For decades, Samsung Electronics and Taiwan Semiconductor Manufacturing Company (TSMC) have waged one of the fiercest technological and commercial rivalries in the semiconductor industry. Yet in a landmark shift reported in October 2026, Samsung’s Device Solutions division agreed to manufacture a critical logic component for Nvidia’s next-generation custom AI memory on TSMC’s 3-nanometer foundry lines.
The component at the center of this unprecedented truce is the NVHBM base die. Modern AI accelerators are increasingly throttled not by raw compute speed, but by how fast data can move between High-Bandwidth Memory (HBM) stacks and the GPU, and how much power that data transfer consumes. By redesigning the bottom logic layer of the memory stack and forcing Samsung, SK Hynix, and Micron onto a single TSMC 3nm manufacturing process, Nvidia is rewriting the rules of memory architecture and supply-chain leverage.
In this technical breakdown, you will learn how an NVHBM base die reshapes GPU silicon budgets, why TSMC’s 3nm yield advantage left Samsung with little choice, and what this split-foundry model means for the broader AI hardware ecosystem.
What is an NVHBM base die, and how does it change GPU architecture?
Understanding why Nvidia pushed rival chipmakers into a foundry alliance starts at the very bottom of a High-Bandwidth Memory stack. While enthusiasts often treat HBM as passive storage, the foundation layer beneath those vertical DRAM slices has evolved into active, high-density processor logic. This architectural shift matters because it directly determines how much compute silicon Nvidia can pack onto its flagship AI GPUs.
An NVHBM base die is the custom logic layer positioned at the bottom of Nvidia’s proprietary High-Bandwidth Memory stack, built upon the seventh-generation HBM4E architecture. Unlike conventional HBM base dies that only handle basic signaling and routing, the NVHBM base die absorbs the GPU’s memory controller directly into the memory stack itself, freeing up GPU die space for additional compute cores.
In traditional GPU architectures, the memory controllers sit around the perimeter of the primary GPU compute die. As AI models demand wider memory buses, scaling from 1,024-bit interfaces in HBM3E to 2,048-bit interfaces in HBM4 and HBM4E, those memory controllers eat up increasingly large tracts of expensive reticle-limited GPU silicon. Nvidia solved this spatial bottleneck by co-designing the NVHBM base die with TSMC and its memory partners, relocating the memory controller circuitry off the main GPU and onto the HBM logic base.
Fitting complex memory-controller logic into the narrow footprint of an HBM base die required aggressive miniaturization. Engineers shrank the physical layer (PHY), the circuit block responsible for translating electrical signals between the memory stack and the interposer, by 60% compared to standard HBM designs. Shortening the physical distance between the memory controller and the stacked DRAM core dies yields immediate electrical benefits:
- Up to 30% Higher Bandwidth: By tightening signal paths and optimizing controller timing directly beneath the DRAM stack, NVHBM boosts peak memory bandwidth by as much as 30% over standard HBM4E, pushing theoretical per-stack throughput to roughly 5.2 TB/s.
- 15% Lower Power Consumption: Shorter interconnects reduce capacitive load and signal attenuation, cutting HBM power draw by 15% and creating extra thermal headroom for the adjacent GPU compute die.
- Expanded GPU Compute Density: Offloading controller logic reclaims square millimeters on the main GPU die, allowing Nvidia to dedicate more silicon area to tensor cores, caches, and matrix-math engines.
Standard HBM4E vs. Nvidia NVHBM architecture
| Architectural Attribute | Standard HBM4E | Nvidia Custom NVHBM | Engineering Impact |
| Memory Controller Location | Integrated on the GPU compute die | Relocated onto the HBM base die | Frees GPU reticle area for more compute cores |
| Base Die Process Node | Supplier-specific (e.g., Samsung 4nm) | Unified TSMC 3nm (N3) across all vendors | Enables dense logic integration and uniform thermals |
| Core DRAM Generation | Standard HBM4E DRAM layers | Standardized sixth-generation 1c DRAM | Ensures consistent bit density and timing across suppliers |
| Physical Layer (PHY) Footprint | Baseline HBM4E PHY area | Reduced by 60% | Creates room on the base die for controller logic |
| Bandwidth & Power Profile | Baseline HBM4E performance | Up to +30% bandwidth, -15% power draw | Higher token generation speed and better rack efficiency |
By transforming the base die from a simple traffic router into an active memory controller, Nvidia turned advanced logic fabrication into a prerequisite for supplying AI memory. That architectural shift sets the stage for why TSMC’s 3nm process became non-negotiable.

Why Nvidia forced Samsung, SK Hynix, and Micron onto TSMC’s 3nm node
Unveiled in August 2026, Nvidia’s NVHBM specification does not merely ask memory suppliers to meet a performance target, but dictates the exact manufacturing recipe. Nvidia instructed all three of its major memory suppliers, Samsung Electronics, SK Hynix, and Micron, to fabricate their NVHBM base dies on TSMC’s 3nm process while standardizing the vertical memory slices above them on sixth-generation 1c DRAM. Understanding why Nvidia enforced this single-foundry mandate reveals the harsh physics and economics of sub-4nm chipmaking.
The first driver behind the mandate is electrical and thermal uniformity. When a memory controller lives inside the HBM base die, subtle differences in transistor threshold voltage, leakage current, or heat dissipation between foundries can cause timing skew across multi-stack GPU packages. If Samsung used its own foundry node, SK Hynix used TSMC, and Micron used another process, Nvidia would have to maintain three separate physical design kits (PDKs) and run exhaustive, multi-month validation cycles for every supplier combination. Routing every NVHBM base die through TSMC’s 3nm lines guarantees identical electrical behavior regardless of which company stacks the DRAM on top.
The second, more decisive factor is manufacturing yield. Packing a shrunk 60% PHY alongside full memory-controller logic demands an ultra-fine leading-edge logic node. According to industry estimates cited in supplier briefs, TSMC’s mature N3 FinFET process achieves yields exceeding 90%. By contrast, Samsung Foundry’s competing leading-edge node, its 2nm SF2 gate-all-around (GAA) process, has tracked at roughly 50% to 60% yield.
In advanced packaging, a 30-to-40-percentage-point yield gap is commercially disqualifying. Because an HBM stack bonds eight, twelve, or sixteen expensive 1c DRAM core dies directly onto a single base die using through-silicon vias (TSVs) and hybrid bonding, a defective base die can ruin an entire multi-hundred-dollar memory stack. Furthermore, Nvidia’s custom controller IP is already heavily optimized for TSMC’s N3 design rules.
Even for a vertically integrated giant like Samsung, the math left little room for pride. Adopting TSMC’s 3nm base die was the fastest, lowest-risk way to meet Nvidia’s qualification timeline without jeopardizing margins on defective stacks.
Inside Samsung’s strategic compromise: Turnkey pride vs. AI supply reality
For years, Samsung’s primary competitive pitch in the AI memory market was its end-to-end “turnkey” capability. As the only company operating a world-class memory business, a leading-edge logic foundry, and an advanced 2.5D/3D packaging division under one roof, Samsung traditionally manufactured both the logic base die and the DRAM core dies in-house. Stepping away from that model to buy base dies from its biggest foundry rival marks a historic strategic pivot.
Samsung’s Device Solutions division approved the TSMC 3nm arrangement only after an extensive internal review. While Samsung and TSMC first collaborated four years ago to verify physical compatibility between Samsung’s HBM stacks and TSMC’s CoWoS GPU packaging lines, the NVHBM agreement is the first time the two companies are co-manufacturing components inside the HBM stack itself.
Two powerful market forces pushed Samsung to make this concession:
- Severe HBM Supply Deficits: Demand for AI memory continues to outstrip global wafer capacity. Investment bank UBS projects total HBM demand in 2027 to surge 91% year-over-year to roughly 61.5 billion gigabits, while total industry shipments are projected to reach only 54.6 billion gigabits, leaving a structural shortfall of approximately 10%.
- Internal Base-Die Bottlenecks: Samsung’s own domestic capacity for fabricating advanced logic base dies is tightly constrained, largely restricted to Hwaseong Line 17 and portions of its Pyeongtaek P3 facility. Tapping TSMC gives Samsung immediate access to high-volume 3nm wafer capacity without waiting years for new domestic fabs to come online. Meanwhile, TSMC is actively expanding 3nm capacity with two new lines at Fab 18 in Tainan—slated for mass production in the first halves of 2027 and 2028, alongside 3nm buildouts in Arizona and Kumamoto, Japan.
Crucially, Samsung is not abandoning its foundry business. The split-manufacturing model applies specifically to Nvidia’s custom NVHBM program.
Samsung’s split memory strategy by product line
| Memory Product Line | Customer Target | Base Die Foundry & Node | DRAM Core Die | Manufacturing Model |
| Custom NVHBM | Nvidia | TSMC 3nm (N3) | Samsung 1c DRAM | Split-foundry (outsourced base die) |
| Standard HBM4 / HBM4E | Broad AI & HPC market | Samsung Foundry 4nm | Samsung 1c DRAM | End-to-end turnkey (100% in-house) |
| Custom HBM (Non-Nvidia) | AMD and hyperscalers | Samsung Foundry (4nm / custom) | Samsung DRAM | End-to-end turnkey (100% in-house) |
| Future NVHBM Overflow | Nvidia (contingent) | Evaluating Samsung 2nm (SF2) | Samsung 1c DRAM | Dual-source backup if TSMC N3 caps out |
By drawing a clear boundary between Nvidia’s custom specification and the rest of its memory portfolio, Samsung preserves its turnkey pitch for major clients like AMD while securing its slot in Nvidia’s supply chain. At the same time, Samsung engineers are actively evaluating whether its 2nm SF2 process can serve as a secondary source for NVHBM base dies if Nvidia’s order volume exceeds TSMC’s 3nm allocation.

Pros and cons of Nvidia’s unified NVHBM foundry mandate
Consolidating custom HBM logic across three competing memory giants onto a single foundry node creates clear winners while introducing new systemic dependencies. Evaluating the pros and cons highlights why this arrangement is both an engineering triumph and a delicate supply-chain balancing act.
Strategic Advantages of a Standardized TSMC 3nm Base Die
- Predictable Multi-Vendor Performance: Nvidia can swap or blend NVHBM stacks from Samsung, SK Hynix, and Micron on the same GPU package without worrying about mismatched controller timing or thermal hotspots.
- Faster Time-to-Market: Sharing a single co-designed TSMC 3nm base die eliminates redundant IP porting and cuts qualification cycles across all three memory partners.
- Higher Memory Profitability: Even with outsourced base dies, memory suppliers like Micron have indicated that NVHBM’s custom value-add and superior yield profile improve overall gross margins compared to commodity DRAM.
- Guaranteed High-Volume Wafer Supply: TSMC captures an extra 3nm logic wafer beneath every NVHBM stack shipped—meaning TSMC wins regardless of which memory vendor wins the DRAM contract.
Supply-chain risks and bottlenecks
- Extreme Concentration Risk at TSMC: Routing all NVHBM base dies through TSMC 3nm stacks puts new logic demand directly on top of TSMC’s already tight N3 and CoWoS packaging capacity. Any disruption in Tainan hits every Nvidia memory supplier simultaneously.
- Reduced Differentiation for Memory Makers: When vendors standardize the logic base die, Samsung, SK Hynix, and Micron compete primarily on 1c DRAM yield, stacking precision, and pricing rather than proprietary controller IP.
- Margin Sharing with a Rival: Samsung must pay foundry wafer fees to TSMC rather than keeping the entire value chain inside its own fabrication plants.
Ultimately, Nvidia’s purchasing power, backed by hundreds of billions of dollars in multi-year supply commitments, allowed it to prioritize architectural performance and yield certainty over vendor autonomy.
Samsung’s decision to tap TSMC’s 3nm lines for the NVHBM base die proves that the boundary between logic foundries and memory makers has permanently blurred. As AI workloads push data transfer rates past 5 TB/s per stack, moving the memory controller off the GPU and onto a leading-edge 3nm base die is no longer an experimental luxury, it is the only way to keep compute cores fed without blowing through data-center power budgets. Even fierce rivals like Samsung and TSMC now find their roadmaps intertwined beneath every next-generation AI accelerator.
Frequently Asked Questions
An NVHBM base die is the bottom logic layer of Nvidia’s custom High-Bandwidth Memory stack (based on HBM4E). Unlike traditional HBM base dies, it houses the GPU’s relocated memory controller in a 60% smaller physical layer, freeing GPU die space for compute cores and boosting bandwidth by up to 30%.
No. TSMC only fabricates the bottom logic base die on its 3nm process. Samsung, SK Hynix, and Micron still manufacture the sixth-generation 1c DRAM core dies in their own fabs and stack them onto the TSMC-made base dies.
Industry estimates show TSMC’s 3nm process achieves over 90% yield, whereas Samsung Foundry’s comparable 2nm SF2 process has tracked around 50% to 60%. Nvidia also required all three memory suppliers to use TSMC’s 3nm node, so every NVHBM stack shares identical electrical and thermal characteristics without separate validation cycles.
Yes. Samsung is keeping its traditional end-to-end turnkey model, using its own 4nm base dies, for standard HBM4 and HBM4E products as well as custom HBM supplied to other customers, including AMD.
While these four answers cover the immediate mechanics of the deal, the broader shift toward logic-heavy memory bases will influence GPU and AI accelerator roadmaps well into 2028.
