Samsung’s reported decision to use TSMC’s 3nm process for NVIDIA’s custom high-bandwidth memory, NVHBM, could help its memory business meet a major customer’s requirements while shifting part of the manufacturing opportunity outside Samsung. The investment question is whether profitable memory volume outweighs the lost opportunity to make those base dies internally. The October 18 report supplies a more specific production plan, but no verified commercial volumes or product economics. We retain our existing valuation assumptions.
News analysis · Samsung Electronics · KRX: 005930 · October 18, 2026. The manufacturing decision below is attributed reporting; the architecture comes from company technical materials. AHNIM’s economic conclusions are conditional analysis.
What is new in the TSMC report
Seoul Economic Daily reported on October 18 that Samsung accepted NVIDIA’s request to use TSMC’s 3nm process for NVHBM base dies. Reuters subsequently carried the report, attributing it to the newspaper. These are not two independent confirmations of a signed supply contract. We have not located a company announcement confirming the specific manufacturing allocation.
The possibility of external production was already public. ZDNet Korea’s September 15 report described Samsung considering both its own foundry and TSMC for custom HBM, depending on customer requirements. The new detail is the reported choice of TSMC’s process for this NVIDIA implementation. It should not be presented as the first disclosure of custom HBM or a completed production ramp.
Standard HBM4E and NVHBM have different manufacturing implications
A base die is the logic layer beneath the stacked memory dies. In its August 26 NVHBM announcement, NVIDIA described moving its custom memory controller from the accelerator into that base die. Its stated objective includes making an implementation available from multiple memory suppliers, reducing the engineering work required to integrate and qualify them.
Samsung’s official COMPUTEX 2026 description of HBM4E pairs 1c DRAM core dies with a base die made on Samsung Foundry’s 4nm process. That documented standard-product configuration is different from the reported NVHBM arrangement. The report does not establish that Samsung is outsourcing all HBM production or abandoning its own foundry across customers.
| Product or question | Manufacturing evidence | Status |
|---|---|---|
| Samsung standard HBM4E | 1c DRAM core dies and Samsung Foundry 4nm base die | Company description at COMPUTEX 2026 |
| Samsung NVHBM | TSMC 3nm base die requested by NVIDIA; Samsung reportedly accepted | October 18 press report, citing industry sources |
| Commercial contribution | No verified shipment allocation, selling price, product margin or revenue schedule in the reviewed documents | Not yet quantifiable |
NVIDIA claims up to 30% more memory bandwidth and 15% lower HBM power consumption than standard HBM4E. These are platform claims from the August announcement, not independently measured results for a particular Samsung shipment. HBM power savings also cannot be read as an equal reduction in total accelerator power, factory costs or Samsung’s manufacturing expense.
The memory opportunity and the foundry opportunity can diverge
Our interpretation: accepting a customer’s manufacturing requirements can protect access to a valuable memory program. Samsung would still need competitive memory dies, stacking, qualification and delivery execution. The potential benefit comes from selling qualified products at an adequate return; the choice of foundry alone does not establish either market share or earnings.
For Samsung’s foundry, external base-die production represents a different outcome from an entirely internal supply chain. It may reduce the work available to Samsung’s own lines for this program relative to an all-in-house scenario. The size of that opportunity cost depends on volumes, alternative factory demand and whether internal capacity can serve other profitable customers.
At the consolidated level, lost internal transfer revenue is not automatically an equal loss of group revenue: intragroup sales are eliminated. The economic comparison is external purchasing cost and the memory profit retained, versus the manufacturing costs, capacity use and investment of an internal route. We cannot infer the net margin effect without those inputs. Outsourcing could avoid some investment while creating external supplier payments; existing equipment and fixed costs do not disappear merely because a new product uses another foundry.
Cash flow would additionally depend on purchase commitments, inventory, yields, customer collection terms and capital spending. A product can become strategically important before it contributes meaningful free cash flow. This is why our Samsung fundamental analysis separates memory earnings durability from evidence of a broad foundry recovery.
What we would change in the model—and what remains unchanged
Our October 12 baseline values Samsung ordinary shares at ₩352,000 in the Base case, using FY2027 earnings for an approximately October 2027 horizon. That is a dated scenario, not a new target issued today. We leave its revenue, margin, share-count and valuation inputs unchanged because the reviewed evidence does not quantify an incremental earnings contribution or a measurable loss.
The first model input to revisit would be the mix and profitability of memory sales, alongside the resulting foundry utilization and spending requirements. We would raise the earnings path only if awarded volumes and realized economics justify it; we would lower it if external sourcing and underused internal capacity materially erode consolidated returns. A manufacturing headline alone provides neither calculation.
We did not verify a broker target or recommendation change issued in response to today’s report. The Stock Analysis/S&P Global page, accessed October 18, displays a one-year mean of ₩478,906 from 36 analysts. Our unchanged Base is 26.5% lower: |352,000 − 478,906| ÷ 478,906. This exceeds our 20% review threshold on an approximate one-year comparison. The page does not show a precise aggregation timestamp or each broker’s endpoint, so it is not an exactly matched-date comparison or evidence of a post-report consensus revision.
The gap was already present in our baseline and chiefly reflected a lower earnings assumption, rather than a large difference in the implied earnings multiple. It is a reason to scrutinize earnings durability, not to lift our valuation to the average. The original model and its historical assumptions remain available in the linked research.
The next evidence that matters
We would first look for Samsung, NVIDIA or TSMC to confirm the manufacturing arrangement and distinguish development from qualified production. Next come customer allocation, shipment timing and sufficient cost or margin disclosure to assess consolidated returns. Finally, cash conversion and investment needs must show that a larger memory opportunity creates value after funding the ramp.
The report strengthens the case for Samsung adapting its manufacturing approach to customer demand. It also weakens a simple assumption that every new HBM program automatically supports Samsung Foundry. Both can be true while the group’s fair-value estimate remains unchanged.
Related coverage
Samsung Electronics: memory earnings, foundry risk and the FY2027 valuation baseline · TSMC Q3: advanced-node growth, manufacturing margins and cash generation. The latter provides foundry context; its reported revenue does not identify Samsung’s NVHBM allocation.
Sources and original documents
Seoul Economic Daily — Samsung to Tap TSMC for Nvidia’s Custom HBM Base Die; October 18, 2026; publisher’s English translation, citing industry sources
https://en.sedaily.com/finance/2026/10/18/samsung-to-use-tsmc-for-base-die-in-nvidias-custom-hbm
Reuters — Samsung to use TSMC technology for Nvidia’s next-generation memory chips; October 18, 2026, 12:08 UTC; secondary attribution to Seoul Economic Daily
https://www.reuters.com/world/asia-pacific/samsung-use-tsmc-technology-nvidias-next-generation-memory-chips-seoul-newspaper-2026-10-18/
NVIDIA — NVLink Fusion Expands With NVHBM Custom High-Bandwidth Memory; August 26, 2026
https://blogs.nvidia.com/blog/nvlink-fusion-nvhbm-custom-high-bandwidth-memory/
Samsung Semiconductor — AI Semiconductor Innovations at COMPUTEX 2026; event June 2–5, 2026; publication date not displayed; accessed October 18, 2026
https://semiconductor.samsung.com/news-events/tech-blog/samsung-showcases-next-generation-ai-semiconductor-innovations-at-computex-2026/
ZDNet Korea — Samsung custom-HBM base-die sourcing strategy; displayed publication date September 15, 2026
https://zdnet.co.kr/view/?no=20260915100431
Stock Analysis / S&P Global Market Intelligence — Samsung Electronics forecasts; accessed October 18, 2026; exact consensus aggregation date not displayed
https://stockanalysis.com/quote/krx/005930/forecast/
Leave a Reply