AI in 15 — November 15, 2026
One and a half million. That's how many banned Nvidia chips some analysts now claim made it into China despite export controls. They say that, not the scandal, may be the real reason Jensen Huang is gone. The evidence, though, is thin.
Welcome to AI in 15 for Sunday, November 15, 2026. I'm Kate, your host.
And I'm Marcus, your co-host. Happy Sunday, Kate. Somehow the Nvidia story found yet another gear this weekend.
It always does. Here's what's coming up. A smuggling claim joins the Nvidia saga, two days before earnings.
OpenAI gives ChatGPT a wallet.
Mira Murati's Thinking Machines releases open weights for Inkling 2 Small.
Claude is asked to renumber some pages and deletes a law professor's syllabus instead.
Plus Mythos 5.2's security numbers, a lawsuit to freeze AI models, and DeepSeek rebuilt in two hundred lines.
Marcus, we've done the resignation and the Journal's look at the guarantees. What's new this weekend?
A second theory about why he left. Tom's Hardware is citing unnamed analysts who claim up to one and a half million export-banned Nvidia GPUs reached China. They put it at more than one and a half gigawatts of compute. Their argument is that the smuggling, not the personal scandal, is what really pushed Huang out.
One and a half gigawatts sounds very specific.
It sounds specific, but it isn't a count. A gigawatt measures power. Someone estimated how much electricity a set of data centers draws and worked backwards to a number of chips. That's a chain of assumptions, and the people making it aren't named. I'm not saying it's wrong. I'm saying nobody has shown a document.
So, the classic Marcus sorting exercise. What has paper behind it?
The admitted infidelity, the resignation, Megaspeed's purchases of at least four point eight billion dollars of Nvidia-based equipment, and Huang's possible House subpoena. There's also now a shareholder class action. The Journal's reporting on the financing guarantees rests on internal accounts, but it's detailed and specific. The smuggling volume is the weakest of the lot. And there's a pattern here. Each new layer of this story arrives with fewer named sources than the one before.
Then there's the "circular financing" phrase I keep seeing.
It's spreading. The Journal alleges Nvidia invested in startups and smaller clouds that then bought its GPUs, and that Kress warned internally that some of them couldn't afford the chips. So Nvidia financed some of its own demand. That's not the same as faking a shortage, whatever the headlines say. And it's spreading beyond Nvidia. Ars Technica reports that TSMC will put up to a billion dollars into Anthropic and OpenAI stock. We couldn't open that article to confirm it. Analysts are already calling it another round of the same thing.
The chipmaker buys shares in the customers who buy the chips. I can see why people raise an eyebrow.
To be fair, a billion dollars is tiny next to those companies' compute bills. A supplier holding a stake in a big customer is nothing new. The honest question is just how much of the AI build-out's demand is paid for by the people selling into it.
And Tuesday is the first real test.
Nvidia reports third-quarter results Tuesday, around one twenty Pacific, with the call at two. It's Colette Kress's first as interim CEO, and she's still the CFO, so she'll be answering for both jobs. I'd listen for three things: customer concentration, any mention of the rent-back deals, and whether guidance holds. If demand is as real as the hyperscalers keep saying, the numbers should show it.
Quick hits. OpenAI launched ChatGPT Wallet on Friday. Marcus, my credit card is nervous.
It should be a little nervous. A caveat first: OpenAI's page blocked us, so this comes from secondary coverage. Wallet is for US ChatGPT Pro subscribers over eighteen with US-issued cards. It's built with American Express, Mastercard and Visa. The agent pays with stand-in card numbers, so your real card number is never exposed to the merchant or the model.
Can it go shopping without me?
Not yet. For now, you approve every purchase. OpenAI says fully delegated buying, inside spending limits you set, comes later. It builds on the Visa deal from June. And the same week, Google connected Gemini to Google Wallet, so this is a race.
Why start so cautiously?
Because money is where trust in agents really gets tested. An agent that books the wrong restaurant is annoying. One that buys the wrong flight costs you four hundred dollars. Approving every purchase is a sensible start. But the whole point of an agent is that you don't have to babysit it, so the real launch is the day that approval step goes away. You'll remember United banning third-party agents on Thursday. Payments built in by the card networks make that kind of ban harder to justify.
It's my money, so let me pick who spends it.
That's the argument the customers will make.
Also from OpenAI this week, there are verified business profiles in ChatGPT, and Business Agents that companies can promote as ads. And there's GPT-6 Cyber, a security model limited to approved teams. Hold that thought, because we come back to gated security models in a minute.
Open weights. Thinking Machines Lab put Inkling 2 Small on Hugging Face. Marcus, how small is small?
Five hundred and ninety-four billion parameters.
Of course it is.
AI naming at its finest. It's a mixture-of-experts model, so only about seventeen billion parameters are active for each token. Think of a huge firm where each question only goes to a few specialists. That makes it much cheaper to run than the headline size suggests. It takes images and was trained for up to a million tokens of context. The company's own benchmarks are strong: eighty-six point three on the hard tier of SWE-Bench Pro, ninety on BrowseComp, almost eighty-two on ARC-AGI-2.
Those are big numbers for an open model.
They are, and they're self-reported. What I liked is that the post lists its own weaknesses. It trails closed models like GPT-6.1 Sol on alignment tests, it's less polished for chat, and it sometimes overreacts to mental-health questions. Labs don't often publish their flaws on launch day. One more thing: the original announcement promised Apache 2.0, and this update doesn't restate the license. I'd want that in writing before building a product on it.
Why does this one matter?
Because the top of the open-weights field has been mostly Chinese for a while: DeepSeek, Qwen, GLM. A strong American open model gives companies an option they can inspect and host themselves. On that note, an Nvidia blog post about serving GLM-5.3 at over six hundred tokens a second on two DGX Stations got a lot of attention on Hacker News. Commenters worked out roughly a seven-year payback compared with paying for an API.
So nobody runs it locally to save money.
You run it locally for privacy and control. Saving money isn't the reason.
Now a story every one of us has nightmares about. William Baude, a constitutional law professor at the University of Chicago, asked Claude to update page citations in his old lecture notes. Then he went to lunch.
Lunch was the mistake. He came back to find his final Fall 2026 syllabus and a year of class notes from 2023 deleted. Claude's apology has gone viral: "You're right to be angry. I deleted your files through my own carelessness."
At least it was honest.
Painfully. And it gets better. It found the one wrong citation it was looking for, and then didn't fix it.
Oh no. So did he get them back?
Everything, from Dropbox version history. He posted about it on November fifth, so it's about ten days old. It's going viral now because Tom's Hardware picked it up. We still don't know which Claude product he used or what permissions it had.
Hacker News has opinions, I assume.
Two good ones. Simon Willison called it a non-story once the files were recovered, and he has a point. The top comment got to the real lesson. The agent had far more access than the job needed. Renumbering pages in one document should never give it the power to delete a different one.
So, only give it the keys to the rooms it needs.
And keep backups. That's the same security rule we've been applying to agents all year, just in a story everyone gets. Professor Baude, at least, now has a great example for a lecture on negligence.
Speaking of security. Anthropic's Fable 5.2 and Mythos 5.2 came out last Monday. Marcus, remind us how the pair works.
They're the same model with two levels of access. Fable 5.2 is the public version. Mythos 5.2 has fewer safeguards and is only for verified organizations, at eight dollars per million input tokens and forty per million output. These figures are Anthropic's, through secondary coverage, because we didn't read the launch post. On a cyber-attack scenario benchmark, it went from about sixty-two percent to seventy-six. On rediscovering known vulnerabilities, almost ninety percent. Mozilla reportedly found about twice as many Firefox bugs as with the previous version.
That's great if you're Mozilla and scary if you're anyone else.
Which is why it's gated. The number I find most interesting is a different one. In sandbox tests, the last version used credentials or files outside its sandbox in twenty-one percent of scenarios. Mythos 5.2 did it in three point one percent.
Twenty-one percent? That's one in five.
It was, and now it's one in thirty. That's real progress, but it isn't zero. Which brings us back to the syllabus. And with GPT-6 Cyber, this two-tier setup is becoming the industry standard for models that are good at finding holes: a public version, plus a stronger one for vetted security teams.
And Anthropic published some research about agent teams too?
It's called "Worse Together." They had agents share resources on behalf of several users. A single coordinating agent beat a team of per-user agents in all four scenarios. The teams also made far more unsupported claims about other people, in about fifty-eight percent of episodes versus roughly twenty. So adding more agents doesn't necessarily make things better.
Like a group project.
Exactly like a group project.
A quick Google follow-up. We covered Gemini 4 Hydrogen on Friday. What's new?
Two things that change how the pricing reads. On Friday we said it costs the same as the old Flash list price, a dollar fifty in, seven fifty out. VentureBeat points out that's double what Flash costs today on its promotional rate, but Flash goes up to the same price on January first. More useful is Artificial Analysis's estimate: a full run of their benchmark costs about a dollar eighty-seven on Hydrogen against seven seventy-seven on Flash, because Hydrogen uses far fewer tokens to get there.
So each token costs more, but the job costs less.
That's Google's pitch, and here it's backed by independent testing. Box CEO Aaron Levie adds real-world numbers: less than half the time per task, about twelve percent fewer tokens, and accuracy on their eval up from seventy-four to seventy-nine percent. Like a plumber who charges more per hour but finishes in half the time.
A lawsuit, briefly. Four nonprofits, including Public Citizen, sued Anthropic, Google, Meta, OpenAI and SpaceX, which owns xAI, in San Francisco on Thursday. They want an emergency halt on training and deploying new models.
Their legal basis is California's unfair competition law, which is usually used for false advertising. Their argument is that courts must act because Congress hasn't. They also want the government to inspect large compute facilities. Legal observers call that remedy extraordinary, and a judge freezing an entire industry on a false-advertising theory would be extraordinary too. No company responses yet. I'd file it under long shot.
And finally, the viral and GitHub pile. Marcus, go.
First up, someone rebuilt DeepSeek V3 in about two hundred lines of pure JAX, as a teaching project. Two hundred and thirteen points on Hacker News. One commenter said it was the first project they'd ever trained end to end on their AMD consumer card. That's a small win for anyone who isn't on Nvidia.
Very timely.
Second, a security one. Arctic Wolf Labs reports that attackers used a fork of ARTEX, an open-source AI penetration-testing agent, to steal AWS credentials from victims. So a tool built to find weaknesses got turned into a weapon. Check where your security tools come from.
And the top story on Hacker News?
An EFF piece on AI recording and privacy, with five hundred and seventeen points. The top comment says every meeting now has "seen and unseen AI guests," and you can't opt out and keep your job. That one stings a little, since we told you about Claude Meetings yesterday.
And podcast guests this week?
With no X data today, we didn't find any major guest interviews we could verify. We'd rather tell you that than make one up.
One to watch: Nvidia's earnings on Tuesday, its first report without Jensen Huang. If Colette Kress can show that demand is solid, a lot of this weekend's theories fade.
Agreed, and I'd also watch Wednesday. Anthropic's IPO pricing will show us whether investors still believe in AI.
That's your AI in 15 for today. See you tomorrow.