AI teams often focus on GPUs first. However, storage now shapes AI performance, cost, and user experience. SNIA, the Storage Networking Industry Association, helps vendors and customers address that challenge through open standards and vendor-neutral education.
For vendors, SNIA creates a shared framework for data movement, memory tiering, cooling, and storage-to-GPU performance. For customers, SNIA provides clearer benchmarks and stronger interoperability signals before major infrastructure decisions.
At AI Infrastructure Field Day 5 on June 10, 2026, SNIA Chair J Metz presented StorageAI and related standards work. Our CTO, Frederic Van Haren, attended as a delegate evaluating SNIA’s approach to AI data infrastructure.
Storage performance directly affects inference cost and responsiveness. Inference workloads need fast access to model weights, vector stores, cached data, and retrieval systems. When storage slows down, expensive accelerators wait for data instead of doing useful work.
This matters more now because inference workloads are scaling faster than training workloads across the industry.
AI workloads change the infrastructure problem. The goal is no longer simple throughput or capacity. Instead, teams need reliable data placement, predictable latency, efficient movement, and strong lifecycle management.
SNIA launched StorageAI in August 2025 with fifteen founding vendors, including AMD, Cisco, Dell, IBM, Intel, NetApp, Pure Storage, Samsung, Seagate, and WEKA.
StorageAI brings existing SNIA specifications into one vendor-neutral framework for AI data services. It coordinates work around memory tiering, data movement, storage efficiency, and latency. As a result, vendors can align around common targets, while customers get a clearer path to interoperable AI infrastructure.
Data often sits in one place and is consumed elsewhere. Every move adds cost, latency, and complexity. Therefore, location matters as much in AI infrastructure as it does in real estate.
SDXI stands for Smart Data Accelerator Interface. It defines a vendor-neutral memory-to-memory data mover and acceleration interface. In AI systems, SDXI can reduce overhead once the software has established the connection. Therefore, hardware can move data across CPUs, GPUs, DPUs, and memory tiers more efficiently. As the real estate saying goes, “location, location, location!”
Many brownfield AI systems use POSIX file systems with RDMA. However, Greenfield AI designs favor object storage for scale, economics, and governance. The challenge is to keep object storage useful without adding avoidable latency.
SNIA’s Object-over-RDMA and file-over-RDMA initiatives target that problem. They aim to shorten the path between storage and accelerators. Direct-to-GPU paths can reduce CPU involvement, lower latency, and improve data delivery for AI workloads.
Important use cases include KV cache movement, faster checkpointing, lower CPU load, and high-throughput data loading for inference and training pipelines.
GPUs draw a lot of power and generate a lot of heat. Yet they are not the only hot components in modern AI servers. Memory, drives, and dense storage devices also need better thermal management.
SNIA’s liquid cooling work for SSDs establishes a standard for an SSD thermal bridge that allows liquid coolant to flow across drive components. Standardized liquid cooling helps vendors reduce engineering costs. It also gives end users a path to high-density, sustainable AI infrastructure. Moreover, this vendor-neutral approach enables components from different manufacturers to work together in a single liquid-cooled server environment.
Traditional storage benchmarks rarely reflect real AI data patterns. AI pipelines include ingestion, preprocessing, training, checkpointing, inference, and retrieval. Each phase creates different block sizes, data structures, and access patterns.
SNIA’s response is to define AI-specific workloads rather than create a single fixed scorecard. That approach keeps performance results reproducible across environments. At the same time, the definitions can expand as AI architectures evolve.
This is relevant because older benchmarks assume predictable traffic. AI workloads often behave differently. They can be random, bursty, and sensitive to latency. Better workload definitions help enterprises compare infrastructure options with more confidence.
SNIA’s work matters because AI infrastructure now depends on coordinated data movement, cooling, benchmarking, and storage-to-compute integration. Open standards give customers a path beyond proprietary point solutions. They also help procurement teams compare options with fewer assumptions.
For organizations building AI infrastructure now, the practical takeaway is simple. Evaluate storage the same way you evaluate compute and networking. Treat it as a first-order design decision, not an afterthought.
If your AI roadmap depends on faster inference, lower GPU idle time, or better infrastructure ROI, start with the data path. Review where data moves, where it waits, and where standards can reduce risk. HighFens can help assess those bottlenecks and turn AI infrastructure into a measurable advantage.