Thirty-three years. That's how long Jensen Huang ran Nvidia, and on Thursday morning it ended with a board announcement saying he had resigned as president, CEO and director, effective immediately. I spent the rest of the day reading AI news through one question: where else does this industry depend on a single thing that nobody planned to lose?
The answer turned out to be: in a lot of places, including some of the stacks I work on.
Nvidia swapped its founder for its CFO overnight#
The facts first, because the gossip is louder than the facts. According to TechCrunch, the board announced Huang's resignation after reports of affairs with a woman U.S. intelligence reportedly suspects of spying for China, and with one of his direct reports. Reuters has since reported a third. CFO Colette Kress, at Nvidia since 2013 and previously a finance executive at Cisco and Microsoft, is interim CEO while the board looks for a successor. Huang has agreed to testify before Congress after senior lawmakers threatened a subpoena.
Kress's statement was the corporate version of keeping calm:
“Our focus remains on serving our customers, partners, developers and shareholders.” — Colette Kress, via Nvidia statement
For anyone buying GPUs, nothing changes this week. Nvidia reports earnings on Tuesday, November 17, and that call will be the first real look at how a finance-led interim chief talks about roadmaps. What I'm more interested in is the stuff that only existed because one founder wanted it.
Open models are the obvious casualty risk. TechCrunch lays out how much of Nvidia's open-source posture ran through Huang personally: the roughly $13 billion Hugging Face acquisition that still isn't closed, the open-weight Nemotron family, bets on Reflection AI and Thinking Machines Lab, and an open letter he led, signed by 25 companies, asking Washington not to over-restrict open-weight models. An insider told TechCrunch they don't expect Nemotron work to change. Maybe so. But Nathan Lambert put the worry bluntly on X: Huang leaving would be “very very bad for the open model cause in the US.”
Then there's the policy thread, which got more awkward on the same day. MIT Technology Review reports that Nvidia representatives privately told Rep. Bill Foster that the chip-tracking measures in the Chip Security Act were technically doable. That was the reaction in technical talks, Foster says, even as the Semiconductor Industry Association, which Nvidia belongs to, called the mandates “potentially infeasible” in March. “They never put anything in writing,” Foster said. Put that next to reports that the alleged spy runs Megaspeed, a Singapore firm under investigation for smuggling GPUs to China, and the argument against location checks on chips gets very hard to make in public.
My read: the open-model ecosystem in the U.S. has leaned on one hardware vendor's founder as its most powerful political ally. That ally is now testifying instead of lobbying. If you've built a roadmap on the assumption that American open weights will keep getting better and stay legal to ship, you should at least have looked at what happens if they don't.
Fireworks shows your inference host is part of your attack surface#
The quieter story is the one that should worry builders more. TechCrunch reports that attackers got into Fireworks AI through its cloud provider, reached a development system, and took billing information and some customer datasets. It's Fireworks' second incident in two months. Notion confirmed it was hit: chat transcripts from Notion AI's Insights feature, including “any input to or output from the model that was written to the chat,” affecting about 1% of its customers. Notion cut its connection to Fireworks on November 3.
Fireworks says the attackers were after money, specifically reselling LLM tokens from breached access. That's a useful detail. It means inference providers are now worth breaking into for the compute alone, and your prompts and completions are what gets taken along the way. Fireworks counts tens of thousands of customers, Cursor and Harvey among them.
Practically: if you route user content through a third-party host, find out today what it stores, for how long, and whether you can turn that off. “We use an open model, so the data stays with us” is not true if someone else runs the GPUs.
Gemini 4 Hydrogen is better, faster, and not a drop-in swap#
Google shipped Gemini 4 Hydrogen as the replacement for 3.8 Flash, across the API, AI Studio, the Gemini app and Search, where it is now the default in AI Mode. Ars Technica notes it skipped the slow, security-conscious rollout Carbon and Argon got. The numbers, per VentureBeat:
$1.50 input / $7.50 output per million tokens, double 3.8 Flash's current introductory price (Flash rises to the same rate on January 1)
2.5× faster than Flash at default reasoning, according to Google
Terminal-Bench 4.1: 57% vs 18.2% for Flash; Claude Sonnet 5.5 scores 63.9% in the same table
About a 1M-token context window, shared between input, output and reasoning
That's a real jump for a mid-tier model, and I'll be testing it on agent workloads where Flash simply couldn't keep up. But the same day, a VB Intelligence survey found that 70% of companies that swapped a production model needed a few weeks or more of extra engineering, and 45% saw performance drop and had to rework the app or roll back. Box's CTO summed up why: “a new model just maybe acts differently.”
So Hydrogen is a good test of whether your team can actually change models. If moving off a retiring Flash takes a month, that's another single point of failure, just a less visible one than a CEO.
Also today#
Anthropic put Claude into Google Meet in beta for Team and Enterprise: it joins calls, drafts docs and whiteboards live, and leaves by default when an external guest joins. In-meeting docs are free until February 10, 2027.
An IP-protection firm found chatbots surfacing risky retail sites in up to 15% of shopping answers (Perplexity), with ChatGPT at 8% and Gemini at 4%. Smaller brands were hit about twice as often.
Here's what I'm changing. Every project I touch gets a one-page dependency check: which model, which host, which vendor's goodwill we're assuming. For each one I want to know how many days it would take to swap it out. Nvidia found out on Thursday that its answer for its CEO was “we don't know.” I don't want mine to be the same for an inference host.