Recommended Articles
Fundraising

Fujifilm's Fivefold Tape Bet Exposes AI's Archiving Bill

By Angeled · November 1, 2026 · Fundraising
Fujifilm's Fivefold Tape Bet Exposes AI's Archiving Bill
Fujifilm's Fivefold Tape Bet Exposes AI's Archiving Bill

On October 29, 2026, Fujifilm announced it will invest approximately 10 billion yen — roughly $63 million — at its Kanagawa Factory in Odawara, Japan, to increase production capacity for data storage magnetic tape by approximately five times. The investment, spread across fiscal 2026 through fiscal 2028, will expand manufacturing facilities and add personnel, primarily for its latest-generation FUJIFILM LTO Ultrium10 cartridge. Fujifilm says it is the largest investment it has made to date in dedicated facilities for LTO tape.

The reason Fujifilm gives is straightforward: the rapid spread of generative AI and cloud services is driving exponential growth in the volume of data generated worldwide, and much of that data needs to be kept. Tape, the oldest mass storage medium in computing, is suddenly a growth business again.

This is not a nostalgia story. It is a window into one of AI's least glamorous and most consequential cost curves — what it costs to keep everything the AI era produces.

Why tape, and why now

Magnetic tape has survived every predicted death for a simple reason: for data that is rarely accessed but must be retained, nothing else comes close on cost. A tape cartridge can sit on a shelf for decades, consuming no power, physically disconnected from any network. Disk arrays and cloud storage tiers, by contrast, consume electricity continuously and cost far more per terabyte at petabyte scale.

Fujifilm says tape use has been expanding particularly among major cloud service providers and data center operators, and points to demand for environments where the enormous amounts of data accumulated through AI training and inference can be stored and used securely and efficiently. That framing matters: AI does not only generate model outputs. Training runs produce checkpoints, logs, and datasets. Inference at scale produces transcripts, images, and interaction records. Much of it must be retained for compliance, for retraining, for auditing, or simply because nobody has decided it is safe to delete.

The numbers behind the bet are striking. The LTO Program, run by HPE, IBM, and Quantum, reported a record 176.5 exabytes of LTO tape capacity shipped in 2024, up 15.4 percent. Shipments dipped 9 percent to 160.3 exabytes in 2025 — still the program's second-best year — then rose 57 percent year over year in the first quarter of 2026, with the program expecting 2026 to set a new record. Fujifilm's fivefold target is measured by shipment capacity against fiscal 2025, so it describes the scale of the planned supply expansion, not a fivefold increase in what a single cartridge holds. Even so, it is a large directional bet from a company that says it is one of only two magnetic data tape manufacturers worldwide and holds the largest market share.

The economics founders should understand

Tape's revival carries three lessons that generalize well beyond storage.

First, AI's cost structure is broader than compute. Headlines focus on GPUs, data centers, and power. But every AI system also creates a data retention obligation, and retention costs compound: data stored this year must still be stored next year, and the year after. Companies that price AI products on compute alone, without modeling storage and compliance costs over the product's lifetime, will find their margins eroding quietly.

Second, air-gapped storage is becoming a security feature, not just a cost saver. Because tape can be physically isolated from networks, it is attractive as protection against ransomware: data that is offline cannot be encrypted by an attacker who has compromised the network. As AI systems accumulate sensitive training data, customer records, and proprietary outputs, the ability to keep an offline copy is shifting from an archival nicety to a board-level control.

Third, boring infrastructure can be a better business than it looks. Fujifilm is expanding capacity in a duopoly market where demand is rising and supply is constrained — exactly the dynamic that worries AI investors elsewhere.

The limits of the bet

Fujifilm's announcement deserves the usual cautions. The company has not said when the new Odawara capacity comes online, only that spending runs through fiscal 2028, and it has not published absolute production numbers — so "five times" tells us the direction but not how many cartridges or exabytes that means. Tape also cannot solve AI's hot-data problem: retrieving data from tape is slow, and it will never replace disk or flash for data that needs instant access.

What founders and investors should do next

Founders building AI products should audit their data retention bill now, not later. Map what data you generate, what you are legally or contractually required to keep, and what it costs to store for five or ten years. If retention costs are buried inside a cloud bill, break them out. Consider tiered storage early: hot data on fast storage, cold data on the cheapest durable tier available. And treat offline backups as a security investment, not just an IT expense — ransomware resilience is increasingly a selling point in enterprise deals.

Investors should look past the model layer for infrastructure exposure. Tape is one example; power delivery, cooling, and archival storage are others. The pattern is consistent: AI's growth creates bottlenecks in unglamorous places, and the companies that own those bottlenecks — especially in concentrated markets — can compound quietly while attention stays on the flagship labs.

Angeled.com connects founders with investors.

Read Next