LONG AI
Home › Blog › NVIDIA's Jensen Huang Steps Down as CEO

NVIDIA's Jensen Huang Steps Down as CEO - First Handover in 33 Years, Market Cap Below $4T

Category:行业资讯 · 16 min read · 2026-11-13

This column aggregates the latest large-model releases and open-source drops from the past one to two days, cross-checked against official announcements and multiple media outlets. Performance figures distinguish between vendor claims and independently verified third-party results; anything not yet replicated is clearly marked. Illustrations are original.

Two silhouettes handing over a torch atop a glowing chip array illustration
A torch handover on a chip array - thirty-three years of single-leader rule comes to an end (illustration)

Headline: NVIDIA's first CEO change in 33 years, Jensen Huang steps down

On Nov 12, NVIDIA's 8-K filing and press release confirmed: the board accepted Jensen Huang's resignation as president, CEO, and board member on Nov 11, effective immediately. The company founded in 1993, with a single CEO for thirty-three years, has handed the reins to someone else for the first time.

Succession (official):

Market reaction (happened, not forecast):

Why this matters more than any model launch - three lines to separate:

Governance: the official statement gave no reason. Days earlier the board opened an independent investigation into Huang's personal matters; Reuters reported Nov 13 a second employee was involved. This is corporate governance and personal matter, unrelated to model tech; this column won't expand on it.

Continuity: both the chair and interim CEO said "continuity" - Jones expressed "full confidence," Kress said she'll "build on Huang's 33-year foundation." NVIDIA plans to hold ~70% gross margin in FY2027. What the successor really faces isn't the roadmap but the pricing structure: NVIDIA's pricing power has been tightly bound to the founder's personal credibility, while customers (Microsoft, Google, Meta, OpenAI) buy at scale while building their own chips. Whether the interim CEO holds that pressure - the Nov 17 earnings call is the first real test.

Spillover: Reuters reported over $3.2B of compute projects in China designed around restricted NVIDIA chips now face uncertainty amid this upheaval and US scrutiny.

The biggest open-source story: a 145B model runs in a laptop, peak 29GB memory

On Nov 12, Edge0 announced open-sourcing Edge0-145B and Edge0-120B, with a unified inference framework and deployment tools, Apache 2.0 code, weights on Hugging Face. It's positioned as "the world's first 145B open on-device model."

How it fits in a laptop - three things: sparse expert-shared large matrix (145B total, only ~10B active per inference); three-tier storage pipeline (hot experts in VRAM, ~40GB expert library in system RAM paged on demand, ~48GB vocab on SSD read live, CPU + GPU in parallel); speculative decoding spreading weight-read cost across more output tokens.

Vendor laptop numbers (self-tested, no third-party replication):

Three discounts to apply: all in-house; peer memory figures are estimates not measurements; "laptop-runnable" assumes 48GB RAM - a normal 16GB machine can't. It solves "a high-end laptop working offline," not "everyone runs a big model locally."

Other releases (5)

Model / ProductDateTypeHighlights
Granite 4.1 LLMON preview (3B/8B, IBM)11-12Open source Apache 2.0Uses a new markup language LLMON (LLM Object Notation) to separate "instruction" and "data" at the vocab level: six reserved tokens as delimiters, unique two-token IDs per block, an exec block stating which instruction applies to which data, unreferenced content read-only. OPI benchmark: text-injection success 3B 100% to 0.1%, 8B 96.3% to 0.02%; tool-output injection (ASB) still 5.6% / 3.1% - the exact tier agents fear. Trained on 4.6B LLMON tokens + 5-round GRPO + 1,000 DPO pairs. Experimental preview, not for production, and only effective if the pipeline renders LLMON.
SOV-Transcribe-1 (Salesforce)11-12Open source Apache 2.0"Sovereign AI" ASR for government and regulated enterprises: Whisper large-v3 encoder + Gemma 4 26B A4B decoder, 26.6B total / 4.6B active, ~53GB BF16. AA-WER 2.28%, first among open, #7 of 44. Hard limits: 7 European languages, no Chinese; non-streaming, no diarization or timestamps; no tech report.
Lara Mini (Translated)11-12Open weights, research/non-commercialA counter-argument experiment: "multilingualism isn't a burden, it's an intelligence source" - trained from scratch on 45 languages and 6T tokens (incl. EU's 24 official languages). Just 600M active params, first among sub-2B-active open bases, claimed to beat models trained on 6x the data. Weights + report + a 110-language-variant benchmark released.
Inkling 2 Small full weights (Thinking Machines)11-12Open source Apache 2.0Delivers the Nov 5 promise: 594B total / 17B active, 1M context, weights on HF after external red-teaming. AA index 31 (prior Small 13), among the highest open scores outside China. Reality check: NVFP4 needs ~500GB VRAM, MXFP8 ~800GB - sparse saves compute, not storage; hosted at $0.30/$1.00 per million.
Gemini 4 Hydrogen (Google)11-12Closed APIThe "workhorse tier" of the Gemini 4 family, distinct from deep-reasoning Carbon/Argon; 1M context at a flat $1.50/$7.50 per million (2x 3.8 Flash, 0.75x Argon); four thinking tiers, default-high AA index 51, xhigh 54, Terminal-Bench 4.1 18.2% to 57%. Antigravity's Orchestrator switches Carbon and Hydrogen in-session. All vendor-reported; separate article on our site.

Industry news

The trend in one line

The two threads today are one thing: the ownership of compute and models is being re-priced. NVIDIA swaps "one person's judgment" for "a system's judgment," and the market priced the handover at a 14% intraday drop; the same day Edge0 squeezes 145B into a 48GB laptop and IBM writes "who may give orders" into the grammar - everyone is turning "must depend on others" into "keep it in your own hands."

(Data from official announcements and public reports, 2026-11-13; vendor claims and third-party results are marked separately, with unverified items noted in the text. Model capability affects selection, but actual pricing follows the model plaza in real time.)

相关文章

GPT-6.2 Cuts Prices 20%: What the AI Price War Means

GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer

What Is Prompt Caching? How Cache Tokens Cut API Bills

Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi

All 12 Claude Models Compared: Price and Best Use

Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload

Which AI Model Writes Code Best? Speed and Price Tested

gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi

One API Key, Every Model: How to Integrate a Multi-Model Marketplace

Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa