Monday Edition · Mon 19 Oct 2026
Samsung Will Use TSMC's 3nm Process for Nvidia's Custom HBM Base Dies, Breaking With Its All-In-House Model
Nvidia has asked its three high-bandwidth memory suppliers to make the base die of its custom NVHBM chips on TSMC's 3-nanometre process, and Samsung has agreed, according to the Seoul Economic Daily and Reuters. The shift moves part of the world's most valuable memory product onto a rival's production line.
Chanan Zevin - Chief Editor and Head of Desks
Published
What was reported
Samsung Electronics will use the manufacturing process of TSMC, the world's largest contract chipmaker, to produce the base die of Nvidia's next-generation custom high-bandwidth memory, the Seoul Economic Daily reported on Sunday, 18 October, citing industry sources. Reuters reported the same development, according to the news aggregator Techmeme. Neither Samsung, Nvidia nor TSMC has published a statement confirming the arrangement, and no supply agreement has been disclosed. The report rests on industry sources, and this article treats it as attributed reporting.
Verified against a Korean primary source
The headline brief for this edition asked for confirmation from a Korean primary source before publication. The Journal has checked the original Seoul Economic Daily report by Lee Suk-jin, published at 16:47 Seoul time on 18 October, and the Korean-language version of the newspaper's coverage. The details in this article come from that reporting and from Reuters' account as summarised by Techmeme and Business Standard. The underlying sources remain unnamed industry officials.
Nvidia's request to all three suppliers
Nvidia asked all three suppliers chosen to produce NVHBM, the custom memory it unveiled in August, to make the base die on TSMC's 3-nanometre process, according to the Seoul Economic Daily. The three are Samsung, SK hynix and Micron. Nvidia also asked that the core die, the part that stores data, be standardised on the sixth-generation DRAM process known as 1c. NVHBM is based on the architecture of HBM4E, the seventh generation of high-bandwidth memory.
Why Nvidia wants one recipe
The reason is consistency. Different manufacturing methods across suppliers make it hard to deliver uniform performance from a single design, the newspaper reported. Standardising the base die on one foundry process and the core die on one DRAM generation also reduces the work of validating products separately for each supplier. For Nvidia, which buys memory from three companies for the same accelerators, a common specification lowers the risk that one supplier's chips perform differently from another's.
What a base die does
High-bandwidth memory stacks several DRAM chips on top of one another, connected by vertical wiring. At the bottom of the stack sits the base die, a logic chip that manages data flow between the memory and the processor. As memory speeds rise, the base die has become more like a processor itself, which is why it is now made on advanced logic processes. In custom HBM, the base die can include functions designed for a specific customer, which is what Nvidia's NVHBM does.
A break with the turnkey approach
Samsung has until now produced both the core die and the base die for all of its HBM supply, the Seoul Economic Daily said. That one-stop, or turnkey, model has been central to Samsung's pitch, because it is the only company that makes memory, operates a contract foundry and offers advanced packaging under one roof. Accepting TSMC for the NVHBM base die means changing that model for its most important customer. The newspaper said Samsung's Device Solutions division decided to accept after an internal review.
Only for Nvidia's custom product
The change is limited in scope. Samsung will keep its turnkey strategy for standard HBM products, the newspaper reported. It uses its own 4-nanometre foundry process for the base die of HBM4, which it began mass-producing earlier this year, and of HBM4E, which is in mass-production validation. Core dies for both are made on 1c lines at Samsung's Hwaseong and Pyeongtaek campuses. Samsung also plans to stay turnkey for custom HBM supplied to other customers, including AMD.
| Product | Base die | Core die |
|---|---|---|
| HBM4 (in mass production) | Samsung 4nm foundry | 1c DRAM |
| HBM4E (in validation) | Samsung 4nm foundry | 1c DRAM |
| Nvidia NVHBM (custom) | TSMC 3nm (Samsung foundry as hedge) | 1c DRAM (standardised) |
A first at the manufacturing stage
Samsung and TSMC have worked together before, but not like this. Four years ago, Samsung worked with TSMC to make sure its HBM was compatible with TSMC's packaging process for connecting memory to graphics processors, the newspaper said. This is the first time the two companies have partnered from the HBM manufacturing stage. TSMC already makes Nvidia's processors, so the change concentrates more of the accelerator's critical parts on TSMC's lines.
Samsung keeps a hedge
Samsung has not given up on its own foundry for NVHBM. "Samsung is validating whether it can also make the NVHBM base die at its own foundry," an industry official told the Seoul Economic Daily. "This is a hedge in case NVHBM demand exceeds TSMC's production capacity." The search summary of other reports said Samsung may validate its own 2-nanometre process as the backup. The Journal could not confirm which node Samsung is testing, so this article does not specify it.
What it means for SK hynix and Micron
SK hynix announced a partnership with TSMC in 2024 to develop HBM4 base dies, so the request aligns with an approach it has already taken. For Micron, the request also points to TSMC. The result is that all three memory makers would build Nvidia's custom memory on the same foundry process. That reduces one area in which suppliers have competed and shifts competition towards the DRAM core dies, yields and the ability to deliver volume on time.
What it means for Samsung Foundry
The decision is a setback for Samsung's contract chipmaking business. Winning base-die production for Nvidia's custom memory would have been a high-profile order for the foundry, which has struggled to attract large outside customers on its most advanced processes. Accepting TSMC for this product signals that, for Nvidia, consistency across suppliers outweighed Samsung's in-house offer. The hedge validation leaves open the chance that Samsung's foundry wins some volume later.
The scale of the market
High-bandwidth memory has become one of the most valuable products in the chip industry because every AI accelerator needs it. Samsung displayed its HBM4E at SEDEX 2026, South Korea's largest semiconductor exhibition, in Seoul on 14 October, the newspaper noted. Demand for HBM has driven record results at memory makers, and the share prices of memory companies have become a barometer of the AI trade. Micron shares rose 3.3% in premarket trading on Monday after last week's selloff, Reuters reported.
Context from the market
The report arrives during a turbulent period for chip stocks. The PHLX Semiconductor Index posted its longest losing streak since 2008 last week, Investing.com reported, amid concerns about how the AI buildout will be financed. Decisions about which foundry makes which component do not change total demand for AI memory, but they do move revenue between companies. For TSMC, more custom HBM base dies add to an order book already dominated by AI processors.
Wider industry signals
The Seoul Economic Daily also reported that China's memory maker CXMT has flagged a large jump in revenue, in a separate article published alongside the Samsung report. Competition from Chinese memory makers in conventional DRAM is one reason Korean companies are focusing on high-bandwidth memory, where technical barriers are higher. Keeping Nvidia as a customer for its custom product is therefore strategically important for Samsung, even if it means giving part of the work to a rival.
Unanswered questions
Several points remain unclear. The companies have not said when NVHBM production will begin, how volumes will be split among the three suppliers or how pricing will change. It is not known whether TSMC has committed specific 3-nanometre capacity to the base dies, at a time when demand for that process is high. And it is not confirmed whether Samsung's hedge validation could lead to a split between TSMC and Samsung Foundry for Nvidia's volumes.
Why packaging ties it together
The change also fits the way AI accelerators are assembled. TSMC's advanced packaging places the processor and several HBM stacks side by side on a shared base, and the chips must work together at very high speeds. When the processor and the memory base dies come from the same foundry family, design rules and testing can be aligned more easily. That is one reason analysts have expected memory and logic manufacturing to converge. Nvidia's request is a concrete step in that direction for its custom product.
The cost question
There is also a question of cost. Making the base die on a 3-nanometre process is more expensive than on older nodes, and adding a third company to the supply chain adds coordination. Nvidia appears willing to pay for consistency and performance. Whether memory makers can pass the extra cost on, or whether it squeezes their margins on NVHBM, has not been reported.
What to watch
Three signals will confirm or qualify the report. The first is any statement from Samsung, Nvidia or TSMC, including in upcoming quarterly earnings calls. The second is evidence of TSMC capacity allocated to memory base dies. The third is whether AMD and other custom HBM customers ask for similar standardisation. If they do, the line between memory makers and foundries will continue to blur.
Journalistic Scoop
Standardising the core die may matter more than outsourcing the base die. Most coverage focused on Samsung using TSMC. The second half of Nvidia's request may have larger consequences. According to the Seoul Economic Daily, Nvidia also asked all three suppliers to standardise the core die, the memory itself, on the sixth-generation 1c DRAM process.
If the base die is made by TSMC for everyone and the core die follows one DRAM generation, the main differences left between Samsung, SK hynix and Micron are yields, cost and delivery. That moves the competition away from design choices, where Samsung has argued its integrated model is an advantage, towards manufacturing execution. It also makes it easier for Nvidia to shift volumes between suppliers, which strengthens its bargaining position on price.
Sources: Seoul Economic Daily · Business Standard
Sources
- Seoul Economic Daily, Samsung to Tap TSMC for Nvidia's Custom HBM Base Die, 18 Oct 2026 · en.sedaily.com
- Seoul Economic Daily (Korean), 삼성·TSMC, 첨단공정 활용 HBM 물량 확대 · sedaily.com
- Techmeme summarising Reuters (Kyu-Seok Shim), 18 Oct 2026 · techmeme.com
- Business Standard, Samsung to tap TSMC technology for Nvidia's custom memory chips: Report, 18 Oct 2026 · business-standard.com
Related
- Yelp, Delta, JPMorgan and Schwab Shut the Door on Outside AI Agents as the Open Web Turns Into a Patchwork of Licences
- Stock Futures Rebound After the Chip Rout, but This Week's Earnings Will Decide the AI Trade
- Anthropic Pledges 5% of Its Shares to a Public Benefit Fund in the Same Week It Pitched Federal Pre-emption of State AI Rules