Annualized Failure Rate by Hard Drive Model, Updated Quarterly

What changed and why

Models with too small a sample to rank

Models below 250 drives or 50,000 drive days are marked: at that sample size one extra failure moves the rate by whole percentage points, so the number is not precise enough to rank on.

Reading this table

Compare the confidence range, not just the rate. Two models at 1.2% and 1.6% whose ranges overlap heavily are not meaningfully different, and treating them as a ranking is reading noise. A model whose range sits entirely above the fleet rate is the one worth paying attention to.

Age matters as much as model. Several of the worst rates in any year belong to drives near the end of a long deployment, where the population is small and old rather than inherently fragile. That is not an argument to ignore a high rate, but it is a reason to check the drive count and the capacity before concluding a model is bad.

Boot devices are excluded. Backblaze runs boot SSDs and small laptop drives alongside its data fleet and the raw snapshots do not label them, so we filter everything below 4 TB. That exclusion is what makes our fleet figure reproduce Backblaze's published rate: the boot devices recorded no failures and only diluted the denominator.

What a failure means in this data

A drive is counted as failed when Backblaze removes it from service because it has stopped working or because its SMART indicators crossed the thresholds they use to retire a drive before it fails outright. That second category matters: a proportion of these failures are drives that were still readable when they were pulled. The rate therefore measures drives that stopped being trustworthy rather than drives that lost data, which is the right measure for anyone running redundancy but a slightly pessimistic one for a single-drive deployment.

It also means the figures reflect one organisation's retirement policy. A more conservative operator would record a higher rate on identical hardware, and a more tolerant one a lower rate. This is a further reason to read the ordering between models rather than treating any absolute percentage as a universal property of the drive.

Who it hits, and how hard

Buyers choosing between two specific models get the most from this table, provided they read the confidence range alongside the rate. Overlapping ranges mean the data cannot separate the drives.

Owners of a model with a high rate should check the drive count and capacity before worrying. Several of the worst rates in any year belong to small, aging populations near the end of a long deployment rather than to inherently fragile drives.

Anyone whose model is absent should not read that as a verdict. Backblaze runs a specific set of high-capacity models bought in bulk; most consumer drives never appear.

Array builders should use the rate to size redundancy rather than to pick a winner. Our MTBF to AFR calculator converts a rate into expected failures across a given drive count.

What to do now

1. Find your exact model in the table and read its drive days before its rate.
2. Compare confidence ranges rather than headline rates when choosing between two drives.
3. Convert the rate into expected failures for your array size with the MTBF to AFR calculator.
4. Check recording technology with the model number decoder; SMR in an array is a bigger risk than a small rate difference.
5. Read the 2025 annual summary for how these figures are computed.
6. Treat a missing model as missing data, not as a good or bad sign.

Outlook

Backblaze publishes quarterly and we re-ingest each period, recomputing rather than copying. The table above covers the most recent annual period; quarterly figures through Q2 2026 are in the database and summarised on our 2025 Drive Stats page.

We make no prediction about future rates. A model's rate in one year is a weak predictor of the next, because the population ages and the fleet mix changes underneath it.