Jensen Huang Resigns as Nvidia CEO as Google Brings Gemini 4 Hydrogen to Everyone and Nonprofits Sue to Freeze Frontier AI
Jensen Huang resigned as Nvidia’s CEO effective immediately, with CFO Colette Kress named interim chief and Harvey Jones as the board’s first chair. Google released Gemini 4 Hydrogen, a fast and cheaper Gemini 4 model now in Search and the Gemini app, plus an Orchestrator that pairs it with Carbon. Plus: OpenAI turns its cyber models into Daybreak Red Team and Defense Factory, four nonprofits sue Anthropic, Google, Meta, OpenAI and SpaceXSI to freeze frontier AI, Mirendil raises $1 billion, and DeepMind opens the datasets behind WeatherNext 3.
Jensen Huang Resigns as Nvidia CEO, and CFO Colette Kress Steps In as Interim Chief
The most important company in the AI hardware supply chain changed leaders on Thursday. Nvidia’s announcement of its CEO transition says its board announced that Jensen Huang has resigned as President, Chief Executive Officer and a member of the board, effective immediately. Colette Kress, Nvidia’s Executive Vice President and Chief Financial Officer since September 2013, has agreed to serve as interim CEO while keeping her CFO role until a permanent successor is named. The board is starting a search and will consider internal and external candidates. It also named Harvey C. Jones, a former CEO and Executive Chairman of Synopsys, as its first-ever Chair of the Board.
“The Board has full confidence in Colette and the leadership team, and we expect the company’s operations, customer commitments and product roadmap to continue without interruption.”
Nvidia’s Form 8-K filing with the SEC states that the board accepted the resignation on November 11 and appointed Kress, who is 59, as interim CEO, with Jones, 73, as chair. The release did not quote Huang and did not give a reason for his departure. The Associated Press profile of Nvidia’s interim CEO reports that Huang stepped down over reports of extramarital affairs; those are press reports, not findings in Nvidia’s filing, and we covered the board’s earlier investigation on Monday. Nvidia says it will still report third-quarter fiscal 2027 results on Tuesday, November 17, as scheduled, with its earnings call at 2 p.m. PT.
For buyers of GPUs and cloud capacity, the practical message is continuity: the board went out of its way to say customer commitments and the product roadmap are unchanged, and the earnings date stands. Still, Huang was the public face of nearly every major Nvidia partnership. Teams negotiating long-term supply or co-development deals should expect slower executive sign-offs during the search and should get key terms in writing rather than relying on relationships.
Google Releases Gemini 4 Hydrogen, a Fast Model for Everyday Agent Work
Days after opening Gemini 4 Carbon to the public, Google added a faster member of the family. Google’s introduction of Gemini 4 Hydrogen, written by Koray Kavukcuoglu, describes the model as built for rapid execution across coding and agentic workflows, while Carbon and Argon handle deep, long-horizon reasoning. Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level, ahead of Argon’s 50, and reaches 54 at its highest xhigh setting. It is priced flat across its full 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens.
“Hydrogen is built to stand on its own as a model for everyday work — or to pair with Carbon and Argon as the high-efficiency execution engine in multi-model workflows.”
Compared with Gemini 3.8 Flash, Google says Hydrogen finishes tasks 2.5 times faster at its default setting, more than triples the Terminal-Bench 4.1 score (57% versus 18.2%) and cuts fabricated claims by more than 5x. Early partners saw it run 2x to 2.8x faster while using 12% to 23% fewer tokens on data-heavy enterprise workflows. On safety, Google reports up to 81% less overreach in agentic tasks, a 96.7% rate on its internal agent honesty evaluation and more than 2.3x better resilience on the Gray Swan indirect prompt injection benchmark. The Gemini API guide to what is new in Gemini 4 Hydrogen adds the model ID gemini-4-hydrogen, availability on both free and paid tiers, a shared 1,048,576-token window for input, thinking and output, and a new xhigh thinking level alongside low, medium and high.
Hydrogen is also reaching consumers immediately. Google’s post on the Gemini 4 Hydrogen upgrade to Search, from Search product VP Robby Stein, calls it the biggest model upgrade to Search yet, available globally for all users, with better handling of multi-part questions, a stronger query fan-out that links to a broader range of sites, and more visual answers. Free Gemini app users get Hydrogen through automatic model selection when it fits the prompt.
The price is the headline for builders. At $1.50 and $7.50 per million tokens with no long-context surcharge, Hydrogen is cheap enough to become the default worker model in agent pipelines while Carbon is kept for planning. Teams on 3.8 Flash should rerun their own evaluations before switching, but Google’s claims of fewer fabricated claims and better recovery from failed tool calls address the exact failures that break unattended agents.
Google Antigravity Adds Orchestrator to Pair Carbon With Hydrogen in One Session
Alongside the model, Google shipped a feature that makes the planner-and-worker pattern automatic. The Google Antigravity blog post on Hydrogen and Orchestrator describes a new model option that hands work back and forth between Gemini 4 Carbon and Hydrogen within one task: Carbon can explore dependencies and build a plan in a read-only phase, Hydrogen writes edits, updates packages and runs tests, and if Hydrogen hits friction it hands control back to Carbon to repair the plan. Google says that on Terminal-Bench, Orchestrator reached near-Carbon task success while consuming 24% less quota on average and finishing 19% faster. The team also says tens of thousands of /swarm runs were launched in the first couple of days after Carbon’s release. Orchestrator is rolling out as Orchestrator (Carbon) and Orchestrator (Argon) for Google AI Ultra subscribers and as Orchestrator (Argon) for paid Google AI Pro subscribers, and enterprises can access both through Gemini for business.
Orchestrator turns a pattern many engineering teams were wiring up by hand into a product default. The useful design detail is the bidirectional handoff, where the fast model returns control when it gets stuck instead of pushing ahead. Teams building their own agent harnesses should copy that rule, and anyone budgeting AI coding seats should note that a quota cut of about a quarter changes the math on Ultra plans.
OpenAI Launches Daybreak Red Team and Daybreak Defense Factory
OpenAI used its Cyber at the Frontier event on November 12 to turn its frontier cyber models into two products. OpenAI’s Daybreak Red Team product page describes AI penetration testing for web applications, in which a coordinated team of agents investigates attack paths, validates findings from code scanners, bug bounty submissions and disclosure reports, and checks whether a patch actually stops the exploit. Scope is enforced through isolated environments and Guardian review, which control which apps and endpoints the agents can reach.
“Agents use source code and app context to investigate attack paths and confirm vulnerabilities.”
SecurityWeek’s report on the two Daybreak products adds that Defense Factory traces vulnerabilities through a codebase and drafts patches that engineers review and test before merging. It can take in findings from Red Team and other tools, supports importing vulnerability data from Wiz, Snyk and Endor Labs, can scan connected GitHub repositories and new commits, and creates draft pull requests; patches are not deployed automatically. Red Team is open for early access requests, while Defense Factory comes as a preview plugin in ChatGPT for workspaces that already have Daybreak access.
The important line is that patches are never deployed automatically. That keeps a human merge step in the loop, which is what most security and compliance teams will require before letting agents near production code. Organizations already paying for scanners like Wiz or Snyk should treat Defense Factory as a triage layer for backlogs, and should define test scopes for Red Team carefully, since offensive agents are a dual-use tool.
Four Nonprofits Sue Anthropic, Google, Meta, OpenAI and SpaceXSI to Freeze Frontier AI
A coalition of advocacy groups asked a California court on Thursday to halt the development of the most advanced AI models. Public Citizen’s announcement of the lawsuit says Irreplaceable Education Fund, Public Citizen, Heat Initiative and The Alliance for Secure AI filed suit in San Francisco Superior Court against Anthropic, Google, Meta, OpenAI and SpaceXSI. The plaintiffs invoke California’s Unfair Competition Law and argue that a freeze on the most advanced systems is necessary to protect public safety, citing risks such as bioweapons and cyberattacks. The complaint as filed seeks injunctive relief under Business and Professions Code section 17200 and names OpenAI Group PBC, the OpenAI Foundation and two SpaceX Super Intelligence entities among the defendants.
“We are asking the court to intervene before the risks AI executives have warned us about cause devastating and irreversible harm.”
That quote comes from former New Jersey Attorney General Matthew Platkin, who filed the complaint for the plaintiffs. The Associated Press report on the lawsuit notes that it follows a midterm election in which Democrats won back control of Congress while stressing the need to regulate the technology. The defendants have not yet answered in court.
A court-ordered freeze on frontier development is a long shot, and nobody should plan around it. The case matters for a different reason: it puts the companies’ own public safety warnings into a legal record under consumer protection law. Enterprise legal teams should watch whether discovery reaches internal safety evaluations, because that could shape what vendors are willing to disclose in security questionnaires and contracts next year.
Quick Hits: Mirendil Raises $1 Billion and DeepMind Opens Its Weather Datasets
Mirendil’s Series A announcement says the company raised $1 billion led by Kleiner Perkins and Clearlake Capital, with participation from Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32. Mirendil says one researcher working with agents pretrained its competitive coding model, Mirendil-Base, from scratch in 3.5 months with less than $10 million in compute, that it post-trained the model into Fern, which holds up to 10 million tokens of context, and that its agent platform Lor, now in private beta, runs hundreds of agents writing code and running experiments for days at a time.
Google DeepMind released two open datasets on Friday. DeepMind’s post on opening the datasets behind its AI weather models introduces SatWeave, an hourly global mosaic of geostationary satellite imagery across 17 channels, and PARDIG, an AI precipitation dataset trained on spaceborne radar observations from the NASA Global Precipitation Measurement Core Observatory, with 15-minute and roughly 10-kilometer resolution. DeepMind says PARDIG improves hourly precipitation detection skill over NASA’s IMERG-Final by more than 30% on light-to-moderate rain. The WeatherNext developer guide to PARDIG lists a CC BY 4.0 license and public access with any Google Cloud account, with no request form.
Both items point at the same trend: agents doing hands-on research work, and big labs releasing the data that makes specialized models possible. Mirendil’s claim of a one-person pretraining run under $10 million is worth testing rather than accepting. For insurers, utilities and agriculture companies, the free PARDIG archive is the more immediate opportunity, especially for flood and drought models in regions without dense ground radar.
What We Left Out
Social posts claiming a $15 billion OpenAI raise at a $1.25 trillion valuation, specific benchmark scores for an upcoming OpenAI cyber model, and an Agile Robots funding round were not confirmed against primary sources in time, so we did not use them. A claim that Chinese labs trained models on restricted Blackwell GPUs remains an unverified allegation. Claims circulating about the circumstances of Huang’s resignation beyond what Nvidia and the AP reported are not repeated here. Claude Meetings was covered in Wednesday’s edition.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy.
Contact SEN-X →