Skip to content

AX Says · Nov 13, 2026Nvidia Leadership Change

Nvidia Replaces Jensen Huang With Its CFO as Interim CEO

Nvidia's board removed Jensen Huang today, with CFO Colette Kress taking over temporarily. It has not named a successor. Google released Gemini 4 Hydrogen the same day, while Anthropic filed its IPO paperwork: new products on one side, financial reckoning on the other. First, the biggest story: Nvidia's timeline.

Nvidia Replaces Jensen Huang, With Its CFO Taking Over Temporarily

Nvidia's board confirmed the rumors today: Jensen Huang resigned immediately as president, CEO and director. CFO Colette Kress became interim CEO, while the board began searching inside and outside the company. TechCrunch's timeline notes that Nvidia confirmed in a November 12 SEC filing that its board had accepted his resignation.

The board says operations, customer commitments and the product roadmap will continue uninterrupted. Third-quarter fiscal 2027 earnings will come out as scheduled on November 17. But treat this as just a personnel change and you miss the loose ends. Kress will also remain executive vice president and CFO, with no timetable for a permanent successor. The board also named Harvey C. Jones chair. Some in the industry are already comparing Huang's departure with Apple's ouster of Jobs in 1985. Lex Fridman called it a huge mistake for America and humanity. Rohan Paul added a list—Nolan Bushnell, Steve Jobs, Travis Kalanick, Sam Altman—and said boards relearn every few years that founders are the strategy. Tomasz Tunguz also posted a five-point analysis of Huang's departure and the AI economy. His conclusion: a big short-term shock to sentiment, but Nvidia's position in chips remains intact over the long term.

There's a wrinkle on the open-source side, too. Nvidia had planned to acquire Hugging Face for $12.9 billion, but the deal is not complete. After Huang's departure, Hugging Face CEO Clément Delangue said he would fight even harder for open-source AI. The US Justice Department has begun asking how the deal might affect competition in the AI market. With Huang gone, open-weight models, agent safety, AI tools, AI PCs and data-center financing have lost the industry's loudest advocate.

Huang Is Gone, but the Reckoning Isn't Over

I'll be more direct here: replacing the CEO does not clear Nvidia of chip diversion allegations. That is Elizabeth Lopatto's assessment: the resignation does not erase the chip-smuggling allegations or the regulatory risk. Brookings Institution researcher Kyle Chan said on a program that too much about the allegations remains unclear. China may want to learn about or influence US chip export controls through Huang, he said.

Democratic lawmakers acted today, too. Rep. Richard Neal and Sen. Ron Wyden wrote to Commerce Secretary Howard Lutnick, asking him to preserve records on Nvidia's chip sales to China. The letter questions whether shipping H200 chips to the US before imposing tariffs amounts to an unconstitutional export tax. It also presses for answers on export licenses for companies including Megaspeed, and on Nvidia's $26 million donation to Trump projects. This reads less like a routine inquiry than an investigation tying together trade policy, export licenses and political donations.

Trump had a suggestion of his own on Truth Social: former Apple CEO Tim Cook should take over Nvidia. Cook ended his 15-year run as CEO this September. He became Apple's executive chair and took charge of global policy affairs, while John Ternus became CEO. If that happened, it would be very strange. An executive from a consumer-products company would run a chip giant, while Apple is building out its own AI chip plans.

The question isn't whether the regulatory fight is coming. It has already reached the CEO appointment.

Google Puts Gemini 4 Hydrogen Everywhere

Google did not let the Nvidia story slow it down today. It released Gemini 4 Hydrogen, a new model focused on fast coding and agent execution. It launched globally today through the Gemini API, AI Studio, the Gemini app and Google Search, among other entry points. The approach assigns deep planning and fast execution to different models. It switches automatically for tasks closer to Carbon, cutting quota use and completion time while maintaining task success rates.

Search got an upgrade, too. Google Search now uses Gemini 4 Hydrogen, available to all users worldwide. It handles multipart questions, hard-to-decipher misspellings and specific requests, while searching related subtopics in parallel. Answers include more images, structured content, comparison tables and custom tools. Every language supported by AI Mode also gets finer-grained language understanding.

My read of these releases: Google wants a cheaper execution model to spread across search and agent use cases, rather than keep piling on parameters. For ordinary users, the test is whether search shows fewer ads and gives more direct answers. We'll see in a few weeks.

Anthropic Files for an IPO and Opens Its Books

Anthropic's IPO filing with the SEC today has numbers more startling than its products. It plans to offer Class A common stock under ANT on Nasdaq. The roadshow began November 11, with pricing and trading expected November 19, subject to regulatory approval. Third-quarter 2026 revenue was $17 billion, up more than 12-fold year over year, with a 15% adjusted operating margin. But its net loss was $23.3 billion, including $21.9 billion in noncash charges from a rising valuation.

That structure means the huge reported loss is driven mainly by valuation and accounting treatment, not cash burned by the business. On the 15% adjusted operating margin alone, Anthropic's operations can already fund themselves. Yet a higher valuation makes the net loss look scarier. The underlying disclosures do not break out net revenue by region or give its specific growth rate. Still, that revenue puts Anthropic near the top among AI companies.

The IPO may affect market sentiment longer than Nvidia's leadership change. It will provide a comparable valuation anchor for AI labs. Every later fundraising round will be measured against Anthropic's price-to-sales ratio.

Claude's Product Line Gets Denser: Default Model, Meetings and Message Threads

Anthropic did plenty on the product side today, not just in finance. Claude Opus 5.5 is now the default model for Pro, Max, Ultra and Team plans. Anthropic recommends it as a daily workhorse for research, analysis, writing and other knowledge work. Then there is the Claude Meetings beta. Claude can join video calls as a visible participant, answer questions, collaborate on documents and whiteboards, and produce notes and action items.

There is a technical thread behind this, too. Message Threads entered public beta today. It cuts the median long-conversation request package from about 800 KB to 1.8 KB. That reduces repeated uploads, validation and tokenization, but not token use or the bill. In other words, chat should feel smoother. It will not get cheaper because of this.

Together, the three moves point in one direction: make Claude the default entry point for longer conversations, meetings and daily workflows. Not just a model called for individual tasks. That matters more than a standalone model release because it changes user habits, not benchmark scores.

The Cost Ledger: OpenRouter's Translation Pipeline, Loop Brakes and Decision Models

OpenRouter's cost breakdown is more detailed than those of many large-model companies. It published a self-correcting website localization pipeline. GPT-6.1 Sol translates 16,918 strings, with code checks, same-model validation and daily AI-review retries. In the worst case, each string takes about 20 model calls. Retranslating the whole site into seven languages costs about $460, roughly 6% of Smartling's list price for machine translation. Error reports become terminology rules, style notes or evaluation cases. OpenRouter also published a guide to stopping AI agents from looping forever. It offers four brakes: doom-loop detection, loopKey normalization, stopWhen step and cost limits, and API-key credit limits.

On another front, Google released Gemini Decision 1 and the Gemini Decisions API. They make decisions among a closed set of options in a single computation, without generating text token by token. The cost logic resembles OpenRouter's: not every question needs a long model response. Some decisions should take one step.

Engineering collaboration is getting a similar cost rethink. A team lead with about 15 years of experience said the team dropped peer PR reviews after adopting Claude. Claude writes code, the task owner reviews it, and the owner merges it once CI passes and Greptile finds no issues. Claude also helps team members quickly learn work that colleagues own independently. This does not save tokens. It cuts the process costs and waiting time the team once depended on.

Small Signals Today

Translated from the Chinese original. Read the Chinese original

Some sources via AI HOT

The day's picks