Who keeps the roadmap moving when the founder leaves? Nvidia announced on November 12 that Jensen Huang had resigned immediately as CEO, president and board member, with CFO Colette Kress taking over as interim chief executive. That is the central development in AI news today: a leadership transition at the company supplying the machinery behind so many AI plans. Nvidia’s board expects continuity. Customers will want evidence of it.
If you run a team, the practical response is neither panic nor a new slide about resilience. It is checking which commitments depend on one supplier, one service or one assumption. Thursday supplied several reasons to do that. Google launched a faster execution-oriented Gemini model. Notion disclosed exposed AI transcripts at a vendor. Nonprofits asked a California court to stop advanced AI development. And Microsoft 365 users encountered the old-fashioned failure mode: the service would not connect. Here is what changed, what remains a claim and what deserves your attention next.
AI news today: Nvidia appoints Colette Kress interim CEO
Nvidia’s November 12 statement confirms Huang’s departure from all three roles, effective immediately. Kress remains CFO while serving as interim CEO, and the board is considering internal and external candidates for a permanent successor. Harvey C. Jones becomes the company’s first board chair.
The board says operations, customer commitments and the product roadmap should continue without interruption. That is an assurance from the company, not an independently established outcome. Nvidia also reaffirmed November 17 for its quarterly results and earnings call.
For customers and investors, the useful next questions concern execution: who owns major account decisions, how delivery commitments are being maintained and whether the leadership search changes strategic priorities. A management change is not evidence that a chip roadmap has stopped. Equally, a continuity statement is not a substitute for checking your own contract and capacity assumptions. Our earlier Nvidia governance briefing provides background to the transition.
Gemini 4 Hydrogen puts model selection inside the workflow
Google introduced Gemini 4 Hydrogen on November 12 for everyday coding and agentic execution. Its launch page, updated November 13, lists a one-million-token context window and pricing of $1.50 per million input tokens and $7.50 per million output tokens. Google positions Hydrogen as a fast execution model that can work alone or alongside its Carbon and Argon models.
The announcement also describes Orchestrator in Google Antigravity, which switches models within a session. Google reports efficiency improvements in its own testing; those are vendor results, not a guarantee for your codebase. Antigravity access to Hydrogen is for paid Google AI Pro and Ultra subscribers, while the rollout also spans the Gemini API, AI Studio, the Gemini app and Search.
The operational question is cost per accepted result. A cheaper worker model is useful only if retries, review and escalation do not consume the savings. Before migrating, replay representative tasks and record completion, correction time and total spend. Treat model routing as something to evaluate, not an automatic discount.
Notion’s vendor breach exposes why data paths matter
Notion’s security advisory, originally published November 12, says the Fireworks AI incident affected fewer than 1% of Notion workspaces. The exposed dataset included Notion AI chat transcripts used for Insights. Most were not labelled with a user or workspace, but the content of a conversation can still be sensitive even without those labels.
Notion’s revised timeline says it disconnected Fireworks on November 3, deleted its data there on November 4 and contacted impacted customers on November 9. It disabled Insights and stopped serving open-weight models through Fireworks. The advisory tells affected customers to treat the relevant data as exposed and rotate passwords, API keys or other credentials pasted into affected chats. The investigation remains ongoing.
This is a November 12 disclosure of an earlier incident, not a claim that the breach happened that day. For buyers, the lesson is to map the complete data route: application, model provider, analytics and stored transcripts. Ask which feature retains what. A reassuring label on the front door does not document every room behind it.
Nonprofits seek an AI-development freeze, not just damages
Irreplaceable Education Fund, Public Citizen, Heat Initiative and the Alliance for Secure AI announced a lawsuit on November 12 in California’s San Francisco Superior Court. They seek a freeze on increasingly advanced AI development at Anthropic, Google, Meta, OpenAI and SpaceXSI, arguing that the systems pose catastrophic public-safety risks.
Those risk assertions are the plaintiffs’ allegations. Filing a complaint does not establish them as facts or mean a court has ordered a shutdown. In Al Jazeera’s reporting, one plaintiff also advocated inspections and enforceable limits on prohibited work.
The distinction matters for anyone planning around a model provider. A request for an injunction is a potential constraint to monitor, not a current service restriction you should invent. Watch for actual court orders, the scope of any requested relief and the defendants’ responses. The commercial question is whether contracts and operational plans can accommodate a genuine availability change without treating every lawsuit headline as one.
Microsoft 365’s outage makes the fallback plan concrete
Cisco ThousandEyes’ independent measurements recorded global Microsoft 365 availability problems on November 12, including connection failures, timeouts, server errors and traffic loss in Microsoft’s network. Its evening update still observed widespread Outlook disruption. These observations establish service failures; they do not by themselves establish the root cause.
A dated follow-up is important for this catch-up edition: at 14:00 UTC on November 13, ThousandEyes reported that Outlook availability had largely recovered, while some locations still experienced packet loss and timeouts. That is a measured recovery update, not a claim that every Microsoft 365 function was healthy everywhere.
If your agent needs email, calendars or documents to complete work, application availability belongs in its operating design. Decide where interrupted tasks wait, how people receive urgent messages and how duplicate actions are prevented after reconnection. The fastest assistant in the world cannot compensate for an unreachable dependency. Test the fallback with the service genuinely unavailable, not merely with a reassuring diagram.
Watch & Learn
“How to Route AI Agent Workloads Across Models with NVIDIA NeMo Switchyard” — NVIDIA Developer (6 minutes 7 seconds). This technical walkthrough explains model pools, routing signals, decision policies, execution and feedback. It is for builders deciding when to use an efficient model and when to escalate to a stronger one. The cascade examples are especially useful alongside Google’s multi-model announcement: watch how evidence changes the routing decision.
AI, Translated: Zero data retention
Zero data retention, or ZDR, is an arrangement limiting a provider’s storage of your inputs and outputs. It is not a universal promise that nothing is stored anywhere. Eligibility, features and exceptions matter: Anthropic’s documentation, for example, distinguishes eligible commercial products from consumer services and describes legal and misuse-related exceptions. Imagine an approved API call feeding a separate analytics log: the API’s retention terms do not automatically govern that log. You should care because “not used for training” and “not retained” answer different questions.
Try This Today: Draft an outage notice with Gemini
Goal: rehearse a useful customer update in ten minutes without inventing a recovery promise. Use a normal Gemini chat; no connected email or paid agent feature is needed. Google’s instructions explain entering a prompt and submitting it. Account, country and workspace access rules still apply.
1. Start a new chat and paste a fictional incident timeline, clearly labelled as an exercise.
2. Ask for a short customer notice and a separate list of unknowns.
3. Challenge every claim: request the exact timeline entry supporting it. Delete unsupported statements before reusing the draft.
Copy-ready prompt: “Exercise only. At 09:10 our fictional booking portal stopped accepting requests. At 09:25 we confirmed existing bookings remain stored. Cause and recovery time are unknown. Draft an 80-word customer notice with one safe next action and a next-update time of 10:00. Do not claim data loss, a cyberattack or restored service. Separately list unknowns and map each factual sentence to my timeline. Prepare text only; do not send anything.”
One thing to remember
Continuity is not a slogan. Know who can make the decision, where the data goes and what happens when the service disappears. Then test those answers before you delegate more work.