Samsung Electro‑Mechanics wins substrate slot
Samsung Electro‑Mechanics will supply FC‑BGA substrates for Groq’s 3 LPU chip and is set to start mass production in Q2 2026 — a direct tie into emerging AI-inference hardware stacks. That contract shows how substrate makers are being pulled into the AI supply chain as customers integrate chips into platforms like NVIDIA’s Vera Rubin, increasing demand for complex substrates and layering up stress across the broader supply chain. For distributors, substrate demand of this kind often precedes tighter lead times and price pressure in adjacent board- and assembly-level categories. (x.com)
The part of an artificial intelligence chip you never see is often the part that breaks the schedule. Samsung Electro-Mechanics has reportedly won the package-substrate slot for Groq’s “3 Language Processing Unit” chip and plans mass production in the second quarter of 2026. (biz.chosun.com) A package substrate is the tiny multilayer board that sits between the silicon chip and the larger circuit board, like a high-rise interchange between a city street and a freeway. The version in this deal is flip-chip ball grid array, a dense design used in artificial-intelligence servers and other high-performance computing systems. (biz.chosun.com) That matters because modern artificial-intelligence chips do not just need more transistors; they need more wiring, more layers, and tighter tolerances under the chip package itself. Samsung Electro-Mechanics said in August 2024 that server flip-chip ball grid array is the most technically difficult category of semiconductor substrate and that only a few global companies can mass-produce high-end server substrates. (samsungsem.com) Groq is one of the companies building chips mainly for inference, which is the step where a trained model answers a prompt instead of learning from new data. Reports on April 8 and April 9 said Samsung Electro-Mechanics would supply the substrate for Groq’s next chip as that chip is prepared for use in NVIDIA’s Vera Rubin generation. (alphabiz.co.kr) (news.nate.com) NVIDIA’s own Vera Rubin NVL72 description shows why suppliers are chasing every layer of this stack. One rack combines 72 Rubin graphics processors, 36 Vera central processors, networking chips, data-processing units, and a sixth-generation NVLink switch fabric inside a single rack-scale system. (nvidia.com) When systems get that dense, the boring-sounding substrate stops being boring. More chips in a rack means more advanced packages, and more advanced packages mean more demand for high-layer-count substrates that can carry power and signals without warping or failing under heat. (nvidia.com) (samsungsem.com) Samsung Electro-Mechanics has been steering hard into that market for a while. The company said in August 2024 that it wanted high-value flip-chip ball grid array products for servers, artificial intelligence, automotive, and networking to exceed 50 percent of its mix by 2026, after already landing a high-performance computing server substrate deal with Advanced Micro Devices. (samsungsem.com) The industry is now behaving like a supply chain that sees the next squeeze coming. Korean business reports in late March and early April said Samsung Electro-Mechanics and LG Innotek were reviewing capacity expansion as artificial-intelligence demand made substrate utilization run hot and pushed more investment toward server-grade package boards. (businesskorea.co.kr) (biz.chosun.com) That is why this contract reaches beyond one chip. When a substrate maker gets pulled into an artificial-intelligence platform build, the pressure usually shows up next in adjacent board, assembly, and component categories, because the same factories, materials, and engineering teams are suddenly serving bigger and more complex server programs. (businesskorea.co.kr) (samsungsem.com) So the headline is not just that Samsung Electro-Mechanics won one more customer. It is that the artificial-intelligence race is now tight enough that even the thin green layers under the chip have become a strategic battleground, and companies that can mass-produce those layers on time are moving closer to the center of the data-center economy. (biz.chosun.com) (nvidia.com)
