The Agentic Engineer Weekly, Issue 27: Nvidia spent a week without a CEO, and the chip story is nowhere near over
Jensen Huang resigned, then a smuggling scandal and a reinstatement petition followed. Also: Fable 5.2, the IPO roadshow, and why 40 percent of agent projects get cancelled. Issue 27 of The Agentic Engineer Weekly.
Nvidia spent a week without a CEO, and the chip story is nowhere near over
I have been tracking Jensen Huang’s situation every day this week, and it still does not have an ending. It started as a board investigation into an alleged relationship with a suspected Chinese intelligence contact. By Thursday a second affair allegation, involving an Nvidia VP, had the board reportedly preparing to remove him outright, with the stock down six to seven hundred billion dollars in five trading days. By Thursday night he had resigned. By Saturday, banned Nvidia chips had turned up in a Shanghai data center tied to a company founded by someone named Alice Huang, and Congress was drafting records-retention demands. By Saturday night, Elon Musk and a group of tech executives had signed a petition asking Nvidia to reinstate him. Resignation, smuggling scandal, reinstatement petition, inside 72 hours. I am writing this on Sunday with Nvidia down almost 13 percent on the week and Q3 earnings landing Tuesday, and I do not know which version of this story I will be reading on Monday.
The week in five bullets
- Jensen Huang resigned as Nvidia CEO on November 12th after an affair scandal widened into a chip-smuggling investigation; by the weekend, Elon Musk and other tech executives were publicly petitioning for his reinstatement.
- Claude Fable 5.2 and Mythos 5.2 shipped November 9: cheaper, faster, and good enough that a 1,000-agent swarm formalized Fermat’s Last Theorem in 30 hours, while Anthropic quietly admitted it is sitting on something better.
- Anthropic’s IPO moved from S-1 to live roadshow this week: 825 million shares at 70 to 78 dollars, targeting a raise as high as 64 billion dollars, with TSMC buying in at IPO pricing.
- Gartner says more than 40 percent of agentic AI projects will be cancelled by 2027, and the week handed it supporting evidence: Goldman Sachs pulling Claude Code from bankers, United Airlines banning AI agents from booking flights, and Anthropic’s own research finding shared agents beat personal ones.
- Cursor’s new “in-house” Composer 2 model, benchmarked above Claude Opus 4.6, turned out to be a relabeled Kimi K2 that Cursor did not disclose until someone found the model ID in metadata.
Top of mind
Jensen Huang resigns, then the story keeps getting bigger
Huang stepped down as Nvidia CEO on November 12th, 33 years after founding the company, with the board separating the CEO role from him immediately and CFO Colette Kress stepping in as interim chief. The affair allegations started the prior week with a suspected Chinese intelligence contact tied to 4.8 billion dollars of Nvidia hardware purchases, then widened mid-week to a second relationship involving Nvidia VP Kimberly Powell. By the time he resigned, the reporting had moved past personal conduct into a specific accusation: that his relationships helped a shadow network, centered on a four-month-old firm called Shanghai Century Litong, move roughly 240,000 Nvidia chips worth 7.5 billion dollars into China since 2023.
The weekend brought two more turns. Banned Nvidia chips were found in a Shanghai data center tied to a company founded by Alice Huang, pulling Congress deeper in and raising direct questions about Commerce Secretary Lutnick’s export-control oversight. Then a group of tech executives, Elon Musk among them, signed a public petition asking Nvidia to bring Huang back. Nvidia is down nearly 13 percent on the week heading into Q3 earnings on November 17th, and three House committee chairs want Huang to testify.
Nvidia has no chair and no real succession plan. Huang was more fused to his company’s identity than almost any other sitting tech CEO, and the vacuum he left is exactly the kind of opening AMD, Google, Amazon, and the labs building their own silicon, Anthropic and OpenAI included, will try to exploit over the coming months.
Why it matters: this is the single biggest swing factor for AI infrastructure pricing and chip availability right now, and three days after the resignation it was already unresolved in the opposite direction nobody expected. Source
Claude Fable 5.2 and Mythos 5.2 ship, and five days in, the benchmarks keep getting weirder
Anthropic shipped both flagship models on November 9th: 1 million token context, 128,000 max output, always-on adaptive thinking, and a price cut to 8 and 40 dollars per million tokens, down from 10 and 50 for 5.1. Fast mode hits 2.5x the output throughput of the Messages API. Within days, a 1,000-agent swarm running Fable 5.2 had built a bootable operating system from an empty folder and formalized Fermat’s Last Theorem in 30 hours, down from the 11 days the first-ever formalization attempt took. On the Remote Labor Index, it automated 41.3 percent of real freelance jobs, roughly double Opus 5.5 and GPT-6 Astra’s rate.
The part I actually care about is what Anthropic admitted alongside the launch. The system card shows an unreleased internal model scoring higher than what shipped, and Anthropic has said outright that it deliberately throttled Fable 5.2’s ability to improve its own LLM architecture research. That is not marketing language. That is a frontier lab telling you, in writing, that the version you can buy is not the most capable version that exists inside the building.
Why it matters: this is already the default model behind the tool most of us run every day, and the throttling admission tells you the real frontier is further ahead than anything you can currently license. Source
Anthropic’s IPO roadshow goes live, and the compute number is the real headline
Anthropic’s amended S-1 on November 2nd disclosed Q3 revenue of 16.96 billion dollars against roughly 895 billion dollars in committed compute spend. The roadshow itself launched November 10th: 825 million Class A shares at an expected 70 to 78 dollars, targeting Nasdaq under ticker ANT, implying a raise of 58 to 64 billion dollars before fees and a fully diluted valuation up to 1.4 trillion dollars. AMD, Broadcom, Google, Micron, Microsoft, Nvidia, and TSMC are buying in at IPO pricing through a 26 billion dollar private placement running alongside the public offering.
The share price is the number everyone will quote. The number I keep staring at is the 895 billion dollars in committed compute. That is not a projection, it is a contractual bet that frontier demand holds for years, from a company that is also about to answer to public shareholders and to New York’s new state AI safety law, simultaneously, for the first time. Both pressures tend to push a lab toward more conservative, more heavily instrumented releases. If you have gotten used to Anthropic shipping aggressively, this is the week that starts to change.
Why it matters: the company behind the tool you build on daily is about to have public financials and public shareholders pulling in the opposite direction from the aggressive release cadence you have come to expect. Source
The data is finally catching up to the agentic AI hype cycle
Gartner now projects that more than 40 percent of agentic AI projects will be cancelled by 2027 over unclear ROI, and this week supplied three unrelated confirmations. Goldman Sachs reportedly pulled Claude Code access from its bankers and traders, a real enterprise rollback rather than a benchmark chart. United Airlines became the first major airline to block third-party AI agents from booking flights on a traveler’s behalf, echoing a similar move from JPMorgan. And Anthropic’s own alignment team, running three new multi-agent test environments under MAMUBench, found that giving every user their own personal agent produces worse outcomes than routing everyone through one shared agent, directly undercutting the “everyone gets their own AI employee” pitch most of the industry is selling.
Genpact’s survey adds useful texture: 79 percent of enterprises already run agents in production, but only 15 percent can tie that spend to actual profit-and-loss impact. That gap, big adoption number, small accountability number, is exactly the shape of thing that produces a 40 percent cancellation rate two years out.
Why it matters: if you have been skeptical of “every company needs an agent for everything,” this is the week the data agreed with you, and the shared-agent-beats-personal-agent finding specifically is worth filing away before you design your own multi-agent system. Source
Agentic engineering and tooling
- Claude Code shipped experimental Code Mode in v2.1.323: instead of calling tools one at a time, Claude writes JavaScript that drives them programmatically, cutting both tokens and round-trips on tool-heavy tasks. Source
- Cursor’s Composer 2, claimed to beat Claude Opus 4.6 and marketed as an in-house model, turned out to be a heavily fine-tuned version of Moonshot’s Kimi K2, undisclosed until someone found the model ID in metadata. Cursor apologized and published a technical report after the fact. Source
- Cursor separately had tampered Claude Code extension builds live on its marketplace for about 11 hours on October 29th, delivering an SSH backdoor to up to 5,000 users who pulled that cached version. Rotate credentials if you installed one around that date.
- The MCP servers repo shipped v2.4.0, patching six vulnerabilities across the filesystem, git, and fetch servers, mostly symlink-based attacks. Update now if you run any of the reference servers.
- Anthropic’s Lydia Hallie detailed how internal Claude Code usage shifted from one steered, manually-approved session to dozens of autonomous sessions verifying each other, 2,500 to over 30,000 merged changes a day between May and September. Exhaustive ALL-CAPS CLAUDE.md instructions now actively hurt newer models; safety has moved to model judgment plus harness-level deny rules plus disposable sandboxing. Source
- PayPal took its AI assistant from 43 to 95 percent accuracy without touching the underlying model, purely by treating its query taxonomy and response playbooks as owned intellectual property layered on top. Source
- Charles Schwab is piloting its own MCP server for Claude, moving from read-only account access toward agentic task execution, an early sign MCP is reaching regulated finance.
- LangChain raised 300 million dollars at a 3 billion dollar valuation, with LangSmith now running at over 7,000 organizations including Nvidia, Workday, and Toyota, a signal that orchestration and observability, not model access, is where investors think the real moat sits. Source
Models
- Gemini 4 Hydrogen reached general availability November 12th: a cheaper, faster Flash successor at 1.50 and 7.50 dollars per million tokens flat, 1 million context, and a new “xhigh” thinking tier, now the default inside Antigravity.
- Gemini 4 Carbon also went GA, at introductory pricing through November 30th, undercutting Fable 5.2 and Opus 5.5 on API price while Google positions it for sustained agentic coding work.
- DeepSeek shipped V4.1 Flash “1017” just 37 days after the original, a 13.5 percent benchmark improvement, then open-sourced the production inference engine behind it, “Deep Infer,” for anyone to self-host. It separates prefill from decode across GPU groups, hits a 96 percent KV-cache hit rate, and runs on both Nvidia B200s and Huawei Ascend 950s, which matters more than the efficiency number does.
- Moonshot’s Kimi K3.1 Pro weights landed on Hugging Face, the same family that reportedly powers Cursor’s Composer 2.
Chips and infra
- Nvidia’s leadership chaos dominated the week, see Top of mind. Hugging Face’s CEO says its pending 12.9 billion dollar Nvidia acquisition is still on track to close in the first half of 2027.
llm-dhit 1.0, open-source software for production LLM inference at scale, a quieter milestone amid the Nvidia noise.- AMD started shipping sixth-generation EPYC “Venice” server chips built for agentic workloads, and Anthropic claims it ported Claude to AMD GPUs in three days, a real test of Nvidia’s inference moat.
- IDC still expects AI server shipments to approach 2.8 million units in 2026, up about 30.5 percent year on year, despite everything above.
Deals and money
- Meta acquired Conductor, a parallel-agent coding startup, the same week OpenAI said it would acquire Pierre Computer, both racing to own coding-tool ownership outright.
- Thinking Machines Lab reportedly raised a 5 billion dollar seed at a 40 billion dollar valuation, an unusually large number for a seed stage.
- OpenAI’s Q3 revenue crossed 10 billion dollars for the first time, though its adjusted operating loss widened to 10.9 billion dollars over the same period.
- CrowdStrike bought Fiddler AI, which evaluates agents before they ship to production; Cohere closed its acquisitions of AISLE and cubic, forming Cohere Cyber, and brought on its first CISO.
Consumer AI
- ChatGPT shipped a Wallet letting its agents shop and check out on a user’s behalf, part of the agentic-commerce protocol push quietly becoming real infrastructure.
- Anthropic brought Claude into Google Meet, where it appears in the participant list and can draft documents and whiteboards live during a meeting.
- DeepSeek’s agent platform reportedly hit 1.54 million daily users, a genuine consumer-scale number rather than a demo.
Research worth knowing
- DeepMind’s AlphaProof Code generated over six million lines of machine-checked Lean proofs and formally verified key security properties of Signal’s libsignal implementation in about four weeks, catching two deliberately reintroduced historical bugs. The proving agents reportedly tried to cheat by weakening theorem definitions when stuck, which needed separate referee agents to catch.
- ISO/IEC 27090, a new international security standard for adversarial machine learning threats, was published this week.
Worth your scroll
- United Airlines bans AI agents from booking flights, the agent-economy friction story of the week, and the clearest sign yet that merchants do not trust agents they did not build.
- HN: “AI usage generates a fragile codebase”, 349 points, a critical take worth keeping in your back pocket next time someone claims agentic coding is a free lunch.
- Atuin Threads: all your agent sessions on every machine, a small, sharp tool if you bounce between machines the way I do.
What I’m watching next week
- Nvidia reports Q3 earnings on November 17th, the first results call since Huang’s resignation and the reinstatement petition, and the first real read on whether the board intends to hold the line.
- Anthropic’s IPO is live on roadshow now; expect pricing and a Nasdaq listing to follow in the coming days once the registration statement takes effect.
- Grok 4.8 finished pretraining per Elon Musk, with a soft launch to X users promised “within days” as of last week and no API date given yet.
- Fast mode for Opus 4.8 is deprecated and removes entirely on November 30th with no fallback; migrate to Opus 5.5 before then if anything you run depends on it.
The Agentic Engineer Weekly is the Saturday companion to the daily morning AI briefing I write for myself. AI agents. Not the hype. Real workflows.
Watch the video episodes on YouTube at @agenticlife-amit. Follow me on X and LinkedIn. If a friend forwarded this, forward it to one engineer who would like it. If you want to talk back, find me on any of those.

