever on Thursday, buying the cloud storage company Tigris Data for three hundred fifty million dollars. Tigris raised a twenty-five million dollar Series A just last year. It's for DBX Agent Cloud, storage for teams building models and agents. And Dropbox files will connect to ChatGPT, Claude and Muse. Storage built for agents is becoming its own market. Theo: Last one, and it's about cost. Microsoft is adding fair-use limits to more Copilot Chat features for workers without a paid Microsoft 365 Copilot license. Smaller organizations already have them. Bigger ones get them on November 9th. Microsoft openly frames hitting a limit as an upsell moment. Big companies are clearly done giving AI usage away for free. Maya: Which brings us to Thursday's chip crash, the worst day since 2020. Up next, we're going deep on whether that was politics, money or just positioning. Now for the Main Episode Maya: So here's the question to hold onto for the next fifteen minutes. When chip stocks had their worst day since March 2020, was the market telling us the AI build-out just got permanently more expensive, because voters and a new Democratic Congress will make data centers harder to build? Or was it mostly hedge funds cashing in their big winners for the year, with the election as the excuse? Theo: And that matters for more than anyone's portfolio. If it's the first one, it changes what compute costs for everybody building on top of it. Maya: Right. So let's take it in order, because the sequence really matters here. Tuesday, November 3rd. Democrats win the House and the Senate. They also flip governorships in Republican states like Ohio and Texas, which happen to host a lot of AI data centers. Theo: And the local results were the part that jumped out at me. Maya: They were brutal. Voters went against data centers in forty of forty-one local ballot measures. That's from a Ballotpedia analysis and results CNBC gathered. Forty of forty-one isn't a close call. It's a near-unanimous vote against having one of these next door. And on top of that, the new Democratic governors in Texas and Georgia have pledged to pause new data-center construction. Theo: Texas and Georgia. Those aren't fringe locations for this industry. That's where a lot of the capacity was supposed to go. Maya: Which is why Wednesday is so interesting. The overall market loved the election. The Dow jumped three percent, its best day after a midterm since 1982. But chip stocks didn't join the party. Nvidia fell about three percent and TSMC fell about two, on a day when the market was going up. Theo: So the warning light was already on before anything crashed. Maya: Exactly. Then after Wednesday's close, Qualcomm reports, and we covered that in the headlines. Profit guidance came in soft, but the part that mattered for this story was the cost warning. Wafers, packaging, memory, all getting more expensive, and Qualcomm passing that on with its own price increases. Theo: Which feeds straight into the cost-of-building story. Even before the politics, the inputs were getting pricier. Maya: And then Thursday morning. Yahoo Finance reported the Nasdaq erased its entire post-election rally in six minutes. Six minutes. By midmorning the S&P 500 had lost more than two trillion dollars in market value. Theo: Two trillion sounds apocalyptic, though. How seriously should we take it? Maya: With some salt. That's a midmorning snapshot, not the close, and it includes plenty of stocks with nothing to do with AI. The number I'd actually anchor on is the chip index. The Philadelphia Semiconductor Index closed down a little over twelve percent. That's the worst day for chips since March 2020. For scale, the Dow fell about three and a half percent. Theo: So chips fell roughly four times as hard as the Dow. Maya: Right, and that gap is the whole story. AP noted that a slight majority of S&P 500 stocks actually rose in the morning. The Nasdaq fell five and a half percent and closed in correction territory. The Dow and the S&P aren't close to a correction. So this was a tech event, not a broad market crash. Theo: Who got hit hardest? Maya: Micron, by far. Down about seventeen and a half percent, closing below eight hundred dollars for the first time since July. And that's notable because Micron had been one of this year's biggest AI winners. AMD fell about fifteen percent, Broadcom about ten, Nvidia just over nine. Theo: And it didn't stay in hardware, from what I saw. Maya: No. It spread into software tied to the build-out. Palantir alone lost about forty billion dollars in market value. CrowdStrike, Cloudflare and Oracle fell sharply. SpaceX and Tesla each dropped more than nine percent on worries about data-center rules and their merger hopes after Texas flipped. And it went global. South Korea's Kospi fell more than five percent. Theo: Here's what I keep coming back to, though. The same day, almost as if on cue, OpenAI publishes what it calls a National Community Compact and a Compute Campus Policy Blueprint. That's my beat crossing into yours, so let me take that bit. Maya: Go. Theo: The headline pledge is that OpenAI will pay the full electricity and grid-infrastructure costs of future data-center campuses it owns and operates. It also says it'll mostly stop using NDAs with public officials, with exceptions for security and trade secrets. It'll publish a water plan and commission independent noise studies for every site, disclose energy and water use, and let residents choose local projects for it to fund. Maya: And there's a real example of that last one. Theo: There is. In Pike County, Ohio, families picked childcare. That led to a joint six million dollar commitment from OpenAI and SB Energy. Which, against a campus budget, is small. But it's exactly the kind of thing a county commissioner can point to. Maya: Read me the part that matters for the economics. Theo: The blueprint asks state utility commissions to make large electricity users cover the grid costs they add. It also wants tax breaks to come with public terms and a way to claw the benefits back. And OpenAI points to its Georgia Power deal for the Effingham County campus. It has three point two gigawatts contracted, and the utility can cut deliveries to OpenAI by up to a thousand megawatts at peak times. Maya: So roughly a third of that contract is interruptible when the grid is stressed. Theo: Right. That's the pitch to neighbors. When it's a hundred degrees and everyone's air conditioning is on, the data center steps back, not your house. Maya: Okay, but the obvious read is that this was a panicked response to forty losses out of forty-one. Theo: That's what I assumed too. But Chris Lehane, OpenAI's chief global affairs officer, said work on the blueprint began in August. So it wasn't written overnight. The timing of the release, sure, that's clearly answering the vote. The substance has been in the works for months. Maya: Which tells you OpenAI saw the local backlash coming well before the market did. That's interesting in itself. Theo: And here's the twist that I think changes how you read the whole day. OpenAI says voluntary pledges are only a starting point. It wants these same rules imposed by law on everyone. Maya: Of course it does. If you've already agreed to eat your own grid costs, you want your rivals forced to eat theirs. Theo: Exactly. A lab is asking for regulation that raises everyone else's costs to match its own. And if utility commissions actually adopt the rule that large users pay their own grid costs, every data center in that state gets more expensive. Not just OpenAI's. Maya: Which is the bull case for the regulatory-repricing argument, right there, written by the industry itself. Theo: So walk me through the people arguing about what Thursday was. Because I'm not sure the market was pricing a utility commission rule in Georgia. Maya: Probably not directly. Let's start with the positioning camp. Charlie McElligott, a cross-asset strategist at Nomura, summed it up as, quote, they're shooting the generals. Theo: Meaning the winners get sold first. Maya: Right. His argument is that the main driver is hedge funds rushing to lock in their profits for the year before year-end, not a sudden collapse in fundamentals. And he says the unwind of momentum trades has been extreme for more than a month. Theo: Is there evidence for the more-than-a-month part, or is that just his read? Maya: There's evidence. The slide didn't start this week. Over the past month, Lam Research fell about thirty-six percent and Applied Materials about thirty, per Yahoo Finance. Nvidia was down about fifteen percent over the month even before Thursday. Theo: So Thursday was a cliff at the end of a long downhill. Maya: That's a good way to put it. And there's a second piece of background I think is underrated. On November 2nd, before the election, Ben Reitzes at Melius Research cut his Nvidia price target from four hundred twenty dollars to three hundred fifteen. He kept a Buy rating. Theo: So he still likes it, he just thinks it's worth a quarter less. Maya: Right. And his reason is the interesting part. He said investor worry has shifted from AI demand to how the build-out gets financed. He cut targets on AMD, Broadcom, Marvell and CoreWeave too. Theo: That's a different question from politics, though. That's who pays for it. Maya: It is, but they connect. If each campus has to cover its own grid costs and gets held up by local opposition, the financing question gets harder. More capital, more delay, more risk per gigawatt. Theo: Okay. Who's in the forced-selling camp? Maya: Viraj Patel, global macro strategist at Vanda Research. He called it a total capitulation by retail investors. He saw no dip-buying in individual stocks, a buyers' strike from institutions, and forced selling by systematic and leveraged strategies. And Vanda has a number behind it. Retail investors sold a net three hundred fifty-eight million dollars of single stocks by about three p.m. Theo: Is that a lot? Retail volumes are huge these days. Maya: It's the most since March 17th, 2020, when retail sold about five hundred million. So the only bigger retail exit in recent memory was the COVID panic. That's the comparison that makes it meaningful. Theo: Retail has been the buy-the-dip crowd this whole cycle. If they stopped, that's a real change in behavior. Maya: That's Patel's point exactly. Then you've got the regulatory read. CNBC framed the slide as fear that a Democratic sweep means more regulation of the build-out. And Dan Ives, who Yahoo Finance identifies as a partner and managing director at Yorkville, wrote that the new political picture in Washington and Texas makes a SpaceX and Tesla merger much harder to get done in 2027. Theo: That's a pretty specific claim to hang a whole market move on. Maya: Agreed. It explains SpaceX and Tesla, not Micron. Then Kevin Mahn, chief investment officer at Hennion & Walsh, says this had been brewing because so many investors were poking at a perceived AI bubble. He expects more choppiness. Theo: And then there's Lisa Su. Maya: There's Lisa Su. AMD's CEO told Yahoo Finance the market can be a little short-sighted and that she sees no slowdown in AI chip demand. While her stock was down about fifteen percent as she said it. Theo: I mean, I believe her on demand. Everything we see on the product side says demand is fine. Four billion images a week from ChatGPT alone, Codex taking off, agents becoming standing jobs. Nobody building is asking for less compute. Maya: But demand was never really the question on Thursday. Reitzes said it before the vote. The worry moved from whether people want chips to whether the build-out can be paid for and built at the planned pace. Su answered a question the market wasn't asking. Theo: Fair. So where do you land? Maya: Mostly positioning, with politics as the trigger. The month-long slide, the momentum unwind, the forced selling, the fact that most S&P stocks rose that morning. That looks like crowded trades getting flushed. And on federal policy, I don't think the risk is near-term. Semafor reports Democrats are planning AI safety bills, curbs on data-center growth, a bigger coalition to ban superintelligence, and subpoenas into the June executive order and export controls. Theo: Which sounds like a lot. Maya: It sounds like a lot. But President Trump has threatened to veto bills he dislikes. A Democratic staffer told Semafor they doubt either chamber broadly backs Greg Casar's pause bill. And Tom's Hardware points out they'd need a veto-proof majority to pass major AI laws in the next twenty-six months. Washington is slow and mostly blocked. Theo: See, this is where I think you're underweighting it. You're looking at Washington. I'm looking at where the shovels go in. Washington doesn't permit data centers. Counties do. Utility commissions do. Governors do. And that's where the signal was forty to one. Maya: I'll give you that. The real squeeze is state and local, not federal. Theo: And none of that needs a veto-proof anything. A governor pausing construction in Texas or Georgia, a utility commission adopting OpenAI's own grid-cost rule. Those just raise the price of a gigawatt directly. For anyone renting compute downstream, the startups and the developers, that eventually shows up in prices and in how fast new capacity comes online. Maya: Here's my pushback, though. We don't know if those pledges become law. And we don't know if resistance actually cuts capacity or just moves it to other states. If it moves, the market overreacted to the politics. If it shrinks, you're right. Theo: That's honest. Nobody knows that yet. But even moving it costs something. Delay, new grid deals, new negotiations. OpenAI just showed what the entry fee looks like. Water plans, noise studies, grid costs, childcare in Pike County. That's the new floor. Maya: So maybe we're both right on different clocks. The crash was positioning. The cost increase underneath it is real but slow. Theo: I can live with that. Though I'd say the market just got a crash course in something builders have known since the summer. Maya: Fair. And there's a very near-term test of which story wins. The jobs report comes out today, and Yahoo Finance flagged it as the next market test. If chips bounce hard on decent data, that leans toward a one-day flush. If they keep sliding, the momentum unwind may not be done. Theo: And then there's the one I'd actually circle. Anthropic's IPO. Maya: Bloomberg reported that its roadshow could start as early as the week of November 9th, aiming to list before Thanksgiving. That's reported, not confirmed. And the amended S-1 still has no price range. Theo: So the biggest AI IPO could price into a Nasdaq correction. That's going to be the cleanest read we get on whether investors still want to fund this build-out. Maya: And it's a company asking public markets for money at exactly the moment Reitzes says the worry is financing. If that book fills easily, the money side is fine. If it struggles, or the price range comes in light, that tells you the repricing is real. Theo: The other thing I'd watch is the hyperscalers. Do they trim capex plans, or keep raising debt and equity to pay for them? Because the Thursday selloff doesn't change a single data center until someone actually cuts a budget. Maya: Right. A stock chart doesn't pour concrete. Theo: So pulling it together. Thursday itself looks mostly like crowded trades getting shot, the generals going first, after a month of sliding. Maya: Agreed. Second, federal AI policy is noisy but slow. The real political cost is in counties and utility commissions, and forty of forty-one says that's not a fluke. Theo: And third, OpenAI is trying to turn that local cost into a rule for everyone. If it works, the price of building AI goes up across the board, and OpenAI has already paid its entry fee. Maya: Which is the part I find most telling. The crash may fade in a week. But the most important company in the space spent Thursday asking regulators to make data centers more expensive. That's not something you do if you think the old price was coming back. Thanks for listening to Super Intelligence News and Report. Please subscribe and share it with your friends.