Introduction

On May 5, 2025, Samsung Electronics officially previewed its next-generation HBM4 memory architecture, promising a 2.5x increase in bandwidth and a 40% drop in power consumption compared to HBM3e. Announced during the Samsung Foundry Forum in San Jose, the new stacked DRAM platform targets high-throughput AI training and inference, signaling a major leap in memory performance for the AI era.

“AI workloads are now bottlenecked by data movement, not computation,” said Siyoung Choi, President of Samsung Foundry.¹ “HBM4 is designed to eliminate that bottleneck and scale with trillion-parameter models.”

Samsung claims HBM4 delivers 1.4 TB/s of bandwidth per stack and supports up to 24 GB per module. Its 12-high die stacks are cooled using enhanced thermal interface materials and through-silicon via (TSV) redesigns. It also integrates error-correcting logic directly into the memory controller to support AI-specific data integrity features.

Why it matters now

• Memory, not compute, is now the primary bottleneck for large model inference.
• AMD, Nvidia, and Intel depend on HBM for their AI accelerators.
• Energy-efficient memory is key to scaling AI without triggering data center sustainability crises.

Call-out: AI infrastructure isn’t just about GPUs—it’s about how fast they feed

Samsung’s HBM4 is expected to power Nvidia’s next-gen Blackwell Ultra, AMD Instinct MI400X, and emerging photonic AI accelerators, which will ship in 2026–2027.

Business implications

AI hardware buyers—particularly hyperscalers and research labs—must begin planning for HBM4-compatible platforms. The expected performance uplift will make HBM4 a critical component for next-gen LLMs and diffusion-based generative models.

Startups working on AI model deployment, inference scaling, or custom accelerators should reevaluate their hardware stack assumptions. Access to HBM4 will likely define competitive advantage in memory-bound AI workloads over the next 18–36 months.

Looking ahead

Samsung confirmed that the volume production of HBM4 will begin in late Q4 2025, with samples shipped to key partners in July. SK hynix and Micron are expected to follow with HBM4 variants in early 2026.

Gartner predicts that by 2028, 85% of AI accelerator sales will require high-bandwidth memory. Memory ecosystems will likely emerge as co-equal with compute ecosystems in defining AI infrastructure economics.

The upshot: Disruption isn’t just about silicon logic, but about the flow of information. Samsung’s HBM4 redefines memory as a strategic enabler of AI progress. Enterprises that anticipate this shift will stay ahead of the AI performance curve.

––––––––––––––––––––––––––––
¹ Siyoung Choi, Samsung Foundry Forum keynote, May 5, 2025.