Samsung Electronics is the only company on earth that fabricates both the DRAM core dies and the logic base die inside its own high-bandwidth memory stack. As of last week, it agreed to stop doing that for its largest customer. Nvidia asked Samsung, SK hynix and Micron to build the base die for its custom NVHBM memory on TSMC's 3-nanometer process, and Samsung accepted after an internal review — the first time Samsung and TSMC have worked together at the HBM manufacturing stage, despite being direct rivals in advanced foundry. Read that again as a supplier being told which of its own components it is no longer allowed to make.
The reported trigger is yield, and the numbers are brutal. Industry estimates put TSMC's 3nm yield above 90 percent; Samsung Foundry's comparable 2nm SF2 process sits at roughly 60 percent. For a part Nvidia intends to buy in volume, a 60 percent base die is not a supply chain, it is a liability. Samsung's own base-die capacity is confined to Hwaseong Line 17 and part of Pyeongtaek P3 — lines the Seoul Economic Daily reports could not absorb custom volumes like NVHBM on their own. Nvidia did not ask Samsung to improve. It asked Samsung to step aside.
What actually moved is the important part, and it is not a memory upgrade. NVHBM takes the memory controller off the GPU and plants it on the base die — the logic chip at the bottom of the stack that decides how fast data moves between memory and compute. The physical layer shrinks by 60 percent, freeing GPU floorplan for compute circuits, and the design delivers up to 30 percent more bandwidth and 15 percent less power than standard HBM4E. Those are the Seoul Economic Daily's figures, not Nvidia's, but the architecture they describe is unambiguous.
The memory controller moves off the GPU and onto the base die, turning the one differentiated logic layer of an HBM stack into Nvidia's design on Nvidia's chosen process.
Once the controller lives on the base die, the chip that governs how memory talks to compute is no longer a memory maker's design. It is Nvidia's design, on Nvidia's chosen process, running Nvidia's controller IP. The memory makers keep the core dies — sixth-generation 1c DRAM, uniform across all three suppliers — and they keep the stacking and packaging. What they surrender is the logic layer, the one piece of an HBM stack that was never a commodity.
This is the difference between NVHBM and the memory it replaces. SK hynix bought its HBM4 base die from TSMC on a 12-nanometer process; Samsung built its own on 4-nanometer lines. Each supplier made a different logic chip on a different node, and Nvidia had to qualify each one separately. NVHBM collapses that: one process, one design, one behaviour from three vendors, fewer qualification cycles, and a GPU with more room for compute because the controller is no longer sitting on it. Nvidia has standardised the top half of its memory subsystem and handed the bottom half to a single foundry.
Here is the irony, and it is structural rather than accidental. Nvidia now has three memory suppliers — a genuine diversification of DRAM core-die capacity across Samsung, SK hynix and Micron. It has exactly one base-die fab. Whichever memory maker wins the larger share of NVHBM orders, the logic chip under every stack comes off a TSMC 3nm wafer, and TSMC is paid per wafer before a single stack ships.
TSMC's 3nm lines are already its biggest business — 33 percent of wafer revenue in the third quarter, with 2nm adding another 8 percent. Two additional 3nm lines at Fab 18 in Tainan reach mass production in the first half of 2027 and the first half of 2028; 3nm arrives at TSMC's second Arizona plant in the third quarter of 2027 and at Kumamoto in 2028. An industry official told the Seoul Economic Daily that allocated volumes "could hit a ceiling." The diversification is cosmetic. The single point of failure sits in Tainan.
The dependency cuts both ways, which is why this is a power grab and not a procurement decision. Nvidia has converted its three "partners" into interchangeable core-die vendors selling into a commodity, differentiated only by DRAM yield and stacking capacity. It has simultaneously concentrated the irreplaceable layer — Nvidia-designed logic on a process only TSMC can yield at volume — in one foundry it does not own. That is leverage over the memory makers purchased with exposure to TSMC. Nvidia clearly judges the trade worth it, because the alternative was three base dies it could not fully control.
For the memory makers, the direction of travel is uncomfortable, and they know it. The base die was the one part of an HBM stack where a memory company could add differentiated logic, own the process, and capture value beyond DRAM density. Nvidia has pushed them toward the commodity half of their own product — the half where Samsung, SK hynix and Micron compete on essentially the same 1c DRAM and the same stacking steps, selling to a customer that has already decided the interesting silicon is someone else's.
Samsung's hedge is the tell. It is keeping its in-house 4nm base die for standard HBM4 and HBM4E and for other customers such as AMD, and it is reviewing whether Samsung Foundry could make NVHBM base dies itself if demand outruns TSMC's capacity. That last clause is not contingency planning. It is a company preserving the option to re-enter a layer it has just been evicted from, in case its rival's 3nm queue proves the bottleneck everyone expects.
The sibling piece on memory as the binding constraint establishes that HBM is scarce. This is what Nvidia is doing with that scarcity. A customer who cannot get enough memory has leverage over the companies that make it, and Nvidia is spending that leverage to rewrite who designs which layer of the stack. The shortage will eventually ease; the design control will not. Once three suppliers are shipping a standardised Nvidia base die on TSMC 3nm, there is no path back to the old arrangement where each memory maker owned its own logic.
Samsung will build the logic half of its own HBM stack on its fiercest foundry rival's process, because its largest customer told it to. That sentence would have been unthinkable eighteen months ago. It is now the architecture of the industry's most important memory product — and the clearest evidence yet that in AI hardware, the company that specifies the component decides who makes it, and on whose lines.