Samsung to use TSMC's 3nm process for Nvidia's NVHBM base dies

A base die is the bottom layer of an HBM stack - a system semiconductor that handles I/O and signal routing, and in Nvidia's design now also holds the memory controller itself ^[5], sitting beneath the stacked DRAM core dies. Until this deal, Samsung built both pieces itself: the DRAM core die and the base die, a turnkey approach that gave it end-to-end control of its HBM supply ^[1]. That changes for Nvidia's custom NVHBM specifically. The reported reason is blunt: TSMC's 3nm process is running yields estimated above 90%, while Samsung Foundry's comparable 2-nanometer (SF2) node is estimated near 60% ^[6]. For a product line Nvidia is staking its next GPU generation on, a roughly 30-point yield gap is not a rounding error - it is the difference between a dependable supply and a bottleneck.

What stands out isn't just Samsung's concession - it's who orchestrated it. Nvidia reportedly asked all three of its HBM suppliers - Samsung, SK Hynix, and Micron - to standardize NVHBM base-die manufacturing on TSMC's 3nm process, and to align their core/DRAM dies on the same sixth-generation 1c process, so one chip design behaves identically regardless of who built it ^[1]. That is an unusual degree of control for a fabless buyer to exert over three competing manufacturers simultaneously: Nvidia isn't just picking a winner, it is forcing convergence. SK Hynix was already most of the way there, having used TSMC's 12-nanometer logic process for its HBM4 base dies and now prioritizing TSMC's 3nm for HBM4E and future custom parts ^[4]. Micron is following the same script, routing both NVHBM and standard HBM4E base-die work to TSMC rather than its own fabs ^[4]. Samsung was the last of the three to be reported making this move, and the only one of the three with its own major foundry business at stake.

Samsung's move reads less like capitulation than a calculated hedge. Even as it hands NVHBM base-die work to TSMC, the company is reportedly still validating whether it could manufacture that same base die at its own foundry, as a hedge ^[1]. The concession is also narrower than headlines suggest: Samsung's turnkey model survives everywhere else. Samsung Foundry still applies its own 4-nanometer process to the base dies of standard HBM4 and HBM4E, the JEDEC-standard products that aren't custom-built for a single customer ^[7]. In effect, Samsung now runs two parallel HBM strategies at once - turnkey for the industry-standard parts it fully controls, and a split build, pairing its own core die with a TSMC-made base die, for Nvidia's custom NVHBM ^[4]. That split arrives as Samsung's foundry division is still posting losses, albeit shrinking ones - down 41% year-over-year to roughly $2.95 billion in 2026 ^[3].

The timing isn't incidental. UBS projects 2027 HBM demand at 61.5 billion gigabits, up 91% year-over-year, against roughly 54.6 billion gigabits of new shipments - a shortfall on the order of 10-11% ^[2]. NVHBM itself is built to squeeze more performance per wafer: Nvidia claims up to 30% higher bandwidth and 15% lower power versus standard HBM4E, partly by moving its custom memory controller directly into the base die rather than the GPU compute die, which also frees up to 25% more area on the GPU itself ^[5]. With Nvidia's reported supply commitments reaching $279 billion in Q2 FY27 - up from $119 billion the prior quarter - and Nvidia's Feynman GPUs, expected in 2028, cited as NVHBM's potential first adopter platform ^[5], Nvidia has every incentive to lock down a predictable, high-yield base-die supply before demand peaks, rather than risk a shortfall spread unevenly across three suppliers each using a different manufacturing recipe.

Not everyone is convinced the standardization fully solves Nvidia's problem. A notable line of pushback circulating on X argues that TSMC can't simply absorb this extra work: TSMC's 3nm capacity is also committed to building Nvidia's own Rubin and Rubin Ultra GPUs, which is why some commentators see Samsung's in-house foundry option as the more critical fallback precisely because TSMC has limited wafer capacity to spare for NVHBM base dies. That tension lines up with the reporting that Samsung itself hasn't abandoned its own-foundry validation ^[1]. Coverage of the story so far mirrors its niche, B2B nature: Korean financial-news channels picked it up quickly and in depth, while English-language video coverage and community discussion - including on Reddit - have not yet caught up, consistent with a supply-chain story that moves through trade press and analysts before it reaches a general audience. Unnamed industry analysts quoted in the broader English-language coverage summed up the ambivalence: the TSMC deal confirms Samsung's pragmatic approach, but it raises real questions about where Samsung's own foundry roadmap is headed ^[7].