Samsung plans to use TSMC process technology for memory chips it is supplying to Nvidia, a report says, in a break from its usual practice of building every part of an HBM stack in-house.

The deal, if confirmed, splits Samsung's high-bandwidth memory production across two of the industry's fiercest rivals. Samsung would keep fabricating the DRAM wafers. TSMC would handle the logic die that sits at the base of the advanced HBM package and connects the memory to Nvidia's AI accelerators.

Why does the logic die matter?

Modern HBM stacks are not just memory. Each stack carries a base die — a logic layer built on a foundry process rather than a DRAM process — that handles signal routing and, in newer generations, adds processing functions.

Nvidia has pushed its memory suppliers toward this TSMC-based base-die approach. SK hynix, the current HBM leader, already produces base dies for Nvidia through a partnership with TSMC. That arrangement is one reason SK hynix has moved faster than Samsung in qualifying HBM3E for Nvidia's flagship AI GPUs, according to industry reporting.

Samsung, by contrast, has insisted on its own in-house base-die path built on its foundry technology. The new report signals that Samsung may now accept TSMC's logic wafer to win Nvidia orders it has so far failed to secure in volume.

What does this change commercially?

For Samsung, the calculus is blunt. Nvidia is the largest buyer of HBM in the world, and every HBM3E stack it doesn't ship to Nvidia goes to a competitor's balance sheet. Adopting TSMC's logic die could shorten Samsung's qualification cycle with Nvidia, even though it means Samsung's foundry arm loses that piece of the stack to its biggest competitor.

For TSMC, the reported arrangement extends its reach further into the AI memory supply chain. The company already fabricates the GPU compute dies and, through CoWoS packaging, much of the module assembly around them. Adding Samsung's HBM base dies would tighten that grip.

The report did not specify which process node TSMC would use, which HBM generation Samsung is targeting, or when volume shipments would begin.

What's confirmed and what isn't?

Only one fact is on the table: a single report, not an official statement from Samsung, TSMC or Nvidia, says Samsung will use TSMC technology for Nvidia-bound memory chips. None of the three companies has commented publicly.

That leaves the details — node, volumes, revenue timing — as open questions. Samsung's memory division reported through 2024 that HBM sales to Nvidia were ramping below its earlier schedule, and executives have said repeatedly that qualification talks continue.

Who wins if this holds?

SK hynix faces a sharper competitive picture. Its TSMC-based base-die formula, once a differentiator unique to its Nvidia relationship, would become the industry standard rather than an edge.

Samsung trades a measure of technological pride for market access. TSMC collects orders at yet another layer of the AI hardware stack. And Nvidia gets a third qualified HBM supplier, which loosens a supply bottleneck that has constrained GPU shipments since the AI boom began.

If the report proves accurate and qualification follows, the most likely next marker is Samsung shipping TSMC-based HBM stacks to Nvidia in commercially meaningful volumes during 2025 — a shift that would redraw the HBM market's competitive map.