AI in 15 — November 14, 2026
"A deer in the headlights." According to the Wall Street Journal, that's what Jensen Huang called his chief financial officer in August, when she questioned the financial guarantees Nvidia was handing out to keep its chips selling. This week, that same CFO took his job.
Welcome to AI in 15 for Saturday, November 14, 2026. I'm Kate, your host.
And I'm Marcus, your co-host. Happy Saturday, Kate. Nvidia had a rough week, and the weekend isn't starting any better.
Here's what's coming up. A new report looks inside Nvidia's finances and the guarantees holding up the AI buildout.
OpenAI breaks up its robotics team, for the second time.
Claude helps decode Kurt Gödel's shorthand notebooks, and finds a proof he claimed he had.
Anthropic puts Claude into your meetings, then goes dark for two hours.
Plus the FAA asks whether its new AI system had anything to do with a nationwide ground stop, and ggml hits version one point oh.
Marcus, we covered Huang's resignation yesterday. What does the Journal add?
It moves the story from the scandal to the balance sheet. I should say up front that we couldn't open the Journal piece ourselves, so this is from secondary coverage. At an August meeting, Colette Kress reportedly questioned the guarantees Nvidia was giving AI customers, and Huang brushed her off. He also reportedly told finance executives that the big cloud providers were getting ready to move to alternative chips. And he described the prepaid wafers and memory coming at Nvidia as an "avalanche."
What kind of guarantees are we talking about?
Three main ones. From July, Nvidia offered to rent back unused GPU capacity from neocloud partners at a fixed rate. Neoclouds are the newer GPU rental companies, like CoreWeave or Nscale. Second, a financing backstop of up to a hundred and twenty-five billion dollars inside a bigger financing platform. Third, a two hundred and fifty billion dollar backstop tied to an OpenAI data center lease in Ohio was reportedly under discussion.
So Nvidia sells you the chips, and if you can't use them, Nvidia rents them back.
More or less. Picture a car dealer promising to lease your car back if you stop driving it. It keeps sales moving, but the risk lands back with the dealer. And there's more. CNBC reports the resignation threatens a five hundred billion dollar plan, agreed in August with Apollo, BlackRock, Blackstone, KKR and others, to turn AI chips into tradable debt. It hasn't produced a single deal yet. Partners wanted Huang to raise his twenty-five percent guarantee on what the GPUs would be worth as collateral. One source told Fortune, "That Nvidia guarantee is not worth what it was worth a week ago."
Ouch. Now give me the caveats, because I know you have them.
Plenty. The "engineered scarcity" headline going around is TechPowerUp's framing, not the Journal's. A Hacker News commenter who read the original says it never uses the word "excess." An avalanche of wafers could mean oversupply, or it could just mean pressure on margins. It all rests on unnamed sources. And every hyperscaler has said recently that demand is still bigger than their capacity. Michael Burry, of course, says the Journal missed the real story.
Naturally. So why does it matter?
Because Nvidia's balance sheet has quietly been the guarantor for a lot of this buildout. Neoclouds, OpenAI's leases, GPU-backed debt. And the person who was skeptical of those guarantees now runs the company. There are stress signals elsewhere too. Nscale posted a one point six nine billion dollar quarterly loss on about three hundred million in revenue. SoftBank took a three point seven billion charge for over-ordering data center power equipment. None of that proves a bust. But Nvidia reports earnings on Tuesday, and that's the first hard look.
Quick hits. OpenAI has dissolved its stand-alone robotics team. Marcus, is robotics dead at OpenAI?
Officially, no. Fortune reported Friday that core robotics researchers now sit in the main research group under Mark Chen, and the applied and hardware people moved under Greg Brockman. A spokesperson says robotics is still central. But Fortune went to the CoRL robotics conference in Austin and found OpenAI's booth, number two hundred and eight, empty.
That's a great detail. What's the reasoning?
That general models can now drive robots directly, with no specialized robotics model. In an Nvidia study, an agent running GPT-6 Astra hit about seventy-five percent success on a simulated manipulation benchmark, against fifty-one for the best specialized robotics models.
So the big brain just learns to use the arms.
That's the claim. Carnegie Mellon's Guanya Shi says robotics looks like it's "no longer a special problem," though you still need small models on the robot itself. The skeptics have points too. That benchmark is simulation only. Streaming commands through a giant model is slow and expensive. And MIT's Russ Tedrake says robot companies keep an edge through real-world deployment data.
And didn't they do this before?
In 2020. They rebuilt the team in 2024, then the Figure partnership ended in early 2025. And the timing matters. OpenAI reportedly lost about eleven billion dollars last quarter, it's raising money, and it's eyeing an IPO, so some see cost-cutting. Meanwhile, the same week, Runway open-sourced Praxis-1, a twenty-seven billion parameter robot-control model under Apache 2.0. Not everyone thinks specialized models are finished. And OpenAI still reportedly wants a humanoid demo in 2027.
This next one made me very happy. Kurt Gödel's notebooks. Marcus, set the scene.
Gödel is the logician who proved that any consistent math system rich enough for arithmetic contains true statements it can't prove. In the early nineteen-forties he filled sixteen notebooks, a thousand and nine pages, much of it in Gabelsberger, a German shorthand that was already dying out in his own lifetime. Jan von Plato's team in Helsinki was transcribing about a page a week.
A page a week. So they'd have finished around the year twenty-forty-five.
Give or take. In September they brought in Claude, and it produced a six hundred and fifty page explanation in a couple of days. It listed six hundred and sixty-three theorems Gödel had proved, sketched or just had an idea for. Then they checked them in Coq, a proof assistant that verifies math by machine. About four hundred checked out as stated, about two hundred needed repairs, and about fifty were false.
And there's a mystery attached.
A good one. In 1966, Gödel said he'd proved the independence of the axiom of choice back in the forties, twenty years before Paul Cohen did it. Nobody had seen the proof. Claude reconstructed one from the notebooks that was a single definition short. Once that was fixed, it passed Coq.
Okay, that's remarkable. Where's the fine print?
Claude didn't read the handwriting. Von Plato taught himself the shorthand and fed in his own transcriptions, plus six of his books on Gödel. He says the output had to be checked page by page. Outside experts haven't reviewed it in detail yet. Notre Dame's Joel David Hamkins called it remarkable if it holds up, and said working out how much is Gödel and how much is Claude is a job for historians. And the coverage doesn't say which Claude model was used.
I still love the workflow.
Me too. The model drafts, the proof assistant checks, the humans decide. And that sixty, thirty, eight split is a useful, honest number for how far to trust a first pass. My favorite Hacker News line: "AI did not comment on the consistency or completeness of the notebooks."
Anthropic. We've covered the IPO all week, so let's talk products. Claude is joining your meetings.
Claude Meetings, in beta. Claude joins a video call as a visible participant. You ask it things out loud or in chat, and it answers in writing by default. With permission it can use your connected tools, build docs and whiteboards live, and send notes and action items before the call ends. It's for Team and Enterprise plans, on Google Meet only for now.
What about privacy? I don't want a bot sitting in on a call with a client.
The defaults are sensible. Claude stays out of calls with outside invitees, and leaves if an outsider joins, unless an admin allows guests. It keeps a transcript and summary, but no audio. The funny part is that integrations with Gong, Granola, Fireflies and Otter are "coming soon," and Hacker News read the launch as a shot across the bow of exactly those note-taking companies.
Partners and competitors at the same time. Very modern. What else?
Self-hosted Claude Managed Agents, in private beta. Companies can run the agent harness, the database and the sandboxes in their own cloud or on their own servers, and send model calls through Anthropic, Bedrock, Google, Microsoft or their own gateway. That matters for banks and anyone else who doesn't want their agent's working files living on somebody else's servers.
And then on Friday, it all went down.
For about two hours. Claude dot ai, Claude Code, Cowork and the Console had an authentication failure. The investigation opened just before ten UTC, and it was resolved just before noon. No root cause yet. People speculated about a link to this week's Microsoft 365 outage, but there's nothing to back that up. On the IPO, pricing is still expected Wednesday, with trading Thursday. One more number: Riot Platforms named Anthropic as tenant for its whole Corsicana, Texas campus. That's a fifteen-year lease worth about twenty-eight billion dollars.
So a workplace platform, a giant data center, and a reminder of how much breaks when one provider sneezes.
Well put. Two hours without Claude Code, and a lot of developers rediscovered their own typing.
Now the FAA. On Friday it halted every departure in the US for more than four hours. Marcus, where does AI come in?
Communications failed at regional air traffic control centers. Transportation Secretary Sean Duffy cited "instability in our radar and our communications." Then Homeland Security Secretary Markwayne Mullin said DHS, through its cyber agency CISA, is checking whether the FAA's new AI system contributed. It's called SMART. It uses live weather, congestion and flight-path data to adjust routes, and it started rolling out in September.
Is there any evidence it did?
None so far. Mullin himself called it a telecommunications issue with no threat to the public. And nobody has explained how a routing tool would knock out radios and radar. Microsoft says it has no sign its outage is related, and Mullin says the two aren't connected. No final root cause on either.
So "was it the AI?" is now a question officials ask out loud.
That's what's new. It's the first major US infrastructure failure with that as an official line of inquiry. If SMART is cleared, fine. If it isn't, expect a lot more caution about AI in safety-critical systems. Either way, I'd rather they publish the investigation than leave it as a rumor.
Quick follow-up on Google and schools from yesterday. Marcus, what's new?
NPR adds that if a school disables AI Overviews, it loses Google Search entirely. New York's mayor, Zohran Mamdani, accused Google of using monopoly power to force AI into schools, and Senator Ed Markey has asked Google to reverse it. The interesting theory comes from Tracy Pizzo, a former Google education employee. Publishers lost an antitrust case over AI Overviews after a judge treated Search and AI as one product. An off switch for schools would suggest they can be separated.
So the legal position is writing the product decision.
Possibly. Antitrust experts weren't fully convinced. Commenters compared it to Microsoft insisting Internet Explorer couldn't be pulled out of Windows. Google says its goal is to give school leaders "control and confidence." An off switch would do that nicely.
Security, briefly. OpenAI is now selling defense.
Greg Brockman launched two agents on Thursday. Daybreak Red Team probes your applications for weaknesses, in early access. Defense Factory proposes patches, which engineers review before deploying. His pitch, per The Register, was that you're only as secure as the frontier model defending you. That's a sales pitch, but given this week's Notion and Fireworks breach, which we've covered, it's not crazy. IBM also released experimental Granite models trained to keep instructions separate from outside data, to resist commands hidden in documents.
And for the local-AI crowd, ggml turned one point oh.
ggml is the tensor library under llama.cpp, which runs a big chunk of the local AI world. Version one point oh drops deprecated interfaces and adds speedups everywhere. Faster quantized math and mixture-of-experts routing on Nvidia and Apple, FlashAttention for Intel GPUs, and even for Qualcomm Adreno phone chips. A one point oh tag means a stable interface, so people building on it can stop chasing changes. One commenter called ggml and llama.cpp "the Linux of the AI world."
High praise. And one viral one?
Jonathan Blow, the game developer, posted a video asking, "If AI solved coding, why is my software still the same?" His answer, roughly, is that companies haven't changed what they optimize for. It pairs nicely with a "No AI Slop" manifesto going around, which says what you shipped and who it helps matters more than tokens spent or agents running.
Measure the result, not the effort. Revolutionary.
Please don't say revolutionary.
One to watch: Nvidia's earnings on Tuesday, Colette Kress's first as interim chief. Listen for anything about those rent-back deals and financing guarantees.
Agreed. And Anthropic prices its IPO the next day, so it's quite a week for AI money.
That's your AI in 15 for today. See you tomorrow.