Samsung and Broadcom Sign $200 Billion AI Chip Deal, Challenging TSMC's Dominance

Samsung Electronics and Broadcom announced on Friday a sweeping memorandum of understanding worth more than $200 billion — spanning next-generation memory, advanced foundry manufacturing, and chiplet packaging — that industry analysts are calling the most significant challenge to TSMC’s grip on AI silicon in a decade.

The deal, signed at an AI summit hosted aboard The Midway in San Francisco during a South Korean presidential visit led by President Lee Jae Myung, runs through 2030 and covers three interlocking pillars of the AI chip supply chain. It is also part of a broader wave: across the two-day summit, South Korean semiconductor firms secured approximately $950 billion in total cooperation agreements with American technology companies.

What the Deal Actually Covers

The MOU is structured around three commitments. First, Samsung will supply Broadcom with HBM4 and HBM4E — the next-generation high-bandwidth memory stacks that power AI accelerators. Samsung shipped the world’s first commercial HBM4 in February 2026 and has already sold out its entire 2026 production capacity, underscoring how tight the advanced memory market has become.

Second, Broadcom’s future networking and AI inference chips will be manufactured using Samsung’s sub-2nm process technology at its massive Pyeongtaek, South Korea fabrication campus. This is the most scrutinized element of the agreement. Samsung’s 2nm yield rates have climbed sharply — from roughly 20% in the second half of 2025 to above 60% today — but remain below the approximately 65-70% threshold that TSMC routinely achieves, and that customers typically demand before committing high-volume production.

Third, the agreement covers advanced 2.5D and 2.3D chiplet packaging, integrating memory and logic dies closer together to reduce energy consumption and improve bandwidth — a critical requirement for trillion-parameter AI models.

Why Broadcom Needs This

Broadcom designs custom AI silicon for hyperscalers including Google (Tensor Processing Units) and Meta (Meta Training and Inference Accelerator). Until now, virtually all of this cutting-edge work has flowed through TSMC, which controls an estimated 90% or more of global leading-edge semiconductor production.

That concentration has created real procurement anxiety in Silicon Valley. When TSMC’s advanced capacity is fully allocated — as it was through much of 2025 and into 2026 — customers have little leverage on pricing or delivery timelines. A credible Samsung alternative changes the negotiating dynamic even if Samsung never captures a majority of Broadcom’s volume.

“Diversification of the supply chain is not just a risk management exercise — it is becoming a competitive requirement,” one semiconductor analyst noted in the immediate aftermath of Friday’s announcement. “Broadcom is telling TSMC: we have options.”

What This Means for Samsung

For Samsung, the stakes are existential for its foundry ambitions. The Korean giant has invested hundreds of billions of dollars in its chip manufacturing business over the past decade, yet its global foundry market share sits at roughly 7%, a distant second to TSMC’s 62%. A marquee, multi-year deal with Broadcom — one of the world’s most demanding chip designers — provides the kind of revenue stability and design-win prestige that Samsung’s foundry unit has been desperate to demonstrate to other potential customers.

The timing matters. Samsung is simultaneously defending its memory business as a newly public company (SK Hynix debuted on the Nasdaq at a $26 billion valuation in July), competing for 2nm volume with Intel Foundry Services, and managing ongoing yield improvement programs. Landing Broadcom shores up investor confidence in what has been a troubled foundry narrative.

The $950 Billion Korean Semiconductor Summit

The Broadcom deal did not happen in isolation. The San Francisco summit produced an extraordinary cascade of Korean-American semiconductor commitments. SK Hynix, for its part, announced an extension of its HBM4 memory supply agreement with Microsoft — part of a broader $750 billion in total cooperation pledges by SK Group with global technology firms, including a long-term HBM4 joint development and supply deal with NVIDIA spanning the Vera Rubin and successor AI chip generations.

In aggregate, Samsung and SK Hynix locked in roughly $950 billion in AI chip supply cooperation with U.S. partners across the two-day summit — a number that reflects not just commercial agreements but Korea’s strategic positioning as an indispensable link in the global AI supply chain at a moment when geopolitical fragmentation is pushing hyperscalers toward supply redundancy.

South Korean President Lee Jae Myung framed the deals in explicitly national terms, arguing that Korea’s semiconductor firms are “not just vendors but co-architects of the AI era.” The summit was designed in part to counter concerns that U.S. chip export restrictions and TSMC’s Arizona expansion could gradually displace Korean firms from the most lucrative layers of the AI ecosystem.

The TSMC Shadow

Analysts are careful to note that the Samsung-Broadcom MOU is not yet a binding production contract. MOUs are intent documents, and the history of Samsung’s foundry business is littered with high-profile design wins that underperformed on volume. The 2nm yield question is real: Broadcom’s chips require consistent, high-quality fabrication, and any meaningful yield gap between Samsung and TSMC translates directly into chip cost and schedule risk.

TSMC, for its part, is not standing still. The company is executing a $265 billion investment in Arizona 2nm fabs announced earlier this year, deepening its U.S. footprint and political insulation against potential export control changes. It has also been developing its own advanced CoWoS and SoIC packaging technologies that directly compete with Samsung’s advanced packaging offer.

What to Watch

The next milestones are concrete. Samsung needs to demonstrate 2nm yield rates in the 65-70% range on real Broadcom test chips — not internal benchmarks — to unlock the production ramp. HBM4E qualification timelines at major AI accelerator vendors will signal whether Samsung or SK Hynix emerges as the dominant supplier for the post-2026 AI infrastructure buildout.

Broadcom reports earnings at the end of August. Analysts will be listening closely to whether CEO Hock Tan characterizes the Samsung relationship as a strategic hedge or a primary production path — the answer will reveal how seriously Broadcom intends to reduce its TSMC dependency over the next five years.

For the AI industry more broadly, a genuinely competitive foundry market — one where Samsung can execute at TSMC-comparable quality — would be a significant structural change. It would put pressure on pricing, expand capacity, and reduce the single-point-of-failure risk that U.S. policymakers have flagged repeatedly since the 2020-2021 chip shortage. Whether Samsung can close the execution gap is the pivotal question the $200 billion agreement leaves open.