Golem reported on 19 October 2026, under the heading "NVHBM", that Samsung is becoming a TSMC customer in service of Nvidia's memory ambitions.
The irony is structural. Samsung and TSMC are rivals in the foundry business, and Nvidia — arguably the most important customer either of them has — sits astride both. Nvidia's accelerators are built from two pillars: high-bandwidth memory stacked from suppliers such as SK Hynix, Samsung and Micron, and TSMC's advanced packaging flow that binds the memory stack next to the GPU die. As HBM advances from one generation to the next, more of the assembly around the memory — including logic components at the bottom of the stack — is built with foundry logic processes rather than pure memory processes. That pulls a memory maker toward exactly the foundry it competes with.
If Samsung's HBM destined for Nvidia passes through TSMC at any point in that chain, revenue that might have stayed inside Samsung's own chip business flows to its largest manufacturing rival instead. Nvidia, for its part, gains a more resilient multi-supplier memory pipeline — the goal it has pursued publicly since the memory squeezes of the AI buildout — and Samsung keeps its seat at the table by accepting a dependency it would rather not have.
Beyond the headline, details such as which process node, which packaging step and what volumes are part of Golem's reporting and should be read in the original article; this write-up was produced during a web-tooling outage and could not independently open it.
