Samsung Lifts HBM4 Capacity to Half of Total Output to Support Nvidia’s Next-Generation AI Chips

High-bandwidth memory, or HBM, is a critical component that gives AI accelerators rapid access to data, and supply has fallen short as data centers expand worldwide. SK hynix currently controls about 60% of the HBM market and is using advanced packaging and next-generation products to strengthen its lead while competing with Samsung Electronics and Micron for orders from major cloud providers.

SK Group Chairman Chey Tae-won recently pledged to sharply increase SK hynix's HBM output and accelerate the latest HBM4 into mass production to narrow the supply-demand gap for AI chips. He said adding wafer capacity takes at least four years and that tight global supplies could persist until 2030. The reports did not disclose the investment amount for the expansion or a specific production start date.

Samsung Electronics is working to narrow the gap with TSMC in semiconductor foundry manufacturing. The 4-nanometer node is an important mature process for high-performance computing chips such as AI accelerators, and yields directly affect costs, delivery schedules and customers’ willingness to adopt it. Crossing the 80% threshold gives Samsung a foundation to take on large orders and improve capacity utilization.

Recent developments show that Samsung’s 4-nanometer process yield has exceeded 80%, helping it secure orders from customers including U.S. AI chip company Groq. Samsung is also preparing for HBM4 mass production, aiming to combine its foundry and high-bandwidth memory strengths and return its non-memory business to profitability as early as the second half of 2026. The reports did not disclose the value of the orders.

High-bandwidth memory is a critical component in artificial-intelligence accelerators, moving data rapidly between processors and memory while limiting power consumption. Samsung Electronics is seeking to leverage its rare combination of memory manufacturing and foundry operations to coordinate memory dies, custom logic base dies and advanced packaging. That vertical integration could help it respond to Nvidia’s increasingly tailored chip requirements and challenge incumbent suppliers for a larger share of the AI-memory market.

Samsung is developing a seventh-generation HBM4E product with an eight-layer stack for Nvidia’s NVHBM initiative, targeting higher transfer speeds and greater customization. Successful qualification could position the South Korean company as a key supplier for Nvidia’s next-generation Rubin Ultra AI architecture. The report did not disclose a sampling date, mass-production timetable, contract value or expected supply volume, leaving the commercial scale and timing of Samsung’s prospective role uncertain.

High-bandwidth memory is a critical component in artificial intelligence accelerators, allowing processors to move large volumes of data at high speed. HBM4 adds more advanced logic-based base dies to improve bandwidth and integration. As spending on generative AI infrastructure accelerates, securing volume production has become central to Samsung Electronics’ effort to win business for next-generation accelerators from customers including NVIDIA and Advanced Micro Devices.

Samsung’s foundry business has allocated more than 50% of its 4-nanometer capacity to HBM4 base dies, with monthly output estimated at about 15,000 wafers. The commitment is aimed at meeting surging demand for the sixth generation of high-bandwidth memory. Large-scale shipments are expected to begin in the third quarter of 2026, supporting a broader supply ramp for NVIDIA and AMD accelerators during the second half of the year.