The quarter in numbers
Backblaze published its Q2 2026 Drive Stats report on September 29, and the headline number moved the wrong way. The quarterly annualized failure rate hit 1.73%, up from 1.24% in Q1 and the highest the report has shown in a while. The lifetime figure barely budged, ticking up to 1.41% across 353,041 drives and 558 million cumulative drive days, which is the first hint that this quarter’s spike is about specific old drives rather than a general decline.
The fleet itself: 359,101 drives monitored during the quarter, with 3,881 boot drives and 705 others excluded for not meeting the analysis criteria, leaving 354,415 drives in the dataset. Those drives logged 31,553,350 drive days and 1,498 failures. All of this is public. The per-drive CSV dataset has been downloadable since 2013, and it remains the largest public hard drive reliability dataset in the industry.
The honor roll and the outliers
The zero-failure list is a Seagate sweep this quarter: the ST8000NM000A (8TB), ST12000NM000J (12TB), and ST14000NM000J (14TB) logged no failures, and the ST16000NM000J (16TB) managed just one. All four are newer additions with small populations, which is worth remembering before concluding Seagate solved reliability.
Three models landed above the outlier threshold of 6.95% AFR: the HGST HUH721212ALN604 (12TB) at 7.63%, the Seagate ST10000NM0086 (10TB) at 9.33%, and the Seagate ST14000NM0138 (14TB) at 8.26%. Two of the three are old drives, nearly seven and eight and a half years in service, and two are Seagate models with populations under 1,250 drives, where a couple of extra failures swings the percentage hard. That is the explanation for the quarterly spike in one paragraph: aging drives and small samples, not a fleet-wide problem. The lifetime AFR holding near 1.41% supports that reading.
Capacity is not a reliability score
The most useful table in the report compares the big drives. The WDC WUH722222ALE6L4 (22TB) ran a 0.55% quarterly AFR across 45,665 drives, the Toshiba MG11ACA24TE (24TB) ran 0.58%, and the WDC WUH722626ALE6L4 (26TB) ran 0.95%. Then there is the Seagate ST24000NM002H, a 24TB model that ran 3.3% with a fleet of 9,631 drives behind it. Two 24TB drives, a sixfold difference in failure rate.
Anyone procuring storage by capacity class alone is leaving money and availability on the table. The pattern in the data is consistent: model-to-model variance within a capacity tier dwarfs the variance between tiers. Qualify by model, weight by drive days, and prefer models with a few qualified quarters of history behind them. A model with 500 drives and no failures is a much weaker signal than the 22TB WDC with 45,000 drives and a year of data.
A fleet mid-transition
Two structural notes stand out. First, no new drive models entered the fleet for the second consecutive quarter, even while deployment continued, and that deployment is concentrated at the high end: 20TB and larger drives keep growing as a share of the fleet while smaller models cycle out through replacement rather than expansion. Second, two long-serving HGST models retired this quarter, the 4TB HMS5C4040BLE640 at nearly nine years old and the 8TB HUH728080ALE600 at nearly eight. Drives running for close to a decade in production storage pods is a quiet marker of how far the reliability floor has moved, and a reminder that the outliers in the previous section were mostly drives of similar vintage.
SMR is the next variable
Backblaze confirmed there are no shingled magnetic recording drives in the fleet today, but the report treats SMR as the coming phase of capacity growth, alongside WD’s announced 40TB UltraSMR drives and the broader HAMR roadmap. SMR overlaps data tracks on the platter to fit roughly 10 to 25% more capacity on comparable hardware, and it does this by making writes more expensive: random writes degrade badly, while sequential and append-oriented workloads barely notice. Backblaze says it will report SMR failure rates separately once those drives enter the fleet, which will be the first large-scale public data on SMR reliability in production storage.
For anyone running their own storage, the actionable part is not the failure rates, it is the workload fit question that arrives with SMR. Rebuild behavior, parity overhead, and zone-aware software all change when the drive under them starts shingling. Testing that before SMR becomes the default high-capacity option is cheaper than discovering it during a rebuild storm.
Takeaways
Buy by model with drive-day weighting, not by terabytes. Treat drives past the five to seven year mark as replacement candidates, since most of this quarter’s outliers came from that age bracket. Download the dataset and run your own analysis before a large purchase; the numbers are public and the per-model detail is exactly what procurement decks usually lack. And keep an eye on the SMR data whenever Backblaze starts publishing it, because that will shape drive qualification for the next several years.