AIC Inc. and SSD maker Solidigm are partnering to build specialized SSD infrastructure positioned as a new storage tier specifically for AI inference workloads. The collaboration targets the growing bottleneck between limited GPU memory and traditional data storage systems, as AI models increasingly require rapid access to datasets that exceed what can fit in GPU VRAM.
This mirrors what I've been tracking since my March piece on storage becoming the new GPU bottleneck — inference is fundamentally different from training, and the storage hierarchy is scrambling to catch up. While training can batch process data efficiently, inference demands low-latency random access to massive datasets for retrieval-augmented generation and real-time model serving. Traditional data lakes are too slow, GPU memory too expensive and limited. SSDs are emerging as the sweet spot for this "warm storage" tier.
What's notable here is how hardware partnerships are crystallizing around this specific use case. Solidigm, SK Hynix's enterprise SSD arm, isn't just selling faster drives — they're co-engineering with platform builders like AIC to optimize for AI workloads. This suggests the market has moved beyond generic NVMe improvements toward purpose-built inference infrastructure.
For developers running production AI, this matters immediately. If you're hitting memory limits with RAG systems or multi-modal inference, specialized SSD tiers could be your scaling path without GPU memory premium. But watch the pricing — purpose-built often means vendor lock-in, and you'll want flexibility as this storage tier matures.