Fujifilm to Boost LTO Tape Capacity 5x by FY2028 as AI Archives Grow

Fujifilm will spend about ¥10 billion to make roughly five times as much data tape by fiscal 2028, and it names generative AI as a reason. The company announced the plan in a press release dated October 29, 2026. The money goes to its Odawara site in Kanagawa, mainly for the latest LTO Ultrium 10 cartridge. This is not a chatbot story. It is a story about where the data behind AI ends up after training, and it is a useful check on the claim that AI demand is reshaping every layer of hardware.

Sources: Fujifilm's own release, a Blocks and Files analysis that cites a paywalled AsiaNikkei report, and coverage by The Next Web.

TL;DR: what is confirmed

What exactly did Fujifilm say?

The release says the investment is the largest Fujifilm has made so far in dedicated LTO tape facilities. It will pay for manufacturing equipment and added staff. The target is a "stable supply system for magnetic tape".

The rationale is short. Generative AI and cloud services are producing fast data growth. Tape offers low-cost, long-term storage, can be isolated from networks, and is gaining attention as protection against ransomware. Fujifilm says demand is rising among major cloud providers and data center operators. The release contains no executive quotes and no absolute capacity figures.

Fujifilm also says it began tape research in the 1950s, launched LTO Ultrium 1 in 2000, and in 2011 became the first to commercialize barium ferrite particles for tape. Ultrium 10, launched in 2025, uses what the company calls "fine hybrid magnetic particles" that combine barium ferrite and strontium ferrite. Fujifilm claims the top global share of the tape market; that is the company's own statement.

Why would AI need tape?

Training and running models produces a lot of data that is rarely read again: raw crawls, earlier dataset versions, checkpoints, logs, synthetic data, evaluation outputs and agent traces. Keep all of it on flash or spinning disk and the bill grows quickly.

Blocks and Files author Chris Mellor argues that AI training data is low-access, that holding it on disk would be prohibitively expensive, and that tape is the only low cost-per-terabyte medium available today. That is an analyst view, not a Fujifilm statement, but the economics are easy to follow. Tape cartridges sit on a shelf and use no power when idle.

The same pressure shows in other parts of the memory and storage chain. Our posts on Anthropic's S-1 supplier risk factors and on Apple's Ternus on the memory crunch cover the hot end, where high-bandwidth memory and DRAM are scarce. Tape is the cold end of the same story.

A green chip in front of server towers, showing storage tiers behind AI data centers

How big is LTO-10, and what comes next?

Per Blocks and Files, an LTO-10 cartridge holds 40 TB raw, or 100 TB compressed at a 2.5 to 1 ratio. LTO-11 is expected to reach up to 70 TB raw, or 175 TB compressed, around 2028 or 2029. Compression ratios depend on the data. Model weights and already-compressed media compress poorly, so plan on the raw number for those.

Is five times really five times?

Read the headline with care.

• Weak base year. TNW reports that LTO shipments fell 9% to 160.3 exabytes across 2025. A multiple measured against fiscal 2025 capacity starts from a soft base.
• Capacity is not sales. The new lines will not be fully running before 2028, and Fujifilm has not said how much of the output is committed to customers, according to HWBusters' coverage.
• No absolute baseline. The announcement does not give a fiscal 2025 capacity figure, so the final tonnage is unknown.
• Demand evidence is partial. Blocks and Files reports a 57% jump in tape shipments in Q1 2026, driven by LTO-10 launches and archiving of AI training data. One quarter is not a trend, and part of the jump may be buyers moving to the new generation.
• The AsiaNikkei report is paywalled. We read the Fujifilm release directly, so the core facts do not depend on it. We did not read the Nikkei piece.

A short history: why tape never went away

Tape is old, and it has been declared dead many times. It survived because the arithmetic of cold storage did not change. Cost per terabyte for tape stays below disk, and a cartridge on a shelf draws no power. Cloud providers quietly kept using it for deep archives while the public conversation moved to flash.

Fujifilm's own timeline fits that pattern. It began tape research in the 1950s, shipped LTO Ultrium 1 in 2000, and moved the industry to barium ferrite particles in 2011. Each generation raised capacity per cartridge, which is the lever that keeps cost per terabyte falling. LTO-10 at 40 TB raw is a large step from the first LTO generations, and the next step to 70 TB raw is the reason buyers keep tape on their roadmap.

What changed in 2026 is the type of data. Old archive demand came from video, medical imaging and compliance records. AI adds training corpora, model checkpoints and agent logs. None of these needs millisecond access, but all of them are large, and many must be kept for audit or reproducibility. A lab that wants to re-run an old training job needs the exact data from months ago. That is a tape-shaped problem.

Questions a skeptical reader should ask

Is this just a company talking up its product? Partly. The release is a company statement, and Fujifilm gains if people believe AI drives tape demand. The investment is real money, though, and a ¥10 billion commitment is harder to fake than a press line. The Q1 shipment figure comes from an independent outlet.

Will AI labs really archive on tape? Fujifilm names cloud providers and data center operators as the main buyers, not the labs directly. Labs that train on rented cloud capacity touch tape only through the provider's archive tier.

Could better compression or deduplication cut demand? Yes. Training data is often duplicated across versions. Teams that deduplicate aggressively need less archive space. That is one reason a five-times capacity target is a bet, not a certainty.

What are the alternatives?

Blocks and Files names several, each far from replacing tape today:

Tape library makers such as Quantum and Spectra Logic are named as likely beneficiaries if library shipments rise, for example to neoclouds.

What does this mean for what you build or pay?

For most builders, nothing changes this week. The practical points are about planning.

1. Tier your data. Hot data (current datasets, checkpoints in use) stays on fast storage. Cold copies belong in an archive tier. Cloud archive services are often tape-backed under the hood.
2. Expect storage prices to stay firm. If hyperscalers buy more tape, disk and flash and memory supply remains tight at the top end, which keeps pressure on prices.
3. Use offline copies against ransomware. An air-gapped copy cannot be encrypted by malware on a connected machine. Our coverage of data center protests over power and water and the public polling on data centers shows why efficient storage matters: every terabyte kept on powered disk adds to the energy debate.
4. Do not over-read it. A single maker's capacity plan, announced by the company itself, is a signal about expectation, not proof of demand.

If you run AI agents against your archives, the safety side matters too. A restore job that an agent can trigger is a powerful action, so put approval steps in front of it.

How it fits the wider AI-hardware picture

AI demand now touches nearly every component: GPUs, high-bandwidth memory, flash, power equipment and cooling. Our posts on the California data center bills and on DeepSeek's KV-cache memory reduction show two ends of the same constraint: policy limits on power and software tricks to use less memory. Tape adds a third answer for the bulk data that never needs to be fast.

What to watch

• Whether Sony announces matching capacity.
• Fujifilm's absolute capacity and shipment figures in later results.
• LTO shipment data for the next quarters, to see if the 57% jump holds.
• LTO-11 timing and capacity.
• Whether tape library suppliers report order growth from AI-focused cloud providers.

Related reading

• Anthropic S-1 supplier risks: TSMC, HBM and memory makers
• Apple's Ternus on the memory crisis
• DeepSeek V4.1 Flash and HBM reduction
• California data center and AI infrastructure bills
• Thane data center protests
• Poll: 66 percent oppose data centers

Figures are as stated by Fujifilm and the cited outlets on November 2, 2026.