Who would have thought that magnetic tape, forgotten by consumers for decades, would find a second life in the AI era? On October 29, Japan's Fujifilm announced an investment of approximately 10 billion yen (about 422 million RMB) to expand its Odawara factory in Kanagawa Prefecture during the 2026-2028 fiscal years, increasing data storage tape production capacity to five times the 2025 level, with a focus on the latest generation LTO Ultrium 10 tapes. According to #Fujifilm, this is the largest investment in #LTO tape-specific facilities in nearly 20 years, driven primarily by the massive data archiving demands of generative AI.
AI uses the most advanced chips, so why does it still need tape? A classic joke circulates in the computer industry. Computer scientist Andrew Tanenbaum once said, "Never underestimate the bandwidth of a station wagon full of tapes hurtling down the highway." It sounds absurd, but when data volumes reach PB or even EB levels, loading storage devices onto trucks can indeed be more efficient than transmitting over networks. In 2016, AWS launched its Snowmobile service, using heavy trucks to transport large storage devices, capable of moving up to 100PB of data at a time. Although Snowmobile did not transport tapes and was discontinued in 2024, it illustrates a long-standing problem: when data scales to a certain point, how to store and move it at a reasonable cost is itself an engineering challenge.
The explosion of generative AI has made this problem increasingly prominent. Training large models requires GPUs, HBM, and high-speed SSDs, but after training, enterprises are left with raw corpora, cleaned datasets, training logs, model checkpoints, and historical versions from continuous iteration. Data actively involved in computation needs high-speed storage, frequently accessed large datasets can reside on HDDs, while materials that won't be used for months or even years but cannot be casually deleted are suitable for long-term archiving systems. The industry typically calls infrequently accessed data cold data. For example, after an AI company trains a new model, the old dataset may have no short-term use, but it may still be needed for reproducing experiments, retraining, or compliance audits in the future. The problem follows: as data accumulates, should all hard drives stay powered on, spending money to maintain files that may never be accessed for years?
Tape happens to excel at this. First, it's cheap and energy-efficient. Tape can be stored offline for long periods without continuous power, and it doesn't require the equipment operation and cooling that online disk arrays need. Fujifilm has estimated that under specific conditions of storing 100PB of data over a ten-year period, automated tape systems can reduce total cost of ownership by about 86% and carbon emissions by about 95% compared to all-HDD solutions. Second, it's secure. Tape can be removed from the system and stored offline; even if servers are hit by ransomware, physically isolated tapes retain an independent backup. Moreover, modern tape is not as slow as imagined. LTO-10 offers raw continuous transfer rates up to 1200MB/s, and automated tape libraries can use robotic arms for loading and reading.
Of course, its drawbacks are equally obvious. Tape requires sequential seeking, and random access to a single file can take considerable time, making it unsuitable for frequent access and real-time computing. If SSDs in AI data centers are responsible for making data instantly available, tape is more like a low-cost, massive digital warehouse. In fact, the scientific community has long used this approach. In December 2025, CERN announced that the Large Hadron Collider had accumulated 1EB of experimental data—equivalent to 1 million TB—most of which is stored on approximately 60,000 tape cartridges. Experimental data collected decades ago may regain research value due to new theories. CERN estimates that the scale of data it needs to preserve and process in the next decade will be about ten times its previous cumulative storage. Similar logic applies to AI companies preserving historical corpora, training records, and model backups.
The growth of the tape market is already supported by actual data. According to figures released by the LTO Consortium in September 2026, global LTO tape shipment capacity reached a record 176.5EB in 2024, up 15.4% year-over-year. Although it declined somewhat in 2025, it still reached 160.3EB. In the first quarter of 2026, with the launch of LTO-10 and growing archiving demand, global tape shipment capacity grew 57% year-over-year. These figures measure the storage capacity provided by tape, not cartridge sales. Enterprise tape has always had a stable market, and generative AI's demand for massive data archiving has given this old technology new growth opportunities.
There is another easily overlooked reason why tape is regaining attention in the AI era: while consumers assumed tape technology had stagnated, engineers have been continuously increasing its storage density. In June 2025, Fujifilm launched LTO-10 30TB data tape, followed by a 40TB version in December, with shipments starting in 2026. The latest product offers 40TB raw capacity per cartridge; at an ideal 2.5:1 compression ratio, theoretical capacity can reach up to 100TB. By comparison, the first-generation LTO tape launched in 2000 had only 100GB raw capacity. In just 26 years, per-cartridge capacity has increased 400 times. However, 100TB is theoretical compressed capacity; for already-compressed video, images, and model weights, actual capacity typically falls short of this figure.
More interestingly, an LTO-10 cartridge is only about 10 cm square and 2.15 cm thick, yet contains 1,337 meters of tape—more than three laps around a standard athletics track. To fit 1,337 meters of tape into a cartridge just 10 cm square, Fujifilm introduced aramid film and new thin-layer manufacturing technology, reducing overall tape thickness to just 4 micrometers, while using ultrafine magnetic particles combining barium ferrite and strontium ferrite technologies to increase data density per unit area. The storage principle is essentially the same as audio tape from decades ago. The tape surface is coated with tiny magnetic particles; when writing data, the head changes local magnetization states to record digital information; when reading, it detects magnetic field changes and restores data through decoding and error correction. Modern LTO tape uses linear serpentine recording, which can be imagined as a highway with tens of thousands of lanes. The head writes from start to finish along a set of tracks, then repositions and writes back in the opposite direction. The latest LTO-10 has 15,104 data tracks, and drives can read and write 32 tracks simultaneously, continuously increasing capacity through more precise materials and servo control technology.
Moreover, tape capacity growth is far from over. In 2020, IBM and Fujifilm achieved a magnetic recording density of 317Gb per square inch in the lab, demonstrating the technical potential for future single cartridges reaching 580TB—14.5 times current 40TB commercial products. Of course, 580TB remains at the lab demonstration stage, with a long road to commercialization. Tape also has an often overlooked long-term cost: compatibility. Tape media can last for many years, but that doesn't mean new drives decades later will still be able to read it. The latest LTO-10 drives can only read and write LTO-10 tape, so enterprises need to retain old equipment or periodically migrate historical data to new media. For data centers managing tens of thousands of tape cartridges, data migration, verification, and equipment maintenance are also significant expenses. Even so, in large-scale cold data archiving scenarios, tape remains highly competitive in cost and energy consumption. Fujifilm's decades of accumulated expertise in precision coating, thin-film materials, and magnetic particle technology have thus regained commercial value in the AI era.
The tape story is also playing out in the HDD industry. In January, Western Digital revealed that its HDD production capacity for calendar year 2026 is essentially sold out, with long-term purchase agreements from some large customers extending to 2028. Seagate's situation is even more extreme. According to its late-October earnings call, the vast majority of its nearline HDD shipment capacity has been allocated through long-term agreements to 2028.
[AgentUpdate Depth Analysis] The revival of tape in the AI era appears to be a commercial story about storage hardware, but at a deeper level, it reveals a severely underestimated bottleneck in the AI Agent ecosystem: data persistence and long-term memory. As AI Agents evolve from single-turn conversations to continuously running, enterprise-grade autonomous systems, the interaction logs, environment states, decision trajectories, and model checkpoints they generate will grow exponentially. Most of this data falls into the "warm-cold" category—not needed for real-time inference, but critical for agent auditing, traceability, compliance, and retraining. Current mainstream Agent frameworks like LangChain and CrewAI rely almost entirely on online databases or object storage, which is costly and energy-intensive. Tape's ultra-low cost, offline security, and physical isolation make it an ideal layered storage paradigm for Agent "long-term memory": hot data on SSD/memory, warm data on HDD, and cold archives on tape. In comparison, cloud providers' archive storage (such as AWS Glacier) is also cheap, but tape still holds an advantage in TCO at EB scale and is not limited by network bandwidth. In the future, we may see Agent-native storage middleware emerge that automatically tiers Agent runtime history to tape libraries based on access frequency. Fujifilm's capacity expansion may look traditional, but it is actually laying a key foundation for the sustainable scaling of the AI Agent ecosystem—without cheap long-term memory, there can be no truly reliable autonomous agents.
