Huang steps down as NVIDIA CEO, company loses more than $1 trillion in market cap
NVIDIA CEO Jensen Huang has resigned after the board opened an investigation into a scandal involving infidelity, affairs with subordinates and spies. The company's market value fell by more than $1 trillion. While the board looks for a replacement, CFO Colette Kress is running the company. If you are planning a large GPU purchase, build extra lead time into your schedule.
AI-processed from TG: @ai_newz (эйай ньюз); edited by Hamidun News
The biggest news of the two November weeks (November 2–15) came not from an AI lab but from NVIDIA's boardroom. Chief executive officer (CEO) Jensen Huang resigned after the board opened an investigation into a scandal involving infidelity, affairs with subordinates and spies. The company's market capitalization, meaning its total market value on the stock exchange, fell by more than a trillion dollars. For anyone building products on NVIDIA GPUs, this is a signal: the leadership of the main supplier of AI hardware has changed. Below is that story and the rest of the releases from AI News digest #137.
What happened at NVIDIA
The sequence was as follows: the board launched an investigation, Huang then left his post, and shareholders reacted by wiping more than $1 trillion off the company's value. While the board searches for a new chief, NVIDIA is run by CFO Colette Kress. The digest's author notes that some of the allegations appear to have been refuted already, and that rumors say Huang is trying to return.
What should you do in practice? If you are planning a large GPU purchase or a long-term cloud compute contract, build extra lead time into your schedule: under an interim leader, strategic decisions may move more slowly. The source gives no reason to rush into rewriting your stack or abandoning NVIDIA over a single resignation; it says nothing on that point.
Language models: weights, prices and speed
Inkling 2 Small. Thinking Machines released the weights of a MoE model with 594B parameters, of which 17B are active for each token. MoE (mixture of experts) means only a portion of the "experts" works on any given response, so it runs faster than a dense model of the same size. The quantized NVFP4 version needs two B300s or four H200s. That is not something you run on a laptop, but for a team with its own cluster it is a realistic task.
Gemini Decision 1. Google released this model; the digest covers it in a single line: "Now Google has its own Jev." The source does not explain what "Jev" refers to, so there are no details about the model's capabilities.
Gemini 4 Carbon. Available in the paid API and in Gemini for Ultra subscribers. Through November 30, it costs $2 per million input tokens and $10 per million output tokens, then rises to $4 and $20. If you have heavy batch workloads, such as bulk document processing, it is cheaper to run them before the end of the month.
Fable 5.2. Generates responses up to 30% faster than version 5.1. In the API it costs $8 per million input tokens and $40 per million output tokens. For comparison, that is four times the price of Gemini 4 Carbon input tokens through November 30, and twice the price after the increase. It makes sense to pay that if response speed matters to your users.
Kimi K3.1 Pro. Moonshot released the model's weights and replaced Gated MLA with Gated DSA, a different variant of the attention mechanism. The aim of the change is more efficient handling of long context. If you run long documents on your own servers, measure memory use and speed yourself.
Images and video
MiniMax H3-Image. Open weights, native 4096×4096 resolution in six generation steps, and up to 16 references. The low step count means fast generation, and 16 references let you keep one style or the same characters across a series of images.
FLUX 3 Max. On the fal platform, users could try video generation with sound for free: up to 15 seconds at 720p from text or an image. It was a promotion, so it was better suited to evaluating quality than to regular work.
Gemini Omni 1.2 Flash. Generates and edits video with sound up to 15 seconds long. In the API it costs about $0.07 per second at 720p, with input tokens billed separately. A maximum-length 15-second clip comes to roughly $1.05, not counting input. That is tolerable for prototypes and ad tests, but for mass generation you need to calculate the budget in advance.
Other: a game console powered by Codex
OpenAI released the ModRetro Chromatic retro console for $200. You can build your own games in Codex and load them onto a cartridge. It is more a demonstration that a coding agent can take a project all the way to a physical medium, but the kit is an interesting option for hobbyists and for learning.
The bottom line
Open weights are getting larger (Inkling 2 Small, Kimi K3.1 Pro), while closed APIs compete on price and speed (Gemini 4 Carbon, Fable 5.2). In practice, two things are worth doing: recalculate your API budget to account for the Gemini 4 Carbon price increase after November 30, and check whether your workload fits on two B300s or four H200s with an open model.
- Nov 18, 2026Google releases Gemini 4 Neon for Chat and Tasks: 1M tokens and $3 per million input tokens
- Nov 18, 2026Google releases Gemini 4 Neon: 3rd place in AI ranking and tasks 32% cheaper
- Nov 18, 2026Devin Bot built an Android app on its own in 44 days: 1 134 PRs from 183 messages
- Nov 18, 2026Google finds AI agents that discover vulnerabilities and hack websites on their own
Key Questions Answered
What happened at NVIDIA?
The sequence was as follows: the board launched an investigation, Huang then left his post, and shareholders reacted by wiping more than $1 trillion off the company's value. While the board searches for a new chief, NVIDIA is run by CFO Colette Kress. The digest's author notes that some of the allegations appear to have been refuted already, and that rumors say Huang is trying to return. What should you do in practice?
What are the key details on Language models: weights, prices and speed?
Inkling 2 Small. Thinking Machines released the weights of a MoE model with 594B parameters, of which 17B are active for each token. MoE (mixture of experts) means only a portion of the "experts" works on any given response, so it runs faster than a dense model of the same size. The quantized NVFP4 version needs two B300s or four H200s. That is not something you run on a laptop, but for a team with its own cluster it is a realistic task. Gemini Decision 1.
What are the key details on Images and video?
MiniMax H3-Image. Open weights, native 4096×4096 resolution in six generation steps, and up to 16 references. The low step count means fast generation, and 16 references let you keep one style or the same characters across a series of images. FLUX 3 Max. On the fal platform, users could try video generation with sound for free: up to 15 seconds at 720p from text or an image. It was a promotion, so it was better suited to evaluating quality than to regular work. Gemini Omni 1.2 Flash.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.