Strategic AI Memory Partnerships Reinforce the Value of HBM and High-End Memory Supply
AI summary card
Strategic AI Memory Partnerships Reinforce the Value of HBM and High-End Memory Supply
Bernstein believes that the large-scale partnerships between NVIDIA, Broadcom, and Samsung and SK hynix further prove that AI growth is elevating the strategic importance of memory, and that the recent pullback in memory stocks can be viewed as a favorable entry point.
- SK hynix signed an MOU with NVIDIA, with the cooperation size described as US$500B+, and involving the 2GW Vera Rubin DSX AI Factory for SK Telecom planned to come online in 2027.
- Samsung signed an MOU with Broadcom involving approximately US$200B over the next 5 years through 2030 for memory, HBM, and foundry services, covering 2nm and below process technologies and advanced packaging.
- The report believes the announced amounts point more to memory demand, showing that NVIDIA and Broadcom need to lock in future AI memory supply.
- Bernstein remains positive on memory, but believes China's long-term threat in NAND is higher, and therefore maintains Underperform on KIOXIA.
- The report believes that even if Broadcom partially shifts to Samsung or Intel, the visible earnings impact on TSMC should be limited because demand for its capacity remains backlogged.
Report interpretation
Overview
This report is Bernstein's rapid commentary on multiple semiconductor strategic partnerships announced by the Korean government at the San Francisco AI Summit. The core events include partnerships among Samsung, SK hynix, NVIDIA, and Broadcom around memory, foundry, and AI data centers. The report interprets these partnerships as a reaffirmation of the AI value chain's dependence on high-end memory, especially HBM and next-generation AI memory supply security.
Core views
The report's core views are: first, AI's demand for and strategic dependence on memory are rising, further strengthening the importance of memory versus logic semiconductors in AI systems; second, the partnership arrangements of NVIDIA and Broadcom help ease market concerns over the memory supply needed for their AI growth; third, the direct impact on TSMC should be limited because its advanced process and advanced packaging capacity still has ample queued demand; fourth, the recent pullback in the memory sector provides a good entry point, but NAND faces long-term competitive risk from Chinese manufacturers, while DRAM is relatively better protected by EUV barriers.
Analysis framework
The report uses event-driven analysis to review deal size, scope of cooperation, supply chain links, and potential beneficiaries, and combines market consensus expectations, target-price valuation methodology, rating framework, and risk disclosures to assess the investment implications for memory makers, AI chipmakers, and foundries.
Methodology notes
Assess value-chain impact through deal size, technology scope, and supply chain links
The report focuses on comparing the size, time horizon, and technology areas of the SK hynix/NVIDIA and Samsung/Broadcom partnerships to judge the importance of AI memory demand and supply lock-in.
Growth in AI training and inference drives demand for HBM, DRAM, NAND, and advanced packaging
The report believes the large-scale partnerships mainly reflect the importance of memory supply, and cites consensus expectations of approximately US$1.3T in annual revenue for the overall memory industry in 2027 and 2028 as context.
Use forward EPS and target multiples to derive target prices
The report separately discloses valuation multiples and target prices for Samsung, SK hynix, Micron, KIOXIA, NVIDIA, and Broadcom based on forward EPS or specific fiscal-year EPS.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Samsung Electronics Co LtdSupplier of memory, HBM, foundry, and advanced packaging; signed an MOU with Broadcom
- Strengths
- Large cooperation scale covering 2nm and below process technologies, HBM, and advanced packaging, which may enhance business opportunities related to AI and networking chips.
- Weaknesses
- The report does not clearly state whether Broadcom will shift AI ASIC wafers or advanced packaging to Samsung, so the actual revenue contribution remains difficult to quantify precisely.
- Comparison
- Compared with TSMC, Samsung showcases solutions such as Cube-S, Cube-E, and Cube-R in advanced packaging, but the report believes the visible earnings impact on TSMC is limited.
- Risks
- If demand weakens, supply increases, or the favorable pricing environment ends early, target prices and earnings expectations may come under pressure.
- SK Hynix IncLong-term AI memory partner of NVIDIA; core supplier of HBM and next-generation AI memory
- Strengths
- Its cooperation with NVIDIA is described as US$500B+ and involves long-term technology development and stable supply of next-generation AI memory.
- Weaknesses
- The specific financial impact of the partnership cannot be precisely quantified and still depends on sustained AI demand and execution of supply.
- Comparison
- Compared with NAND, DRAM has higher EUV technology barriers, and the report believes the competitive threat from China is lower.
- Risks
- Downturn in the memory pricing cycle, weaker-than-expected demand, or supply expanding too quickly.
- Micron Technology IncGlobal memory beneficiary
- Strengths
- The report maintains Outperform and believes AI memory demand provides positive support to DRAM/HBM suppliers.
- Weaknesses
- Valuation in the table is relatively high, and near-term business volatility may affect market multiples.
- Comparison
- Along with Samsung and SK hynix, it is part of the DRAM/memory beneficiary theme favored by the report.
- Risks
- Competitors cutting prices or taking share, revenue growth in key end markets slower than expected, customers shifting to in-house chips, and technology export regulation risks.
- KIOXIA Holdings CorpNAND-related name
- Strengths
- If NAND demand is stronger than expected, pricing improves, Japanese government capex subsidies materialize, or costs decline more than expected, there is upside risk.
- Weaknesses
- The report maintains Underperform primarily because the long-term competitive threat from China in NAND is higher.
- Comparison
- Compared with DRAM, NAND's technology barriers and competitive landscape are viewed by the report as more vulnerable to pressure from Chinese manufacturers.
- Risks
- Chinese NAND competition, weakening pricing environment, and weaker-than-expected demand.
- NVIDIA CorpAI accelerator demand-side player and memory supply lock-in party
- Strengths
- The data center opportunity is massive and still in its early stages, and the report believes the partnership with SK hynix can ease supply concerns amid strong future AI growth.
- Weaknesses
- Highly dependent on continued delivery of AI infrastructure demand and supply chain execution.
- Comparison
- Like Broadcom, NVIDIA is viewed by the report as a core AI growth company that needs to lock in high-end memory supply.
- Risks
- Near-term business trend volatility, slowing revenue growth in key end markets, customers migrating to in-house chips, and export regulation risks.
- Broadcom IncDemand-side player for AI ASICs, networking chips, and memory/foundry partnerships
- Strengths
- The report believes the AI trajectory in 2026 is strong and may continue to accelerate in 2027 and beyond, with support from software, cash deployment, margins, and free cash flow.
- Weaknesses
- The specific coverage of the Samsung partnership in AI ASIC wafers or advanced packaging remains unclear.
- Comparison
- Similar to NVIDIA, Broadcom's strong AI growth path requires more stable memory and advanced packaging supply.
- Risks
- Unexpected weakness in AI demand, loss of share or sockets at key customers, M&A synergies falling short, insufficient cash returns, and management changes.
- TSMCPotentially affected advanced foundry and advanced packaging benchmark competitor
- Strengths
- The report believes that because TSMC's capacity still has queued demand, even if a small amount of share flows to Samsung or Intel, its visible earnings should not be affected.
- Weaknesses
- The Samsung-Broadcom MOU shows competitors are strengthening capabilities in sub-2nm processes and advanced packaging.
- Comparison
- Samsung Cube-S is similar to TSMC CoWoS-S, while Cube-E and Cube-R are similar to TSMC CoWoS-L and CoWoS-R.
- Risks
- If future migration of advanced packaging or AI ASIC orders expands, long-term share risk will need continued monitoring.
Key data
- Scale of SK hynix-NVIDIA partnershipUS$500B+The cooperation covers memory and the 2GW Vera Rubin DSX AI Factory for SK Telecom planned to come online in 2027.
- Additional reported memory supply partnership by SK GroupUS$250B, over the next 5 yearsThe report says SK Group will also work with other global technology companies to provide related memory supply.
- Scale of Samsung-Broadcom MOUUS$200B, over the next 5 years through 2030The scope includes memory, HBM, foundry, 2nm and below process technologies, and advanced packaging.
- Consensus expectation for memory industry revenueApproximately US$1.3T in both 2027 and 2028Used to help investors understand the context of the above partnership amounts within the overall market.
- Background of NVIDIA AI growth guidanceApproximately US$1T in aggregate for CY25-27, implying nearly US$500B next yearThe report believes the related memory partnerships may alleviate supply concerns for NVDA.
- Background of Broadcom AI growth guidanceUS$100B+ next yearThe report believes the related partnership helps support AVGO's AI growth trajectory.
- NVIDIA target priceUS$315Based on approximately 25x FY28/CY27 non-GAAP EPS estimate of US$12.52.
- Broadcom target priceUS$550Based on approximately 25x FY2027/28 pro-forma EPS average estimate of US$22.17.
Impact & implications
For investors, this report reinforces the theme that AI infrastructure investment is extending from GPUs and ASICs to memory, advanced packaging, and supply assurance. If AI chip demand continues to materialize, HBM and high-end DRAM suppliers may gain stronger pricing power and higher visibility; at the same time, companies with advanced packaging and foundry capabilities may also gain incremental opportunities. However, the report also warns that Chinese competition in NAND, slowing demand, oversupply, and regulatory restrictions may still weigh on valuations.
Risks
- The favorable pricing environment for memory may end early due to weaker demand or increased supply.
- Investor sentiment and the valuation multiples assigned by the market may change.
- China's progress in memory, especially NAND, poses downside risk.
- NVIDIA- and Broadcom-related names face risks from weaker-than-expected AI demand, customer migration to in-house chips, and technology export regulation.
- Broadcom also faces risks of losing share or sockets at key customers, insufficient execution of M&A synergies, inadequate cash returns, and management changes.
What to watch
- Formal orders, delivery pace, and the HBM4 and subsequent product technology roadmap in the SK hynix-NVIDIA partnership.
- Whether the Samsung-Broadcom MOU translates into meaningful revenue from AI ASIC wafers, advanced packaging, or HBM.
- Construction and commissioning progress of SK Telecom's 2GW Vera Rubin DSX AI Factory in 2027.
- Whether global memory industry revenue in 2027 to 2028 approaches the consensus expectation of approximately US$1.3T.
- The backlog situation for TSMC CoWoS capacity and customer adoption progress for Samsung Cube-S, Cube-E, and Cube-R.
- Capacity expansion, yield, pricing competition, and customer penetration of Chinese NAND manufacturers.