How dense media, smart fabrics, and sovereign architectures are reshaping where AI lives.

The Next Bottleneck in AI Isn’t Compute

For the first wave of generative AI, the infrastructure story sounded simple: whoever controlled the GPUs would control the future. By late 2025, that narrative has shifted. After accelerators, storage is quickly becoming the critical constraint in AI data centers.

NetworkWorld recently described storage capacity as “the next major constraint” after GPUs, with nearline hard-drive lead times stretching beyond a year and enterprise flash heading into shortages and price spikes as inferencing workloads explode. Network World IDC, meanwhile, expects AI infrastructure spending to keep climbing at over 40% CAGR into the second half of the decade, with accelerated servers absorbing most of that budget—but only if storage and networking can keep those accelerators fed. my.idc.com

The new question for AI operators is no longer just “How many GPUs can we buy?” but “Where do all the bits go—and how fast can we move them?”

Why AI Workloads Break Traditional Storage Models

Most legacy storage architectures were designed around relatively predictable patterns: OLTP databases, file shares, VMs, backup windows. AI workloads behave very differently.

John Metz of SNIA (the Storage Networking Industry Association) points out that ML pipelines include distinct phases—ingestion, preprocessing, training, checkpointing, archiving, and inference—each with different data structures, block sizes and access methods. Forcing all of that through a single, generic architecture creates “network detours” and idle GPUs. Network World

Recognizing this, SNIA launched Storage.AI in August 2025, an open standards effort backed by AMD, Cisco, Dell, IBM, Intel, NetApp, Pure Storage, Samsung, Seagate, WEKA and others. The goal: define AI-specific data services—placement, caching, tiering, metadata, and data movement—so that storage platforms can be tuned for ML pipelines rather than forcing ML to pretend it’s a legacy workload. Network World+2Business Wire

We’re watching storage evolve from a passive pool into an active participant in the AI lifecycle.

Dense Media: From 30 TB HAMR Drives to 122+ TB SSDs

On the media side, density is exploding—driven as much by energy and space pressure as by data growth.

Solidigm’s D5-P5336 SSD is a poster child for this shift. The drive scales to 61.44 TB today, with a roadmap to 122.88 TB per SSD and marketing material explicitly aimed at AI, analytics and content data lakes. Solidigm A 2025 lab cluster built around 192 of the 122 TB versions reached 23.6 PB of usable capacity in just 16U of rack space, hitting over 116 GB/s per node in MLPerf Storage benchmarks. TechRadar

Even more aggressive densities are on the way: Solidigm has confirmed plans for SSDs exceeding 245 TB by the end of 2026, arguing that high-capacity QLC SSDs can increasingly displace HDDs except for deep archival tiers. TechRadar

On the spinning side, Seagate’s Exos HAMR drives crossed the 30 TB line in early 2024 using heat-assisted magnetic recording, with roadmaps pointing toward 40–50 TB models. Blocks and Files For exabyte-scale archives and colder training corpora, these nearline drives remain compelling on cost per terabyte—even as flash eats into higher tiers.

In practice, AI data centers are converging on a simple, if extreme, tiering pattern:

• Dense QLC SSDs near the GPU fabric for hot data, small-batch training, checkpointing and high-QPS inference indices.

• HAMR nearline HDDs for bulk datasets, logs and regulatory archives that need to stay in-region but aren’t touched every minute.

The economic and environmental pressure is the same: pack more data into fewer devices to reduce floor space, cooling overhead, cabling and maintenance.

Fast Paths to GPUs: 400G Fabrics, RDMA and GPUDirect

Capacity alone doesn’t win; GPUs need a firehose, not a garden hose. That firehose is built from three layers: high-speed networking, RDMA and GPU-aware storage stacks.

On the network side, 400 GbE and 400 Gb/s InfiniBand are becoming baseline for new AI pods, with 800 Gb/s technologies already in standards and early deployments. These fabrics provide tens of terabits per second across a rack or small cluster, enough to saturate parallel file systems and object stores feeding thousands of accelerators.

RDMA (Remote Direct Memory Access) is the second pillar. By letting NICs move data directly between memory spaces without involving the CPU, RDMA cuts latency and CPU overhead for data-hungry jobs.

NVIDIA’s GPUDirect Storage (GDS) then goes one step further. GDS creates a direct data path between NVMe (local or NVMe-over-Fabrics) and GPU memory, bypassing the host DRAM “bounce buffer” and avoiding extra copies and context switches. NVIDIA Docs+2NVIDIA Developer NVIDIA’s docs emphasize that this approach increases system bandwidth while reducing CPU utilization—a double win in AI clusters where CPUs are already busy orchestrating jobs. NVIDIA Docs

Storage vendors are responding with AI-tuned systems. Dell, recently named IT Brand Pulse’s 2025 Market and Innovation Leader for file and object storage for AI, highlights its PowerScale and ObjectScale platforms as central to the “Dell AI Factory,” explicitly designed to feed modern GPU-driven workloads from edge to core. Dell Others, from NetApp to Pure and WEKA, are leaning into similar “AI file/object” narratives built around parallelism and GPU integration.

Training vs. Inference: Two Very Different Storage Problems

“AI storage” is often lumped together, but training and inference behave differently enough that they deserve separate designs.

As Edgecore and other infrastructure guides note, training is dominated by large, mostly sequential reads of huge datasets, along with periodic bursts of heavy writes for checkpoints and snapshots. It can tolerate some latency because jobs run for hours or days—but it punishes insufficient throughput and parallelism. edgecore.com

Inference, by contrast, tends to:

• Run continuously in production, often 24/7.

• Care more about latency and fan-out than raw bulk throughput—think millions or billions of small lookups against vector indexes, feature stores and embeddings. Aivres

Clarifai and others emphasize that while training is largely offline and happens per model version, inference is where users live: every user request, every personalization event, every real-time decision depends on fast, predictable access to models and associated data. Clarifai

Architecturally, that means:

• Training storage favors wide striping, huge I/O sizes and bulk movement from nearline tiers into flash.

• Inference storage favors hot flash caches, vector databases, key-value stores and edge-resident SSDs close to users or devices. edgecore.com

By 2030, many analysts expect inference to drive the majority of AI infrastructure demand, making “latency-optimized storage for inference” at least as important as “throughput-optimized storage for training.” my.idc.com

Hyperscale vs. Edge: Two Ends of the Storage Fabric

On one side of the spectrum are hyperscale AI campuses—multi-gigawatt sites where exabytes of data are staged for pretraining, fine-tuning and global services. On the other are edge facilities: small modular data centers in factories, cell towers, vehicles and smart-city infrastructure. Both need Storage.AI, but in different flavors.

IDC expects edge infrastructure spending to accelerate as enterprises push more inference to the edge for latency, privacy and bandwidth reasons. my.idc.com Edge inference guides from Mirantis, for example, describe architectures where regional micro-data centers host compact GPU or NPU clusters backed by ruggedized NVMe storage and minimal local HDD, often synchronizing with cloud or core sites during off-peak windows. Mirantis

Hyperscalers, meanwhile, are building massive “AI fabrics” where storage is split across tiers:

• Flash-heavy clusters directly cabled into GPU pods for high-intensity training and fine-tuning.

• Large HAMR archives and increasingly dense QLC-based SSD tiers holding petabytes of pretraining corpora and user history. Blocks and Files+2TechRadar

In both cases, the logical fabric must span cloud, private data centers and edge while respecting data governance and cost constraints—an increasingly brutal balancing act.

AIOps: Letting AI Run the Storage That Runs AI

The complexity of AI storage fabrics is outpacing what human operators can manage manually. That’s where AIOps—AI-assisted operations—enters the picture.

Enterprise Storage Forum describes AIOps as a fusion of AI and advanced analytics with traditional IT operations, aimed at automating capacity planning, performance management and anomaly detection. Enterprise Storage Forum IBM’s 2025 Redpaper on Intelligent Storage Management with AIOps goes further, detailing how IBM Storage Insights uses AI to spot performance deviations, recommend workload rebalancing and predict capacity shortfalls before they hit. IBM Redbooks

In practice, AI-assisted storage management is increasingly used to:

• Predict when specific volumes, tenants or clusters will exhaust capacity or IOPS, and recommend tiering or expansion.

• Detect “noisy neighbor” behaviors and unusual access patterns that might indicate misconfiguration, bugs or even security issues. Enterprise Storage Forum

As fabrics span multiple locations and vendors, these kinds of predictive and prescriptive tools become less a convenience and more a necessity. Without them, even well-designed Storage.AI architectures risk devolving into sprawling, opaque complexity.

Sovereign AI and the Geography of Storage

Another force reshaping AI storage is geopolitics. “Sovereign AI” has become a rallying cry for governments that want local control over critical data and models.

Dell’s 2025 white paper defines Sovereign AI as the ability of nations to maintain control over their AI infrastructure, algorithms and data in order to ensure security and align with local values. Dell That has direct implications for storage:

• Certain datasets—health, finance, citizen identity—must remain in-country, sometimes even in-region, with strict residency and access controls.

• Multi-tenant global storage fabrics must be logically segmented and auditable so that cross-border replication or access can be proved—or disproved—to regulators. Dell

Sovereign AI architectures are increasingly built around:

• Region-bound object and file stores with encryption and key management under local jurisdiction.

• “Data gravity” strategies that move compute to the data wherever possible instead of replicating sensitive data globally. Dell

For AI operators, this means storage decisions are no longer purely technical or economic—they’re also legal and diplomatic.

Glimpses of the Future: Glass, DNA and Beyond

While today’s AI storage is built on NAND and HAMR, research labs are working on exotic media that could eventually serve as deep archives for models and training corpora.

The University of Southampton’s “5D memory crystal” uses femtosecond lasers to encode data into nanostructured quartz glass, yielding up to 360 TB per disc with thermal stability to 1,000°C and theoretical lifetimes measured in billions of years. University of Southampton+2University of Southampton TechRadar notes that recent experiments have stored an entire human genome on one of these “eternity crystals,” underscoring their potential for long-term scientific and cultural archives. TechRadar

Microsoft’s DNA Storage project, developed with the University of Washington, has already demonstrated automated systems for writing digital data into synthetic DNA strands and later reading it back—showcasing densities orders of magnitude higher than magnetic or solid-state media. Dell+2Dell

Neither technology is anywhere near ready for high-throughput AI training, but they point toward a future where “eternal archives” of key models and datasets sit on glass or DNA in secure vaults, while active AI work happens on more conventional SSD and HDD tiers.

Closing Thoughts and Looking Forward

Storage has quietly become the central nervous system of AI data centers. The organizations that navigate this transition most effectively will be those that:

• Combine ultra-dense media (122 TB SSDs, 30+ TB HAMR drives) with GPU-aware fabrics (RDMA, GPUDirect Storage) so that capacity, bandwidth and latency scale together. StorageReview.com+2Blocks and Files

• Design separate but integrated patterns for training and inference, across both hyperscale and edge. edgecore.com

• Embrace AIOps so that AI helps manage the storage that feeds AI—closing the loop on predictive, self-optimizing infrastructure. Enterprise Storage Forum

• Build storage with sovereignty and sustainability in mind, treating data residency, energy density and water use as first-class requirements. Dell+2Network World

In the near term, the challenge is pragmatic: keep today’s GPU clusters from starving. In the longer view, it’s about building a Storage.AI fabric that can hold, move and preserve the world’s knowledge for decades. The GPU wars may define the headlines—but the quiet battle over where all the data lives, and how gracefully it flows, will decide who actually wins.

Reference Sites

1. “SNIA launches Storage.AI to address AI data infrastructure bottlenecks” – NetworkWorld
   https://www.networkworld.com/article/4033580/snia-launches-storage-ai-to-address-ai-data-infrastructure-bottlenecks.html Network World

2. “Solidigm 122.88TB D5-P5336 Review: High-Capacity Storage Meets Operational Efficiency” – StorageReview
   https://www.storagereview.com/review/solidigm-122-88tb-d5-p5336-review-high-capacity-storage-meets-operational-efficiency StorageReview.com

3. “Seagate unveils 30 TB+ Exos HAMR disk drives” – Blocks & Files
   https://blocksandfiles.com/2024/01/17/seagate-hamr-drives/ Blocks and Files

4. “GPUDirect Storage Overview Guide” – NVIDIA Documentation
   https://docs.nvidia.com/gpudirect-storage/overview-guide/index.html NVIDIA Docs+2NVIDIA Docs+2

5. “Powering the Global Movement for Sovereign AI” – Dell Technologies Blog
   https://www.dell.com/en-us/blog/powering-the-global-movement-for-sovereign-ai/ Dell

Author: Serge Boudreaux – AI Hardware Technologies, Montreal, Quebec
Co-Editor: Peter Jonathan Wilcheck – Miami, Florida

AI storage, Storage.AI, 122TB SSD, HAMR drives, GPUDirect Storage, AIOps, sovereign AI, edge inference, AI data centers, high-density storage