In a move that reshapes the landscape of AI semiconductor manufacturing, Samsung Electronics and Broadcom have signed a memorandum of understanding (MOU) valued at more than $200 billion. The agreement, announced during an AI summit in San Francisco, spans memory chips, foundry production, and advanced packaging through 2030. It is among the largest semiconductor partnerships ever disclosed, though as an MOU it represents a statement of intent rather than a binding contract. However, the scale and scope signal a strategic realignment in the AI chip supply chain, which has been heavily dominated by a single manufacturer.
The deal covers three critical layers of AI chip production: memory, logic manufacturing, and packaging. On the memory side, Samsung will supply Broadcom with next-generation high-bandwidth memory (HBM), specifically HBM4 and HBM4E. These are the latest iterations of HBM, a technology that stacks memory dies vertically and connects them through a silicon interposer to achieve extremely high bandwidth. HBM is essential for AI accelerators because large language models and other neural networks require massive amounts of data to be fed into processors rapidly. Without HBM, the computational throughput of accelerators would be severely bottlenecked. Samsung currently holds a significant share of the HBM market, competing closely with SK Hynix, which has been the primary supplier for Nvidia's GPUs. By locking in Broadcom as a customer for HBM4 and HBM4E, Samsung secures a long-term revenue stream in a market that is projected to grow from roughly $10 billion in 2024 to over $50 billion by 2030, according to industry estimates.
The foundry component of the agreement is equally significant. Samsung will manufacture Broadcom's custom AI chips using two-nanometer and smaller process technologies at its advanced facility in Pyeongtaek, South Korea. Two-nanometer (2nm) refers to the transistor gate length, a key measure of process technology sophistication. Smaller nodes allow more transistors to be packed onto a chip, improving performance and energy efficiency. Samsung had announced its 2nm process, called SF2, with plans for mass production in 2025. However, the company has struggled with yield rates—the percentage of functional chips per wafer—on its most advanced nodes. Low yields have historically deterred major customers like Nvidia and AMD, who instead rely on Taiwan Semiconductor Manufacturing Company (TSMC) for leading-edge production. Broadcom's commitment, valued at billions of dollars in foundry revenue over five years, could help Samsung refine its process and demonstrate improved yields to the broader market.
Advanced packaging forms the third pillar of the MOU. Packaging involves connecting the memory, logic, and other components into a single module. In AI chips, this is increasingly critical because memory and logic must be placed as close together as possible to minimize data travel distance, reducing latency and power consumption. Samsung offers a range of packaging technologies, including its proprietary I-Cube and X-Cube 3D stacking solutions, as well as the more advanced Interposer-Based High-Bandwidth Memory Integration. These techniques allow multiple HBM stacks to sit alongside a logic die on a single silicon interposer, creating a giant system-on-chip (SoC). As AI models grow larger, the demand for such integrated packaging solutions is skyrocketing. Broadcom, which designs custom chips for hyperscale cloud providers, requires cutting-edge packaging to differentiate its products. By collaborating with Samsung on packaging, Broadcom gains access to a second source of advanced packaging capacity, reducing its reliance on TSMC's CoWoS (Chip-on-Wafer-on-Substrate) technology, which is currently oversubscribed.
Broadcom designs custom AI accelerators for hyperscalers such as Google, Meta, and Amazon. These companies prefer tailor-made chips over buying off-the-shelf GPUs from Nvidia because they can optimize the architecture for their specific workloads, achieving better performance per watt and lower total cost of ownership for massive data center deployments. Broadcom's custom chip business has grown rapidly; in April 2025, the company extended its partnership with Meta through 2029. CEO Hock Tan has noted that surging AI revenue has made organic growth more attractive than acquisitions. Diversifying manufacturing to Samsung gives Broadcom supply chain resilience at a time when geopolitical tensions have heightened concerns about TSMC's concentration in Taiwan. TSMC currently controls more than 90% of leading-edge chip production below 7nm, and any disruption could cripple the global AI supply chain. The Samsung partnership provides an alternative, geographically dispersed source for the most advanced manufacturing processes.
For Samsung, the deal addresses a persistent credibility gap. The company's foundry division has struggled to gain traction against TSMC, holding roughly 7% market share in the fourth quarter of 2024. Samsung has invested heavily in its foundry business, spending over $100 billion in the past decade to build new fabs and develop advanced nodes, but yields and customer trust have lagged. The Broadcom MOU, if executed, would become one of the largest customer commitments in Samsung's foundry history. It could shift investors' perception of Samsung's competitiveness at the leading edge. In the second quarter of 2025, Samsung Electronics posted a record operating profit of 89 trillion won (approximately $66 billion), driven almost entirely by AI memory demand. The foundry segment, however, contributed a relatively small portion of that profit. Analysts have long argued that Samsung must secure a marquee customer like Broadcom to validate its process technology and bring in the revenue needed to continue R&D at a scale comparable to TSMC.
The MOU was signed alongside a broader package of semiconductor deals worth roughly $950 billion between South Korean and American companies, announced during the AI summit hosted by South Korean President Lee Jae Myung. SK Hynix signed memory partnerships worth $750 billion with Nvidia and others, and AI startup Anthropic signed supply agreements with both Samsung and SK Hynix. These agreements underscore the enormous scale of investment flowing into AI infrastructure. They also reflect the strategic alliance between South Korea and the United States to secure supply chains for critical technologies. The summit follows Seoul's $880 billion domestic investment plan unveiled in June 2025, which committed Samsung and SK Hynix to building four new chip fabrication plants in the country's southwest. That plan aims to turn South Korea into a global semiconductor manufacturing hub, reducing dependency on a few regions and ensuring that AI chip production can meet exploding demand.
From a technical standpoint, the HBM4 and HBM4E memories that Samsung will supply represent a generational leap. HBM4 is expected to double the bandwidth per stack compared to HBM3, reaching speeds of up to 2 Terabytes per second. It will use a 2048-bit interface, up from 1024-bit in HBM3, and move to a more advanced memory cell technology. HBM4E is an even faster enhanced version, likely arriving in 2027. These memories are fabricated using extreme ultraviolet (EUV) lithography and advanced chip stacking that requires a combination of through-silicon vias (TSVs), microbumps, and hybrid bonding. Samsung's experience in memory manufacturing gives it an edge in producing HBM at scale, but the yield challenge is shared across all manufacturers. The Broadcom deal incentivizes Samsung to ramp up HBM capacity aggressively; the company has already announced plans to triple HBM production by 2027 to meet customer demand.
On the foundry side, Samsung's 2nm process (SF2) is based on a gate-all-around (GAA) transistor architecture, which replaces the previous FinFET design. GAA offers better electrostatic control and lower leakage, enabling higher performance at lower power. Samsung was the first to mass-produce GAA transistors, starting with its 3nm node in 2022. However, early yields were low, and TSMC's competing 3nm FinFET process won most customer orders. With 2nm, both companies are using GAA, and Samsung has a year of learning ahead. The Broadcom partnership could accelerate that learning by providing a steady stream of design tape-outs and production runs, helping Samsung identify and fix yield-killing defects. Broadcom, meanwhile, gets access to a process that may eventually offer performance and cost advantages if Samsung's yields improve.
The partnership also includes co-development of advanced packaging solutions. Samsung's packaging division has invested heavily in 3D IC and hybrid bonding technologies that allow multiple chips to be integrated vertically with extremely fine pitch connections. This is crucial for AI accelerators because it enables the memory to be stacked directly on top of the logic die, reducing the distance data must travel. For example, Broadcom's next-generation accelerator could use a 3D stack of HBM4E memory on top of a 2nm logic die, all packaged together in a single module. Samsung's advanced packaging plant in Cheonan, South Korea, recently opened a dedicated line for high-volume manufacturing of such modules. The Broadcom deal will likely utilize that capacity, providing a further boost to Samsung's foundry-ecosystem integration.
Market reactions to the MOU have been positive. Samsung's stock rose 3% in trading on the Korea Exchange following the announcement, while Broadcom shares gained 2% on the Nasdaq. Analysts have noted that the deal is non-binding, but the sheer size of the commitment—$200 billion—lends it considerable weight. Breaking such an agreement would damage both companies' reputations and foreclose future business opportunities. Moreover, the deal is structured with milestones and technology roadmaps that will be updated annually. Broadcom has a track record of honoring long-term MOU with manufacturing partners; its relationship with TSMC spans decades, and the Samsung agreement is likely to be similarly durable.
The MOU also highlights shifting dynamics in the AI chip market. Nvidia currently dominates AI computing with its GPUs, but companies like Broadcom, Marvell, and a host of startups are designing custom accelerators that offer higher efficiency for specific tasks. As the market matures, these custom chips are expected to take a larger share. The total addressable market for AI semiconductors is projected to exceed $400 billion by 2030, and Broadcom's custom chip business alone could generate $30 billion in annual revenue. Securing manufacturing capacity for that growth is essential, and Samsung's offer of comprehensive memory, logic, and packaging services simplifies Broadcom's supply chain. Instead of managing multiple vendors for memory, logic, and packaging, Broadcom can work with Samsung as a one-stop shop, reducing complexity and time-to-market.
Looking ahead, the partnership will face challenges. Samsung must deliver on its technology promises, especially in yield improvement for 2nm and advanced packaging. Broadcom will need to manage the transition from design to high-volume manufacturing without significant delays. Both companies operate in a fast-paced industry where any setback can erode competitive advantage. However, the alignment of interests is clear: Samsung needs a flagship foundry customer, Broadcom needs a reliable second source, and the AI industry needs more manufacturing capacity to avoid bottlenecks. The $200 billion MOU is a bet that these needs will converge into a long-term symbiotic relationship.
In the broader context, this deal is part of a wave of investments in semiconductor capacity outside Taiwan. The United States has channelled $52 billion in subsidies through the CHIPS Act, but that has not yet resulted in leading-edge fabs. Samsung itself is building a $17 billion facility in Taylor, Texas, which will produce 4nm and 3nm chips starting in 2025. However, the Broadcom deal centers on Samsung's Korean fabs. South Korea has also grown as a hub for advanced semiconductor packaging, with Samsung and SK Hynix expanding their packaging capacities. The global chip supply chain is gradually becoming more multi-polar, reducing the risks of a single point of failure. The Samsung-Broadcom MOU accelerates this trend by tying together two industry giants in a multi-year collaboration that spans the entire AI chip production spectrum.
With the summit concluding, the focus shifts to execution. Samsung and Broadcom will establish a joint steering committee to oversee progress, with quarterly reviews. The first HBM4 samples are expected in early 2026, with volume production by mid-2027. The 2nm tape-outs for Broadcom's design are slated for the second half of 2025. If these milestones are met, the partnership could redefine the competitive dynamics in the semiconductor industry and provide a new blueprint for AI chip manufacturing alliances.