← Home AI in 15

AI in 15 — November 13, 2026

November 13, 2026 · 15m 43s
Kate

"Effective immediately." With those two words, Nvidia's board ended the thirty-three-year run of the only chief executive the company has ever had. Jensen Huang is out, and nobody at Nvidia is saying why.

Kate

Welcome to AI in 15 for Friday, November 13, 2026. I'm Kate, your host.

Marcus

And I'm Marcus, your co-host. Happy Friday the thirteenth, Kate. Somebody at Nvidia is not going to enjoy that date.

Kate

Not one bit. Here's what's coming up. Jensen Huang resigns, and Nvidia's finance chief takes over.

Kate

Anthropic opens its books ahead of next week's IPO, and the numbers are enormous in both directions.

Kate

Google ships Gemini 4 Hydrogen, a cheap model that outscores a flagship.

Kate

A researcher says Chinese open models carry the fingerprints of banned Nvidia chips.

Kate

Plus a hundred and forty-five billion parameters on a laptop, a rebel embroidery machine, and Google versus Chicago's schools.

Kate

Marcus, yesterday we said the board might remove him. Today it's done. What exactly happened?

Marcus

On Thursday the board announced that Huang has left as president, chief executive and board member, effective immediately. The company gave no reason. Colette Kress, the chief financial officer since 2013, becomes interim CEO and keeps the finance job too. Director Harvey Jones becomes the first board chairman Nvidia has ever had. And the board is searching for a permanent chief, inside and outside the company.

Kate

Four days ago a spokesperson said he'd continue "without interruption."

Marcus

That lasted about as long as a New Year's resolution. Look at the sequence. On October twenty-ninth, Huang told Bloomberg the relationship with a chip buyer called Megaspeed was "normal and not unusual." On November eighth, the board opened its investigation and he admitted being unfaithful to his wife. Then came the reports of a relationship with a direct report, and Reuters reported a second employee. On the twelfth, he resigned.

Kate

Let's do your favorite exercise. What's established, and what isn't?

Marcus

Established: the admitted infidelity, the resignation, and the fact that Megaspeed bought at least four point eight billion dollars of Nvidia-based equipment in under two years. What's not established is the espionage part. The claim that Alice Huang, a former Megaspeed executive and no relation, was working for China rests on unnamed intelligence officials. Nvidia's board says it has no reason to believe secrets were compromised. And there's no public evidence Jensen Huang knew of any intelligence link.

Kate

And Singapore actually looked into her.

Marcus

They did. Authorities there questioned her, confiscated her passport, found insufficient evidence of a crime, and then barred her from returning. Singapore's Home Affairs ministry also disputes parts of the press account. Beijing calls the whole thing a smear, which tells us nothing, because it would say that in any case.

Kate

And Washington?

Marcus

The President said the FBI is looking into it, praised Huang's "fantastic job," and floated Tim Cook as a successor. CNN reports Huang has been summoned to testify before Congress, and lawmakers from both parties want investigations.

Kate

Tim Cook running Nvidia. That's quite a casting call. The stock closed down more than four percent. Honestly, that seems mild.

Marcus

The market had already priced in a lot on Wednesday night. The harder question is what Huang took with him. The relationships with the big cloud companies, with TSMC, with the memory suppliers, with governments, were his personally. Yale's Jeffrey Sonnenfeld says Nvidia never prepared a crisis succession plan. Kress is highly respected, but the interim chief is also doing the old job. I wouldn't want that calendar.

Kate

One caution for listeners. The Hacker News thread on this has three thousand comments, and a lot of confident theories.

Marcus

Lots of theories and no evidence for any of them. Enjoy it as entertainment. What I'd actually expect is pressure for tighter export controls, whatever the investigations find.

Kate

Quick hits. Anthropic's IPO, which we've tracked all week. The roadshow is underway, and now we have real third-quarter financials. Marcus, give me the numbers.

Marcus

These are from Anthropic's own IPO page. Third-quarter revenue was seventeen billion dollars, more than twelve times the same quarter last year. More than eleven thousand nine hundred customers each spend over a hundred thousand dollars a year. And the net loss was twenty-three point three billion.

Kate

Hang on. They lost more than they made?

Marcus

On paper. Twenty-one point nine billion of that loss is a non-cash charge, driven by the rise in the company's own valuation. It's an accounting entry, with no cash going out the door. The cleaner figure is the operating margin: minus three percent as reported, plus fifteen percent adjusted. And the company says compute efficiency has improved more than fourfold since early 2025.

Kate

So the underlying business is close to breaking even.

Marcus

Close, with a big asterisk. Earlier coverage of the filing described a five hundred and eighteen billion dollar compute plan, with commitments that can't be cancelled. That's a business with twelvefold growth and a bill that's due whether or not the growth continues.

Kate

And the timetable?

Marcus

Pricing is expected November eighteenth and trading on the nineteenth. CNBC reports the order book is already covered at the seventy to seventy-eight dollar range. There's a retail site, including access for European and Swiss investors through banks like Deutsche Bank, UBS and BNP Paribas.

Kate

There's one awkward detail in the range.

Marcus

The May private round valued Anthropic at nine hundred and sixty-five billion. On shares outstanding, seventy dollars works out to about nine hundred and sixteen billion. So the bottom of the range would be below what private investors paid. The trillion-plus headlines use a fully diluted share count.

Kate

And the comparison with OpenAI?

Marcus

We gave you OpenAI's reported ten billion dollar quarter yesterday. That one's unaudited and from unnamed sources. Anthropic's audited seventeen is about one point seven times that. So the company that started this whole market is now second on revenue. Also this week, the Economist ran a leader titled "Why you should expect an AI crash." It's paywalled and we haven't read it, so I'm only giving you the headline.

Kate

Lovely timing for a roadshow.

Kate

Google. Two days after Carbon, here comes Gemini 4 Hydrogen. Marcus, I need a periodic table to follow this lineup.

Marcus

You and everyone on Hacker News. Hydrogen is the fast, cheap tier, and effectively the successor to Gemini 3.8 Flash. Google says it's about two and a half times faster than that model. And on the Artificial Analysis Intelligence Index, Google reports fifty-one at the high setting and fifty-four at a new extra-high setting. Argon, the flagship from six weeks ago, scores fifty.

Kate

So the budget model beats last month's best.

Marcus

On that index, by Google's account. The jump on Terminal-Bench, which tests command-line agent work, is fifty-seven percent against eighteen for the old Flash. Google also claims first place on a tax agent benchmark and a tie for first on legal research. Those are Google's own figures, so wait for outside testing.

Kate

What does it cost?

Marcus

A dollar fifty per million input tokens and seven fifty per million output, flat across a million-token context. Commenters pointed out that's the same as the old Flash list price. So you get more for the same money, but it isn't cheaper.

Kate

And who gets it?

Marcus

Everyone, right away. It's in the API, in the Gemini app for free users, and in Google Search. That last one matters most. The AI that most people meet without ever choosing a model just got a big upgrade.

Kate

There's also something called Orchestrator.

Marcus

That's in Antigravity, Google's coding tool. A big model, Carbon or Argon, writes the plan and decides when to re-plan. Hydrogen does the legwork: edits, package updates, test runs. When it gets stuck, it hands back to the big model. Think of an architect and a site crew. Google says it gets near-Carbon results with about twenty-four percent less quota, finishing about nineteen percent faster. Anthropic's advisor mode works along similar lines.

Kate

So the tool picks the right model for each step, and I don't have to.

Marcus

Which is how it should work. People shouldn't have to keep a model lineup in their heads, especially one named after elements.

Kate

Now a story I found fascinating. A report called "Checkpoint forensics" claims you can look at a published AI model and tell which generation of Nvidia chip it started life on.

Marcus

It's by Will DePue, published Thursday. The idea is clever. When training starts, a model's weights are filled with random numbers from a seed. In PyTorch, the pattern of those numbers depends on a hardware detail of the GPU, a count of processing units that differs between chip generations. Hopper, the older generation, has one count. Blackwell, the newer export-controlled generation, has others. The report says traces of that starting pattern survive training, like a watermark in paper.

Kate

And what did the watermarks show?

Marcus

Six models from four Chinese labs read as Blackwell. That's DeepSeek's two V4 models, Moonshot's Kimi K3 and K3.1 Pro, MiniMax M3, and one from Ant Group. Earlier models from the same labs, and models from Alibaba, Tencent and others, read as Hopper. The method matched all twenty-four cases where developers had disclosed their hardware.

Kate

That sounds like a smoking gun. Is it?

Marcus

No, and the author says so. The fingerprint shows what initialized the weights, which isn't necessarily what trained them. It can't tell you who owns the chips, where they sit, or whether any rule was broken. A Chinese company renting Blackwell capacity in Singapore or Malaysia is legal. The estimate of about twelve thousand chips is low-confidence. And no code has been released, so nobody else can rerun it yet.

Kate

And the messenger?

Marcus

Commenters noted DePue is associated with OpenAI, which lobbies for export controls. That doesn't make the math wrong. It's one more reason to want the code published. What I like is that this is a testable claim, in a debate that usually runs on anonymous officials. It landed the same day as the Huang news, but that's a coincidence of timing and nothing more.

Kate

Open models. A company called Edge0 says it fits a hundred and forty-five billion parameter model into twenty-nine gigabytes of memory. Marcus, how?

Marcus

It's a sparse mixture-of-experts design. The model is split into many specialist sub-networks, and only a few run at each step. So of a hundred and forty-five billion parameters, roughly ten billion are active at once. Groups of sixty-four experts share two large matrices, which saves a lot of space. The software then spreads the model across the graphics card, ordinary memory and the SSD.

Kate

So it's a huge library where you only pull a few books off the shelf at a time.

Marcus

And some of the shelves are in the garage. Edge0 claims over eighty tokens per second at about a hundred and twenty watts, on Apple M-series machines and high-end Nvidia cards. On its own benchmarks, it beats a recent Qwen model on coding and comes within a point of Claude Opus 4.6 on a science test.

Kate

Opus 4.6 is not exactly this week's model.

Marcus

No, that's an older model, and every benchmark is internal. Also, the post says "open source" without naming a license. And Hacker News users say the weights ship in a format that for now only runs on Nvidia's newest hardware. If independent tests hold up, it's a real result for people who want capable models on machines they own. I'll wait for those tests.

Kate

Two quick ones to finish. A developer bought a three-hundred-dollar embroidery machine and found the software wanted a subscription.

Marcus

So the developer built an open-source replacement with AI help, called zigzag. Four hundred and thirty-four points on Hacker News. The thread filled up with similar stories. One commenter said a model wrote card-printer software in twenty minutes. Charging a subscription for the software on hardware you've already bought gets harder when the customer can write a replacement over a weekend.

Kate

I love that. You bought the machine, it's yours. And the second: Google and schools.

Marcus

Per Chalkbeat and Axios, Google first told schools that opting out of its AI products would mean losing Search. It then dropped the Gemini requirement, but it still won't switch off AI Overviews in school Search. Parents in Chicago are protesting, and Chicago Public Schools dropped its planned Gemini rollout for high schoolers.

Kate

I saw a headline saying Chicago would rather lose Search entirely.

Marcus

We couldn't open that article, and other coverage doesn't show a formal decision. So treat it as unconfirmed. The simple fix is an off switch that schools control.

Kate

One to watch: Nvidia under Colette Kress. The first public statement to customers and investors will tell us how steady the company is without Huang.

Marcus

Agreed. And if Huang's congressional date gets set, that's the next headline.

Kate

That's your AI in 15 for today. See you tomorrow.