Samsung, TSMC Co-Develop Nvidia's Custom HBM Stack

Samsung Electronics and TSMC are co-developing a custom High Bandwidth Memory stack for Nvidia, per a SammyGuru report published this week. The partnership pairs the chip industry's two largest manufacturers around its most demanding customer.

That is the headline. That is also the body. SammyGuru carried no executive comment, no process node, no volume figure, no tape-out date. Everything past the headline is context — and the context is loud enough to make the rumor worth dissecting.

What does "custom HBM" actually mean?

HBM is the vertically stacked DRAM that rides next to Nvidia's accelerators — the H100, the H200, the B100, the B200. It is the AI buildout's tightest supply link. It is also the memory industry's richest margin pool.

Each HBM stack bonds multiple DRAM dies through-silicon-via to a base logic die. That base die handles the PHY, the test functions, and increasingly the buffering that lets the GPU controller talk to the stack at multi-terabit-per-second rates. It is the part of the package where memory meets logic, and where the foundry's process node matters most.

"Custom" almost certainly refers to that base die. JEDEC-standard HBM stacks use an off-the-shelf base die from the memory vendor. Nvidia, by public reporting, wants its own — silicon tuned to its memory controller, fabricated on a leading-edge logic process, packaged by the foundry ecosystem the GPU giant already runs on.

That is TSMC's specialty. Samsung brings DRAM stack expertise and a memory business eager to claw back share. Nvidia writes the spec.

Why does Nvidia want a second source?

SK hynix has been Nvidia's primary HBM supplier across the H100 and H200 generations. Samsung won qualification on certain H100 SKUs after a multi-quarter push but remains the underdog in Nvidia sockets.

The commercial case is plain. Nvidia wants a credible second source. SK hynix wants to keep its incumbent share. Samsung wants the wedge. The AI customer has watched one Korean supplier dominate the slot for two product cycles and is hedging with the other.

The technical case is sharper. Nvidia has reportedly pushed HBM4 timelines, signaled openness to non-JEDEC form factors, and explored Samsung's 4nm base-die process. A Samsung-TSMC partnership would put Samsung's DRAM on a logic die fabricated by Nvidia's incumbent foundry partner. It removes the base-die question that has dogged Samsung's HBM pitch since 2023.

What isn't in the report?

The SammyGuru filing leaves more questions than it answers. No announced geometry. No volume commitment. No tape-out date. No executive on-record. No mention of which generation — HBM3E, HBM4, or beyond — sits at the heart of the program.

Whether the engagement is a sampling deal, a production order, or a roadmap handshake remains unspecified. Read it as signal. Not as shipment.

When will the picture sharpen?

Look for detail at Samsung's late-January earnings call and at Nvidia's spring GTC — the traditional venues where memory roadmaps surface. If a custom HBM4 base-die tape-out lands publicly, the Samsung-TSMC partnership graduates from rumor to line item.

That is the moment the chip industry's most-watched whisper becomes its most consequential supply contract of the cycle.