Fabrizio Mazzei, home
AI BlogAI NewsAI LabBooksAboutPortfolio
How can I help?
How can I help?
SundAI Blog

What weighs more this week: Nvidia without Huang, or the cost of a task?

#Nvidia/#Costi AI/#Sicurezza AI/#Infrastruttura AI/#OpenAI/#Google

Fabrizio MazzeiFabrizio MazzeiAI Solutions Architect11/15/20269 min readINSIGHT#50
What weighs more this week: Nvidia without Huang, or the cost of a task?. AI-generated image

Image generated entirely with AI.

TL;DR

"Huang leaves Nvidia and the stock loses about 13% on the week, with market value under $4 trillion. In the same stretch a Linux exploit costs $1,279, and Google switches on an endpoint that answers without generating text."

Loading audio player...
  1. 01What happens to Nvidia without Jensen Huang?
  2. 02Has closing a flaw become slower than finding one?
  3. 03Has the moat moved to the dataset?
  4. 04Does a judge that only answers closed questions actually help?

On Thursday, November 12, Jensen Huang resigned as president, chief executive, and a member of the board of Nvidia, effective immediately. Thirty-three years after the founding. Colette Kress, chief financial officer since 2013, takes over as interim chief and keeps the CFO role. Harvey C. Jones, a director since 1993, becomes the board's first chair. The stock closed Thursday at -4.3%, after a drop of as much as 16% during the session. For the week it lost about 13%, a sixth straight down week, the longest streak since 2012. On Friday market value closed below $4 trillion. Since early October, according to CNBC, the decline was already about 30%.

In the same stretch the labs kept shipping, and kept putting a price on the finished operation. OpenAI priced vulnerability research, Google switched on an endpoint that answers a thousand questions without generating text, and Gemini 4 Hydrogen entered Search as a default. Two different price lists: one is moved by whoever sells the chips, the other by the cost of a task.

What weighs more this week: Nvidia without Huang, or the cost of a task?. AI-generated imageImage generated entirely with AI.

What happens to Nvidia without Jensen Huang?

There was no designated successor. Huang leaves the board as well: he was the fourth-largest shareholder, with about 3.6% of the company. The board said it has full confidence in Kress and in the team, and it opened an internal and external search.

The release promises continuity. This week's price put a number on the vacancy.

On Tuesday, November 17, fiscal third-quarter 2027 results stay on the calendar, at 1:20 p.m. PT, with the call at 2 p.m. Analysts compiled by LSEG expect revenue of $108.9 billion, up 91% year over year. Listen for whether guidance holds. If it holds, the market had priced the person more than the business. If it comes down, the premium on the person was a real slice of the price.

The board moved because of reports about private life, and the thread that stays industrial is the chips. On November 8 Huang wrote to employees that he had been "unfaithful", and he tied the apology to the national-security concerns in the press. The same day the board opened an investigation: it has no reason to believe Nvidia secrets or national security were compromised. The relationship that matters for the industry is the one with Alice Huang, a Megaspeed executive and no relation to the CEO, described by the New York Times in late October. Megaspeed, a Singapore neocloud, reportedly bought about $4.8 billion of Nvidia chips over two years. The US and Singapore have been investigating it since October 2025 on suspicion of routing those chips to China. US officials believe she was tied to Chinese intelligence. Beijing denies it, and the Justice Department has not yet concluded whether she was.

On November 11 the Times reported a second relationship, with a vice president who reported directly to Huang. The board met on Wednesday, and on Thursday morning he resigned. Reuters also reported a third relationship, with another employee, unnamed. In the letter opening an inquiry, Senator Ron Wyden and Representative Richard Neal write that in December 2025 the administration still authorized H200 sales to China, in exchange for a cut of the exports. Huang has agreed to testify in Congress before November 26. JPMorgan, still before the resignation, raised the risk of tighter rules on shipments to China. For anyone buying APIs or reserving GPUs, if the licenses tighten then part of the price stops being a slide. It is the same pattern as when the token costs half as much: watch the endpoints, how often 429s show up, and fallbacks onto more expensive models.

Insight tecnico. AI-generated imageImage generated entirely with AI.

Has closing a flaw become slower than finding one?

OpenAI introduced GPT-6 Cyber, a model specialized in computer security. The stated results: 81 previously unknown Firefox vulnerabilities, judged serious by Mozilla and largely already fixed, and 948 working exploits against the Linux kernel, able to gain administrator privileges from a limited account. The model reads the code, looks for the flaw, checks whether it is exploitable, and proposes a fix. On ExploitGym it reaches 48.5%, against 32.1% for the previous version.

The number that weighs is the bill for the Linux campaign, told in the study A thousand fewer paths to root: about $1.2 million in API credits, or $1,279 per exploit, across 2,308 signals examined and 1,724 reproduced. Vulnerability research becomes a metered service with a readable unit price.

Outside checks cut the picture down. On 288 trials against real software and hardware run by Irregular, GPT-6 Cyber passes 90, against 96 for GPT-6 Astra, and neither solves the 15 hardest challenges. Access stays closed to selected organizations, through Daybreak Red.

The operational point is the asymmetry of time. Someone attacking with this tool takes hours. Someone defending takes weeks to ship a patch and get it into production. In practice the metric to watch inside a company is mean time to close, not the count of vulnerabilities found. And for basic hygiene, stale dependencies and open configurations, a previous-generation model is cheap and enough. Thinking in terms of cost per completed task here is a spreadsheet: $1,279 against the cost of an incident.

Has the moat moved to the dataset?

Andrej Karpathy stood up HackerTube in a weekend. Claude Opus 5.5 went through 43 million Hacker News comments and distilled them into 6,454 videos. The product is spare: no login, no optimized thumbnails. TristanTube followed, built in three minutes with Claude Code, with 47,000 videos. Three steps stayed manual: the API key, the server rental, the domain.

The frontend now costs almost nothing. The moat, if there is one, is the filtered corpus and the selection criterion. Rebuilding yet another pretty interface builds nothing.

On the same line, from the commercial side, OpenAI opened Businesses in ChatGPT: verified profiles, a panel of aggregate insights on mentions and clicks, and a self-serve product feed where a company uploads prices and availability. Shopify merchants connect the store and products pass through Shopify Catalog. The stated reach is 1.2 billion people a week. Whoever claims the profile and keeps the catalog clean decides what the model can say about them. Whoever stays out gets described by whatever the model collected elsewhere.

The same week Anthropic updated Claude Fable 5.2 and Claude Mythos 5.2, available on Amazon Bedrock, with the expected Wall Street IPO and a $3.4 billion gas-turbine purchase in the background. In tests reported by AI Parabellum, Fable 5.2 supposedly designed three working electronic devices with no circuit corrections. That claim needs the schematics, not the post. What can be judged now is the price per token on Bedrock, the latency, and the stability of the endpoint.

AI News, every morning

SundAI comes out once a week, but AI news doesn't wait for Sunday. Every morning I pick the best one and comment on it.

TodayNovember 16, 2026

Claude joins your meeting and your spreadsheet: Anthropic targets knowledge workers

Go deeper

Does a judge that only answers closed questions actually help?

Google announced the Gemini Decisions API, an endpoint built on gemini-decision-1. You pass a single multimodal input, up to 256K tokens, together with at most 1,000 independent questions, and you get back probabilities and confidence scores without generating a single text token.

The economics change shape. Instead of decoding token by token, the model scores the options in a single forward pass: 150-250 ms latency, $0.10 per 1M input tokens, zero output tokens billed, latency almost flat from 1 to 1,000 questions. Google says the cost is about five times lower than Gemini 3.5 Flash-Lite with the Interactions API. Supported types are yes/no, choice from 2 to 255 options, and score on ordered rubrics, each with confidence from 0 to 1. Internal uses cited: model routing in Gemini Notebook, fast wireframes in Stitch, emoji selection.

The value is in the contract. A judge that answers in prose forces you to parse the output and pay for generation. Here the model does the parsing and the code receives a number: above the threshold it proceeds alone, below it lands in a human queue. The hard part moves to the vocabulary of the questions. If the confidence is honest, a 150 ms classifier replaces whole intent routers. If it is optimistic, it stays a fast classifier. The test to run is calibration error on your own questions, not on Google's benchmarks.

Gemini 4 Hydrogen arrived inside Google Search globally, the same day as the DeepMind post. A frontier model inside search means anyone searching uses it without picking a plan. It is not an API to integrate: it is a default already on. The release promises more visual answers and is silent on latency, cost per query, and usage limits. For publishers, a search that answers in place of the links moves traffic. For people building products, it is another case in which the model under the product changes without a release of their own.

Tool watch

  • Langfuse: tracing and evaluation of LLM calls, open source and self-hostable, with prompts, cost, and latency per feature.
  • llama.cpp 1.0: a stable API and semantic versioning, the starting point for models on your own hardware.
  • AQuA: Google's open-source agent that rereads production sessions, finds recurring failures, and traces the cause back into the code.
  • CodeBurn: a local count of tokens spent with Claude Code and of plan limits, no account.
  • LiteLLM: a gateway with an OpenAI-compatible API in front of cloud models, with quotas, cost, and observability in one place.

On Tuesday Nvidia says whether guidance holds without the person the market was paying a premium for. The price list you can use on Monday is the other one: $1,279 per exploit, $0.10 per million tokens for a judge that answers yes or no. Products are still compared on the cost of a finished task, even in the week the chief executive of the company that sells the shovels changes.

Sources

  1. GPT-6 Cyber finds 81 Firefox flaws and 948 Linux exploits: security becomes an industrial processilsole24ore.com
  2. Karpathy filters 20 years of Hacker News with Claude and builds a better YouTubethe-decoder.com
  3. Anthropic ships Fable 5.2 and Mythos 5.2 on the eve of its IPOtechcrunch.com
  4. OpenAI opens ChatGPT to businesses: verified profiles, product feeds and commerce agentsopenai.com
  5. Google's Decisions API: a thousand answers in one pass, zero output tokensblog.google
  6. Gemini 4 Hydrogen lands inside Google Searchblog.google
  7. Half a trillion dollars gone as Nvidia chips turn up in Shanghai serversilmanifesto.it

Text created with AI assistance and reviewed by me.

Share this insight
  1. 01What happens to Nvidia without Jensen Huang?
  2. 02Has closing a flaw become slower than finding one?
  3. 03Has the moat moved to the dataset?
  4. 04Does a judge that only answers closed questions actually help?

Share this insight

Copertina: Dal CV al Colloquio con l'Intelligenza ArtificialeNewDal CV al Colloquio con l'Intelligenza ArtificialeDiscover the book

Found it useful? I have more like this

Every week I pick the most interesting and high-impact AI news and share them in an email recap. Subscribe so you don't miss the next one.

One email a week. Unsubscribe in one click.

Before you go, I recommend you also read these insights.

How much does the autonomy of our AI agents really cost?

How much does the autonomy of our AI agents really cost?

OpenAI's agents breach government sites, Nvidia responds with hardware guardrails and the Pentagon rewrites its supply chain. It was the week the bill for autonomy came due.

Read more
Are Chinese open-weight models and multi-agent swarms redefining artificial intelligence infrastructure?

Are Chinese open-weight models and multi-agent swarms redefining artificial intelligence infrastructure?

From GPT-5.6 solving historical theorems with 64 parallel agents, to the rise of Kimi K3 slashing corporate costs. Less apocalyptic hype, more focus on productivity, security, and prompt engineering.

Read more
Are dynamic routing and low-cost models the real solution for scaling autonomous agents?

Are dynamic routing and low-cost models the real solution for scaling autonomous agents?

Hype gives way to engineering: from dynamic routing to reduce API costs, to new hardware architectures where CPUs once again dominate to orchestrate complex workflows.

Read more
Copertina: Dal CV al Colloquio con l'Intelligenza Artificiale
New

Dal CV al Colloquio con l'Intelligenza Artificiale

From reading to your next application: a practical path to choosing job ads, showing your experience and preparing for difficult interview questions with AI. Book in Italian.

Discover the book
Discover my first book too

Listen to the Insight

AI Audio Version

Listen while driving or coding.

Ready
Fabrizio Mazzei, AI Solutions Architect e consulenza AI
Author

Fabrizio Mazzei

AI Solutions Architect

As an AI Solutions Architect I design digital ecosystems and autonomous workflows. Almost 10 years in digital marketing, today I integrate AI into business processes: from Next.js and RAG systems to GEO strategies and dedicated training. I also write practical guides to AI. My new book, "Dal CV al Colloquio con l'Intelligenza Artificiale", helps people find or change jobs: eight chapters and over 50 prompts to show what they can do throughout their application. The book is in Italian. My superpower? Looking at a manual process and already seeing the automated architecture that will replace it.

Need help with AI?Need a hand?Discover my books (in Italian)Let's Connect