For a technology that most people assumed had gone the way of the floppy disk, magnetic tape is enjoying a remarkable second act. Fujifilm, one of only a handful of companies still making the stuff, has signalled it will ramp up production dramatically, and the reason is the same force reshaping almost every corner of computing: artificial intelligence.
According to a report from Ars Technica, the Japanese imaging giant intends to lift its tape manufacturing capacity by roughly five times to keep pace with a surge in demand for cheap, long-term data storage. The driver is the enormous volume of information that AI systems both consume and generate, from the training datasets behind large language models to the logs, checkpoints and outputs that pile up once those models are in production. All of that has to live somewhere, and increasingly, it is landing on tape.
Why a format from the 1950s still has a job
Magnetic tape has been written off repeatedly over the decades, yet it keeps surviving because it does two things better than almost anything else. It is extraordinarily cheap per terabyte, and when a cartridge is sitting on a shelf rather than plugged into a drive, it is completely offline. That second quality has become unexpectedly valuable in an era of ransomware and state-sponsored intrusion. A tape that is not connected to any network cannot be encrypted by an attacker or quietly exfiltrated, which is why the industry talks about an “air gap” as a genuine security feature rather than a limitation.
The modern incarnation of the format, known as Linear Tape-Open or LTO, is developed by a consortium that includes Hewlett Packard Enterprise, IBM and Quantum, with Fujifilm and Sony doing much of the actual media manufacturing. Each new generation roughly doubles capacity, and a single current-generation cartridge can hold tens of terabytes of compressed data. The trade-off has always been speed. Tape is slow to access because a drive has to physically spool to the right spot, which makes it hopeless for anything you need in a hurry but close to ideal for data you want to keep for years and rarely touch.
That profile fits the AI build-out neatly. Training a frontier model produces mountains of intermediate data that organisations are reluctant to delete, partly because it may be useful again and partly because regulators and auditors increasingly expect a trail. Cold storage, the industry term for archives that are seldom read, is exactly where tape shines, and the economics become compelling at the scale hyperscale operators now work at.
Two ways to read the move
One view is that Fujifilm’s expansion is a straightforward bet on a structural shift. If AI keeps generating data faster than anyone can afford to store it on spinning hard drives or flash, then the cheapest durable medium wins the overflow, and tape is that medium. On this reading, the company is simply positioning itself to supply a shortage that the AI boom has made inevitable.
A more sceptical take notes that storage predictions have a poor track record, and that every few years someone declares tape either dead or resurgent depending on the cycle. Hard drive makers continue to push capacities higher, and the gap between “cold” and “warm” storage keeps blurring as cloud providers invent new intermediate tiers. A fivefold capacity increase is a serious capital commitment, and it assumes demand holds up even if the current AI investment frenzy cools. If model builders pull back, or if they get better at deleting data they do not need, the archive mountain may not grow quite as fast as the bullish case suggests.
There is also a concentration question worth sitting with. The world now relies on a very small number of suppliers for a format that underpins a surprising share of its long-term memory. That makes the supply chain efficient but fragile, and it hands significant leverage to whoever controls manufacturing capacity. A fivefold expansion by one dominant player is good news for buyers chasing supply, but it does little to broaden a market that is already uncomfortably narrow.
What it means for Australia
Australia is in the middle of a data centre construction wave, and the same AI demand pushing Fujifilm to expand is reshaping local infrastructure plans. Operators in Sydney, Melbourne and increasingly Western Australia are weighing how to store colossal volumes of data affordably, and tape is quietly part of that calculation even when the headlines are all about GPUs and power supply.
The offline security angle lands hard here too. Australian organisations have spent the past few years absorbing a brutal run of cyber incidents across health, telecommunications and financial services, and boards have learnt that a backup an attacker can reach is not really a backup at all. An air-gapped tape archive offers a last line of defence that survives even a total network compromise, which is precisely why many enterprises and government agencies never fully abandoned the format. For sectors bound by long data-retention rules, including health records, financial compliance and government archives, the combination of low cost and genuine offline isolation is hard to beat.
There is a sovereignty dimension as well. As Canberra and industry continue to debate what sovereign AI capability should actually mean for the country, where and how Australian data is stored becomes part of the conversation. Tape does not solve the question of who owns the compute, but it does make it more affordable to keep large national datasets onshore for the long haul rather than shipping them to cheaper storage offshore.
What is next
The practical test will be whether Fujifilm’s expanded output actually eases the supply pressure that has frustrated large buyers, and whether pricing stays low enough to keep tape attractive against falling flash costs. The LTO roadmap continues to promise bigger cartridges, so the capacity story is unlikely to stall any time soon. For Australian technology leaders, the lesson is less about one company’s factory plans and more about a quiet truth of the AI era: the glamorous end of the stack still depends on an unglamorous foundation, and someone has to pay to remember everything the machines produce.
Sources: Ars Technica.
