Samsung Reportedly Taps TSMC 3nm for Nvidia HBM Base Dies

Samsung has reportedly adopted TSMC's 3nm process to build the logic base dies for its Nvidia-bound HBM stacks, according to TradingKey. If the report holds, Samsung would route a critical layer of its high-bandwidth memory through the world's largest foundry rather than its own fabs. For memory suppliers, that changes the economics of who owns the most advanced logic inside an HBM package.

The base die isn't a small detail. In modern HBM, the logic die at the bottom of the stack carries the interface circuitry that ties the DRAM layers to the host GPU, and its process node increasingly determines speed and power draw. Building it on TSMC's 3nm class instead of an in-house node is a bet that foundry-grade logic beats integrated manufacturing at this layer.

What does the switch change for Samsung?

It signals where Samsung thinks the performance ceiling sits. TSMC's 3nm family — GAA transistors, denser logic — gives the base die a path to higher bandwidth and lower standby power than most memory makers can deliver internally. Samsung keeps its DRAM stack and packaging; TSMC supplies the logic slab underneath.

That division of labor matters for Nvidia's roadmap. HBM demand is driven almost entirely by AI accelerators, and the base die is where suppliers compete on interface speed. A 3nm base die is, functionally, the difference between a memory supplier and a memory-plus-logic partner.

Will SK Hynix and Micron follow?

That's the open question TradingKey raises, and it's the one practitioners should watch. SK Hynix has been the HBM leader on Nvidia programs and has historically leaned on TSMC for advanced base-die logic — Samsung's reported move brings it to parity in approach rather than differentiating it. Micron, the third supplier, now faces a decision: invest in advanced logic internally or follow the same TSMC route.

There's also a capacity angle. Every HBM supplier sourcing base dies from TSMC adds load to the same 3nm-class lines that already serve AI GPU customers. More HBM volume through TSMC means tighter allocation for everyone else queuing at that node — GPU designers, custom silicon programs, and now memory makers competing for the same wafers.

Why now?

AI accelerator demand has turned HBM from a niche DRAM product into the decisive bottleneck for GPU shipments. The logic base die is the part of the package closest to the processor's requirements, so the supplier with the best logic story wins sockets. Samsung choosing TSMC's 3nm says, in effect, that the base die has become a foundry problem, not a memory problem.

For buyers and system integrators, expect the HBM market to sort into suppliers that control advanced logic — directly or through TSMC — and those that don't. The reported deal narrows Samsung's gap with SK Hynix on the logic layer while raising the bar for Micron.

Nothing in the report is confirmed by the companies involved. But the direction is clear: HBM competition is migrating from the DRAM stack down to the silicon beneath it, and TSMC just picked up another seat at that table.