← älter | neuer →
Nvidia Drama Nears Climax: New Affair, Plunging Stock, Succession SearchSynthszr
synthszr #318 from Thursday, November 12, 2026

Nvidia Drama Nears Climax: New Affair, Plunging Stock, Succession Search

  • • Anthropic introduces new features for Claude, including enterprise apps.
  • • Nvidia CEO Jensen Huang could be ousted by the board following affairs.
  • • Treasury Secretary Bessent confronts Lutnick over Nvidia customer Megaspeed.

Anthropic Brings Full Enterprise Apps Into Claude

Anthropic is expanding Claude so that the assistant can open complete enterprise applications directly within the chat, retrieve information, and make changes. This was unveiled on Tuesday at an event in San Francisco with several hundred guests, alongside new tools for meetings, spreadsheets, collaborative canvases, and enterprise data. “For years, we’ve learned to work around software. Now, Claude works around you,” said Chief Product Officer Ami Vora. According to her, the goal is to extend the benefits that programmers have enjoyed with Claude Code for over a year to knowledge workers across organizations.

Previously, interactive connectors only showed a small preview of the linked app. Moving forward, the entire application can be opened within the conversation, launching first with Adobe, Atlassian, and HubSpot. Claude can now, for example, edit a Confluence page or adapt an Adobe design using elements from a company’s Brand Kit. Users must approve such changes before Claude executes them. Technically, this is built on MCP Apps, an extension of the Model Context Protocol that allows servers to deliver interactive user interfaces instead of pure tool calls. Developers can embed existing app interfaces without rebuilding them. The feature is in public preview and is also open to other AI clients that support MCP Apps.

According to the company, more than 9,000 connectors and plugins are available in the Claude Marketplace, including Salesforce and Workday. For applications without a connector, Claude relies on a built-in browser. Anthropic states that existing employee access permissions remain in effect: anyone who cannot open a Google Drive document will not be able to access it via Claude either. Administrators can restrict individual capabilities, such as allowing Slack messages to be read without permitting sending. New in beta is “Auto-connect,” allowing admins to enable connectors organization-wide in about ten minutes. “Smart Reports” displays usage, completed work, and spend per team.

Claude Dashboards is now named Claude Data and is generally available for paying customers. The tool connects Claude to data warehouses and CRM systems. Newly introduced are certified queries approved by data teams to ensure different departments pull identical figures from the same data. Through live sharing, employees can view updated dashboards without direct access to the underlying data source. Both capabilities are currently in beta. Anthropic did not detail how versioning and schema migrations are handled.

Claude Meetings is scheduled to roll out first for Team and Enterprise customers in Google Meet, serving as a visible participant to deliver plans, timelines, and answers, as well as fetch data from connected apps. → Fortune, The New Stack, Engadget

Synthszr Take: Zapier, Make, and UiPath spent years thriving on the fact that apps don’t talk to each other. Claude turns this integration layer into a mere afterthought of the chat window. With Auto-connect, an admin can enable connectors for the entire organization in ten minutes, down to granular permissions like reading Slack without sending. This exact governance was the core selling point of the automation providers. In the future, their workflows will run invisibly in the background, and invisible vendors don’t dictate pricing anymore.

Stock Plummets: Nvidia Board Reportedly Prepares Ouster After Second Affair

Nvidia’s board of directors is reportedly preparing to replace CEO Jensen Huang. According to two people familiar with the matter, the board met Wednesday evening to discuss the 63-year-old’s future. Just days earlier, Huang had admitted to an affair with a suspected Chinese spy. In response, the board announced an internal investigation, which is still ongoing.

Immediately before the meeting, a report emerged about a second relationship. According to it, Huang had an affair with Kimberly Powell, 48, Vice President for Healthcare at Nvidia, who reports directly to him. Both are married to other partners. During the annual developer conference in San Jose in March, the two were allegedly seen acting intimately in a hotel. A bystander filmed the scene. Huang’s security team subsequently offered the person a non-disclosure agreement. Nvidia declined to comment on the report. Powell did not respond to a request for comment.

The stock reacted sharply. By 7:00 PM Eastern Time, Nvidia lost nearly 7 percent in after-hours trading. Since the initial report on October 29 about an “intimate relationship” of Huang’s, the company has shed roughly one trillion dollars in market value.

The ownership structure is crucial for the balance of power. The bulk of Huang’s Nvidia shares are held jointly with his wife Lori Huang and are housed in trusts. These trusts hold approximately 800 million shares. Huang personally owns around 70 million shares, leaving him with limited direct economic control over the company.

This leadership shakeup extends far beyond a single company. Nvidia supplies the chips that power a vast portion of the AI industry, and Huang has led the company for decades as its founder. As things stand, there is no obvious successor. → Wall Street Journal, Washington Post, Axios

Synthszr Take: The most critical figure sits at the bottom: around 800 million shares are held in trusts owned jointly by Huang and his wife; he personally holds about 70 million. His corporate power base thus hinges on the very person he admitted to hurting before the staff on Sunday. If the board names a successor by Wednesday, that power base is visibly gone. At that point, the market will price Nvidia without a founder premium for the first time—and that 9 percent was merely the down payment.

Bessent Confronts Lutnick in Situation Room Over Nvidia Customer Megaspeed

According to Axios, U.S. Treasury Secretary Scott Bessent confronted Commerce Secretary Howard Lutnick during a meeting in the White House Situation Room. The confrontation was prompted by Nvidia customer Megaspeed still not being placed on the Commerce Department’s Entity List. Megaspeed is accused of illegally diverting U.S. chips to China. According to sources, Lutnick initially hesitated and then responded that he didn’t know and would look into the matter. Axios reports that the White House and intelligence agencies have been pushing for action against Megaspeed since autumn 2025. Lutnick had reportedly stalled these efforts for a year out of concern over escalating trade tensions with China. → Axios

Synthszr Take: A year without a single addition to the Entity List comes at a price, and Nvidia booked it as revenue. As long as Megaspeed remains unlisted, every diverted chip is a grey-market transaction with predictable margins. Lutnick weighed this trade-off against trade tensions. What truly shakes it up is the CEO’s affair, despite intelligence warnings dating back to autumn 2025. Export controls only kick in when things become embarrassing; until then, business as usual.

United Airlines Blocks AI Agents, Resisting a New Middleman

United Airlines plans to block personal AI assistants from accessing its ticket booking systems in the future, the Wall Street Journal reports. The airline announced the move on Wednesday. As justification, United cited the risk that third-party agents could expose sensitive customer data or trigger unintended purchases. According to the carrier, automated actions could also violate regulations from the FAA and the U.S. Department of Transportation, which govern what information airlines must present to customers. → Wall Street Journal

Synthszr Take: The alleged explosive risk in the baggage is just a pretext; the line about the middleman is the real motive. Airlines have spent years fighting to pull bookings back from portals and reservation systems onto their own websites. A travel agent comparing prices and clicking on the customer’s behalf would be the next portal—just without a contract. For providers of such assistants, this means their product depends on interfaces they don’t control. Other carriers can simply copy United’s rationale, FAA citation included.

Anthropic IPO Multiple Times Oversubscribed, Banks Eye Price Above Range

Anthropic’s IPO is multiple times oversubscribed, according to Bloomberg. Underwriting banks are now considering a valuation of $1.5 trillion or more. Demand was reportedly propelled by large, Ankerinvestoren-style orders. According to the Financial Times, banks expect to price the offering above the $70 to $78 per share range set just this week, which corresponded to a valuation of around $1.4 trillion. Bloomberg reports that some investors expecting large allocations are already reselling portions before trading begins at a 3 to 5 percent markup over the offering price. → Capital Brief

Synthszr Take: An order book where allocations are flipped before trading starts at a 3 to 5 percent premium measures the expectation of being scaled back. Large investors therefore order more than they intend to hold. Banks interpret the oversubscription as a bullish signal nonetheless, as their fees grow with the offer price. The flippers' premium will ultimately be paid by retail investors, for whom Vanda expects a record $250 million in purchases. Every future AI lab will inherit this benchmark.

Anthropic Ports Claude to AMD Chips in Three Days, Chipping Away at Nvidia’s CUDA Lead

Anthropic reports that it got its Claude model running on AMD GPUs in just three days, with almost no human intervention. According to the IPO investor presentation, the work was handled by its proprietary frontier model, Mythos, which reportedly ported or rewrote nearly 100 percent of the Kernel autonomously. Previously, introducing a new chip platform engaged dozens of engineers over six to twelve months. These figures refer to prompt processing for an Anthropic model on AMD’s MI355X. Co-founder and Chief Compute Officer Tom Brown explained that Anthropic now trains and runs Claude across all four major chip families: AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs. → Techpresso

Synthszr Take: That chart comes from an IPO investor pitch and measures prompt processing on the MI355X. That is a narrow slice of inference, not training or continuous production load. AMD spent years demonstrating with ROCm that the first functional kernel is cheap, while production operations are expensive: debugging, profiling, and stability sustained over months. Nvidia’s 70-plus percent gross margin is anchored in precisely that stretch; it will only decline once Mythos takes over ongoing maintenance too.

Adobe Opens Creative Cloud to Third-Party AI Models and Builds Its Own Agents

At its Adobe Max conference, Adobe introduced new AI capabilities for Photoshop, Illustrator, Premiere, and After Effects, opening project files in these applications to third-party AI models for the first time. The integration relies on local connections built on the Model Context Protocol, initially available as a preview. This allows users to let a model of their choice make direct edits to their files, with layers, vectors, and sequences remaining fully editable according to The Deep View. A marketer who has already trained ChatGPT on brand assets and tone of voice won’t have to re-explain that context in another tool, the newsletter notes. → The Deep View

Synthszr Take: Adobe opens the door while fortifying the house. When ChatGPT operates directly within Photoshop files, no one needs a standalone, budget tool with its own image model. Layers, vectors, and sequences stay editable, and that is where a project’s true context resides. Migrating that context is far harder than canceling a subscription. Canva, Midjourney, and others are left competing on price alone. Adobe can lose the foundational model race as long as the file stays inside Adobe.

Product Managers in Unprecedented Demand as US Firms Hire Half as Many Engineers

Product and career writer Lenny Rachitsky has published his second job market report of 2026, his fifth overall. According to his analysis, product managers are winning in the AI era, while demand for engineers has stagnated. Design has hit an all-time low, as have remote roles. Tech jobs are increasingly concentrating back in the Bay Area. AI proficiency is now expected across virtually all roles. While the tech job market is expanding rapidly by his metrics, growth is occurring primarily outside of software development.

The report draws on two datasets. TrueUp tracks open positions worldwide across more than 9,000 tech companies. Live Data Technologies monitors job transitions, hires, and promotions for 124 million professionals, with hiring metrics focusing on the US. Comparing job postings against actual hires forms the methodological core of the report.

For product managers, the volume of open postings is at its highest level in nearly four years. Compared to the low point in March 2023, postings are up 75 percent. Since early 2023, PM job listings grew by 56 percent—faster than any other core role; engineering grew by 43 percent. By contrast, actual hires across all five core roles are flat or falling. PM is holding up best and is the only role whose share of total hires has increased, rising from 16 percent in 2021 to 21 percent today. The author attributes this divergence to growing demand combined with increasing hiring difficulty.

In engineering, job listings have ground to a halt for the first time in the report’s history. Postings grew by 29 percent in 2024, 15 percent in 2025, and just 2.5 percent over the past year. There are currently fewer open engineering roles than in March. US companies hired around 113,000 software engineers between May 2025 and April 2026. That is less than half of the 2022 peak of 239,000, and the number has declined every year since. Since January 2024, postings rose 44 percent, while actual hires dropped 17 percent.

The author flags two trends as positive for engineers. Postings for tech recruiters sit at record highs, up nearly 300 percent since mid-2023. He views this as a potential leading indicator of re-accelerating engineering hires. → Lenny’s Newsletter

Synthszr Take: Following the dot-com and financial crises, recovery followed a simple chain: first job postings, then recruiters, then hires. In engineering, that chain is broken. Since January 2024, listings rose 44 percent, but actual hires fell 17 percent. Companies post openings but leave them unfilled because agents fill the gap. The nearly 300 percent jump in recruiter postings hardly works as an early indicator; it reflects instead the hunt for the few individuals capable of formulating intent.

Google Follows Jev and OpenAI: Gemini Decision 1 Evaluates 1,000 Questions Per API Call

Google DeepMind released the Gemini Decisions API on Wednesday. Powering it is the new Gemini Decision 1 model, which answers up to 1,000 questions on the same input material in a single pass. Pricing is set at 10 cents per million input tokens, with Google charging nothing for output tokens. This matches the pricing OpenAI introduced for its own recently unveiled decision model. According to Google, the model is nearly five times cheaper than its entry-level Gemini 3.5 Flash-Lite.

The model supports three types of questions. A boolean check outputs the probability that a condition is met. A multiple-choice question selects among 2 to 255 named options and returns a probability for each. A rubric question produces a weighted score along a developer-defined scale. Instead of generating text token by token, the model scores predetermined options and returns probabilities alongside confidence metrics. Applications can use this to escalate ambiguous cases to a human or route them to a secondary model.

As input, the model accepts text, images, audio, video, and PDFs bundled into a single request. The context window of 262,144 tokens covers the shared material along with instructions and response choices across all questions. Each question is evaluated independently. If decisions depend on previous outcomes, the application must handle that logic in code or via a follow-up call. For explanatory reasoning or extracting arbitrary values like an invoice number, a separate generation step remains necessary. Google points to an insurance claims workflow as an example, where a claim form, damage photos, and call notes are submitted together in one request to determine routing queue, missing documentation, and urgency level.

Billing counts the shared material once per call while charging separately for the tokens in each individual question. For a dataset of 4,000 tokens with 20 checks of 50 tokens each, a bundled request costs about 50 cents per 1,000 records. Twenty individual calls resending the dataset each time would total $8.10 per 1,000 records. Google reports median compute latency between 50 and 100 milliseconds, with end-to-end latency ranging from 150 to 250 milliseconds. For contexts up to 16,000 tokens, latency is said to remain virtually constant across 1 to 100 questions per call. Jev clocks in at 70 to 500 milliseconds, while OpenAI’s alternative ranges from 100 milliseconds to three seconds.

Startup TypeSafe AI opened the market for these decision models in mid-September with Jev. Jev attracted one million users within days, according to the company; three weeks after launching, TypeSafe raised $870 million at a $7.5 billion valuation. → VentureBeat, SiliconANGLE, The Deep View

Synthszr Take: Rules engines have spent years selling deterministic, explainable decisions, charging primarily for maintaining if-then logic project by project. Gemini Decision 1 checks 20 conditions for around 50 cents per 1,000 records. In doing so, it reads damage photos and call notes that no static rule ever captured. Confidence scores handle escalation straight to a claims adjuster. Legacy vendors are left fighting only for cases where regulators strictly mandate determinism.

Mentioned in this article

Can you build magical companies with magical technology?

3RD EDITION. 452 PAGES, 20 EUR (German and English).

Can you build magical companies with magical technology?

The core thesis of CODE CRASH is: when code becomes cheap, intent becomes scarce — the clarity about what should be built. Now also becoming scarce is the ability of organisations to absorb what the agents deliver. The third edition of CODE CRASH describes the layers of agentic AI — intent, context, tools, checks and human judgement — in the interplay of teams and organisations.

codecrash.ai →

Subscribe free. Unsubscribe the second it sucks.

High-signal news across AI, business, UX, and tech. Every morning.