Export Controls Reach Your Base Image, Not Just Your GPU

A walkthrough of a November 2026 round of reporting — Nikkei Asia, Businesskorea, BusinessToday Malaysia and BigGo Finance, on 19 and 20 November — on new US chip export rules taking effect and TSMC tightening restrictions on Chinese AI customers, including a reported halt on orders at 16nm and below. The post is explicit about what that reporting does not establish: no rule citation, no ECCN, no named customer, no named SKU. It then does the part a build engineer can own, mapping the six layers between a process node and a pushed container — the fab node, the SKU and its export classification, the cloud instance and its physical region, the driver and CUDA base image lineage, the CI runner and its build cache, and the artifact with its SBOM and provenance. From there: why a node threshold catches far more than accelerators (networking silicon, management controllers, edge SoCs, older inference parts), why a CUDA base image is a redistributable rather than a thin Ubuntu wrapper, the deemed-export question for container layers, and how to inventory the chain with digests, manifest inspection and attestations so that a change in the rules is a policy-file edit rather than a rewrite. The framing is defensive and lab-scoped throughout: record vendor claims, verify what you ship, and put a fail-closed gate in front of the chain. No exploit, no real target, no credentials. Read the full post: https://hackyjs.com/posts/from-gpu-sku-to-base-image-mapping-where-export-controls-reach-your-ai-build-pipeline

export-controlsgpuai-infrastructuresupply-chaincompliance

Read the post →