🌾 Welcome to StableBread’s Newsletter!
It's no secret that AI's massive growth is being fueled by the data its models create and consume. That means every new dataset, training run, and inference log has to live somewhere, which is where storage sits in the AI infrastructure value chain.
International Data Corporation (IDC), one of the most widely cited research and advisory firms in technology (alongside Gartner and Forrester), expects that 394 zettabytes (1 zettabyte = 1 billion terabytes) of data will be generated in 2028, representing a 2023-2028 CAGR of 24%.
But where exactly is all of this data stored? Most of the AI stack lives on a mix of flash/NAND and hard disk drives (HDDs):
Flash/NAND: Sits closest to the GPUs and handles the hot, high-throughput workloads. Beneficiaries include Micron (MU), SanDisk (SNDK), and Pure Storage (PSTG).
HDDs: Handle the warm and cool tiers, where data still gets read regularly but doesn't need flash speeds. Beneficiaries include Seagate (STX) and Western Digital (WDC).
HDDs carry the bulk of that data in practice. In a March 2026 study by Scality and Freeform Dynamics of 504 enterprises running private AI in production, 91% reported meaningful use of object storage (which typically runs on HDDs), the highest adoption of any storage architecture.
But with AI requiring petabytes (1 petabyte (PB) = 1,000 terabytes) of training data to be kept available for multi-year retention and retraining, flash and HDD become cost-prohibitive once you're storing that much for that long.
Magnetic tape is the exception. It's been around since the 1950s, and for most of the last forty years it lived quietly behind overnight backups and compliance archives.
The modern enterprise tape standard is Linear Tape-Open (LTO), an open format jointly developed by IBM, HPE, and Quantum (QMCO) that celebrated its 25th anniversary in 2025.
Magnet Tape’s Value
According to Qualstar, ~80% of the world's data is cold storage, meaning it's rarely accessed but still has to be retained. That's the slice tape was built for, and the cost gap over flash and HDD widens as archive volumes grow.
Horison's March 2024 analysis, sponsored by Fujifilm (one of only two manufacturers of LTO tape media, alongside Sony), cites Brad Johns Consulting's updated LTO-9 TCO model.
Storing 100 PB for 10 years on all-HDD runs $17.7M. A 60% tape / 40% disk hybrid drops the bill to $9.5M, a 46% reduction once media, power, cooling, and floor space are factored in. Moving everything to tape cuts costs by 78%.
And cost is just one of several advantages tape has over flash and HDD when you're managing petabytes of long-retention data:
Power efficiency: A tape cartridge sitting in a library draws zero power, except when actively reading or writing data (~1% of the time). In contrast, HDDs spin continuously regardless of whether data is being accessed, which is why tape has become a fixture in data-center sustainability and carbon reporting.
Longevity: LTO media built on Barium Ferrite (BaFe), the magnetic material used in modern tape cartridges, is rated by Fujifilm and the JEITA industry body for 50+ years of archival life (vs. HDD of 6-7 years).
Reliability: LTO-10 is rated at 1 uncorrectable bit error per 1020 bits read, 100,000x more reliable than enterprise SATA HDD.
Air-gap security: A cartridge physically disconnected from the network cannot be remotely encrypted, which is why cyber insurers increasingly require immutable storage and air-gapped or offline backup copies before writing a ransomware policy.
LTO-10 began shipping in mid-2025 at 30 TB native per cartridge, with a 40 TB version slated for Q1 2026, and the published LTO roadmap extends through Generation 14 (updated November 2025):
This gives archive operators a committed, multi-vendor growth path over the next decade without format replacement.
Market Opportunity
Cost, power efficiency, longevity, reliability, and air-gap security are not new advantages to magnet tape. What's new is that shipments, pricing, and deployment scale are finally reflecting them.
Shipment growth re-accelerated
The LTO Program's annual shipment report showed 176.5 EB of tape capacity shipped in 2024, up 15.4% y/y and the fourth consecutive year of growth.
The report cited unstructured data growth from AI/ML workloads as the driver.
Component prices squeezed the alternatives
Quantum (QMCO) is a tape drive and library maker and one of the three LTO technology provider companies alongside HPE and IBM.
On its Q3 FY26 earnings call on February 17, 2026, Quantum CEO Hugues Meyrath described an industry-wide supply shock in competing storage:
“Critical components, particularly memory, disk, and flash, are becoming increasingly difficult to procure, and prices are rising as a result. In just the last 10 days, we've seen pricing double and in some cases triple. This is not unique to Quantum; it's affecting the entire industry. In addition to pricing volatility, lead times are extending into weeks and sometimes even months.”
Meyrath said the pricing and lead-time volatility was "similar to the early COVID period," with the path to stabilization unpredictable. He then connected the shock directly to tape demand:
“As flash and disk become more expensive and harder to secure, we believe tape provides customers with a practical way to offload massive volumes of data to a cold tier, freeing up primary storage while meaningfully reducing power, cooling, and operating costs.”
Quantum said tape sales doubled quarter-over-quarter as a result. Backlog ended Q3 FY26 at over $20M, well above the historical $8-10M run rate.
Deployments went from pilot to production
On the same call, Meyrath disclosed a seven-figure deal with a large multinational production studio running ActiveScale Cold Storage integrated with Scalar tape libraries.
The customer was repatriating content from the cloud to on-premise archive, with an initial deployment of 100 petabytes and plans to scale to 400 petabytes over time.
That's not a pilot. It's a production-scale commitment with a 4x scaling path.
Competitive Threats
The question for an archive buyer and investor is whether something newer is about to displace it.
These are the candidates I found, but none are shipping:
Ceramic-on-glass: Data is laser-etched into a thin ceramic layer coated on a glass substrate, then read back optically. The glass tablets last longer and can hold more data than a tape cartridge. Cerabyte is the lead company. It secured a strategic investment from Pure Storage in July 2024 and a follow-on strategic investment from Western Digital in May 2025, terms undisclosed on both. One of the world's two largest HDD makers backing a ceramic archive startup is not a trivial signal. However, it’s still in pilot production at 1 PB per rack, with the 100 PB per rack product targeted for 2030.
Quartz glass: A femtosecond laser writes data as three-dimensional marks inside a block of quartz glass. The pitch is millennia-scale durability with no media migration ever required. Microsoft's Project Silica published an end-to-end system demonstration on February 18, 2026, but has not disclosed a product release date, pricing, or a mass-manufacturing partner. Current write throughput is ~8 MB/s, roughly 50x slower than LTO-10's 400 MB/s native, which keeps it in write-once, read-rarely territory.
Synthetic DNA: Binary data is encoded into synthetic DNA strands stored in vials. The density is extraordinary in theory, but write/read throughput is slow and the cost curve is nowhere near commercial scale. Atlas Data Storage, spun out of Twist Bioscience in May 2025 with $155M in seed funding, is the most capitalized pure-play.
None of them are shipping commercial product at any price today, let alone at a $/TB that beats tape. None have a committed multi-vendor roadmap comparable to LTO's cadence through Generation 14.
That does not mean these technologies go nowhere. But the investable question today is not what wins in 2030. It's who's getting paid to store archive data right now. That’s tape.
Who’s Actually Using Tape
The biggest misconception about tape is that it's a legacy medium kept around
by laggards. The opposite is true. The largest data owners in the world run
tape at scale, and most have said so publicly:
CERN: Archives its Large Hadron Collider physics data on magnetic tape. LHC Run 3 is planned to exceed 600 PB of stored data, with tape designated as the main long-term storage medium.
Meta: Runs exabyte-scale cold storage with tape in the coldest tier, behind HDD and flash. Training datasets are projected to migrate to that tier as models age.
Microsoft: Uses tape for archival storage inside Azure. A Microsoft Principal Developer Manager has publicly walked through the architecture.
IBM: Built the Diamondback tape library specifically for hyperscalers at 1.8x the petabytes per square foot of competing libraries, targeted explicitly at Google, Microsoft, AWS, and Meta.
AWS: Has not publicly disclosed the media behind S3 Glacier Deep Archive. Treat as a probable tape user, not a confirmed one.
Financial institutions sit in their own bucket. SEC Rule 17a-4 historically
required WORM (write-once, read-many) media for broker-dealer records. The
2023 amendments added an audit-trail alternative, but WORM tape and optical
remain valid paths and most of the installed base stayed on tape.
The tape question is where those datasets live after training is done. That decision belongs to the customer (Meta, Anthropic, OpenAI, Microsoft) or the customer's hyperscaler, not the neocloud. Tape sits downstream at the archive layer, where cost per terabyte over a decade is what wins.
The Pure Plays
So which companies are best-suited to capture the upside from renewed tape demand?
The investable universe is smaller than it looks. Most names that touch tape treat it as a side business. Two actually qualify.
Subscribe to StableBread's Research to keep reading!
Full access to this write-up and every future one. Cancel anytime.
View PlansSubscribe to unlock:
- 1-3 write-ups per week
- Periodic industry, market, and thematic research
- Financial models behind most write-ups
- Full archive of all past research
