Samsung Electronics and Broadcom signed a $200 billion memorandum of understanding on July 24 in San Francisco covering HBM4 memory supply and 2nm foundry services through 2030, the largest bilateral chip deal announced as part of a sweeping $950 billion South Korea-US semiconductor cooperation package.
The numbers are hard to ignore. At a San Francisco AI Summit tied to South Korean President Lee Jae Myung's visit to the United States, Samsung and Broadcom formalised a partnership spanning five years and covering two of the most contested battlegrounds in AI hardware: high-bandwidth memory and advanced foundry manufacturing. Samsung will supply Broadcom with HBM4 and HBM4E chips - the next-generation memory that AI accelerators require - and will manufacture Broadcom's AI chips using its 2nm and below process nodes at its Pyeongtaek campus in South Korea. The deal is structured as a "turnkey" arrangement, covering foundry manufacturing and advanced packaging together rather than splitting them across vendors.
The broader package frames the Samsung-Broadcom MOU in context. SK Group agreed in parallel to pursue $750 billion in long-term memory supply agreements with Nvidia and other US technology firms, with SK Hynix's $500 billion commitment to Nvidia anchoring the larger figure. Combined, the two South Korean chipmakers have pledged to become indispensable infrastructure partners for the US AI industry through the end of the decade. As the Korea Herald reported, these are commitments closer in structure to long-term purchase agreements than loose letters of intent: if Samsung and SK Hynix produce the memory, their American partners are committing to buy it in volume. That is a binding kind of ambition.
Samsung shipped the world's first commercial HBM4 in February 2026, and has already sold out its entire 2026 HBM4 production capacity. More than 50% of Pyeongtaek's 4nm capacity is currently allocated to internal HBM4 logic base-die production. Broadcom is locking in a memory supply chain that is genuinely scarce. That scarcity is real, not manufactured.
The memory side is the easier half #
The foundry side is harder. Samsung's 2nm yield rates have improved, rising above 60% from roughly 20% in the second half of 2025, according to reporting by Digitimes. But more recent data from TrendForce paints a less tidy picture: some sources place Samsung's 2nm yields in the mid-50% range, still short of the roughly 60% threshold the industry typically requires for stable mass production. TSMC, by comparison, is reportedly achieving 60 to 70% yield consistency at 2nm, and its foundry revenue share hit a record 70.2% in the second quarter of 2025. Samsung's share sat at 7.3%. The gap is not narrowing fast.
Where the MOU does its most speculative work #
Samsung Foundry is not expected to reach profitability until around 2027, when its Taylor, Texas plant begins ramping in earnest. The Broadcom deal gives Samsung a reason to push harder on yield improvement and a headline client to point to - but converting an MOU into consistent 2nm volume production for AI chips is exactly the test Samsung has repeatedly struggled to pass. Qualcomm, which had considered Samsung for 2nm work, reportedly shifted toward TSMC earlier this year over yield concerns. Broadcom knows this history.
What the deal does accomplish is structural. Broadcom makes the networking chips, custom AI accelerators, and switching silicon that sit inside nearly every major hyperscaler's AI cluster. Locking Samsung in as a foundry partner, even at MOU stage, gives Broadcom a credible alternative to TSMC as the company scales its custom AI chip volume for clients including Google, Meta, and Apple. It also gives Samsung something it has badly needed: a marquee customer willing to bet publicly on its 2nm roadmap.
The AI memory supply chain context matters here too. HBM4 demand is outpacing the industry's ability to produce it. SK Hynix currently leads the HBM market, and Samsung has been working to close that gap. A committed, multi-year supply agreement with Broadcom at HBM4 and HBM4E volumes locks in revenue and, more importantly, justifies the capital expenditure required to expand production. Samsung reportedly invested $73 billion in semiconductor capacity in 2026 alone. Deals of this scale are how that investment earns a return.
Whether this reshapes foundry competition depends entirely on execution. TSMC's lead at advanced nodes is real and earned over years of yield discipline. Samsung has the facilities, the memory expertise, and now a high-profile foundry customer. What it needs is process consistency to back the headline number. The $200 billion is an ambition. The yield charts will tell you whether it becomes a business.
Also read: Etched raises $300M at a $10.3B valuation on a bet that transformer-only chips will displace Nvidia at the inference layer • Paper raised $34M on the exact thesis that wiped out half of Figma's market cap • Alphabet posted its biggest revenue quarter ever and Wall Street erased $255 billion in market cap anyway