Beyond SSD: SK Hynix AIN Family — Redefining Storage as the Core Enabler of AI at Scale

CS Kim, SVP of eSSD PD at SK hynix, presented the AIN Family at OCP Global Summit 2025, arguing that AI infrastructure has entered an era where token throughput — not compute FLOPs — is the primary bottleneck. He introduced three AI-optimized NAND product lines (AIN P, AIN B, AIN D) targeting distinct inference workloads: ultra-high IOPS vector search, HBM-complementary bandwidth flash, and high-capacity low-cost SSD. Key metrics include AIN P delivering 50M IOPS at 512-byte access (7x conventional Gen6 eSSD), with a 100M IOPS product targeted for end of 2027.

Key Takeaways

Reasoning models (e.g. Grok 4) produce 110M output tokens versus 11M for traditional LLMs, creating ~10x more pressure on memory and storage infrastructure in 2025.
AIN P targets 50M IOPS at 512-byte access on PCIe Gen6 — 7x more than a conventional Gen6 eSSD — with E3.x PoC samples by end-2026 and 100M IOPS production by end-2027.
AIN B stacks up to 16 NAND dies on a base/logic die to deliver HBM-like bandwidth with SSD-like capacity, enabling batch size and token length to scale simultaneously without GPU scale-out.
AIN B addresses multi-turn inference TTFT degradation by retaining KV cache context across turns without spilling to CPU memory or conventional SSD.
The AIN D product targets NL-HDD replacement using a JEDEC NL-SSD form factor, addressing data center power and front-space constraints.
SK hynix's future direction is a tightly integrated architecture combining AIN P, AIN B, xPU, and HBM — moving from loosely connected storage/memory components to a unified AI memory platform.