Global memory leaders and chip designers signed multi-hundred-billion-dollar long-term agreements at a Korean government-hosted AI summit in San Francisco, signaling the AI computing supply chain is shifting from on-demand procurement to ultra-long-term capacity locking.
SK hynix signed a letter of intent with NVIDIA valued at over $500 billion, covering memory supply and a planned 2GW Vera Rubin DSX AI factory project for SK Telecom set to launch in 2027.
Samsung Electronics reached a $200 billion strategic memorandum with Broadcom, locking in high-bandwidth memory (HBM) supply and advanced foundry services below 2 nanometers through 2030.
Following the announcement of these deals, prominent firm Bernstein reaffirmed its bullish stance on the memory sector, maintaining outperform ratings on Samsung Electronics, SK hynix, and Micron Technology, and noted that recent sector pullbacks offer a favorable entry point.
Two Mega Deals Locking Long-Term Capacity
The letter of intent between SK hynix and NVIDIA is the largest single agreement disclosed at the summit.
Bernstein analysts, led by Mark Li, pointed out that the memory partnership covers long-term technology development and stable supply of next-generation AI memory, aiming to help SK hynix expand its growth base.
The SK Group also plans to seek an additional $250 billion in global memory supply cooperation over the next five years, further solidifying its core position in the AI memory supply chain.
The memorandum between Samsung and Broadcom covers both memory (including HBM) and foundry. The foundry portion will focus on 2nm and more advanced process technologies, utilizing advanced packaging solutions that analysts compare to TSMC's CoWoS.
The agreement extends through 2030, providing ultra-long-term demand visibility for Samsung's foundry business.
Bernstein: Memory Is More Important to AI Than Logic Chips
Bernstein explicitly stated in its report that the disclosed amounts mainly point to the memory segment, reflecting the urgent need for NVIDIA and Broadcom to secure memory supply.
Analysts provided a reference data point: market consensus expects global memory annual revenue to be around $1.3 trillion in both 2027 and 2028. While Bernstein said it cannot precisely quantify the impact of these agreements, the magnitude of the numbers relative to industry revenue expectations further supports the critical weight of memory chips in AI infrastructure.
Bernstein believes memory chips are more important to AI than logic semiconductors. Based on this judgment, the firm maintains outperform ratings on Samsung, SK hynix, and Micron.
For TSMC, Bernstein views the impact of these agreements as negligible, as it is uncertain whether Broadcom will actually produce AI ASICs at Samsung, and TSMC's capacity demand queue remains full.
In the NAND field, long-term competitive pressure from China cannot be ignored. Based on this assessment, Bernstein gives Kioxia an underperform rating.
Deals Ease Supply Anxiety for NVIDIA and Broadcom
For the two chip designers, NVIDIA and Broadcom, downstream AI-related growth guidance is strong, but supply anxiety over upstream memory capacity is equally real.
Bernstein noted that the new partnerships help alleviate supply-side concerns for these two companies.
The agreements disclosed at the summit essentially represent a form of risk hedging at the supply chain level—AI computing demand parties lock in future capacity through ultra-long-term contracts, thereby eliminating supply uncertainty of key materials in their expansion strategies. For investors, this trend means the revenue visibility of memory giants is undergoing a qualitative improvement.
Comments