How SMR Actually Stores Data

We've covered this in depth in Stop Murdering Your SMR Drives, so here's the short version. A conventional (CMR) hard drive writes data in concentric tracks with gaps between them — the write head can update one track without touching its neighbors. Shingled Magnetic Recording removes those gaps and overlaps tracks like shingles on a roof, because the physical write head is wider than the physical read head. That overlap is where the extra capacity comes from — typically 20–25% more per platter than an equivalent CMR drive, though the gain shrinks as areal density climbs with HAMR-based platters, since there's simply less room left to shingle into.

The cost of that density is that you can no longer update a single track in isolation. Rewrite track 3 and you clobber the edge of track 4, which now has to be read, merged, and rewritten, which clobbers track 5, and so on. Drives manage this by dividing the platter into zones (commonly ~256MB) so a single small write doesn't cascade across the entire disk — but a write anywhere in a zone still means the drive effectively has to re-lay that whole zone. This is why SMR drives handed to a normal filesystem full of small random metadata writes — journal updates, inode changes, directory entries — spend their life thrashing.

The one thing SMR does well

Sequential append. Write once, from the start of a zone, moving only forward, and there's no overlap penalty at all. That's exactly the write pattern an archive produces.

How LTO Tape Actually Stores Data

LTO tape doesn't spin. A cartridge holds a single reel of half-inch magnetic tape — commonly a metal-particle or barium-ferrite (BaFe) formulation — that a drive physically pulls across a multi-element read/write head at a controlled linear speed. There's no platter, no random-access arm sweeping across concentric tracks. There is only forward and backward.

To use the tape efficiently, LTO drives record in a linear serpentine pattern (not the helical-scan approach older formats like DLT and AIT used). The head writes a set of parallel tracks in a "wrap" running nearly the full length of the tape, then the head assembly steps to the next set of tracks and the tape reverses direction, writing back the way it came — back and forth, like a lawnmower cutting a field, until every track across the width of the tape has been used. LTO-9 cartridges hold 18TB native this way (up to 45TB with 2.5:1 hardware compression), on tape rated for a 30-year archival shelf life.

The practical consequence: reading a specific file that lives three-quarters of the way down a cartridge means winding the tape there first. LTO-9 drives are fast at this by tape standards — full-cartridge seeks land in well under a minute — but "well under a minute" is still an eternity compared to an HDD's single-digit milliseconds. Tape is not slow at moving bytes (LTO-9 streams at 400 MB/s native); it's slow at moving to a specific location.

The Distinction That Actually Matters

This is the point most tape-vs-disk comparisons skip, and it's the one that should actually drive your architecture decision:

Disk Media

SMR: Sequential-Write, Random-Read

Once data is written, you can read any file in normal HDD seek time — a few milliseconds
The shingling penalty only applies to writes, never to reads
Still needs a catalog, so you know which drive a file lives on and can spin it down when idle
Isn't fighting a physical seek penalty once the right disk is spinning
Tape Media

Tape: Sequential-Write, Sequential-Read

There is no such thing as a fast random read on tape — every read is a physical wind operation
A system like HuskHoard leans heavily on an off-media catalog and fanotify demand-loading
Without an index telling you which cartridge and roughly where, tape is close to unusable except for full sequential restores
Why this matters for architecture

"SMR vs tape" isn't purely a cost question. If your archive gets browsed and spot-checked often — a media library people scrub through, a compliance archive that gets partial audits — SMR's random-read behavior is worth something on its own, independent of price. If your archive is genuinely written once and read in full rarely (backups, legal hold, deep cold storage), tape's read penalty basically never gets exercised.

2026 Pricing, For Real

Here's the part that actually changed since the last time this comparison got made casually. 2025–2026 has been a rough stretch for HDD pricing — AI datacenter demand has soaked up capacity across the industry, Western Digital has reportedly sold out its 2026 hard drive allocation, and even consumer external drives have seen list prices climb before falling back somewhat on sale. Tape hardware has not been immune either.

ItemSpecStreet Price (mid-2026)Cost / TB
LTO-9 cartridge18TB native / 45TB compressed~$105~$5.80/TB native
LTO-9 drive, half-height SAS (desktop-ready)18TB native~$4,500–$5,900
LTO-9 drive, Thunderbolt/USB4 desktop18TB native~$8,000–$10,500
Enterprise CMR HDD (e.g. Exos X24)24TB~$480~$20/TB
Consumer SMR external (shuckable, e.g. Expansion/Elements)20–22TB~$320–$400 (sale: ~$16–18/TB)~$16–24/TB

The tape drive, not the media, is the expensive part

At ~$5.80/TB, LTO-9 cartridges are the cheapest storage medium in the table above by a wide margin — cheaper than either flavor of disk. But you don't get to buy cartridges without first buying a reader for them, and that reader now costs more than it did even a year ago.

The SMR-vs-CMR discount has narrowed

The original SMR deep-dive on this blog cited SMR at roughly $14–16/TB against $20–25/TB CMR. In today's tighter market both have moved up, and current shuckable-SMR street pricing (~$16–24/TB depending on sale timing) sits closer to CMR than it used to. SMR is still consistently the cheaper disk option — it's just a smaller gap than a couple of years ago.

LTO-10 exists now

30TB and 40TB cartridge variants have been shipping in limited quantities since early 2026, but pricing and drive availability haven't matured enough to be a realistic budget option yet. LTO-9 remains the practical generation for a cost comparison today; the used LTO-8 market (interoperable for reads and writes with LTO-9 drives) is also worth a look if you're trying to get into tape cheaply.

The Breakeven Math

Set up the comparison as two cost functions of total archive size x (in TB):

// cost as a function of archive size x, in TB cost_disk(x) = disk_$_per_TB * x cost_tape(x) = drive_cost + cartridge_$_per_TB * x

Tape wins once cost_tape(x) < cost_disk(x). Solving for the crossover point:

breakeven_TB = drive_cost / (disk_$_per_TB - cartridge_$_per_TB)

Plugging in the 2026 numbers above — using a blended $20/TB for disk (CMR/SMR are close enough now that it barely matters which) and $5.80/TB for LTO-9 cartridges — here's where tape actually starts winning on raw media cost, at different drive price points:

Tape drive you buyBreakeven pointNotes
Used LTO-8 (12TB native, ~$1,500–$2,000 on the used market)~110–140 TBClosest to the "100TB" rule of thumb — but it's a used, previous-gen drive
New LTO-9, half-height SAS (~$4,500)~320 TBRealistic new hardware budget option today
New LTO-9, Thunderbolt desktop (~$10,500)~740 TBConvenient for a desktop workflow, expensive to justify on media cost alone

About "~100TB" of data storage is the right order of magnitude for a used, previous-generation drive. But if you're pricing a new LTO-9 drive at 2026 rates, the honest breakeven on pure media economics is closer to 300–700+ TB, not 100TB. The storage crunch pushed the tape hardware side of this equation up faster than it pushed disk prices up, which is a real shift worth knowing about before you commit budget to a drive.

What the Simple Math Leaves Out

The crossover formula above only prices the first purchase of media and hardware. Several things push the real-world calculation further in tape's favor than the raw numbers suggest.

You only buy the drive once

Every TB of archive you add after the breakeven point costs $5.80 on tape versus $20+ on disk. If your archive keeps growing — and cold archives generally only grow — the marginal economics matter more than the one-time crossover. An archive that's below breakeven today but growing steadily will cross it eventually; the question is really "when," not "if."

Disks don't last as long as tape

LTO media carries a 30-year archival rating under normal conditions. Spinning HDDs are typically rated in the 1–2.5 million hour MTBF range on paper, but in practice archive operators plan for a 3–5 year practical service life before meaningful failure rates show up, even sitting idle — bearings and lubricant age, and the electronics on a 10-year-old drive are themselves a supportability risk. A disk-only archive held for 20 years means buying that capacity roughly 4–6 times over. A tape archive held for 20 years means buying the cartridges once. This is the single biggest thing the simple breakeven table understates.

Power draw is a wash if you manage disks properly

The classic argument for tape's "0W idle" advantage assumes disks are spinning 24/7 in a NAS. But if you're running SMR the way our earlier piece describes — catalog on NVMe, drives spun down until something is actually requested — idle SMR pulls close to nothing too. Power stops being a tape-exclusive advantage the moment your disk tier is managed instead of just mounted.

The air gap is real and it isn't priced in at all

A tape cartridge sitting in a box has no network path, no USB path, no bus connection — it is physically unreachable by ransomware or a compromised host. An ejected external SMR drive is close to this but not identical: it still has to be plugged in and mounted to write or read, and a compromised host with the drive attached can reach it. For records under legal hold or compliance retention, tape's WORM cartridges give you a physical write-once guarantee that's hard to replicate as cheaply on disk. This has real value, but it isn't a number that belongs in the $/TB column.

Enclosure and controller overhead scales differently

A 100TB disk-only archive across shuckable externals is genuinely simple — a handful of USB or SATA drives. Beyond a few hundred TB, disk-only archives start needing JBOD enclosures, SAS expanders, and rack space that tape's "one drive reads every cartridge you own" model sidesteps entirely.

30 yrs
Archival Shelf Life

Rated tape lifespan versus a practical 3–5 year disk replacement cycle. A 20-year hold means buying disk capacity 4–6 times over.

$5.80/TB
Marginal Cost Past Breakeven

Once the drive is paid for, every additional TB of archive costs a fraction of what the equivalent disk capacity would.

0 paths
True Air Gap

A shelved cartridge has no network, USB, or bus connection at all — physically unreachable by ransomware or a compromised host.

A Decision Ladder

Putting all of that together, here's a reasonable way to think about it by scale:

Under 50TB

Just use SMR (or CMR, the gap is small right now). No tape drive purchase is defensible at this size — you'd be years away from breakeven even on the cheapest hardware, and the operational simplicity of "it's a drive I can plug into anything" is worth a lot at this scale.

50–150TB

Genuinely gray. A used LTO-8 drive can pencil out here, but you're buying previous-generation hardware and betting on the used market staying healthy for parts and support. If you don't already have a strong reason to want an air gap (compliance, ransomware concerns, offsite rotation), disk is still the simpler answer.

150–400TB, or an archive you know will keep growing for years

This is where a new LTO-9 drive starts to make sense — not necessarily because you're past the pure media-cost breakeven yet, but because the trajectory matters. If you're confidently headed past 300–400TB over the life of the archive, buying the drive now and starting to bank $5.80/TB cartridges instead of $20/TB disks is the right call even before the crossover point on day one.

400TB+, or anything under legal hold / compliance retention regardless of size

Tape. The media economics are decisively in tape's favor at this scale, and for regulated data the air-gap and WORM properties are frequently a requirement, not a nice-to-have, independent of cost.

If you already own a tape drive

None of the above applies — your marginal cost is just the cartridge price. At $5.80/TB against $20+/TB disk, tape wins at any scale the moment the drive is already a sunk cost.

Where HuskHoard Fits

The reason this comparison matters to us specifically: HuskHoard doesn't ask you to pick one. Both LTO tape and archive-tier SMR disk are, under the hood, the same kind of citizen in a HuskHoard archive — and the reason that's possible comes down to one deliberate design choice: the Archive Worker never writes anything except the next frame in an append-only stream.

When a file gets cold enough to migrate, HuskHoard doesn't just gzip it and drop it on the volume. It's compressed into a sequence of Zstd frames, each one landing at the current end of the archive stream and each one indexed in the catalog with its own byte offset. That per-frame indexing is what makes the format seekableStreamGate can jump straight to the frame containing byte 4.2GB of a 10TB file, decompress just that frame, and start handing bytes to mpv or ffmpeg without touching anything written before or after it. That's a meaningfully different shape than a single compressed blob per file, which would force a full decompress-from-the-start just to get to the middle.

Append-only-forever is also exactly the write pattern both of this article's media types are built for, for the same underlying reason and in slightly different ways:

Implementation

On Tape

The Archive Worker packs frames into the 256KB blocks LTO drives stream most efficiently, separated by filemarks the drive can navigate directly
The native SCSI tape driver talks to /dev/nst0 through real hardware positioning commands, not as a dumb sequential file
A steady, correctly-sized, forward-only stream keeps the drive from "shoe-shining" — the stutter-and-rewind that happens when input can't keep up with tape speed
Catalog byte-offset lookup plus filemark navigation turns a tape seek into "wind to roughly the right spot and decode from there" — seconds, not minutes
Implementation

On SMR

The same append-only frame stream is written as a strict log-structured format
Every write lands at the current tail of the log; nothing already on disk is ever rewritten
That's precisely the constraint that makes shingled tracks a non-issue — the overlap penalty only exists for drives asked to update data in place
This format structurally never does that, so the shingling penalty simply never gets triggered

The catalog itself is the other half of the "offline but still browsable" story. It's a SQLite database — the README calls it the Catalog, or informally "the Brain" — that lives on the hot tier alongside the stub files, not on the archive media. It tracks every file, its full version history, and its exact byte offset on whichever volume holds it. That separation is why ls and du against an archived directory keep working correctly even with the tape cartridge sitting on a shelf or the SMR drive spun all the way down: the stub reports its real logical size from the catalog while occupying next to nothing on the hot tier. Nothing has to touch the actual media until a process opens the file for real, at which point fanotify intercepts the read, checks the catalog for the offset, and recalls it — from tape, from SMR, or from a cloud bucket, transparently to whatever asked for it.

That matters for the decision ladder above, too: it isn't a one-time, all-or-nothing choice. It's entirely reasonable to run SMR as your primary cold tier today, at whatever scale you're at, and add a tape drive later once your archive's growth trajectory or compliance requirements justify the hardware — without restructuring anything or moving a single file. The stub stays at the same path, the catalog just gains a new volume to point at.