← all AI news
TechCrunch · 15 Nov 2026 · 6 MIN READ

The case for Jensen Huang's return ignores the security team Nvidia lost in August

policysafetyhardwareagents

By Saturday afternoon, more than 115,000 people had signed a petition asking Nvidia's board to bring Jensen Huang back. On the same day, TechCrunch published a piece with a much less comfortable question in the headline: where was Nvidia's security team? I think the second story matters more, and it's the one the petition skips.

Weekends are usually quiet in AI. This one wasn't, but nearly all the noise came from one company's fallout. Under it is a theme I keep coming back to this month: the systems we depend on assume that a few key people won't fail, and nobody built a plan for when they do.

5,965 Nvidia employees signed a petition that asks the wrong question

The petition at standwithjensen.com says it was signed by 5,965 current Nvidia employees, 1,108 former employees and 63,063 shareholders. Elon Musk, Replit's Amjad Masad, Epic's Tim Sweeney and Y Combinator's Garry Tan back it. It wants Huang reinstated while an independent investigation establishes the facts, and it argues he was removed "without meaningful due process."

"Jensen deserves due process, not mob justice." — the petition's slogan, via TechCrunch

I understand why people are angry. Huang built the most important company in AI, and a resignation four days after a confession feels abrupt. But the petition treats this as a fairness dispute about one man, and a board's job isn't fairness to its CEO. Its job is the company's risk. Huang has admitted an affair with an executive whom U.S. intelligence reportedly suspects of spying, and the reported affairs with Nvidia employees include one with a vice president who reported directly to him. That makes the risk the story. Due process means letting the investigation finish. It doesn't mean putting the subject of that investigation back in charge of it.

The security chief left in August and nobody connected the dots

TechCrunch's reporting adds a detail I hadn't seen anywhere else. David Reber, Nvidia's chief security officer, stepped down in August after four years in the role and seven at the company, for health reasons he disclosed publicly. His replacement, Saša Zdjelar, spent 17 years at ExxonMobil before stints at Salesforce and ReversingLabs. TechCrunch is careful to say that nothing public ties either change to the Megaspeed matter, and I'm not suggesting anything did. Still, a leadership change in security at the company that was quietly in a federal inquiry about chip diversion deserves more attention than it got.

Nicholas Eftimiades, a retired senior intelligence officer, told TechCrunch that CEOs often answer security warnings with "not a nickel more than compliance," and that intelligence services mostly work through ordinary conversation, not hacking. He also put a number on the damage that I can't verify and won't repeat as fact. His estimate is "trillions" in lost IP.

The lesson for anyone running an AI company is plain. Your threat model includes your executives' social lives, and a security team that reports to the CEO can't treat the CEO as a threat.

Seoul called an emergency meeting while Taipei shrugged

Governments across Asia reacted within a day, and how they reacted says a lot about who is exposed:

If you're building national AI capacity on one vendor's roadmap, one CEO's departure is enough to call an emergency meeting. That isn't a surprise. It's what a single point of failure looks like.

The FAA put AI in all 22 centers without asking the controllers

The other story that caught my eye has nothing to do with Nvidia, but it rhymes. The FAA is expanding its SMART advisory system from the airspace over Washington, DC to all 22 en-route centers in the country. The agency says the $875 million system has rerouted 180 flights since September 21, saving an average of 10 minutes and nearly 80 gallons of fuel per flight. Airlines love it.

The air traffic controllers' union, according to Ars Technica, says it wasn't involved in designing or testing the system. That's the part I'd fix first. An advisory tool is only as good as how much the person receiving the advice trusts it. Controllers who never got to probe its failure modes will either ignore it or lean on it too heavily, and both are bad at 35,000 feet. The FAA has run a 180-flight pilot, which is a promising start but small for a system now headed nationwide.

Helen Toner's long Decoder interview with The Verge works as a frame for both stories. She recorded it before Huang resigned, so she doesn't discuss him. Her argument, as Nilay Patel sums it up, is that if our hopes depend on any one person being trustworthy, we're in big trouble. She also says the election-season money aimed at AI-safety candidates mostly didn't work, because voters care and the issue has become bipartisan. Nvidia's board and the FAA's rollout are both tests of whether institutions can carry weight that individuals can't.

Octen wants to be the cheap search layer under your agents

One item for builders. Octen, a startup founded by Kuan Zou (who previously led AI search at Alibaba Cloud), is selling search built for agents. Its Pulse model returns URLs in a median of about 10 ms in lite mode and 20 ms in standard mode, measured server-side. Pricing starts at $0.10 per 1,000 requests, and there's a $20/month personal plan with unlimited web search. On the Artificial Analysis Search Index it scores 77, behind Exa and Perplexity at 80, but it's faster and cheaper per task.

Read the fine print before you plan around it. Latency figures leave out fetching and model time, the personal plans can queue during busy periods, and the self-hosted weights carry a noncommercial license. I'll benchmark it against my current search tool on a retrieval-heavy agent loop this week, because search is where most of my agent latency hides.


Here's my prediction. Nvidia's board won't reinstate Huang, and the petition will fade by the time he testifies before Congress. The question that stays is the one TechCrunch raised: what did Nvidia's security function know, and who did it report to? Every AI lab heading for an IPO this year should be asking the same thing about itself.

Source: TechCrunch ↗


Working on something similar?

Say hello — I read every email.