A report from Digitimes claims that Nvidia's demand for custom high-bandwidth memory is pulling Samsung onto TSMC's 3nm process. If the report holds, Samsung-designed silicon would be manufactured by its fiercest foundry rival. The memory giant would become, in effect, a customer of the plant it most wants to beat.

The claim concerns HBM, the stacked DRAM that feeds data to AI accelerators. Nvidia's appetite for that memory now appears to be reshaping who builds what, and where. Custom HBM variants, tailored to a single customer's accelerator, place tighter demands on the logic and packaging around the memory stack. According to Digitimes, meeting those demands routes part of the work to TSMC's 3nm node.

Why would Samsung fabricate at TSMC?

The irony is structural. Samsung Foundry competes directly with TSMC for advanced logic customers, and it has marketed its own 3nm-class nodes, built around gate-all-around transistors, as an answer to TSMC's FinFET-based N3 family. A Samsung product flowing through TSMC's 3nm line would concede, at least for that chip, that the Taiwanese fab holds the process the customer wants.

Digitimes attributes the arrangement to Nvidia's requirements rather than to any stated Samsung strategy. The report frames Samsung's foundry arm and its memory arm as separate businesses with separate incentives. The memory division wins if it sells HBM into Nvidia's supply chain, wherever the surrounding logic is built. The foundry division loses a showcase slot either way.

What does this mean for the HBM market?

The report lands amid a three-way race in HBM supply. SK Hynix has led recent Nvidia HBM qualifications; Samsung has worked to catch up; Micron has queued behind both. Custom HBM changes the calculus because a bespoke variant binds the memory supplier to a specific accelerator design and, apparently, to a specific foundry process.

For TSMC, the reported arrangement adds a marquee Samsung-linked product to its 3nm customer list. For Samsung, it secures a position in Nvidia's memory stack at the cost of a foundry embarrassment. That is the sort of trade a memory company with weak logic leverage tends to accept. The transaction, in plain terms, swaps pride for volume.

Nvidia sits at the center of every leg of this triangle. It buys the HBM, it dictates the custom specification, and its foundry relationship with TSMC apparently decides where the supporting logic lands. No other customer currently commands that kind of leverage across memory and logic at once.

How solid is the claim?

It is a single-outlet report, and Digitimes does not name its sources. Samsung, TSMC, and Nvidia have not confirmed the arrangement publicly. Readers should treat the specifics — the 3nm routing, the custom HBM scope — as report-grade information rather than announcement-grade.

The wider context is not in dispute. Nvidia's accelerator demand has strained HBM supply for successive product generations, and suppliers have reorganized packaging and qualification pipelines around that demand. Reports of reshuffled supplier roles have followed each cycle. This one simply cuts closer to the foundry pecking order.

What comes next?

Watch for confirmation signals: a Samsung HBM qualification in a future Nvidia accelerator family, or evidence of Samsung logic silicon on TSMC's 3nm-class wafers. Until one arrives, the Digitimes account stands as a plausible map of where Nvidia's custom memory requirements are dragging the supply chain — toward TSMC, and toward a Samsung that ships through its rival's front door.

The competitive dynamic it describes, if accurate, will shape HBM sourcing decisions well beyond the current generation.