Samsung and Broadcom Just Signed the Largest Semiconductor Deal of 2026

Broadcom has spent the last three years quietly becoming one of the most important companies in AI infrastructure — not by building AI models or running cloud platforms, but by designing the custom silicon that some of the largest AI workloads in the world run on. Its application-specific integrated circuits power the TPUs that Google trains its models on, the AI accelerators that Meta deploys at scale, and a growing list of hyperscaler customers who want silicon tuned to their specific workload rather than a general-purpose GPU. That business model makes Broadcom’s supply chain decisions extraordinarily consequential. On July 24, 2026, the company made the most consequential one in its history: a $200 billion-plus partnership with Samsung Electronics, covering everything from the memory that feeds its AI chips to the factories that produce them.

The memorandum of understanding, signed at an AI Summit in San Francisco and confirmed by Samsung’s official newsroom, spans five years through 2030. Its scope — HBM4 and HBM4E memory supply, 2-nanometer foundry production at Samsung’s Pyeongtaek campus, and 2.3D and 2.5D advanced packaging — makes it a turnkey manufacturing commitment rather than a single supply agreement. Samsung will, in effect, be responsible for producing the complete physical package from which Broadcom’s next-generation AI accelerators emerge.

What the Deal Covers

Samsung confirmed the agreement on its Global Newsroom, describing it as establishing ‘a turnkey system spanning HBM4 and HBM4E, 2-nanometer foundry, and advanced packaging to support production of next-generation AI semiconductors.’ Han Jin-man, President and Head of Samsung’s Foundry Business, and Broadcom CEO Hock Tan both attended the signing ceremony in San Francisco alongside senior executives and Korean government representatives. Jun Young-hyun, Vice Chairman and Head of Samsung’s Device Solutions division, framed the deal as the product of a relationship that has been building since 2007: ‘Our partnership with Broadcom has been built on decades of trust and mutual success, and this strategic expansion marks a new chapter in our shared journey to accelerate next-generation AI innovation.’

The memory component of the deal centres on two HBM generations. HBM4, Samsung’s sixth-generation high-bandwidth memory product, began mass production as a world first in February 2026, built on 10-nanometer-class 1c DRAM and a 4nm base die. It delivers over 200 percent more bandwidth than previous generations, shrinks the die area by up to 70 percent, and cuts power consumption by half compared to planar designs — numbers that matter enormously inside an AI accelerator where memory bandwidth and power efficiency are primary performance constraints. HBM4E, the seventh-generation product for which Samsung supplied samples first in the industry in May 2026, extends that performance further: 16Gbps pin speed, 16-layer stacking, and 64GB per stack capacity. The foundry component covers 2nm and below processes for Broadcom’s AI chips and Wireless Broadband Communications solutions, with production planned at Samsung’s Pyeongtaek campus in South Korea.

What Advanced Packaging Adds

2.3D and 2.5D Integration

Advanced packaging is the element of this deal that makes it genuinely different from a conventional memory supply agreement. 2.5D integration — placing memory and logic dies side by side on a silicon interposer, connected by extremely dense interconnects — is the architecture that allows HBM’s thousand-plus pins per stack to connect to a processor die at the bandwidth rates AI workloads require. 2.3D integration, a more recent development, goes a step further by directly bonding memory to logic dies through copper-to-copper hybrid bonding, achieving even higher connection density and lower power consumption than interposer-based 2.5D approaches. By covering both integration schemes in the partnership, Samsung positions itself to package Broadcom’s AI chips using whichever architecture best suits the specific performance, power, and cost requirements of each product generation rather than being locked into a single approach. For Broadcom, whose custom AI ASICs are already among the most performance-sensitive chips in production, this packaging flexibility is as commercially important as the underlying process node.

Why Turnkey Changes the Dynamic

Samsung’s framing of the deal as a ‘turnkey system’ — connecting chip design, HBM supply, foundry production, and advanced packaging in one relationship — is deliberate and significant. The conventional model for AI chip production involves separate vendors for each layer: a memory supplier for HBM, a foundry for the logic die, and a packaging house for integration. Each handoff between vendors is a potential source of delay, yield variation, and communication friction. A turnkey approach consolidates those layers under one relationship, allowing Samsung and Broadcom to optimise chip design and manufacturing from the earliest design stages, shortening development timelines and improving overall yield. That development compression matters commercially: the AI accelerator market moves fast enough that a chip that arrives six months earlier than a competitor’s has a meaningful revenue advantage over its entire production lifetime.

Why Broadcom and Why Now

Broadcom’s Position in Custom AI Silicon

Broadcom’s AI accelerator business has grown from a niche to one of the most structurally important segments in the AI hardware market. Google’s Tensor Processing Units, the AI chips that handle a significant fraction of the world’s AI training and inference at scale, are designed by Google but manufactured in a relationship where Broadcom provides key intellectual property and design expertise. Meta’s custom AI accelerator programme, Amazon’s Trainium and Inferentia lines, and a growing list of undisclosed hyperscaler clients all involve Broadcom-designed silicon. The common thread is that these companies want AI chips tuned specifically to their workloads — their model architectures, memory access patterns, and inference serving requirements — rather than general-purpose GPU designs that accommodate a broader range of use cases. Broadcom is the company that designs those custom chips at the highest scale in the industry, which means its HBM and foundry decisions are proxies for the infrastructure plans of the hyperscalers whose names do not appear in the supply agreements.

Samsung’s Foundry Recovery Play

For Samsung, the Broadcom deal is the flagship example of what its foundry chief Jun Young-hyun recently described as the business being ‘primed for a great leap forward.’ Samsung Foundry has trailed TSMC in both process technology maturity and customer roster for the past four years, losing Apple’s most advanced chip orders, struggling with yield issues at 3nm, and watching NVIDIA deepen its exclusive TSMC relationship. A $200 billion MOU with one of the most important custom AI chip designers in the industry — covering 2nm and below processes — provides both revenue visibility and the production volume scale that drives process technology maturity on new nodes. Combined with a previously reported $16.5 billion agreement with Tesla for chip manufacturing and the broader context of Samsung’s established HBM partnership with NVIDIA through SK Hynix and competing supply agreements, the Broadcom deal positions Samsung’s foundry business with enough committed demand to justify the capital expenditure required to bring 2nm processes to yield-capable mass production.

The Memory Shortage Backdrop

The deal arrives during what industry analysts are describing as a severe structural memory shortage. TrendForce has reported that contract DRAM prices rose 90–95 percent quarter-on-quarter in early 2026, with NAND flash up more than 50 percent, driven by AI data-center demand and the shift of wafer capacity toward premium products like HBM. Samsung and SK Hynix have both publicly cautioned that AI-driven HBM shortages could persist into the next decade. In that supply environment, a long-term committed supply agreement with a defined production plan — covering two HBM generations across five years — provides Broadcom with supply certainty that is commercially valuable independent of whatever pricing mechanisms the MOU specifies. The deal also locks Samsung into producing increasingly complex HBM variants for a customer whose roadmap it now knows in detail, which shapes Samsung’s own HBM capital allocation decisions for the duration. The broader AI infrastructure energy and resource demand trajectory that is driving HBM demand is not self-correcting: as model sizes and inference throughput requirements grow, the memory bandwidth demands on AI accelerators grow proportionally, which means HBM’s role as the enabling constraint in AI infrastructure is deepening rather than easing.

Geopolitical Context: Korea’s $950 Billion Semiconductor Push

The Samsung-Broadcom deal does not sit in isolation. The Korean government has framed the broader set of semiconductor cooperation agreements being signed at and around the San Francisco AI Summit — of which Samsung-Broadcom is the largest single example — as part of a roughly $950 billion total semiconductor cooperation push with global technology firms. That framing reflects a deliberate Korean government strategy to secure long-term semiconductor manufacturing relevance as the US-China technology competition intensifies. Korea’s position — home to Samsung and SK Hynix, the world’s two largest memory chip producers, and to Samsung’s foundry operation — makes it structurally important to the AI hardware supply chain regardless of geopolitical tensions. The concentration of agreements at San Francisco in the same week as Korea’s President Lee Jae-myung attended the AI Summit confirms that the semiconductor partnership announcements are as much a diplomatic and strategic exercise as a commercial one.

Reactions

Hock Tan, Broadcom’s CEO, described the partnership as enabling Broadcom to ‘deploy our accelerators more effectively than ever, with better performance and greater energy efficiency.’ The framing is commercially precise: Broadcom’s competitive position in the custom AI ASIC market depends on its chips delivering better performance-per-watt than general-purpose GPUs for the specific workloads its hyperscaler customers run. A turnkey partnership with Samsung that covers both HBM generation transitions and 2nm foundry production gives Broadcom a cleaner path to achieving that differentiation at each product generation than a fragmented multi-vendor supply chain would allow.

What Happens Next

The MOU’s five-year duration through 2030 spans at least two more HBM generations beyond what has been announced: HBM4E, which Samsung has already produced samples of, and the as-yet-unannounced HBM5 generation that industry roadmaps suggest will be required for the AI accelerator products of 2028 and beyond. Samsung’s commitment to 2nm and below processes for Broadcom’s chips also implies that 1.4nm technology — Samsung’s next planned node after 2nm — will be under serious consideration for Broadcom’s post-2028 product generations within this partnership framework. The specific production timelines, yield commitments, and pricing structures within the MOU have not been disclosed; those details are where the commercial substance of a partnership of this scale either delivers or disappoints.

Why It Matters

The Samsung-Broadcom deal matters at the level of the global AI hardware supply chain rather than just as a bilateral commercial agreement. Broadcom’s custom AI ASIC business is the mechanism through which the AI compute infrastructure of the largest technology companies in the world — Google, Meta, Amazon, and their peers — gets built. A committed, five-year, turnkey supply relationship between Broadcom and Samsung means that a significant fraction of the AI compute that will power the world’s largest AI workloads from now through 2030 has a defined production pathway: Samsung HBM at 4nm base die, manufactured on Samsung’s 2nm process, assembled using Samsung’s advanced packaging. The deal will not eliminate Samsung’s competitive challenges in foundry relative to TSMC, but it provides the committed demand and the technical roadmap alignment that give those challenges a credible path to resolution.

Sources

Samsung Electronics Global Newsroom (news.samsung.com), July 24, 2026. Seoul Economic Daily (en.sedaily.com), July 25, 2026. SammyFans, July 25, 2026. The Week India, July 25, 2026. Korea Economic Daily via Bloomingbit, July 25, 2026.