Fujifilm will raise its magnetic tape production capacity fivefold, according to Nikkei — the company's largest investment in the medium in roughly two decades. The reason is not nostalgia. It is the archive tier of AI infrastructure.

What changed

Tape spent the 2010s as the format everyone predicted would die. It survived in bank vaults, broadcast archives and research labs precisely because it is bad at the thing it is not for: it reads sequentially, so a cartridge takes minutes, not milliseconds, to reach a file. But the workloads AI produces are mostly the opposite of interactive. A training corpus is written once, read once, and then sits untouched for years while regulators, auditors and reproducibility requirements insist it stay. Model checkpoints accumulate the same way, and so do inference logs.

A shelf of cartridges draws no power. A rack of idle disks does — and needs cooling. That difference turns tape from a legacy line item into the cheapest place to park a petabyte, and it is why the same technology that lost the consumer market keeps winning the bottom of the storage hierarchy.

The numbers

The LTO consortium's own figures show the reversal. Tape capacity shipments reached 160.3 exabytes in 2025, down 9% year on year — then jumped 57% year on year in the first quarter of 2026, which the consortium attributed to AI data growth, archival demand, air-gapped backups and the start of the LTO-10 transition.

LTO-10 holds 30TB natively, or 40TB in the extended-length cartridge that shipped later. The density jump is real: 15,104 data tracks against 8,960 in LTO-9, enabled by a strontium-doped barium ferrite recording layer and a redesigned servo. The 40TB version adds an aramid base film thin enough to fit roughly 1,335 metres of tape into the same shell, against about 1,035 metres before — and it works in the same drives. The roadmap, deliberately slowed in late 2025 to favour reliability and cost per terabyte over doubling every generation, now targets 70TB for LTO-11, 120TB for LTO-12, 210TB for LTO-13 and 365TB for LTO-14. A 2020 IBM and Fujifilm demonstration at 317 Gb/in² suggested 580TB per cartridge is physically reachable.

What has not improved much is throughput. LTO-10 still runs at 400 MB/s native, the same as LTO-9; writing a 40TB cartridge sequentially takes about 28 hours before overheads. And LTO-10 dropped backward compatibility entirely, so anyone adopting it must migrate LTO-8 and LTO-9 archives or keep the old drives alive.

Why one supplier matters

Fujifilm and Sony dominate global tape media production. A fivefold expansion by one of the two effectively raises the ceiling for the whole supply chain, at a moment when the constraint has been media rather than drives, libraries or robots. Fujifilm supplies the particulate coatings that determine density and also builds finished cartridges sold under its own name and packaged for others, which is why its coating and slitting capacity is the number to watch.

The risk is concentration in the other direction too. Tape revenue depends on a small set of hyperscale buyers; if AI infrastructure spending slows, the new lines idle. Fujifilm's bet is simply that the archive tier grows with the training tier — that every model produces data which outlives the hardware that trained it, and that the cheapest place to keep it will still be a cartridge on a shelf.

Sources