Samsung has launched commercial shipments of its most advanced high-bandwidth memory chip, escalating a three-way contest with SK Hynix and Micron Technology to supply the components that underpin NVIDIA’s next generation of AI accelerators.
The South Korean electronics giant said it has started delivering its HBM4 products to customers, though it did not identify specific buyers. The move is part of a aggressive push to regain momentum in a market where it has trailed domestic rival SK Hynix in recent product cycles.
High-bandwidth memory, or HBM, has become a critical enabler of AI infrastructure. These stacked memory chips sit alongside powerful graphics processors, feeding them data at extraordinary speeds to train and run large AI models. As hyperscale data centers expand globally, demand for the fastest HBM variants has intensified.
Aiming for Competitive Advantage
Samsung said its HBM4 achieves a base transfer speed of 11.7 gigabits per second, with the ability to reach 13 Gbps under certain configurations. The company also reported a substantial increase in overall memory bandwidth compared with its previous HBM3E generation. By boosting throughput and efficiency, Samsung aims to ease the data bottlenecks that emerge as AI models consume ever more data.
Samsung framed the milestone as evidence that the company has stabilized production yields (the share of usable chips produced on manufacturing lines) by leveraging advanced DRAM processes and a 4-nanometer logic base die. Achieving consistent yields is crucial in the memory business, where small inefficiencies can erode margins.
The announcement is well-timed for competitive advantage. NVIDIA’s upcoming Vera Rubin AI accelerators are expected to rely on HBM4 to support more demanding training and inference workloads. Securing a position in that supply chain would strengthen Samsung’s standing in a segment that commands premium pricing and long-term contracts.
Samsung shares rose sharply following the announcement, reflecting investor optimism that the company can close the gap with competitors. SK Hynix has dominated recent HBM generations and has indicated it would begin HBM4 shipments later this year, with broader commercialization to follow. The company has also pledged to maintain high yields comparable to its earlier HBM3E production.
Meanwhile, Micron Technology has told investors it has entered high-volume production of its own HBM4 chips and begun shipping to customers. The U.S.-based firm is increasingly seen as a leading beneficiary of the AI infrastructure boom, as memory suppliers work to align output with accelerator roadmaps.
Forecast Revenue to Triple
AI model developers, from hyperscalers to developers like OpenAI and Anthropic, are scaling up compute clusters that require massive pools of high-speed memory. Shortages of GPUs in recent years highlighted how supply limits can limit AI progress. Memory, though less visible than processors, is an essential part of this process.
To prepare for projected demand, Samsung said it is expanding HBM4 production capacity and expects revenue from its HBM business to more than triple in 2026 compared with 2025. The company also plans to introduce enhanced variants, including HBM4E samples later this year, as customers look for incremental performance gains.
Beyond raw speed, energy efficiency and heat management are emerging as differentiators. AI data centers face mounting power and cooling pressures, and memory modules must operate reliably in dense configurations. Samsung claims it has upgraded its designs to improve power efficiency and heat dissipation relative to prior generations.
The competitive stakes are high. HBM products carry higher margins than conventional DRAM, and long-term supply agreements with GPU makers can boost revenue streams for years. With NVIDIA’s accelerator roadmap advancing rapidly, memory vendors that meet performance and yield targets stand to capture significant share.
