Nvidia (NVDA) intends to begin small-batch shipments of a China-tailored LPU by the end of 2026, reports The Information citing sources

• The LPU, which utilises technology licensed from Groq, is designed to accelerate AI chatbot response times and complies with current US export control regulations.

China-tailored silicon has become an established playbook since the first wave of US export controls: vendors ship deliberately derated variants of flagship parts engineered to sit just inside the performance thresholds, Washington tightens the thresholds, and the cycle repeats. The precedent pattern is that such parts carry lower ASPs and thinner margins than the unrestricted equivalents, but preserve the revenue line and the software ecosystem lock-in in a market that has historically been a meaningful share of data centre sales. The novel element here is the licensing of external architecture, which signals the compliance-grade segment is now worth dedicated engineering rather than simple binning of existing product. The long lead time to small-batch shipments is itself a tell: volumes of that scale test regulatory reaction rather than serve material demand, and past episodes show the controlling variable is not the part but whether regulators move the goalposts before ramp. Worth watching are any Commerce Department signals on the applicable thresholds, the margin disclosure around the part once it ships, and whether domestic Chinese accelerators close the gap in the interim, which has been the persistent bear case on the tailwind from compliant parts.