Will AI Stay Dumb for the Lack of Memory and Storage?
Santa Clara Ballroom
Tue Sep 17 | 11:20am
Abstract
The rapid evolution of artificial intelligence (AI) technologies is precipitating a profound transformation in data storage requirements, highlighting a potential bottleneck in AI advancement due to insufficient memory and storage capacities. This presentation examines the interplay between AI development and data storage technologies, focusing on the growing disparity between their respective growth rates.
Current AI clusters are experiencing a doubling in computing speed approximately every two months, a pace that starkly contrasts with the 3-5 year doubling time of contemporary data storage technologies. This discrepancy is generating significant challenges, as the demand for data storage surges in tandem with the proliferation of AI applications. Notably, recent trends indicate that the price per terabyte (TB) for solid-state drives (SSD), hard disk drives (HDD), and tape storage has increased by 15-35%, driven by the burgeoning appetite for data storage solutions spurred by language-based generative AI models such as ChatGPT.
The demand for data storage is further exacerbated by AI's capacity to generate vast quantities of images and videos, necessitating even greater storage capabilities. AI systems, particularly those involved in training complex models, require fast-access SSDs for efficient data processing and HDDs for mid-term data retention (up to five years) to continuously refine and improve these models. Additionally, long-term cold storage is becoming increasingly critical for AI applications that rely on extensive historical data, such as autonomous driving. Here, training data must be preserved for decades, encompassing development, production, and operational phases.
Beyond autonomous driving, other AI applications, including drug design, healthcare, and aviation, also necessitate long-term data storage solutions. These fields require the retention of vast datasets over extended periods, underscoring the critical need for advancements in storage technology to keep pace with AI's accelerating computational demands.
This presentation aims to shed light on the urgent need for innovative storage solutions to sustain AI's growth trajectory and explores potential avenues for bridging the gap between AI's computational power and data storage capabilities. By addressing these challenges, we can ensure that AI continues to evolve and unlock its full potential without being hindered by storage limitations.
Learning Objectives
Upon completion, participants will be able to understand the potential throttling impact that lack of memory or storage can have on the rise of AI.
Upon completion, participants will be able to understand the use cases for AI training and datasets that drive data storage requirements.
Upon completion, participants will be able to understand the use cases for AI inferencing and governance that drive data storage requirements.
Related Sessions
Keynote
SNIA Awards Presentation
Join us as SNIA presents annual member recognition honors, highlighting major technical contributions, task force advancements, and leadership.
Keynote
StorageAI - 1 Year of Progress
After only 1 year, SNIA's StorageAI has made incredible progress.
From workload performance to accelerator I/O to efficiency and sustainability to data movement to much more, the rapid pace of development has driven the industry towards greater open standardization. When the industry collaborate to solve foundational infrastructure problems, everyone scales faster - vendors, customers, end users, and entirely adjacent industries.
Join the Chair as we discuss the incredible work that's been done in a very short period of time and project the need for collaborative work as we move into the future.
Keynote
Divide, Then Conquer: A New Frontier for Storage Optimization
Hyperscale storage systems face extreme aggregate demands for throughput, IOPs, power, and endurance from concurrent workloads. Maximizing every metric simultaneously across all tenants is impossible due to conflicting design requirements. However, the scale and diversity of these platforms enable vertical integration and specialization—from high-level services down to the media—driving unmatched aggregate efficiency. This keynote will present counterintuitive examples and a novel case study.
Keynote
Lessons From the Front Lines of Enterprise AI: Real World Impact, Infrastructure Realities, and What Engineers Must Understand
Artificial Intelligence is often positioned as the next phase of hyperscale computing—but most Enterprises operate under a very different set of constraints. Unlike cloud native companies, Enterprises face a complex intersection of legacy systems, economic pressures, and real-world operational constraints—factors often invisible to engineers building the underlying technologies. This session focuses on the realities of deploying AI at scale across the hybrid cloud with environments that do not resemble hyperscale architecture.
In this session we will examine how three forces are reshaping Enterprise architecture. First, the economic model: inference, retrieval, and orchestration create ongoing non-linear costs making efficiency—not scale—the primary design principle. Second, the technical reality: retrieval-augmented generation (RAG) requires new compute with latency and security constraints, making hybrid cloud more essential to balance the tradeoffs of performance, cost, and control. Third, the operational considerations: legacy systems, regulatory requirements, and the inability to trust AI with sensitive data fundamentally shape what can be deployed into production.
This session will focus on lessons learned from managing AI workloads constrained by database bottlenecks, cloud capacity shortages, inconsistent supply of hardware components, and the operational challenges of model drift, quantization choices, and cloud-specific performance variability. It reveals why Enterprise buyers prioritize reliability, predictable supply, and cost efficiency over cutting edge performance—and how misaligned industry innovation can unintentionally disadvantage non-hyperscale customers.
Ultimately, this keynote aims to bridge the gap between AI system and component builders and AI operators—illustrating the practical consequences of design decisions, the constraints faced by real customers, and the opportunities for the industry to build more durable, accessible, and economically sustainable AI infrastructure.
