Report: Samsung Taps TSMC for NVHBM Base Die Production

Samsung has tapped TSMC to produce the base die for its NVHBM memory, TechPowerUp reports. The decision sends a core component of Samsung's next high-bandwidth memory generation to its foundry rival's process lines.

The report names the product as NVHBM, the follow-on stack in Samsung's HBM roadmap aimed at the AI accelerator market. The base die is the logic layer at the bottom of every HBM stack. It handles the interface and data routing between the DRAM layers above it and the customer's GPU.

What is changing hands?

Under the arrangement, TSMC would fabricate the NVHBM base die while Samsung continues to build the DRAM layers on its own memory lines. Samsung then stacks and packages the finished modules. The report does not state which TSMC process node, wafer size, or fab site will carry the work.

The move matters because the base die is the most logic-intensive part of an HBM module. Routing it to TSMC hands the rival foundry a direct role inside a Samsung product family sold against HBM supplied to the same AI customers.

Why would Samsung split the stack?

The report gives no stated reason from Samsung. The arrangement nonetheless follows a pattern in the HBM segment, where the leading supplier of advanced HBM has already paired its stacks with TSMC-made logic dies to satisfy demand from Nvidia.

For Samsung's memory division, the stakes are straightforward. Samsung has trailed its chief HBM rival in qualifying advanced stacks with Nvidia. Winning that qualification is the commercial gate to the AI accelerator memory market. A TSMC-made base die appears in the report as part of Samsung's push to close that gap with NVHBM.

For TSMC, the work adds a second major HBM base-die customer to its logic portfolio. That deepens the foundry's position at the intersection of AI memory and AI compute, where it already builds the GPUs that consume the stacks.

What remains unconfirmed?

TechPowerUp does not cite shipment volumes, pricing, production start dates, or capacity commitments. The report also does not say whether Nvidia has approved NVHBM for its accelerator platforms. Samsung and TSMC have not confirmed the arrangement on the record in the source material.

Readers should treat the deal as a single-outlet report rather than a vendor announcement. No contract value, node assignment, or fab allocation appears in the account.

The competitive dynamic is nonetheless clear from the facts reported. Samsung would use its largest foundry competitor as a manufacturing partner to strengthen a memory product aimed at the same AI customers both firms court. If the NVHBM base die reaches volume production at TSMC, the arrangement would bind the two rivals more tightly across the AI supply chain.