Bernstein has reiterated its positive outlook for memory chipmakers following a set of large-scale agreements announced at an AI Summit hosted by the South Korean government in San Francisco on Friday. The accords, disclosed by Samsung, SK Hynix, Nvidia and Broadcom, span memory supply, foundry services and AI data-center buildouts.
At the center of the headlines was a letter of intent between SK Hynix and Nvidia describing a partnership valued at more than $500 billion. The agreement covers memory supply and facilities tied to AI data centers, including a 2GW Vera Rubin DSX AI Factory that is planned to be operational for SK Telecom in 2027.
Bernstein analysts, led by Mark Li, highlighted the SK Hynix-Nvidia arrangement as containing "long-term technical developments & stable supply of next-gen AI memory" meant to help SK Hynix "expand the foundation for growth." The firm also noted reports that SK Group is seeking roughly $250 billion more in memory supply collaborations with other global technology companies over the next five years.
Separately, Samsung and Broadcom signed a memorandum of understanding that Bernstein said covers about $200 billion of memory, including high-bandwidth memory (HBM), plus foundry services through 2030. The foundry portion will concentrate on 2-nanometer and below process nodes and advanced packaging approaches that analysts likened to TSMC’s Chip-on-Wafer-on-Substrate (CoWoS) technology.
Bernstein's team described the wave of announcements as reinforcing memory's increasing importance within AI infrastructure. The analysts wrote that they "believe the announced amounts are more for memory and indicate the need for NVIDIA & Broadcom to secure memory supply."
On quantification, the analysts acknowledged limits: they said they "can’t precisely quantify the impact of the announcements," while noting consensus forecasts that show roughly $1.3 trillion in annual memory revenue in both 2027 and 2028.
Regarding potential effects on leading foundry TSMC, Bernstein judged any impact to be "negligible" for now. The analysts pointed out uncertainty over whether Broadcom will actually manufacture AI ASICs at Samsung and highlighted the existing queue of demand for TSMC’s capacity.
Overall, Bernstein maintained a constructive stance on the memory sector and characterized the recent pullback in memory stocks as "a good entry point." The firm emphasized that memory may be more critical to AI than logic semiconductors. At the same time, Bernstein flagged a longer-term competitive threat from China in NAND products, which led the broker to assign an Underperform rating to Kioxia. The analysts qualified that the China-related risk appears substantially smaller in DRAM, where they said China would likely struggle to compete without access to EUV lithography.
In its ratings, Bernstein lists Samsung, SK Hynix and Micron as Outperform. The analysts also noted the new partnerships could help ease supply concerns for Nvidia and Broadcom, both of which have guided toward substantial AI-related growth.
Implications for markets and infrastructure
- Large contractual commitments for memory capacity underscore how integral memory supply is becoming to AI systems and data-center planning.
- Foundry commitments tied to advanced nodes and packaging may affect capacity planning across the semiconductor supply chain, though Bernstein views immediate impacts on TSMC as limited.
- Investor sentiment toward memory names strengthened with the brokerage's reiteration of an overweight stance across several major producers.
The announcements at the San Francisco summit tie commercial memory supply directly to AI deployment plans and foundry roadmaps. For real assets tied to AI data centers, the agreements suggest more defined supply lines for memory modules and possibly greater clarity around capacity used to meet AI workloads. Yet the analysts' caution about quantifying the precise market effects and the identified geopolitical competitive risks in NAND reflect ongoing uncertainties that market participants will watch closely.