Week in review: Nov 8 – Nov 15
The AI-money beat came for NVIDIA and the exchanges — a tech-stock crash, a discounted Anthropic IPO and Gamers Nexus's Jensen Huang files — while the local-model lane kept its own scorecard with Qwen-versus-Kimi, a 120B MoE on one 24GB card, and Dave Farley on why inference is memory-bound.
A heavier week than usual: twenty-one items across five rollup days (Nov 8–13), and they sort cleanly into two lanes — the money side of AI falling over in public, and local model work quietly carrying on.
The money lane had a lot to work with. Pivot to AI ran the week’s longest streak: tech stocks and data-centre loans crashing together, AI Overviews eating Google’s ad revenue for a 12% stock drop, and Sam Altman wanting your RAM for the compute campuses. Better Offline put the political frame on it — Ed Zitron reading the election’s Blue Wave as an ill omen for the AI trade, and Paris Marx’s Hyperscale book tour documenting the community groundswell against data centers. Gamers Nexus went after the figurehead directly, first with the RTX 5090 as evidence that AI and workstation demand is squeezing gamers out of the high end entirely, then with a full investigation into Jensen Huang’s abrupt downfall — greed, corruption and export-control evasion. The capstone was Anthropic’s IPO, which priced well below its requested near-$2 trillion valuation after a messy S-1; the market is starting to mark its own homework.
The local lane stayed practical. No Place Like Localhost staged a proxy war for the ecosystem — Qwen 3.8 Flash Next’s 125 billion local parameters against Kimi K3.1 Pro’s 2.8 trillion in the cloud — and got Nemotron 1.5 Super VL 120B running at usable speeds on a single 24GB card. Dave Farley explained why all of this lives and dies on memory bandwidth rather than raw arithmetic, and Pivot to AI supplied the week’s best myth-busting: a head-to-head test of the claim that AI chatbots are great at updating legacy code. They aren’t.
Notable items
- JensenGate and Anthropic’s IPO w/ Ed Elson | Better Offline — Anthropic’s IPO priced around $1–1.4 trillion after a confusing filing, a down payment on the market repricing AI labs.
- 20261106 - Tech stocks crash, and so do data center loans — the sell-off reached the debt funding the buildout, not just the share prices.
- The Downfall of Jensen Huang: NVIDIA’s Greed & Corruption — Gamers Nexus frames the fall as greed, export-control evasion and geopolitical leverage, not personal scandal.
- The NVIDIA RTX 5090 Problem & Future Generation Price Creep — the gaming flagship is being squeezed out by workstation and AI demand plus rising component costs.
- Monologue: The Future Is Ours — Ed Zitron reads the Blue Wave as an ill omen for the AI trade, with voters rejecting data centers state by state.
- The Truth About Data Centers w/ Paris Marx — the Hyperscale author on the AI buildout and the community backlash politicians and the media keep missing.
- Qwen 3.8 Flash Next vs Kimi K3.1 Pro — a local-versus-cloud benchmark: 125B on your hardware against 2.8T in a data center.
- Nemotron 1.5 Super VL 120B A12B — a 120B MoE with 12B active, hitting 50–70 tokens/s on a single 24GB card.
- The Hidden Bottlenecks Slowing Down Large Language Models — Dave Farley on mechanical sympathy: inference is memory-bandwidth bound, it’s moving data that costs, not math.
- ‘But the AI is great at updating code!’ — no it isn’t — a chatbot test of legacy-code upgrades punctures the refactoring claim.
From the blog
- Week in review: Nov 1 – Nov 8 — last week’s digest, where DRAM rationing and the first tech-stock sell-off set up this week’s reckoning.