THE SHORT VERSION
- Fujifilm will invest about 10 billion yen, roughly $63 million, to raise magnetic tape production capacity fivefold against fiscal 2025 levels.富士胶片将投入约 100 亿日元（约 6300 万美元），把磁带产能提高到 2025 财年的五倍。
- The company attributes the demand to generative AI's exponential growth in data and to cloud operators looking for cheap, air-gapped archives.公司称需求来自生成式 AI 带来的数据指数级增长，以及云厂商对廉价、可物理隔离档案的需求。
- LTO cartridge shipments fell in 2025 to 160.3 exabytes from 176.5 exabytes in 2024, though the LTO Program expects a record year in 2026.LTO 磁带盒出货量在 2025 年降至 160.3 EB（2024 年为 176.5 EB），但 LTO 计划预计 2026 年将创纪录。
The size and shape of the bet这笔投资的规模与形状
Fujifilm said it will invest roughly 10 billion yen, about $63 million, to increase its production capacity for magnetic tape fivefold. Ars Technica reported that the increase is measured against fiscal 2025 production levels, that the money goes to the Kanagawa Factory's Odawara site, and that the project runs from fiscal 2026 through 2028.富士胶片表示将投入约 100 亿日元（约 6300 万美元），把磁带产能提高五倍。Ars Technica 报道称，这一增幅以 2025 财年产量为基准，资金投向神奈川工厂小田原厂区，项目期从 2026 财年到 2028 财年。
The company called it the largest investment it has made to date in dedicated LTO tape facilities. The output is aimed primarily at the latest generation of the format, the FUJIFILM LTO Ultrium10, which is the product line that a fivefold expansion would need to sell in volume to make sense.公司称这是其迄今在专用 LTO 磁带设施上最大的一笔投资。新增产能主要面向最新一代产品 FUJIFILM LTO Ultrium10——五倍扩张只有在产线产品能大量卖出时才说得通。
Why AI creates a tape customerAI 为什么需要磁带
Fujifilm's explanation is about volume rather than speed. The company said the spread of generative AI is producing exponential growth in the amount of data generated worldwide, and that building environments where the data accumulated through AI training and inference can be stored securely and efficiently is becoming more important. The Register reported the same statement and the same number: five times more LTO production, on a 10 billion yen plan.富士胶片的解释关乎数据量，而不是速度。公司称生成式 AI 的普及正带来全球数据量的指数级增长，建立能安全高效存储训练与推理所积累数据的环境，正变得越来越重要。The Register 报道了同一份声明与同一个数字：LTO 产能提高五倍，投入 100 亿日元。
The buyers are not consumers. Fujifilm said major cloud service and data-centre operators have been expanding their use of tape because it stores large volumes at relatively low cost, and because cartridges can be disconnected and stored in a vault. The company also pointed to air-gapping as a defence against ransomware, which is the argument that has kept tape in data centres long after the consumer market moved on.买家不是消费者。富士胶片称，主要云服务与数据中心运营商一直在扩大磁带使用，因为它能以相对低的成本存储大体量数据，而且磁带盒可以断开连接、存进保险库。公司还强调物理隔离可作为勒索软件的最后一层防线——这正是消费市场远离磁带多年后，它仍留在数据中心的原因。
The market shrank last year这个市场去年是缩小的
The expansion is a bet against the most recent number. The Register reported that LTO cartridge shipments went backwards in 2025: buyers took 160.3 exabytes of tape, against a record 176.5 exabytes in 2024. A company adding fivefold capacity is not extrapolating that line; it is assuming the next few years look different.扩张是对最近一个数字的反向下注。The Register 报道称，2025 年 LTO 磁带盒出货量出现倒退：买家购买了 160.3 EB 的磁带，而 2024 年为创纪录的 176.5 EB。把产能提高五倍的公司并不是在延续这条曲线，而是在假设未来几年不一样。
The LTO Program, the industry group behind the format, attributed the 2025 result to difficult macroeconomic conditions, noted that sales picked up strongly in the second half, and predicted record sales in 2026. That forecast is the assumption the investment rests on, and it is a forecast rather than a reported order book.LTO 计划——推动该格式的行业组织——把 2025 年的结果归因于艰难宏观环境，并指出下半年销售强劲回升，预计 2026 年将创纪录。这个预测正是投资所依赖的假设，而它只是预测，不是已披露的订单簿。
What the format promises next这个格式接下来承诺什么
The roadmap explains why an archive product can justify new factories. The Register reported that tenth-generation LTO tapes come in variants with native capacity of 30TB or 40TB and transfer data at 400MB/s, and that many users compress data to more than double each cartridge's capacity. The trade-off is that decompression takes time and moving data off tape is slow.路线图解释了为什么一个慢速介质值得投新工厂。The Register 报道称，第十代 LTO 磁带原生容量有 30TB 与 40TB 两种，传输速率 400MB/s，许多用户还会压缩数据，把每盒容量翻倍以上。代价是解压需要时间，把数据从磁带取出也不快。
Further capacity is already scheduled. The LTO Program has defined four more generations with updates roughly every four years: LTO-11 around 2029 at 70TB, LTO-12 at 120TB, LTO-13 at 210TB and LTO-14 at 365TB per cartridge. A buyer planning a ten-year archive can see the next four steps, which is not true of every storage medium.更大的容量已在计划中。LTO 计划已定义后续四代，约每四年更新一次：LTO-11 约 2029 年达 70TB，LTO-12 达 120TB，LTO-13 达 210TB，LTO-14 达每盒 365TB。规划十年档案的买家能看到未来四步，这并非每种存储介质都能提供。
What tape is still bad at磁带仍然不擅长什么
Tape is an archive, not a working disk. Access times are not the medium's strong point, which is why tape users employ it for offline or nearline archives and expect to wait. Any workload that needs a file now, or needs it repeatedly, is a poor fit, and AI training data that is re-read constantly is exactly that kind of workload.磁带是档案，不是工作盘。访问时间不是它的强项，因此用户把它用于离线或近线档案，并接受等待。任何需要立刻取文件、或反复读取的工作负载都不适合它——而训练数据恰恰是会被反复读取的那类负载。
Compatibility has also narrowed. The Register reported that the LTO Program used to insist drives could read the current and past generations of cartridges, and that this ended with LTO-10 devices, which can read only the current generation. A fivefold expansion in LTO-10 output therefore assumes buyers are ready to move their archives onto the newest generation rather than keep reading older media.兼容性也在收窄。The Register 报道称，LTO 计划过去坚持驱动器可读取当前与过去数代磁带，但这一点在 LTO-10 设备上终结：它们只能读取当前一代。因此，LTO-10 产量提高五倍，前提是买家愿意把档案迁到最新一代，而不是继续读旧介质。
What is established, and what is not哪些已经确定，哪些还没有
Established: Fujifilm will spend about 10 billion yen to lift tape capacity fivefold against fiscal 2025; the project runs from fiscal 2026 to 2028 at the Odawara site; the company attributes demand to generative AI data growth and to cloud operators' archive needs; LTO shipments fell to 160.3 exabytes in 2025; the LTO roadmap runs to 365TB per cartridge. Not established: whether AI demand keeps growing fast enough to absorb the new capacity, whether buyers adopt LTO-10 at the pace the plan needs, and what the air-gap argument is worth once archives are written once and rarely read.已经确定的：富士胶片将投入约 100 亿日元，把磁带产能提高到 2025 财年的五倍；项目 2026–2028 财年在小田原推进；公司把需求归因于生成式 AI 数据增长与云厂商的档案需求；2025 年 LTO 出货量降至 160.3 EB；LTO 路线图到每盒 365TB。尚未确定的：AI 需求能否增长到吸收新增产能、买家是否会按计划采用 LTO-10，以及当档案写一次、极少读取时，物理隔离的理由还值多少。
For readers, the useful way to read this is as a physical consequence of the AI build-out. Data centres are usually discussed in terms of chips and power; this is the shelf at the end of the pipeline. This site's coverage of OpenAI's reported plan to buy Google chips and of Meta's $97.5 billion AI equity raise describes the spending that creates the data, and Proton's data-access dispute shows how much of the argument is about where that data is allowed to sit.对读者而言，有用的读法是把这件事看作 AI 建设的物理后果。讨论数据中心时通常讲芯片与电力，而这是管线末端的货架。本站关于OpenAI 拟采购谷歌芯片与Meta 975 亿美元 AI 股权融资的报道，描述的是制造这些数据的支出；而Proton 的数据访问争议则说明，争论的很大一部分在于这些数据被允许存放在哪里。
A slow medium with a long tail慢介质的长尾生意
Tape economics are about cost per terabyte held over years, not about speed. A cartridge sitting in a vault costs almost nothing to keep, which is why operators use it for the copy they rarely read: a backup, a compliance archive, a record of what a model was trained on. The AI build-out produces exactly that kind of data at a scale that makes cheap shelves attractive.磁带的经济学讲的是每 TB 多年持有的成本，而不是速度。一盒放在保险库里的磁带，保存成本几乎为零，因此运营商把很少读取的副本放在上面：备份、合规档案、模型训练数据记录。AI 建设正好以这样的规模生产这类数据，让廉价货架变得有吸引力。
The same roadmap that makes the plan plausible also contains its risk. If LTO-14 reaches 365TB per cartridge, the same archive needs roughly a tenth of the cartridges it takes at 40TB, so a fivefold increase in production capacity is a bet that total archived data grows faster than each cartridge grows. That is a reasonable bet in an AI boom and a poor one if the boom cools.让计划显得合理的那张路线图，也包含着风险。如果 LTO-14 每盒达到 365TB，同样的档案所需的磁带盒数量只有 40TB 时代的约十分之一；因此把产能提高五倍，是在押注档案总量的增长快于单盒容量的增长。在 AI 热潮中这是合理的下注，热潮降温后则未必。
Sources & reporting limits来源与核查范围
Follow Industry to find related stories on your next visit.关注产业动态，下次访问时阅读相关文章。
