AI in 15 — November 16, 2026
From four billion dollars to two hundred and sixteen billion. That's how much Nvidia's annual revenue grew while Colette Kress was its finance chief. Tomorrow she takes her first earnings call as chief executive, and revenue is the one thing nobody plans to ask her about.
Welcome to AI in 15 for Monday, November 16, 2026. I'm Kate, your host.
And I'm Marcus, your co-host. Morning, Kate. This is the biggest week for AI money all year, and it's only Monday.
Here's what's coming up. Nvidia reports tomorrow, and analysts say the bottom isn't in yet.
Anthropic's IPO filing shows Nvidia is a ten billion dollar anchor buyer.
Cursor confirms a backdoored Claude Code extension sat on its marketplace for eleven hours.
Alibaba's Qwen4-Max arrives with no weights and no benchmarks.
Plus the weekend's biggest developer argument, real numbers from a seventy-five thousand dollar AI school, and a price rise at Google.
Marcus, we've covered the resignation, the Journal's financing report and the smuggling claims. So let's just set up tomorrow. What should people expect?
The odd thing about this call is that the numbers matter least. Yahoo Finance points out that this quarter's revenue and next quarter's are largely locked in, because the orders are already placed. So analysts will go after three other things: who leads the company, what happens to the financing deals, and whether there are write-downs.
Write-downs on what?
Mainly the H200, the chip Huang pushed to sell into China. If those sales are now politically impossible, there's inventory that has to be marked down. Baird's Tristan Gerra said the episode "hasn't hit its bottom yet." He's pointing at those charges, and at the chance that more senior people leave.
And Kress is in a strange spot. She's the interim chief, she's still the finance chief, and by the Journal's account she's the one who argued against these deals in August.
That's the question for tomorrow. Again, we're getting that Journal reporting second-hand, through outlets like Wccftech. But if it's right, she wanted to scale back the deals where Nvidia funds a customer who then buys Nvidia chips, including one with OpenAI. Now it's her decision. If she signals a pullback, that's sensible housekeeping for Nvidia, and a colder winter for every startup whose compute budget ran through Nvidia's balance sheet.
So Nvidia has been the bank of the AI industry, and the new manager doesn't like the loan book.
That's a fair summary. I'd add that she has earned the benefit of the doubt. She's been in that seat since 2013, and the revenue went up more than fiftyfold. She knows where every dollar is.
One listening tip. The All-In podcast put out an episode on Saturday called "Jensen Ousted as Nvidia CEO." It's an hour and twenty-five minutes on the spy allegations and export restrictions.
It's worth your commute. Just remember what Al Jazeera's reporter said: there's still no established evidence that Huang knowingly helped Beijing, or that company secrets got out.
Quick hits. Anthropic prices its IPO on Wednesday, and the amended filing has fresh detail. Marcus, what's new?
Quite a lot. It's eight hundred and twenty-five million shares at seventy to seventy-eight dollars, trading on Nasdaq from Thursday under the ticker A-N-T. Revenue for the nine months to the end of September was thirty-three point two billion dollars. Yahoo Finance puts the valuation at one point three to one point four five trillion. That's the fully diluted count we talked about on Friday, and it's down from early talk of two trillion.
And there's a side deal.
A twenty-six billion dollar private placement, alongside the IPO, at the IPO price. The buyers are Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC. Google and Nvidia each put in about ten billion. At the midpoint, Nvidia gets a little over a hundred and thirty-five million shares. They're non-voting, so these companies get a stake but no say in how it's run.
So the chip suppliers and cloud providers are buying shares in the customer. We were just talking about that.
We were, and here it is in an SEC filing, which is at least the transparent way to do it. The timing is awkward, though. Huang's Nvidia agreed to that ten billion. Kress's Nvidia has to write the check. The IPO doesn't depend on the placement, but if Nvidia's part changed, we'd see it this week.
What number stood out to you?
Eight hundred and ninety-five billion dollars. That's total committed spending, mostly on compute and infrastructure. Against net proceeds of about eighty-five billion from the offering. So the IPO, possibly the largest tech listing ever, covers under a tenth of the bill.
That's like raising a record down payment and still having thirty years of mortgage.
With a mortgage that assumes your salary keeps growing twelvefold. To be fair, so far it has. Wednesday is the first time public investors put a price on a frontier lab, and I'm glad they get to. Private rounds have told us what a small group of insiders think. On Wednesday everyone else gets a say.
Security. We've mentioned the Fireworks breach before. On Friday, Cursor filled in its side of the story, and it involves a fake Claude Code extension. Marcus, walk me through it.
It's a chain, so I'll take it link by link. Fireworks is an inference provider, a company that runs AI models for other companies. Cursor, the coding tool, had issued Fireworks an access key. An attacker got that key and copied a storage bucket Cursor used to share data with them.
What was in the bucket?
Coding-agent traces. Those are logs of what the AI agent did during people's sessions, mostly from 2025, from users who weren't in Privacy Mode. And some of those logs contained live credentials, because people paste passwords and API keys into agent sessions all the time. Cursor says fewer than one in ten thousand customers had keys exposed. But that was enough to reach archived source code, support emails, and eventually the cache behind Cursor's extension marketplace.
And that's where the fake extension comes in.
On October twenty-ninth, the attacker planted trojanized builds of Anthropic's Claude Code extension, versions two point one point three one two and three one three. They tried to install a persistent SSH backdoor, meaning a hidden remote door into your machine. They were live about eleven hours, and at most five thousand users could have received them.
Does uninstalling fix it?
Reports say maybe not, because a backdoor like that is designed to outlive the extension. If you pulled those versions from Cursor's marketplace in late October, treat the machine as compromised. Cursor says Anthropic's genuine extension, the VS Code marketplace, Privacy Mode users and customer repositories weren't touched. It has moved all inference off Fireworks, which is also linked to breaches at Uber and Notion.
What should everyone else do?
Two things. Companies should ask their AI vendors what they keep, and who else gets a copy. And developers should rotate any key they've ever pasted into an agent session. Those logs are a new kind of attack surface, and most people don't know they exist.
Alibaba. Qwen4-Max hit the top of Hacker News overnight. Marcus, is it good?
Nobody can tell you, and that's the story. It's the first of the Qwen4 family, pitched at coding agents. It plugs into Claude Code, Cursor, Codex and the rest, with a subscription starting around three dollars. But Alibaba published no model card, no parameter count and no weights.
Wait, no weights? Qwen is the open-weights family.
It has been the leading one. A closed, subscription-only flagship is a real change of strategy. One Hacker News commenter asked whether the gap between open models and the frontier is widening. I'd put it more bluntly. When a lab has always shown its work and suddenly stops, I want to know why.
Are there any outside numbers at all?
One, second-hand. A commenter cited Artificial Analysis as scoring Qwen4-Max at forty-three, at an average cost of three dollars twenty-seven. GPT-6.1 Sol at medium scores forty-six for thirty-six cents. If that holds, it's slightly weaker and about nine times the price. That's a commenter quoting a tracker, so hold it loosely. Also, the million-token context everyone's citing looks like a client default, not a confirmed limit. And it's unclear whether there's a pay-as-you-go API yet.
So the pelican test is the best data we have.
Simon Willison did run his pelican-on-a-bicycle drawings at four reasoning levels, yes. Until Alibaba publishes evals or weights, a pelican is about as rigorous as it gets.
The biggest thread of the weekend: more than eleven hundred points and six hundred comments on a blog post titled "All AI-generated code should be open source." It's by Jakub Jankiewicz, who wrote jQuery Terminal. Marcus, what's his case?
He makes three arguments. First, ethics: the models learned from open-source code under licenses that require attribution, so closing up the output is what he calls "license laundering." Second, copyright: code with no human author shouldn't be copyrightable at all. He cites the monkey selfie case, where a photo taken by a monkey couldn't be owned. Third, the rules differ by country, so an open license keeps things simple.
I have sympathy for the open-source authors. But "should be" is doing a lot of work. Who's going to make me?
That's where the thread went. The law he's leaning on cuts both ways. Thaler versus Perlmutter says no copyright without a human author. But Doe versus GitHub found that model output doesn't automatically inherit open-source licenses. So purely machine-written code may be owned by nobody, which isn't the same as being open source. And several people argued that AI is a tool, like a compiler, and nobody thinks the compiler owns your program.
And most code is a mix anyway.
He admits that. The pointed bit came from an author of AGPL software, who said early models reproduced his code almost verbatim, and a company shipped it without credit. That's a real grievance. I just think the fix is enforcing attribution where copying happened, rather than a blanket rule over everything a model touches. But it's a live question for businesses. If your product is mostly agent-written, you may not have the copyright you assume you have.
Education. Alpha School charges up to seventy-five thousand dollars a year, and promises kids learn twice as fast with two hours of AI tutoring a day. Fortune now has the internal numbers, first reported by the New York Times. Marcus?
Forty-two percent of students hit the twice-as-fast goal in both math and reading. About half of the K-to-eight private-campus students missed it in math, and about one in five missed it in reading. And among kids who started at or below the national median in math, only thirty-seven percent got there.
So is that a failure, or is it just less than the brochure promised?
Less than the brochure promised, I'd say. Hacker News made the fair point that doubling the pace in a third of the class time, for four in ten kids, would still be impressive. The caveats are reports that the two-hour cap isn't always kept, and that prizes are used as incentives. And it worked least well for the kids who started behind, which is the group you'd most want it to help.
And it's moving into public schools.
More than twenty-six hundred public-school students in Texas and South Carolina use the platform, and Houston is piloting it at two schools. I like that parents and districts get to try something new. I'd like it more if they got these numbers before signing up, not from a leak.
Two short ones. First, a reminder for developers on Gemini 4 Carbon, Google's top model.
It's on introductory pricing: two dollars per million input tokens and ten per million output, same as Argon. After November thirtieth that doubles, to four and twenty. So if you're running evaluations, the next two weeks are half price. It also says something about the market. The frontier labs are raising prices on their best models, not cutting them.
And the Hacker News "Who is Hiring?" thread?
Two hundred and forty-four postings this month, down from three hundred and eighty-seven, and near a twelve-year low. The tracker says about half are AI roles. But that's one forum's job thread, not the labor market. Commenters noted broad unemployment has barely moved. It's a mood reading and I wouldn't lean on it.
One to watch: Nvidia's earnings call tomorrow afternoon. If Colette Kress signals a retreat from financing customers, that lands on Anthropic's pricing the very next day.
Agreed. I'll be listening for any mention of that ten billion dollar commitment.
That's your AI in 15 for today. See you tomorrow.