On July 25, 2026, Samsung Electronics and Broadcom announced a memorandum of understanding covering memory, foundry and advanced packaging. The companies estimate that the collaboration will exceed $200 billion over five years through 2030.
That number needs a qualification: the announcement describes an estimate attached to an MOU, not a disclosed binding purchase contract or guaranteed spending commitment. Samsung did not publish annual amounts, order volumes, capacity reservations, exclusivity terms or delivery schedules. What the announcement does provide is a clear technical scope spanning memory, logic manufacturing and multi-die integration.
On memory, Samsung says the companies plan to collaborate on high-bandwidth memory for Broadcom’s next-generation AI accelerators. Samsung’s HBM technical overview describes HBM as vertically stacked memory dies connected through through-silicon vias and a wide interface to provide high data throughput. The MOU does not identify an HBM generation, capacity, qualification status or shipment date for Broadcom.
The foundry portion covers Samsung’s 2-nanometer-and-below processes for Broadcom products, including wireless broadband communications solutions. Samsung’s current logic-node roadmap says SF2 entered stable volume production in late 2025 and schedules the performance-oriented SF2P process for mass production in late 2026. Those are Samsung roadmap statements; the MOU does not say which Broadcom designs will use SF2 or SF2P, where they will be fabricated, or when they will reach production.
The packaging work includes 2.3D and 2.5D integration built around Samsung’s 2nm process. Samsung’s advanced-packaging technical overview describes 2.5D Cube-S as a silicon-interposer architecture, while its 2.3D options use an embedded silicon bridge or redistribution-layer interposer. These approaches bring logic and memory closer within one package, but the announcement does not assign a particular package architecture to a Broadcom product.
AI accelerators increasingly depend on more than logic performance. Memory bandwidth, data movement, interconnects, power delivery, thermals and packaging all affect how much useful work a system can sustain. Bringing HBM, advanced process nodes and packaging into the same collaboration gives the companies a framework for coordinating those layers, although the public announcement does not disclose a finished architecture or measured system benefit.
Reuters reported that technology companies’ growing use of custom AI accelerators is increasing demand for specialised chip-design and manufacturing relationships. That is useful market context for Broadcom, which designs custom silicon, but it does not establish future demand, volumes or revenue from this MOU.
Reuters also assessed that long-term Broadcom work could improve utilisation at Samsung’s advanced manufacturing facilities. That is analysis, not a disclosed outcome. Actual utilisation will depend on signed orders, design progress, qualification, yields, production allocation and customer demand—none of which the companies quantified in the announcement.
The breadth of the MOU is still notable. It places memory supply, leading-edge foundry work and advanced packaging inside one five-year framework rather than presenting them as unrelated component deals. If that framework becomes production business, Samsung could participate in more layers of Broadcom’s silicon supply chain. The announcement alone does not show how much of the estimated value belongs to each layer or whether all contemplated work will proceed.
The next meaningful evidence will be more specific than the headline estimate: definitive supply or manufacturing agreements, named Broadcom platforms, HBM qualification milestones, confirmed process assignments, selected packaging variants and production disclosures. Samsung’s late-2026 SF2P schedule is another checkpoint, but it should not be treated as a Broadcom delivery date.
For now, the MOU signals an ambitious expansion of the relationship across three technically connected parts of AI infrastructure. It does not yet provide enough information to judge volumes, yields, performance, capacity, delivery timing or financial returns.
