Why Magnetic Tape Is Booming in the AI Era: LTO-10 and Cold Storage

Why Magnetic Tape Is Booming in the AI Era: LTO-10 and Cold Storage

Magnetic tape is supposed to be obsolete. Yet, as generative AI pushes data center storage requirements to unprecedented levels, one of computing’s oldest storage technologies is attracting billions in new investment.

On October 29, 2026, Fujifilm announced an investment of approximately ¥10 billion (roughly $65 million) to expand its Odawara manufacturing facility in Kanagawa Prefecture, Japan. The project, scheduled to run from fiscal 2026 through fiscal 2028, aims to increase magnetic tape production capacity to approximately five times its FY2025 level, with a particular focus on the latest-generation LTO Ultrium 10 media.

According to Fujifilm, this is its largest investment in dedicated LTO tape manufacturing facilities to date. The immediate driver is growing demand for long-term data archiving as generative AI and cloud services generate increasingly large volumes of information.

The interesting part is not simply that an old storage medium is making a comeback. It is that AI infrastructure is exposing a problem that faster SSDs and larger HDD arrays cannot economically solve on their own: storing enormous quantities of data that must remain available for years but rarely need to be accessed.

🧠 Why Does AI Still Need Magnetic Tape?

#

There is an old joke in computer science circles that remains surprisingly relevant.

Computer scientist Andrew Tanenbaum famously observed:

“Never underestimate the bandwidth of a station wagon full of tapes hurtling down the highway.”

It sounds ridiculous in an era of high-speed networking, but the underlying engineering principle is valid. At petabyte (PB) and exabyte (EB) scales, physically transporting storage media can be more practical than transferring the same amount of data across a network.

Amazon demonstrated a related idea in 2016 with AWS Snowmobile, a heavy-duty transport solution designed to move massive storage appliances between customer facilities and AWS data centers. A single vehicle could carry up to 100PB of data.

Snowmobile was not a magnetic tape service and was eventually retired in 2024. However, it illustrated a persistent infrastructure challenge: moving enormous datasets is not always a networking problem that can be solved by purchasing faster links. Cost, transfer time, available bandwidth, and operational constraints all matter.

Generative AI has made this problem considerably more visible.

AI generates more than model weights

#

Training a large language model requires GPUs, high-bandwidth memory (HBM), fast networking, and high-performance storage. Active training datasets and frequently accessed model files need to remain close to the compute infrastructure.

But an AI training pipeline produces considerably more data than the final model.

Enterprises may need to retain:

• Original text, image, audio, and video datasets.
• Cleaned and normalized training corpora.
• Data preprocessing outputs and intermediate artifacts.
• Training logs, evaluation results, and experiment metadata.
• Model checkpoints from different training stages.
• Previous model versions and fine-tuning datasets.
• Compliance records and historical datasets used for auditing.

Not all these files need immediate access.

A dataset used yesterday might not be touched again for six months. An older model checkpoint may remain important for reproducing an experiment, even if the production model has already moved through several generations.

Deleting this information is not always an option. Keeping everything on expensive, continuously powered storage is not particularly attractive either.

This is where cold data becomes an infrastructure problem.

Understanding hot, warm, and cold data

#

Enterprise storage architectures generally classify information according to access frequency and performance requirements.

The objective is not to move every file onto the cheapest possible medium. It is to match storage cost and performance to actual access patterns.

For example, a dataset actively feeding GPUs during training belongs on high-throughput storage. Once the training run finishes, older copies of the dataset may be moved into a lower-cost archival tier.

A similar approach applies to historical model checkpoints, experiment logs, and regulatory records.

Keeping all of these files on an HDD array that remains powered on around the clock can create unnecessary electricity, cooling, and maintenance costs.

Magnetic tape addresses precisely this category of storage demand.

📦 Why Magnetic Tape Still Makes Economic Sense

#

Magnetic tape has several characteristics that remain difficult for HDD and SSD infrastructure to match when the primary requirement is long-term, infrequently accessed storage.

1. Lower operating costs and energy consumption

#

A magnetic tape cartridge does not consume electricity while sitting on a shelf.

This is a significant distinction from conventional HDD arrays, which typically require continuous power, cooling, monitoring, and maintenance even when their data is rarely accessed.

Automated tape libraries introduce additional operational costs because robotic mechanisms, drives, and control systems require power. However, individual cartridges can remain offline for extended periods.

Fujifilm estimates that, under a specific scenario involving 100PB of stored data over a 10-year lifecycle, an automated tape library can reduce total cost of ownership (TCO) by approximately 86% and carbon emissions by approximately 95% compared with an all-HDD architecture.

These figures are vendor estimates, not universal guarantees. Actual savings depend on drive utilization, cartridge handling, media replacement, data retrieval frequency, facility costs, and the design of the competing HDD system.

Still, the underlying economics are straightforward: continuously powering storage devices to retain data that nobody accesses is expensive.

2. Air-gapped security against ransomware

#

Magnetic tape also offers a security advantage that online storage cannot fully replicate.

A tape cartridge can be physically ejected from its drive and stored offline. Once disconnected, it is no longer directly accessible to malware operating inside the production network.

This creates an air-gapped backup layer.

If ransomware compromises a company’s servers, network-attached storage, and backup infrastructure, physically isolated tape copies can remain outside the attacker’s reach.

That does not make tape backups automatically secure. Organizations still need access controls, encryption, secure cartridge handling, off-site storage, inventory management, and tested restoration procedures.

Immutable backup policies and Write Once Read Many (WORM) media can add another layer of protection by preventing stored information from being overwritten during its retention period.

For enterprises dealing with ransomware and long-term compliance requirements, this combination makes tape a useful component of a broader backup strategy.

3. LTO-10 is faster than many people expect

#

Magnetic tape is often dismissed as extremely slow. That assumption comes from confusing random access latency with sustained sequential throughput.

Tape is not designed to retrieve arbitrary files instantly. It is designed to read and write large continuous data streams efficiently.

The latest LTO Ultrium 10 generation supports:

• Up to 40TB of native capacity per cartridge.
• Up to 100TB of compressed capacity at a 2.5:1 compression ratio.
• Up to 400MB/s of native sustained transfer throughput.
• Up to 1,200MB/s under specified compressed-transfer conditions.

The compressed figures depend on the compressibility of the source data and the supported drive configuration. Already-compressed datasets, encrypted files, and many media formats may deliver substantially less compression.

The distinction between native and compressed capacity matters when planning storage infrastructure. A company should not assume that every 40TB cartridge will reliably hold 100TB of arbitrary AI datasets.

Tape also relies on sequential access. Locating an individual file may require mounting the cartridge, positioning the tape, and reading through the relevant data region.

This makes it unsuitable for latency-sensitive workloads and frequent random reads.

A useful way to think about the architecture is:

• NVMe SSDs act as high-speed working storage.
• HDDs provide relatively inexpensive online capacity.
• Magnetic tape serves as a high-density archival warehouse.

Each tier solves a different problem.

🏭 Fujifilm’s LTO-10 Investment Signals Growing Demand

#

Fujifilm’s October 2026 announcement provides a concrete example of how the storage industry is responding to AI-driven data growth.

The company plans to invest approximately ¥10 billion in its Odawara facility between FY2026 and FY2028. The expansion will increase magnetic tape manufacturing capacity to approximately five times its FY2025 level, measured by shipped storage capacity.

The primary focus is its latest-generation LTO Ultrium 10 product.

One interesting aspect of LTO-10 is the magnetic material technology behind it. Fujifilm combines technologies involving Barium Ferrite (BaFe) and Strontium Ferrite (SrFe) magnetic particles through its proprietary fine hybrid magnetic particle approach.

Higher recording density requires more than simply making tape thinner or increasing the number of cartridges produced. Improvements in magnetic media, recording heads, track density, and signal processing all contribute to increasing usable capacity.

LTO-10 represents another step in that process.

LTO-10 capacity and compatibility considerations

#

The LTO-10 generation supports native cartridge capacities of 30TB and 40TB, depending on the media configuration.

Its compressed capacity can reach 75TB or 100TB, respectively, assuming a 2.5:1 compression ratio.

However, there is an important deployment consideration: LTO-10 does not maintain the traditional backward read and write compatibility with earlier LTO generations.

Organizations upgrading existing tape libraries should therefore verify drive and media compatibility before purchasing new cartridges.

This is particularly relevant for enterprises maintaining large historical archives. Moving to a newer generation may require a migration strategy rather than simply replacing the drives and continuing to use the existing media.

For long-term infrastructure planning, media compatibility and migration costs should be included alongside cartridge capacity and drive throughput.

🔬 CERN Demonstrates Why Long-Term Data Preservation Matters

#

Scientific research has relied on magnetic tape for decades, and CERN provides one of the clearest examples of why archival storage remains necessary.

The Large Hadron Collider (LHC) generates enormous quantities of experimental data. Not all of that information can be kept on high-performance storage, especially when researchers need to preserve historical observations for future analysis.

By December 2025, CERN’s accumulated experimental data had reached approximately 1 exabyte, with the vast majority preserved across roughly 60,000 tape cartridges.

The value of this archive extends beyond immediate research requirements.

Experimental data collected years ago can become useful again when scientists develop new theoretical models, improve analysis techniques, or discover previously overlooked relationships in existing measurements.

Discarding historical datasets would make some future research impossible or require expensive experiments to be repeated.

CERN’s approach demonstrates an important principle for AI infrastructure as well: data can lose immediate operational value without losing its long-term research or business value.

An older training dataset might become relevant when a company develops a new model architecture. Historical checkpoints may help reproduce a previous experiment. Archived records may also be necessary to demonstrate how a model was trained or evaluated.

Long-term retention is therefore not simply about keeping files. It is about preserving the ability to reproduce, investigate, and validate previous work.

📈 LTO Tape Shipments Show That the Market Is Still Growing

#

Fujifilm’s investment is not happening in isolation. Industry shipment figures indicate that enterprise demand for tape storage has remained substantial.

According to the LTO Program, global LTO tape shipments reached a record 176.5 exabytes of compressed storage capacity in 2024, representing a 15.4% increase over 2023.

The figure refers to the total storage capacity shipped during the year, not the amount of data actively stored on tape worldwide.

The LTO Program’s annual reports also show that demand continued at a high level in 2025, when shipped capacity reached 160.3EB. Although that represented a decline of approximately 9% from the 2024 record, it remained the second-highest annual shipment figure on record.

In September 2026, the LTO Program also reported that tape capacity shipments in Q1 2026 had increased by 57% year over year.

These figures suggest that the market is not relying on nostalgia for an old storage technology. Enterprise demand is being driven by practical requirements involving data retention, cybersecurity, energy consumption, and storage economics.

For organizations managing hundreds of petabytes or more, even a small reduction in the average cost per stored terabyte can translate into substantial savings.

⚙️ Tape Is Not a Replacement for SSDs or HDDs

#

The resurgence of magnetic tape should not be interpreted as a sign that SSDs and HDDs are becoming less relevant.

AI infrastructure needs multiple storage tiers because different stages of a data pipeline have very different performance requirements.

During model training, datasets must be delivered to GPUs at sufficient throughput to prevent expensive accelerators from sitting idle. Moving active training data directly from tape would introduce unacceptable delays.

The same applies to inference systems, databases, vector indexes, and frequently accessed model files.

Tape becomes attractive after data moves out of the active compute path.

A practical AI storage architecture might look like this:

This architecture allows enterprises to avoid paying premium storage prices for information that rarely needs immediate access.

It also introduces operational complexity. Data must be classified, migrated, indexed, and tracked across different storage systems. Retrieval workflows need to account for tape mounting delays, media availability, and the possibility of multiple cartridges being required to reconstruct a dataset.

A well-designed archival system must therefore consider more than raw capacity.

Data integrity verification, encryption key management, metadata preservation, retention policies, cartridge lifecycle management, and periodic restoration testing are all part of the system.

For additional context on how storage tiers fit into modern AI infrastructure, KAD’s guide to AI/ML data storage architecture examines the relationship between workload characteristics, storage performance, and infrastructure costs.

🔐 The Future of AI Storage Depends on Managing Data Growth

#

The AI industry tends to focus on GPUs, HBM, high-speed networking, and increasingly powerful processors. These components receive most of the attention because they directly influence model training and inference performance.

But compute is only one part of the infrastructure equation.

Every training run, evaluation cycle, fine-tuning operation, and production deployment generates information that organizations may need to preserve. As AI adoption expands, the volume of historical data will continue to put pressure on storage budgets and data center energy consumption.

Magnetic tape offers a practical answer for one specific part of that problem: retaining enormous quantities of information at relatively low cost without keeping every storage device continuously powered.

It is not fast enough to replace NVMe storage. It is not convenient enough to replace online HDD arrays for frequently accessed files. And it cannot eliminate the operational work involved in maintaining large archival systems.

But it does not need to.

Fujifilm’s fivefold production expansion, the continued demand for LTO media, and the growing archival requirements of AI infrastructure all point toward the same conclusion: as data becomes more valuable, inexpensive long-term preservation becomes increasingly important.

In an industry obsessed with making computers process information faster, magnetic tape is quietly solving a different problem—making sure that the information is still there when someone needs it years later.