Chips & semiconductors · Evidence update
From the Aug 7, 2026 daily brief
Our July 25 issue covered the semiconductor research firm SemiAnalysis's change of heart on AMD, one core claim of which was: AI agents that automatically write and tune GPU kernels (the low-level compute routines a GPU runs) are eroding the CUDA moat NVIDIA built over a decade-plus of software accumulation. At the time, the entire item stood on that one firm's word. Today we completed targeted verification: AMD's official engineering blog (July 20) confirms the industrialization half — GEAK really is an agent-driven optimization framework spanning three kernel languages at repository scale, and the anti-gaming design is officially visible: candidate patches must clear correctness and benchmark gates before their gains count — guarding against an agent cheating for a higher score, for example by quietly modifying the benchmark itself (AMD ROCm official, Jul 20; SemiAnalysis's original, Jul 25). Verification: What upgraded is only the "AMD has industrialized this" leg; SemiAnalysis's own account of a 2.5-person team doing chip enablement with agents, and the +21.8% end-to-end gain figure, remain single-source — that number appears nowhere in AMD's official post. Judgment update: The judgment moves from "only an analyst firm claims it" to "the industrialization half has an official primary source"; credibility adjusts up, but the claim as a whole is not confirmed — the boundary sentence from two weeks ago holds today exactly as written.
Just an email address, unsubscribe anytime. This is the only thing we ask of you.
The most structurally valuable post in this Teortaxes batch. On the software side, Chinese…
The same commentator argues the real story in Chinese advanced process is diffusion — not …
Starting from a proof of concept he himself calls toy-grade, Teortaxes offers a structural…
In late April, Gooaye aired a piece of information that cut against the host's own holding…