Samsung Electronics and Broadcom have signed a $200 billion memorandum of understanding that locks in HBM4 memory supply and 2nm foundry services for AI chips through the end of the decade. The non-binding agreement, announced Monday, is one of the largest chip supply pacts ever made public, signaling both companies' intent to dominate the exploding market for artificial intelligence hardware.
A multi-year pact for cutting-edge chips
The MOU covers two critical pieces of the AI chip equation: high-bandwidth memory (HBM4) and advanced logic fabrication at the 2-nanometer node. HBM4 is the next-generation memory stack that AI accelerators depend on to move data quickly between compute units and memory. Samsung is one of only three companies capable of producing HBM4, alongside SK Hynix and Micron.
On the foundry side, the deal reserves capacity at Samsung's 2nm fabrication lines. That process node is expected to deliver significant performance and power efficiency gains over current 3nm chips, making it attractive for large-scale AI models. Broadcom, which designs custom AI chips for hyperscalers like Google and Meta, will be a key customer for those 2nm wafers.
The agreement runs through 2030, giving both sides a long planning horizon. Financial terms beyond the headline $200 billion figure were not disclosed.
Why this deal matters for AI hardware
AI chip demand is growing faster than supply, especially for advanced memory and cutting-edge logic. Samsung has been investing heavily to catch up with TSMC in the foundry business and to fend off SK Hynix in the memory race. This deal gives Samsung a guaranteed revenue stream and a prominent customer in Broadcom, which already works with several of the biggest names in AI.
For Broadcom, the MOU ensures access to both leading-edge memory and next-generation logic from a single supplier. That simplifies supply chain management at a time when AI chip orders are measured in the hundreds of millions of dollars per customer. The company's custom ASIC business has grown rapidly as cloud providers design their own accelerators to reduce dependence on Nvidia's GPUs.
The pact also puts pressure on TSMC, which has been the dominant foundry for AI chips. Samsung is offering a combination of memory and logic that TSMC can't match, since TSMC doesn't produce memory. That bundling strategy could help Samsung win more foundry contracts in the AI space.
The MOU is non-binding, meaning specific orders, pricing, and delivery schedules still need to be negotiated. Both companies said they expect to finalize a definitive agreement in the coming months. Samsung's foundry division has been ramping its 2nm process, with pilot production expected later this year and volume production targeted for 2025.
Broadcom is likely to be among the first customers for those 2nm chips, alongside other large AI chip designers. The company has not yet announced which specific products will use Samsung's 2nm node or HBM4 memory. Those details will emerge as the partnership moves from MOU to signed contracts.
One open question is how the deal affects Samsung's ongoing legal and business disputes with other chipmakers. The company is locked in a bitter rivalry with SK Hynix over HBM technology and with TSMC over foundry customers. This Broadcom pact could shift the balance in both battles, but the real test will be whether Samsung can deliver the promised yields and performance at 2nm.
