Nvidia's post-Huang turmoil tests trust across the AI stack

Ramit KoulFounder, Software Engineer & Innovator · Published 5:04 AM ET

Nvidia's stock has lost nearly a quarter of its value since late October reports linked Jensen Huang to a suspected Chinese spy, and a petition signed by Elon Musk now asks the board to reinstate him.^{1, 2} The fallout reaches chip supply and pricing, from a report that Nvidia used customer financing to keep GPU prices high to new research suggesting Chinese open-weight models were trained on Blackwell chips.^{3, 4} Developers face a closer risk: Cursor says a malicious Claude Code extension in its marketplace started with an access key exposed at Fireworks AI.^{5, 6} Here is the story that matters most, then 12 briefs. Numbers in the text link to the references at the end.

Industry

AMD's MEXT and Google's new TPUs take aim at the AI memory squeeze

AMD has given more detail on MEXT, a predictive memory engine that moves rarely used memory pages to flash storage and uses a transformer model to bring them back before they are needed.⁷ AMD says servers can run on half the DRAM and keep 95% to 99% of their performance, depending on the workload.⁷ It already runs MEXT on 145 of its own servers and plans to reach 2,000 next year.⁷

Google added two chips to its eighth-generation TPU lineup.^{7, 8} TPU 8i Max puts two TPU 8i chips on one package for long context windows, with twice the compute, SRAM and HBM bandwidth of TPU 8i.⁹ TPU 8tx handles confidential inference and uses AMD EPYC hosts instead of Arm hosts.^{7, 10}

On safety, the Frontier Model Forum published version 1.0 of a Jailbreak Rating Framework for Frontier Cyber Threats.¹¹ It scores a successful jailbreak on four components that cover capability gain, scope and how easily the jailbreak can be deployed.⁷

All of this played out in the week Jensen Huang stepped down as Nvidia's chief executive.⁷ Tom's Hardware also reports that Nvidia is tightening rules on overseas cloud GPU rentals to Chinese customers, with partners facing a 51% non-Chinese quota.⁷

Why it matters for builders

DRAM shortages and accelerator design now shape what it costs to serve long-context agent workloads, so flash-backed memory and denser TPU packages could change how teams size inference servers.⁷ A shared jailbreak severity scale also gives teams a common way to rank safeguard bypass reports.¹¹

In brief

Policy & legal

Voters rejected data centers in 40 of 41 midterm ballot measures, Axios reports. Axios says the result is part of a wider loss of trust in AI, and that Democrats plan a wave of AI bills and testimony from AI executives once they take control of Congress in January.¹

Industry

More than 115,000 people, including Elon Musk, have signed a petition asking Nvidia to reinstate Jensen Huang. The petition site says signers include 5,965 current Nvidia employees and 63,063 shareholders, and it asks the board to bring Huang back while an independent investigation runs.² Colette Kress is serving as interim CEO while Nvidia searches for a permanent replacement.²

Industry

Cursor says the malicious Claude Code extension in its marketplace started with an access key exposed at Fireworks AI. Attackers copied old coding-agent traces that held unredacted credentials and used them to get into Cursor's internal tools.^{5, 6} On October 29 they posted tampered versions 2.1.312 and 2.1.313 for about 11 hours, reaching up to 5,000 users. Cursor has since moved all inference off Fireworks and says Privacy Mode users, customer repositories and production systems were not affected.^{5, 12}

Products

CopilotKit released OpenBot 1.0, an MIT-licensed agent workspace that teams can host themselves. Every bot action passes through a gateway that checks a policy and writes an audit record before it runs, and the workspace works with any AG-UI agent and with OpenAI, Claude, Gemini, Grok or open models.^{13, 14} Personal use is free, and OpenBot for Business costs up to $29 per active user per month.¹³

Research

Stanford's Memorilla gives a frozen LLM an entire document collection as just 16 memory tokens. A 194M-parameter module reads the collection for each question, and with a frozen Qwen3-8B it scored 84.33% on FactKG against 80.43% with every document in context, though it trailed on PubMedQA.^{15, 16} The code and checkpoints are public under Apache 2.0.^{15, 17}

Dev tools

Vercel Labs released UI SDK, a beta library of interface behaviors that coding agents put together into components. It ships more than 50 behaviors, such as focus handling, typeahead and popups, in one core with adapters for React, Vue, Svelte, Solid and Web Components, and agents can read and act on UI state directly through a bridge.^{18, 19} In Vercel's own test on a single task, the bridge cut tool calls by 69% and time by 61% compared with browser tools.¹⁸

Products

Gigabyte announced the AI TOP AERO, a one-liter RTX Spark desktop with up to 128GB of memory, but has not given a price. It pairs Nvidia's N1X Superchip, with a 20-core Grace CPU and a Blackwell RTX GPU, with Windows 11 for local AI work.²⁰ TechRadar notes that rival 128GB Spark systems cost between $5,999 and $7,999.²⁰

Industry

China's leading AI labs have started limiting the cyber capabilities of their most advanced open-weight models. The South China Morning Post says the move follows a series of high-profile AI hacking incidents, but because anyone can download and modify open weights, it is unclear how long such limits will hold.²¹

Industry

Alphabet's third-quarter call showed Google Cloud revenue up 103% to $30.7 billion. Cloud backlog reached $561 billion, and Alphabet said its own model APIs processed nearly 32 billion tokens per minute, four times the level a year earlier.²²

Industry

A Wall Street Journal profile of interim CEO Colette Kress says Nvidia used customer financing deals to keep GPU prices high. According to TechSpot's account of the report, Nvidia pushed for more deals in mid to late 2026 as TSMC wafers piled up, then slowed or paused some, including an OpenAI deal that had included a $250 billion backstop.³

Industry

SpaceX has reportedly laid off about 60 Grok Imagine staff to shift resources toward coding tools and agents. The Information reported the cuts, which SpaceX has not confirmed, and the image and video tool is still running with only a few staff left to maintain it.²³ SpaceX is reportedly prioritizing coding tools and Grok Bot after closing its $60 billion Cursor acquisition in August.²³

Infrastructure

Researchers say weight statistics show DeepSeek V4, Kimi K3.1 and other Chinese models were trained on Nvidia Blackwell chips. UnemploymentLabs matched how the models' weights were randomly initialized to the core counts of specific Nvidia chip families, but it cautions that the method cannot show where the chips were or whether export controls were broken.⁴

AMD's MEXT and Google's new TPUs take aim at the AI memory squeeze

In brief

References