← all AI news
TechCrunch · 12 Nov 2026 · 6 MIN READ

Nvidia has no plan B for Jensen Huang, and that is the real scandal

hardwareresearchagentsfunding

Nvidia's stock fell about 8% in after-hours trading on Wednesday, and the reason was not a product, a quarter or an export rule. According to the Wall Street Journal, via TechCrunch, the board is preparing to remove Jensen Huang, its founder and CEO of 33 years, and could meet as soon as Wednesday evening to decide.

Read today's news next to each other and a pattern shows up. The parts of AI that can be checked with math are getting checked. The parts that run on trust, in a CEO, in a funding loop, in a bot holding your credit card, are wobbling. I find that split more useful than any single headline.

Sixty direct reports and no plan B

The facts first. On Sunday Huang admitted to an extramarital affair with a woman U.S. intelligence reportedly believes is a Chinese spy. On Wednesday the New York Times reported a second affair, with Nvidia's vice president of health care, Kimberly Powell, who reports to him. Bloomberg reported that Nvidia is working with an outside firm to find CEO candidates. House Republicans separately gave Huang until the end of Wednesday to agree to a closed-door briefing on the first affair. The stock is down about 11% over five days.

The detail I keep coming back to is structural, not personal. Huang has about 60 direct reports, "famously delegates very little", and the company has no obvious successor or succession plan. That's the most important supplier in AI, and it was run as a single point of failure.

For builders, nothing changes on Monday. CUDA doesn't stop compiling because a board meets. But Nvidia's roadmap, its China positioning and its pricing power all ran through one person's judgment and relationships. If he goes, the next CEO will be learning the job in the middle of an export-control fight and a congressional inquiry. I'd expect more hedging toward alternative accelerators in 2027 plans, not because the GPUs got worse but because the company behind them just looked a lot less predictable.

What got proven

AlphaProof Code checks Signal and finds nothing, which is the point

Google DeepMind used a new agentic framework called AlphaProof Code to formally verify a number of the Signal protocol's key cryptographic guarantees in a four-week proof-of-concept. It found no new vulnerabilities in the parts it examined. The more convincing test: run against older Signal versions with known, long-fixed bugs, it flagged them and showed the protocol could not be verified.

The limits matter. This was not the whole Signal app, DeepMind had a head start from Signal's own earlier verification work, and Google won't say what it cost or whether it will become a product. The specification and proof are being released. What I like most is the researchers' insistence on humans at the key steps:

"We don't want this sort of vibe world where AI is generating the goals and AI is generating the proofs and basically taking tests that are self-written." — Swarat Chaudhuri, DeepMind

Anyone who has watched a coding agent "fix" a test by rewriting the assertion knows exactly what he means.

Google sells calibrated answers by the thousand

Less grand, more immediately usable: Gemini Decision 1 answers up to 1,000 independent yes/no, multiple-choice or rubric questions about the same material in one API call, and returns probabilities instead of prose. The numbers that stood out:

That last point is why I wouldn't switch blindly. Google leads on its own benchmarks; the independent index is less flattering, and VentureBeat notes a single planted field cut a competing model's block probability from 0.76 to 0.48. The useful idea here is the category: a classification layer you can threshold and audit. Keep access control in your own code, though, not in a probability.

What still runs on faith

TSMC buys into its own best customers

TSMC's board approved up to $500 million of OpenAI preferred stock and $500 million of Anthropic shares at the IPO price. The company that makes the chips is now an investor in the labs whose cloud providers buy the chips. The Journal's analysts said the quiet part: this kind of circular financing makes it hard to see what real demand looks like.

There is real demand, to be fair. OpenAI's quarterly revenue reportedly topped $10 billion for the first time, up almost 50% from $6.7 billion, with an adjusted operating margin still at -109%. And CoreWeave is asking customers to prepay, borrowing against specific contracts and leaning toward shorter deals. When the GPU landlord wants cash up front, that tells you who it thinks is carrying the risk.

United won't let a bot tick the explosives box

United told the Journal it won't let personal AI agents complete bookings. Agents can still show you flights; they can't buy them. The reasons are privacy, unintended purchases and federal rules, such as passengers attesting that their bags don't contain explosives. CEO Scott Kirby was more candid: "To be honest, I don't want to create another middleman."

This follows eBay and several banks. Agents like Meta's Muse and OpenAI's Dots are hitting the same wall: the companies at the end of the transaction don't trust them, and plenty don't want them between themselves and their customers. The protocol for an agent to prove who it acts for and what it is allowed to do is the missing piece, and nobody has standardised it yet.


My bet: the gap between these two halves closes from the proof side. DeepMind's process for verifying Signal and Google's calibrated decision API are both early versions of the same thing, AI that produces evidence and not just output. The first agent platform that can hand United a verifiable record of user consent will get the checkout button back. Nvidia's board, unfortunately, has no such shortcut.

Source: TechCrunch ↗


Working on something similar?

Say hello — I read every email.