The New Hot Spot in AI Infrastructure: Storage, Not Just GPUs

In 2023–2024 the story of AI infrastructure was almost entirely about GPUs: who could buy them, who couldn’t, and how many megawatts of power they’d swallow. By late 2025, a different picture is emerging. After accelerators, storage capacity and throughput are quickly becoming the next big constraint for AI data centers.

NetworkWorld reports that hard-drive lead times for nearline HDDs have stretched beyond a year, while enterprise flash is heading into shortages and price hikes as inferencing workloads explode. Network World+1 IDC, meanwhile, expects AI infrastructure spending to keep growing at over 40% CAGR, with most of that money flowing into accelerated servers and the storage and networking needed to keep them fed. my.idc.com+1

When you step back, the question becomes simple and brutal: if GPUs are the engines of AI, where do we put all the data, and how fast can we move it?

High-Capacity, Low-Latency Media: 122TB SSDs and HAMR Drives

For years, data center architects assumed that spinning disks would own capacity and flash would own performance. The AI era is muddying that neat divide.

On the flash side, vendors are racing to build gigantic SSDs. Solidigm’s D5-P5336 SSD, for instance, reaches 122.88 TB per drive and is explicitly targeted at AI and analytics workloads that need massive datasets online without spreading them across hundreds of small devices. Solidigm+2Network World Industry trackers now count multiple suppliers with 60 TB-plus models and several with 122 TB devices announced, pushing an argument that high-capacity SSDs are on a path to overtake HDDs for many cold and warm workloads in dense facilities. Blocks and Files

At the same time, hard drives have not stood still. Seagate’s HAMR (heat-assisted magnetic recording) Exos platform has already crossed the 30 TB mark in production, and roadmaps point toward 40–50 TB nearline drives later in the decade. Horizon Technology HAMR squeezes more bits per square inch by briefly heating the media with a laser as it writes—allowing smaller magnetic grains and far higher areal density. For hyperscalers building exabyte-scale archives, these drives remain compelling on cost per terabyte, especially for colder training corpora and compliance data.

In practice, AI data centers are adopting a tiered approach:

• Enormous QLC and TLC SSDs for hot “GPU-adjacent” data and small-batch training or fine-tuning.

• HAMR nearline HDDs for bulk datasets, logs, and rarely accessed historical data that still needs to live in the same region or sovereign domain.

The capacity numbers are eye-catching, but they also reflect a more subtle reality: keeping data dense is an energy and cooling strategy, too. Packing 20+ PB into a fraction of a rack with 122 TB SSDs, as Solidigm recently demonstrated in an AI storage lab cluster, means less floor space, fewer enclosures, and tighter cooling design. TechRadar

Performance for AI: Parallel Filesystems, 400G Fabrics and GPUDirect

Raw capacity is only half the equation. GPUs will happily sit idle if the storage fabric can’t deliver training data or model weights fast enough. That’s why the storage stack for AI now looks a lot like HPC—with a few twists.

Modern AI storage architectures lean heavily on parallel filesystems and NVMe-based object stores that can stripe data across dozens or hundreds of nodes. NetApp’s AFX architecture and Dell’s Project Lightning, for example, scale parallel file systems across large flash clusters to keep thousands of GPUs saturated—Dell reports up to 97% network saturation with its PowerScale F910 all-flash system serving AI workloads. HPCwire+2SiliconANGLE

Underneath, high-speed networking is becoming non-negotiable. NVIDIA’s Quantum-2 InfiniBand switches deliver up to 64 ports of 400 Gb/s per 1U chassis, pushing more than 51 Tb/s of bidirectional throughput, while 400G/800G InfiniBand optical modules are showing up in DGX-class superpods to keep GPU-to-GPU traffic and storage I/O in lockstep. NVIDIA Docs+2NVIDIA+2 On the Ethernet side, 400 GbE fabrics are standard for new AI pods, with 800 GbE already in standards and early products to follow. Network World+2Voltage Park

A key software ingredient is RDMA (Remote Direct Memory Access), which lets NICs move data directly between memory spaces without waking the CPU. NVIDIA’s GPUDirect family goes a step further: GPUDirect Storage allows NVMe drives and NICs to DMA directly into GPU memory, bypassing the host RAM “bounce buffer” and dramatically cutting latency and CPU overhead. NVIDIA Developer+2Solidigm+2 For data-hungry training jobs and streaming inference pipelines alike, this tight coupling between storage and accelerators is becoming table stakes.

One under-appreciated nuance is that AI training and AI inference place very different stresses on storage:

• Training tends to be throughput-bound: large sequential reads, massive checkpoints, and parallel streaming across many workers.

• Inference is increasingly latency-bound and metadata-heavy: billions of small reads against vector indexes, embeddings, and feature stores that power personalized and real-time applications. edgecore.com+2Hammerspace

By 2030, some forecasts suggest that around 70% of total data center demand will come from inference, not training—forcing storage architects to optimize not just for big reads, but for microsecond-sensitive query patterns at the edge and in the core. edgecore.com

Storage as the Next Supply-Chain Shock

Just as the industry began to adapt to GPU shortages, a new choke point appeared: memory and storage.

In late 2025, Samsung raised prices on some DDR5 server memory by up to 60% amid a worsening global shortage driven by AI data centers. Reuters Phison’s CEO recently confirmed that NAND flash prices have more than doubled in six months, that all 2026 NAND production has effectively been sold out, and that enterprise SSD customers are being prioritized over consumers. Tom’s Hardware Silicon Motion’s leadership went further, warning that HDD, DRAM, HBM, and NAND are all headed into simultaneous shortage in 2026—something they say they’ve never seen before. PC Gamer

NetworkWorld reports that nearline HDD lead times for AI-oriented data centers have already ballooned past a year, and that flash shortages are likely to worsen as hyperscalers shift more workloads from disk to SSD. Network World

For operators, the message is harsh but clear: storage is no longer a commodity that can be ordered just-in-time. Procurement and architecture strategies now need to anticipate multi-year volatility, diversify suppliers, and design with flexibility—mixing tiers of flash, HDD, and potentially cloud-adjacent storage services so that no single bottleneck can stall an AI rollout.

Power, Sustainability and the “Green” Storage Tier

Storage doesn’t draw as much attention as GPUs when it comes to power, but at hyperscale it’s far from trivial. High-performance NVMe SSDs, dense HDD trays, and the networking that ties them together all contribute to the multi-megawatt profiles of AI campuses.

Vendors are trying to turn capacity density into a sustainability story. Solidigm argues that its 122 TB SSDs can dramatically reduce rack count and floor space for petabyte-scale deployments, translating into lower cooling and infrastructure overhead per terabyte. forbes.com HAMR HDDs, for their part, aim to deliver ever more terabytes per spindle, lowering watts per terabyte even if each drive uses slightly more power than its predecessor. Horizon Technology

These efforts dovetail with broader “green data center” ambitions: AI campuses are experimenting with liquid and immersion cooling for both compute and dense storage, pairing them with renewable energy, adiabatic heat rejection, and sometimes even district heat reuse. NVIDIA+2NVIDIA As PUE (Power Usage Effectiveness) improves, energy-efficient storage designs—fewer enclosures, smarter spin-down and tiering policies, heavy use of all-flash caches—become critical levers.

The architectural split between hyperscale and edge compounds this. IDC projects edge computing spend reaching ~$378 billion by 2028, driven by real-time analytics and localized AI. my.idc.com That means thousands of small or micro data centers scattering storage into factories, roadside cabinets, and 5G sites. Each one needs efficient, ruggedized, often flash-centric storage that sips power but can still support time-sensitive inference and caching close to where data is generated.

AIOps for Storage and the Rise of “Sovereign AI”

AI isn’t only the workload running in these data centers—it’s also becoming the tool that runs them.

AIOps platforms are increasingly applied to storage: using ML models to predict capacity needs, detect anomalies in I/O behavior, and automatically rebalance workloads across tiers or clusters. EnterpriseStorageForum notes that predictive analytics is now a core tool for anticipating storage demand and mitigating issues before users notice them. Enterprise Storage Forum DataCore and others describe AIOps-driven storage as a way to continuously tune performance and placement based on real-time access patterns rather than static policies. DataCore Software

In the background, a geopolitical layer is forming around “Sovereign AI.” NVIDIA, Oracle, and others talk about sovereign AI as the ability for nations and enterprises to build and operate AI stacks—compute, networking, and storage—within their own legal and geographic boundaries. JFrog+3NVIDIA Newsroom+3NVIDIA

Storage is central to that idea. Sovereign AI platforms emphasize:

• Data residency and locality: keeping certain datasets in-country or even in-region, with storage clusters mapped to legal domains.

• Unified access across file, object, and table formats without moving data between jurisdictions. ddn.com

Architecturally, that can mean mirrored but logically separate storage fabrics per jurisdiction, sovereign object stores, and encryption plus key-management schemes that ensure cross-border replication can be audited or disabled as regulations evolve.

Beyond NAND and Rust: Glass, DNA… and Orbit?

Today’s AI data centers are built on NAND and magnetized platters, but research labs are working on far stranger media—motivated by a simple reality: data growth will continue to outpace conventional storage manufacturing.

At the University of Southampton, researchers have advanced “5D optical data storage,” fusing nanostructures into quartz glass using ultrafast lasers. Their so-called “eternity crystals” can theoretically store up to 360 TB on a single disc-sized crystal and retain data for billions of years even at high temperatures. University of Southampton+2phys.soton.ac.uk It’s archival, not high-throughput, but it hints at a world where AI training corpora could be etched into glass for essentially permanent preservation.

Microsoft and the University of Washington are exploring DNA as a molecular storage medium. Their DNA Storage project has already demonstrated automated systems that encode digital bits into ACTG base sequences, store them in synthetic DNA, and later decode them back—successfully writing and reading “hello” and larger datasets. UW Homepage+3Microsoft+3UW Homepage DNA is astonishingly dense and long-lived, making it a candidate for deep-archive tiers decades from now.

Then there are ideas that sound like science fiction but are edging into feasibility. A recent study highlighted in Singularity Hub explores orbital data centers: facilities in space, powered by abundant solar energy and cooled by the vacuum of space, beaming data via high-bandwidth links back to Earth. SingularityHub The concept raises enormous engineering and legal questions, but it underlines a growing willingness to rethink where we store planetary-scale AI data—and what physical environment best suits high-density, high-power infrastructure.

None of these technologies will replace NAND or HAMR disks this decade. But as AI systems become long-lived societal infrastructure—storing culture, science, and institutional memory—the idea of “eternal” archives on glass or DNA will only become more appealing.

Closing Thoughts and Looking Forward

For all the attention paid to GPUs and networking, storage is quietly becoming the real long-term differentiator for AI data centers. The organizations that thrive in this new era will be those that:

• Design tiered, GPU-aware storage architectures that combine dense SSDs, HAMR HDDs, and archival media in coherent workflows rather than as siloed products.

• Invest early in high-speed fabrics, RDMA, and GPUDirect-style pathways so that storage never becomes the bottleneck for training or inference.

• Treat procurement and capacity planning as strategic disciplines, hedging against years of supply volatility in NAND, HDDs, and memory.

• Embrace AIOps to let AI manage storage as dynamically as AI workloads consume it.

• Align storage design with sovereign AI and sustainability goals—thinking in terms of jurisdictions, carbon budgets, and water use as much as IOPS and latency.

In the short term, the job is to keep today’s GPU clusters properly fed. In the longer view, it’s about building a storage fabric capable of holding, moving, and preserving the world’s knowledge in an AI-driven civilization. The GPU wars may grab headlines, but the question of where all the bits live—and how gracefully we can move them—will define who actually wins.

Reference Sites

1. “Solidigm Path to 122TB SSD” – Solidigm
   https://www.solidigm.com/products/technology/solidigm-path-to-122tb-ssd.html Solidigm+2Network World+2

2. “Hard Drive Capacity and the Road to 50TB” – Horizon Technology
   https://horizontechnology.com/news/hard-drive-capacity-and-the-road-to-50tb/ Horizon Technology+1

3. “NVIDIA GPUDirect” – NVIDIA Developer
   https://developer.nvidia.com/gpudirect NVIDIA Developer+2Solidigm+2

4. “Storage Constraints Add to AI Data Center Bottleneck” – NetworkWorld
   https://www.networkworld.com/article/4076565/storage-constraints-add-to-ai-data-center-bottleneck.html Network World+1

5. “DNA Storage” – Microsoft Research
   https://www.microsoft.com/en-us/research/project/dna-storage/ Microsoft+2Source+2

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

#AI storage, #122TB SSD, #HAMR HDD, #GPUDirect Storage, #400G InfiniBand, #AIOps, #sovereign AI, #edge data centers, #DNA data storage, #5D optical storage