Sunday, October 11, 2026

TL;DR: ERCOT's interconnection pause is the grid telling hyperscalers to bring their own power and flexibility.

McEntush unpacks what a grid connection physically commits a utility to, why that makes Texas's queue pause rational, and how behind-the-meter generation plus curtailable load becomes the price of admission. The 'getting to hundreds of gigawatts' section is a useful frame for judging every power-adjacent AI announcement.

TL;DR: The AI boom was built by a small sect of believers who were right — and now own the upside.

Roose walks through his history of OpenAI, Anthropic and DeepMind as a story about faith rather than technology: sci-fi as scripture, scaling laws as creed, and what happens to true believers once the prophecy pays out in equity. The portraits of Altman, Amodei and Hassabis as rival theologians are the best way in to understanding why these labs behave as they do.

TL;DR: Two ex-OpenAI/DeepMind researchers left to build an AI lab that runs physical experiments, not just benchmarks.

Fedus and Cubuk explain why they walked away from frontier-model work to pursue materials synthesis — semiconductors to superconductors — where the bottleneck is wet-lab iteration rather than tokens. The forward-deployed-engineering angle shows how an AI-for-science startup actually has to operate when the ground truth is a furnace, not an eval.

TL;DR: A yield-engineering veteran builds a scenario model for how a wafer-scale chip can be manufactured and tested at volume.

A test-and-yield engineer with a decade in fault diagnosis asks the question Cerebras bulls skip: what does 'yield' even mean for a chip the size of a wafer, and how do you screen it? The piece turns redundancy and repair into actual numbers, which matters as Cerebras moves from boutique volumes toward public-company production commitments.

TL;DR: A ban on Chinese transceivers doesn't swap supply chains — the lasers and chips underneath are shared.

Markets traded the draft FCC rule as a clean win for US optics names, but Ennis follows the module down to the laser and the InP chip to show how entangled the dependencies remain. The piece also tests how far the cybersecurity rationale actually reaches, a template for reading the next several hardware restrictions.

TL;DR: At The Curve, the frontier's leading minds admitted the plan for transformative AI is still no plan.

Zvi's Chatham-House-filtered report from the conference captures the gap between how fast capabilities are moving and how little institutional response exists — 'the plan is no plan' is his verdict on both labs and governments. It's a mood piece with teeth, useful for anyone trying to gauge how insiders really feel versus what they say on stage.