TSMC will build the base dies for Nvidia's custom high-bandwidth memory (HBM) chips, according to a report by Crypto Briefing. The deal moves a critical layer of AI accelerator production into the world's largest contract chipmaker's hands.

The headline fact is short but consequential. Nvidia is not simply buying standard HBM from memory makers anymore. It wants custom HBM, and the base die — the silicon layer that sits beneath the stacked memory cells and routes data in and out — will come from TSMC.

What is a base die, and why does it matter?

An HBM module is a stack. DRAM dies sit on top of a logic die. That logic die, the base die, carries the wiring that connects the memory stack to the processor it serves. Until recently, memory vendors built this layer themselves as part of a finished HBM product.

Moving base-die production to TSMC changes the division of labor. The foundry already builds the GPU dies for Nvidia's accelerators. Building the memory base die as well puts both sides of that critical interface inside one manufacturer's process technology.

Why is Nvidia pursuing custom HBM?

AI accelerators are constrained by memory bandwidth, not only by compute. The faster a GPU can pull data from HBM, the more of its raw compute it can actually use. Custom HBM lets Nvidia tune the memory subsystem to its own architecture rather than adapt its designs to a commodity part.

The report does not specify which Nvidia products will use the custom HBM, the timeline for production, or the financial terms between the two companies. Those details remain open questions for now.

What does this change for the memory makers?

For SK hynix, Samsung and Micron — the three main HBM suppliers — the shift carries a clear signal. If the base die moves to TSMC, the memory vendors' role narrows toward supplying the stacked DRAM layers themselves. That compresses the share of value they capture per module and hands the logic-intensive part of the stack to the foundry.

It also deepens TSMC's position in advanced packaging, the set of technologies that assemble chips and memory into single packages. Packaging has become a bottleneck for AI hardware. Demand for capacity there now rivals demand for the leading-edge wafers themselves.

What does this change for buyers of AI hardware?

In the near term, nothing visible. Custom HBM serves Nvidia's internal product roadmap, and end customers buy finished accelerators, not memory components. The effects show up later, in product cycles where memory bandwidth is tuned to the chip rather than shared across a vendor's whole catalog.

The longer-term question is supply concentration. One foundry would then sit at the center of both logic manufacturing and a growing share of memory logic. Customers planning multi-year AI infrastructure builds will watch how that dependency develops.

Crypto Briefing reported the arrangement; neither company has publicly detailed volumes, dates, or pricing. Those specifics will determine how much of the HBM market's economics actually shift toward TSMC.