[AINews] Jensen Huang resigns after 33 years leading NVIDIA

NVIDIA announced on November 12 that Jensen Huang has resigned as CEO, president and board member, ending a 33-year run at the company. The SEC filing dates the board's acceptance of the resignation to November 11. Colette Kress, CFO since 2013, keeps the CFO role while serving as interim CEO. Harvey Jones becomes the board's first chair, and the board has opened a search for a permanent successor. Separate reports of personal conduct had circulated in the days before the announcement; the source notes that any causal link between those reports and the resignation is debatable.
The board had been moving toward a transition before the announcement. On November 10 the independent directors named Joon H. Kim and Wachtell, Lipton, Rosen & Katz to run the investigation they announced on November 8. An earnings call is expected next week. Discussion in developer communities focused on what the change means for CUDA, the GPU roadmap and Hugging Face, with readers debating whether access to model weights and datasets would be preserved after the proposed acquisition; some said they were backing up assets locally. A reported employee petition to bring Huang back circulated without independent confirmation. Clem Delangue and NVIDIA's Nader Khalil both reiterated support for open source.
Google DeepMind shipped Gemini 4 Hydrogen, a speed-first model slotted alongside Carbon and Argon. Google claims it completes tasks 2.5x faster than Gemini 3.8 Flash and edges Argon on the Artificial Analysis Intelligence Index, 51 versus 50 at default thinking and 54 at a new xhigh level. It has four thinking levels (low, medium, high, xhigh) and a 1M context window, priced at $1.50/$7.50 per million input/output tokens with a 95% discount on cached input. It is available on all Gemini API tiers including a free tier in AI Studio, is the default in Search AI Mode, reaches free Gemini app users through auto-selection, and is open to paid Pro and Ultra subscribers in Antigravity. Google reports 57% on Terminal-Bench 4.0, more than 3x Gemini 3.8 Flash on Terminal-Bench 4.1, and more than 5x fewer fabricated claims, plus leading results on Vals Tax Agent Bench, AutomationBench-AA and Agents' Last Exam.
Independent measurements paint a more mixed picture. At xhigh, Artificial Analysis scores Hydrogen at 54, one point behind Claude Sonnet 5.5 and GPT-6.1 Sol at max. Its cost is roughly 32% of Sonnet's but about 2.5x Sol's, at $3.16 per task at xhigh and $1.87 at high, with output speed around 200 tok/s. Its 8% hallucination rate on AA-Omniscience compares with 43% for Sonnet and 52% for Sol, and it shares the top score on GDP.pdf visual reasoning. On Terminal-Bench 4.0 it lands at 54% at both high and xhigh, below Sonnet's 64% and Sol's 56%. On Agent Arena it debuts fifth with a 5.3-minute median task time and $1.6 median cost; Code Arena WebDev puts it tenth at 1703 points; Vals Index has it sixth at 68.5%, two points behind Carbon at 38% of Carbon's price. Users report runaway reasoning at xhigh, including 131K thinking tokens and about $1.08 spent without answering an image question, and a 1% score on MazeBench. Simon Willison notes the price is only a 25% discount on Carbon/Argon, which reportedly double on December 1, and Theo disputes the speed framing from his own timings. Google is also rolling out an Antigravity mode where Carbon plans while Hydrogen writes and tests code, claiming near-Carbon Terminal-Bench 4.0 success with 24% less quota and 19% faster completion.
Thinking Machines released Inkling 2 Small under Apache 2.0: a 594B-total, 17B-active mixture-of-experts model with text and image input and up to 1M context. Its architecture adds attention residuals with a q/k/v split, gated attention, a dynamic attention sink, dynamic shortconv and value norm; experts use NVFP4 in layers 0–39 and MXFP8 in the last eight, with sliding-window attention widened to 1k. On vLLM it fits one GB200 or GB300 node and reaches up to 630 tok/s per user at TP4. It ranks first among US open models on Code Arena WebDev at 1647 and second among open models on Agent Arena at a $0.26 median cost per task. Other releases include Perplexity's open-weight Decider v1.3 with a span-offset "locate" capability, MiniMax H3-Image, Salesforce's SOV-Transcribe-1 at 2.3% AA-WER, and vendor claims of 86.5% Terminal-Bench 2.1 for the device-native Edge0-145B/120B models.
A separate thread concerns checkpoint forensics. Will Depue says initialization patterns in released weights reveal which GPU produced them, reporting that six Chinese models, including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3, were initialized on Blackwell systems, four on Grace Blackwell, with a 24-for-24 match on models with disclosed hardware. He estimates 2–3K Blackwells each for the largest runs. Critics counter that this is roughly 6MW against China's roughly 24GW of installed compute, arguing the finding may show export controls working, and note that his analysis also places some Chinese models on Ascend chips.
On the tooling side, Anthropic opened a private beta of self-hosted Claude Managed Agents that runs the harness, sandboxes and session history in a customer's own cloud or on-prem, adding agent identity, OpenTelemetry, kill switches and an evals product. Claude Meetings is in beta on Team and Enterprise, and OpenAI launched Daybreak Defense Factory and Daybreak Red Team.
Why it matters: NVIDIA is the substrate most AI teams build on, so a founder-to-interim-CEO handoff injects real uncertainty into GPU roadmaps, CUDA continuity and the Hugging Face acquisition — this is inference, but procurement and platform teams may want contingency plans. Gemini 4 Hydrogen's aggressive pricing and low hallucination rate give developers a cheaper default for high-volume and agentic workloads, provided they account for runaway thinking costs. Open-weight releases like Inkling 2 Small let teams self-host strong models on a single node.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
November 12, 2026 — Colette Kress takes over as interim CEO. What happens to open source/American AI leadership?