LOG-036 ·
Skeptical Entry: I Priced Decentralized Storage for My Backups
Thirty days, three decentralized storage networks, one 87 GB encrypted backup set, and a spreadsheet. The doctrine says I should love this. The measurements say not yet, and I trust the measurements.
- Words
- 754
- Est. read
- 3.3 min
- Confidence
- 0.78
- Topics
- skepticism, storage, backups
On paper, decentralized storage networks are built for a person like me. Files sharded, encrypted, and spread across strangers' hardware in dozens of jurisdictions, with no single company able to lose, read, or ransom my data. That is my whole doctrine, rendered as a product category.
I have learned to be most careful exactly when a product flatters my priors. So for the month of July I put my real offsite backup set (87 GB, client-side encrypted, ~2,100 files after chunking) onto three decentralized storage networks in parallel with my incumbent, a boring centralized object store. I will not name the networks, because my sample is one household and a name would let the numbers travel further than they deserve. Patterns, not verdicts on brands.
The measurements
Uploads ran nightly from the same machine on the same fiber connection. Restores were tested weekly: 12 full-restore drills per service, timed, with hash verification on every file.
| Metric (median of trials) | Incumbent | Network A | Network B | Network C |
|---|---|---|---|---|
| Upload throughput | 41 MB/s | 19 MB/s | 8 MB/s | 23 MB/s |
| Full restore time (87 GB) | 39 min | 1 h 48 min | 4 h 12 min | 2 h 3 min |
| Worst restore of 12 | 44 min | 3 h 27 min | 9 h 51 min (1 failure, retried) | 2 h 40 min |
| Restores with any failed chunk | 0 of 12 | 1 of 12 | 3 of 12 | 0 of 12 |
| Effective cost, all-in | $1.87/mo | $1.12/mo | $0.64/mo | $2.31/mo |
Cost needs an asterisk the size of a barn. Two networks price in their own tokens, so my July bill depended on exchange rates that moved 22% during the month I was measuring. I have redenominated everything to dollars at month-end. A backup bill that is itself a speculative position is a novel dependency, and not one my threat model asked for.
What the failures were
Network B's three failed restores were all the same shape: a handful of chunks stored on peers that had gone offline, requiring repair-and-retry cycles the client handled, eventually, with retries. Nothing was lost. Everything was verifiable, which I genuinely credit: every network delivered bit-perfect files or a clean error, never silent corruption, because content addressing (see 2026-08-13) makes silent corruption structurally hard. The integrity story is real and it is better than the centralized default.
The availability story is not there. My backup restore drill has a rule: if the drill fails when I am calm, the incident will fail when I am not. A worst-case observed restore of nearly 10 hours, on a 87 GB set, extrapolates brutally to the 341 GB photo archive.
The economics, from first principles
The pitch is that spare disks in a million basements undercut data centers. My measured reality is that the discount is modest (and in one case negative) once you include redundancy overhead, because the networks must erasure-code aggressively precisely because basements are unreliable. Storing my 87 GB actually consumed between 2.3x and 3.4x that in raw network capacity. The unreliability of the substrate eats most of the price advantage of the substrate. This is not a scandal. It is physics plus incentives, honestly accounted. But it caps the "10x cheaper" story at something closer to 1.5x, sometimes 0.8x.
Verdict, scoped precisely
For the specific job of third-copy offsite backup, held to my drill standards, the decentralized networks do not measure up in 2026. The incumbent survives on restore speed and twelve-for-twelve reliability, and my third copy stays where it is.
Two scoping notes, because a skeptical entry that overreaches is just hype with a minus sign. First, my incumbent is itself the thing this log warns about: one company, one account, one deletion away. That risk is real and I hold it deliberately, priced, as the third copy of three, not the only copy of one. Second, for censorship-resistant public publishing, rather than private backup, the calculus inverts completely, and none of today's numbers apply.
I will re-run this benchmark in twelve months, same set, same drills. The failure rates I measured are the kind that engineering grinds down year over year, and I would rather be the person with a before-and-after table than the person with a settled opinion.
Confidence 0.78 in the central claim: that in the August 2027 re-run, at least one of the three networks will complete 12 of 12 restore drills with a worst case under one hour. If that happens, my third copy moves, and I will say so here, with the table. The doctrine bends to the drill, never the reverse.