伯恩斯坦解读：7000亿美元合作背后，AI最缺的可能不是GPU

TL;DR

• SK Group, NVIDIA, Samsung, and Broadcom have successively announced AI supply chain collaborations, with the total framework scale exceeding $700 billion.
• The collaborations cover HBM4, AI factories, sub-2nm foundry services, and advanced packaging, primarily aimed at securing AI server supply in advance.
• These arrangements are mainly based on partnerships and MOUs, with actual purchase volumes, prices, and delivery schedules yet to be disclosed.

The AI supply chain is extending from a race for GPUs to securing memory and advanced packaging in advance.

On July 24th, SK Group and NVIDIA announced an expanded comprehensive collaboration at the AI Summit in San Francisco, with a total scale exceeding $500 billion, covering AI factories and next-generation memory. Subsequently, Samsung Electronics signed an MOU with Broadcom, anticipating a collaboration scale exceeding $200 billion over the next five years until 2030 in the memory and foundry sectors.

In a July 27th report, Bernstein focused on memory rather than another round of AI chip orders. AI server expansion relies on more than just GPUs. For computing chips like GPUs and ASICs to effectively enter data centers, they require HBM, DRAM, NAND, advanced packaging, and stable delivery. If these links fall behind, the shipments of AI chip companies like NVIDIA and Broadcom could be slowed down.

The two massive figures also need to be examined closely. The 'over $500 billion' is not a memory procurement contract solely for SK hynix. It is a comprehensive partnership between SK Group and NVIDIA, where SK hynix takes responsibility for HBM4 and next-generation memory collaboration, while SK Telecom is the primary entity for the AI factory. Samsung and Broadcom's $200 billion MOU has a clearer scope, covering HBM and other memories, support for next-generation AI accelerators, 2nm-related manufacturing, and 2.3D/2.5D integration.

NVIDIA Locks In HBM4, Broadcom Negotiates Memory, Foundry, and Packaging Together

The SK Group and NVIDIA collaboration places AI data centers and next-generation memory within the same framework.

According to official information from NVIDIA and SK hynix, the collaboration includes SK Telecom's up to 2GW AI factory, NVIDIA DSX, the Vera Rubin platform, and SK hynix's HBM4. The first AI factory is planned to go live in 2027.

The market implication of this is quite direct. AI systems beyond Vera Rubin will continue to increase demands on HBM and system-level supply. Memory is no longer just a standardized procurement item but a part of whether the AI chip platform can be delivered on schedule.

Samsung and Broadcom's MOU is more akin to a bundled supply chain solution. Samsung official stated that the collaboration scale between the two parties over the next five years until 2030 exceeds $200 billion, encompassing memory and foundry services, supporting Broadcom's next-generation AI accelerators, and involving 2nm process-related advanced packaging, 2.3D, and 2.5D integration.

This goes beyond just whether Broadcom will entrust some of its AI ASICs to Samsung for foundry. Samsung aims to package memory, advanced processes, and advanced packaging for AI chip clients, striving to offer an alternative deliverable solution alongside TSMC.

In technical terms, Samsung's Cube-S, Cube-E/R fall under 2.5D/2.3D Cube packaging. Bernstein draws an analogy with TSMC's CoWoS-S/L/R, the core purpose being to solve the high-bandwidth connection and multi-chip integration between logic chips and HBM.

A comparison of Samsung's Cube-S, Cube-E, Cube-R with TSMC's CoWoS-S/L/R, illustrating how 2.5D and 2.3D packaging connect HBM with logic chips.

Memory Annual Revenue Eyes ~$1.3 Trillion, AI Clients Start Queuing Up Early

Bernstein believes the most noteworthy aspect of these announcements is that memory supply is being locked in ahead of time.

In the expansion of AI servers, computing chips themselves are certainly important, but the supply elasticity of HBM and high-end memory is smaller, customer qualification cycles are longer, and advanced packaging can also become a bottleneck for full system delivery. Major clients are unwilling to wait until capacity becomes tight to enter the spot market; instead, they secure future supply through multi-year framework agreements.

The market consensus cited in the report indicates global memory industry annual revenue is approximately $0.9 trillion in 2026, and around $1.3 trillion in both 2027 and 2028. TrendForce's previous public forecast also anticipated the global memory market size to be about $1.28 trillion in 2027. Key players include Samsung, SK hynix, Micron, KIOXIA, and Chinese memory manufacturers.

Bar chart of consensus revenue for global memory vendors, approximately $0.9 trillion in 2026, and about $1.3 trillion in both 2027 and 2028.

These figures cannot be directly equated to incremental revenue generated by the collaborations between SK Group, NVIDIA, Samsung, and Broadcom. The announcements did not disclose specific purchase volumes, product mix, pricing formulas, delivery schedules, nor did they specify how much of the framework value comes from new capacity versus extending existing collaborations.

However, the signal is clear enough. AI clients are treating HBM and advanced packaging as strategic resources, and the pricing power of leading memory manufacturers is consequently under continuous market scrutiny. Traditional memory cycles were more influenced by PC, mobile, and server inventory. Under the AI cycle, HBM is deeply integrated with advanced packaging, making supply-demand dynamics more likely determined by a few large clients and a few key suppliers.

Samsung Aims for Packaging and Foundry, TSMC's Impact Seen as Muted

With Samsung and Broadcom's MOU covering sub-2nm processes, advanced packaging, and AI accelerators, the market naturally wonders about the impact on TSMC's market share.

Bernstein's assessment is relatively cautious. Even if Broadcom shifts some AI ASIC production to Samsung in the future, the near-term impact on TSMC's earnings may be limited because demand for advanced capacity remains very high. Samsung gaining more qualification opportunities from AI clients does not automatically mean TSMC loses orders immediately.

The real test for Samsung is its delivery capability. AI chip clients need not just point manufacturing capabilities, but stability across HBM, logic chips, packaging, substrates, yield rates, and delivery timelines simultaneously. If any link is delayed, the technology roadmap outlined in the MOU will struggle to translate into actual shipments.

Structural diagram of TSMC's three CoWoS variants, showing differences in interposers, RDL, and HBM stacking for CoWoS-S, CoWoS-L, and CoWoS-R.

For Samsung, this is an opportunity to extend its memory strengths towards the AI system-level supply chain. Samsung is large-scale in the memory space, but in terms of HBM leadership and trust from advanced foundry clients, the market has historically focused more on SK hynix and TSMC. If the long-term framework with Broadcom executes smoothly, Samsung can re-emphasize its 'Memory + Foundry + Packaging' combined capability.

Ratings Tilt Positive, but MOUs Are Not Orders Yet

In its report, Bernstein maintained Outperform ratings for Samsung Electronics, SK hynix, Micron, NVIDIA, and Broadcom, while rating KIOXIA as Underperform. Regarding target prices, it set Samsung common stock at KRW 440,000, SK hynix at KRW 3,300,000, NVIDIA at $315, and Broadcom at $550. The relevant ratings and target prices are based on the source report's methodology.

Table of key stock ratings and target prices: Samsung KRW 440,000, SK hynix KRW 3,300,000, NVIDIA $315, Broadcom $550, all rated Outperform.

This set of ratings does not mean all collaborations have been converted into confirmed performance. A more prudent understanding is that AI memory supply is becoming more strategic. Leading memory manufacturers and major AI chip clients are using long-term arrangements to mitigate future supply risks.

The risks are also clear. MOUs and partnerships are not final procurement contracts; prices, volumes, and delivery timelines remain undisclosed. Whether the Samsung-Broadcom collaboration yields significant AI ASIC foundry or advanced packaging share depends on customer qualifications, yield rates, and capacity arrangements. Even with a large overall value, it is difficult to gauge its incremental contribution relative to the ~$1.3 trillion annual memory market projected around 2027.

Longer-term pressure comes from the dynamics of the storage industry. Progress in China's storage industry, especially competition in the NAND segment, will impact industry profit margins. DRAM and HBM face higher barriers due to EUV, specific processes, and customer qualification, so near-term pressure is relatively lower. However, the current supply tightness cannot be simply extrapolated as a permanent advantage.

The most definitive signal from these collaborations is not that "$700 billion in orders have landed," but that AI giants are beginning to lock up memory and packaging resources through multi-year frameworks. Whether this materializes ultimately depends on capital expenditure, HBM shipments, customer prepayments, capacity expansion, and actual delivery cadence reported in financial results.