Samsung Builds a Custom Memory Base for Meta’s Chips

  • AI
  • November 17, 2026
  • 0 Comments

At the bottom of a high-bandwidth memory stack sits a sliver of silicon that used to be an afterthought. On the version Samsung Electronics is building for Meta Platforms Inc., it will be made on a 2-nanometer line.

The Korean outlet The Bell reported Tuesday that Samsung has begun developing custom high-bandwidth memory for Meta’s in-house AI chips. The product is based on HBM4E, the seventh generation of the standard, and it moves functions Meta needs out of the accelerator and into the memory’s base die. That base die will be manufactured on Samsung’s foundry 2nm process.

Neither company would comment on the report, which did not give the number of stacked dies, the capacity of each stack, or a date for mass production.

The base die is the logic chip at the bottom of an HBM stack. It carries the interface that talks to the processor, the test circuits, and in newer generations a controller that manages how data moves in and out of the stack. Until HBM4, most of that logic was built on a memory process. HBM4 pushes it to a foundry node, which turns a memory product into something closer to a systems component.

That is why the customer now has an opinion about it. A standard stack is designed for the widest possible set of processors. A custom one is designed for a specific chip, with the pin assignments, power delivery and timing tuned to that chip’s design. Samsung will supply both versions, according to the report, with the custom base die built on either 2nm or 4nm depending on the customer.

Early versions will integrate the HBM controller and the die-to-die interface into the base die to cut the power spent moving data between memory and processor. Power is the binding constraint in AI data centers, where the cost of a system is increasingly set by what it takes to cool it.

Meta is a plausible customer for this. It designs its own accelerators, sold under the MTIA name, and uses them for ranking and recommendation work where the model architecture is stable and the volumes are large. A company that designs its own silicon can specify its own memory, and Meta has the additional motive of wanting a second source in a market where one supplier has held most of the volume.

The strategic read is about Samsung, not Meta. Samsung is the only company in the world that makes both memory and logic at scale, and the HBM4 generation is the first where those two businesses have to be sold together. If the base die is a foundry product, then a memory customer is also a foundry customer, and Samsung can offer a bundle that SK Hynix and Micron cannot match on their own.

Its rivals have solved the same problem differently. SK Hynix and Micron both have their HBM4 base dies manufactured by Taiwan Semiconductor Manufacturing Co. That works, and it puts the most advanced part of their product in the hands of a competitor’s foundry.

Samsung’s version keeps the whole stack inside the company. Whether that is an advantage depends on yield. Its 2nm process is newer than the nodes TSMC is running for the same generation, and a base die with a controller and a wide interface is a large, defect-sensitive chip. Custom silicon also means small volumes, at least at the start, and small volumes are where foundries lose money.

The commercial logic on Meta’s side is about power and bandwidth per dollar. HBM4 stacks carry wider interfaces than their predecessors, which raises the number of wires between memory and processor and the energy each bit costs to move. Putting the controller in the base die shortens those paths. Whether the saving is large enough to justify a custom part is the question Samsung’s engineers have to answer with silicon.

There is a broader pattern here. For twenty years the memory industry competed on cost per bit and the logic industry competed on performance per watt, and the two supply chains barely touched. HBM4 puts them in the same package. That is why memory makers now talk about design wins and why foundries now talk about memory customers.

It also changes what a memory contract looks like. A standard HBM deal is a volume and a price. A custom deal is a development program with agreed checkpoints, engineering teams exchanged, and a design that cannot be resold to anyone else. The supplier gets a customer locked in for the life of the platform. The customer gets a part nobody else can buy, and a supplier it cannot easily replace.

Meta’s volumes will not strain Samsung’s capacity the way Nvidia’s orders do. What the deal buys Samsung is a reference customer for a product line it wants to sell widely: custom HBM with a Samsung-foundry base die, sold as a bundle to every company building its own accelerator. Several of those companies are American, and several are Chinese, and all of them have the same problem, which is that the memory they need is designed around someone else’s chip.

That is the business case, and it rests on a report neither company has confirmed. What would confirm it is a qualification. Custom HBM has to pass the customer’s own testing before it ships in volume, and that process takes quarters. If Meta’s accelerators appear with Samsung base dies, the report was right, and the memory business has acquired a design-services arm.

If they do not, Samsung will have spent a year building a custom part for a customer that changed its roadmap. That has happened before in this industry, and it is the reason memory makers prefer to sell standards.

Related Posts

  • November 20, 2026
  • 11 views
Sandisk Completes Eight-for-One Split as NAND Outlook Darkens

Eight for one. Sandisk’s stock split took effect on Nov. 19, and the shares began trading on the adjusted basis on Nov. 20, a change that divides the price by…

  • November 20, 2026
  • 10 views
A Lawmaker’s Privacy Suit Puts Model Design on Trial

For three years British courts have sidestepped the question of whether the company that builds an AI model answers for what its users produce with it. A written defence filed…