Krueger's New Paper: 'Hardwire' AI Models Into Chips Instead of Dismantling the Compute Supply Chain
DavidSKrueger · x · 2026-10-10
Instead of dismantling the AI compute supply chain, Krueger's new paper proposes repurposing it: chips can be 'hardwired' so they can only run specific AI models and nothing else. This captures most benefits of dismantling while being far more palatable — existing models can keep being deployed at scale, so the immediate economic hit of a hardwired pause would be much more limited.
More from Infra
- SGLang-Diffusion serving framework for diffusion models to be unveiled at PyTorch Conference 2026 — PyTorch · 2026-10-10
- Andrew Ng: one of my agents makes 5,000–10,000 web searches a day, data centers will need to scale far beyond current plans — DeepLearningAI · 2026-10-10
- SGLang Summit 2026 lineup: Intel CEO, Lilian Weng, Perplexity CEO, Nov 12-13 in SF — ying11231 · 2026-10-10
- AI inference could hit $350B next year, pushing software margins from 72% down to 30-50% — Sethwinterroth · 2026-10-10
- Nebius vs CoreWeave: net debt 0.2x vs 2.9x revenue explains the stock divergence — Beth_Kindig · 2026-10-10
- Devon Eriksen explains why raw elements are easy but chipmaking knowledge is the real barrier — Yamapama · 2026-10-10