AI Agent UpdateSamsung wins Nvidia's next-gen HBM — and TSMC builds the part that matters
Samsung landed a place in Nvidia's next-generation custom HBM program — on the condition that the base dies come off TSMC's 3nm line. Nvidia is standardizing the HBM supply chain around its favorite foundry, and Samsung's own fab gets cast as the backup.
The News
Samsung Electronics will use TSMC's 3nm process to produce the base dies for Nvidia's next-generation custom HBM — the NVHBM program built on the HBM4E architecture — according to a Seoul Economic Daily report on Oct 19. The arrangement was made at Nvidia's request, which is the detail that turns a supply-chain footnote into a statement about the foundry hierarchy.
The standardization goes further. Nvidia has reportedly asked all three memory suppliers — Samsung, SK Hynix and Micron — to build their custom HBM on TSMC-made base dies paired with 6th-generation DRAM core dies. The logic is industrial, not political: a common base die across suppliers simplifies validation of the finished stack and guarantees consistent performance characteristics, so Nvidia qualifies one logic platform and swaps memory vendors on top of it.
For Samsung, the structure is double-edged. This is the first time it has used TSMC for fabrication at the HBM stage — a genuine breakthrough in a relationship where Samsung has historically insisted on end-to-end in-house production. Samsung keeps its own foundry as the backup option and retains in-house manufacturing for standard HBM sold to AMD and other customers, so the concession is specific to Nvidia's custom program — but it is a concession, made to the industry's most demanding customer.
The base die is not a commodity layer. In HBM4-generation designs, the base die carries logic functions that increasingly determine the stack's performance and power profile — which is precisely why Nvidia wants it built on the most reliable advanced process available, and why who fabricates it matters more than who stacks the DRAM on top.
Samsung won the HBM4E race and lost the fab — Nvidia just standardized the industry's most strategic component around TSMC.
Why It Matters
Nvidia is quietly standardizing the HBM supply chain around TSMC — and every layer of standardization is foundry demand that compounds. Base dies for three memory vendors' custom HBM programs is incremental 3nm wafer volume with multi-year visibility, on top of TSMC's existing accelerator and CoWoS packaging dominance. For $TSM, this is the moat widening by specification rather than by capacity: when the industry's reference design names your process, competitors aren't bidding against your price — they're bidding against the standard.
For Samsung the signal is harder. Winning a place in the NVHBM program is real progress after its HBM qualification struggles — but the condition attached is a public ranking of foundry yields. Nvidia looked at the most performance-sensitive die in its next memory generation and chose TSMC's 3nm over Samsung's in-house line. That admission will follow Samsung's foundry sales team into every customer meeting this year, even as its memory division celebrates the design win.
The read-through for the other two memory makers cuts both ways. SK Hynix and Micron get the same standardized base-die platform, which lowers their qualification risk on custom HBM programs — a real operational benefit. But it also means the logic layer of their most advanced products is no longer theirs to differentiate on. Over time, standardization commoditizes everything stacked above the base die and concentrates the scarcity — and with it the pricing power — in the one layer TSMC now owns.
Analyst's Take
Samsung won the HBM4E race entry and lost the fab — both halves matter. The memory business gets its Nvidia custom program; the foundry business gets told, in writing, that it is the backup plan. Investors who treat Samsung as one company will misprice this: the right frame is a conglomerate where one division's win is another division's demotion, and the demoted division is the one Samsung has spent a decade and tens of billions trying to build.
The sharper watch is TSMC's margin gravity. $TSM already captures the accelerator wafer and the CoWoS packaging; if standardized base dies pull HBM logic into its fabs too, does CoWoS-L pricing power let TSMC capture a slice of the HBM margin pool that memory makers considered theirs? $TSM is down 4.1% over the past month in the sector wash — the tape is pricing the cycle, while Nvidia keeps quietly re-specifying the industry around one supplier. We know which signal we'd follow.
By The Numbers
FAQ
What is NVHBM?
Nvidia's next-generation custom HBM program, built on the HBM4E architecture. Unlike commodity HBM, the custom design uses a logic base die tailored to Nvidia's accelerators — which is why the choice of foundry for that base die is strategically significant.
Why is Samsung using TSMC to make part of its HBM?
At Nvidia's request. Nvidia has reportedly asked Samsung, SK Hynix and Micron to standardize on TSMC-made 3nm base dies (with 6th-generation DRAM core dies) to simplify validation and ensure consistent performance across suppliers. It's the first time Samsung has used TSMC fabrication at the HBM stage.
Does Samsung still make HBM in-house?
Yes — for standard HBM sold to AMD and other customers, manufacturing stays in-house, and Samsung's own foundry remains the backup for the Nvidia program. The TSMC arrangement is specific to the custom NVHBM base dies.
Why does the base die matter so much?
In HBM4-generation stacks, the base die carries logic that increasingly determines performance and power efficiency — it's becoming a small processor in its own right. Whoever fabricates it holds a strategic position in every stack, which is why Nvidia wants it on the most reliable advanced process available.
Source: Seoul Economic Daily via Seeking Alpha / TradingView