The pub question has been the same for twenty years: which drive brand lasts longer? A peer-reviewed paper covering 442,992 drives and more than 1.66 million drive-years has just answered it, and the answer is blunt: some fail at less than half the rate of others. That is the headline doing the rounds this week. Section VI of the same paper has travelled a good deal less, and it is where the authors work out how much you should be willing to pay for that difference: on their own assumptions, $11.80 per drive over ten years. And in the very table the headline comes from there is a row that has barely been quoted which, in our experience running hardware, decides more than the brand does.
What they measured, and why it is not another quarterly table
The paper is called Are There Manufacturer Differences in Hard-Drive Reliability?, by Christoph Siemroth (University of Essex) and Yeomyung Park (Sungkyunkwan University, Seoul). It is published in IEEE Transactions on Cloud Computing, volume 14, issue 2, pages 1015 to 1024. The University of Essex repository holds the accepted version in the open, and that is the one we read end to end for this piece. The data is what Backblaze publishes every day since 2013: the state of each drive, SMART attributes, serial number and model. The window runs to the second quarter of 2025.
The interesting part is not the size, it is the method. The usual quarterly tables compare whatever drives were in the fleet that quarter, and each manufacturer's fleet has a different age, a different capacity and a different installation year. Here they use duration regression — survival models, Cox and Weibull — which know how to handle right-censoring: a drive still spinning on the last day of the data does not tell you when it will die, only that it got that far alive. And they compare while holding age, capacity, form factor and average temperature constant. Drive against comparable drive, not fleet against fleet.
A limit the authors put up front, worth repeating: only 0.52% of the drives have more than ten years in the data, so this estimates short- to mid-term differences, not fifteen-year ones.
The row everyone quoted
Table I takes Seagate as the baseline — value 1 — and measures the rest against it. Both models give practically the same answer, which is exactly what you want to see:
The gap between HGST and WD is not noise: the authors' own test returns p = 0.000. So far, the headline is correct and well made. The trouble starts when someone reads it as a buying recommendation, because the winner is not for sale. The paper says so in its first footnote: WD bought HGST in 2012 and later stopped making drives under that brand. The HGST units in the dataset are mostly hardware installed in the early years of the sample. The best actual purchase in the table is the second row.
The row underneath: 1.021
Below the manufacturers, in that same Table I, there is a line reading Average HDD Temperature and a number: 1.021, identical in both models. That is the drive's average daily temperature, taken from SMART attribute 194, with 20 °C as the baseline. It means each extra degree multiplies the failure rate by 1.021. That is 2.1% per degree, holding everything else constant.
On its own that number looks small. Let us do the arithmetic with their own coefficient. Four extra degrees is 1.021 to the fourth: 8.7% more failures. The distance between Seagate and Toshiba, the two manufacturers the table leaves almost level, is 7%. In other words: a rack running four degrees hotter than it should costs you more than the difference between those two manufacturers. And ten degrees, which is roughly what separates a properly contained aisle from one recirculating hot air, works out to 23% more failures. Worth stating the obvious, or the comparison cheats: the Seagate-to-WD jump really is large — 49% fewer failures — and four degrees does not cover it. What these numbers say is that temperature plays in the same league as brand, not that it replaces it.
The 2.63 that comes out of Table II
The authors repeat the analysis adding the drive's physical position inside the datacenter — the cluster, in Backblaze's terms — to check that the manufacturer ranking is not just one brand getting the bad slots. It is not: the order does not move. But along the way a number appears that they underline themselves.
With a caveat the authors do not write down, and which belongs on the table: of the two numbers making up that ratio, the weak one is the denominator. The 0.489 for cluster 60 carries a standard error of 0.203 and a single asterisk, meaning it is significant only at the 10% level; the 1.287 for cluster 20, by contrast, holds at 1%. Translated: the direction of the finding is robust and the exact magnitude is not. That the worst slot kills drives faster than the best one is well established; whether it is 2.63 times rather than 1.8 is another argument.
2.63 times. Between two places in the same building. With the same drive inside. For scale: the whole spread between the most and least reliable manufacturer in the study is 2.7 times, and the good end is not for sale.
The levers, ranked by how much they move
Putting together what the paper measures across its four tables, this is the full map. The last column is ours, and it is the one that changes the conversation.
Capacity deserves a pause. The coefficient is 0.966 per terabyte: a drive with 1 TB more carries a 3.4% lower failure rate. Going from 4 TB to 16 TB is twelve steps, which works out to 34% fewer failures using their own number. One of the three explanations they offer is not magic: big drives are the new ones, and over the study window more and more enterprise drives moved to helium-filled enclosures, which cuts aerodynamic drag on the platters and with it heat and spindle-bearing wear. The honest caveat: fewer and bigger drives also means longer rebuilds when one dies, and this paper does not measure that. That decision is not about component reliability, it is about architecture, and we picked it apart in replica 3 versus erasure coding, with the full math.
The twelve dollars
Section VI of the paper is literally titled "When does it pay to buy the more reliable HDD?". The authors build the whole calculation: a drive's total cost is its price plus its annual failure rate times the cost of each replacement times the years you will keep it. Out of that comes the maximum premium worth paying.
And they apply it themselves. With a ten-year life, a $100 replacement cost and the failure-rate difference from their own regression, the maximum acceptable premium for the more reliable drive comes out at $11.80. Eleven dollars eighty, for a drive that will spend a decade spinning.
Before taking that figure to the shop, two things deserve a look, because both parameters in the example are Backblaze's and not yours. First: the rate difference they use is HGST's 41% — the drive we have spent half a page saying is no longer made. Redo the sum with the only jump you can actually buy, Seagate to WD, and you get $9.72. Second, and this one runs the other way: the paper defines that $100 replacement cost as the labour and logistics cost of replacement plus expected costs due to downtime. For an operator with technicians on staff and drives on the shelf, a hundred dollars is reasonable. For an SME with nobody on call and no spare, that number is higher — and since the acceptable premium is that cost times the failure difference times the years, the worse your replacement story, the more it pays to buy the good drive. At $300 replacement cost over ten years the margin rises to $29.
The paper closes reasoning in exactly that direction: if datacenter technician salaries rise faster than price differences between drive models, they say, it may be increasingly worth shifting to more expensive, more durable drives. The formula does not tell you "do not pay". It tells you to do your own sum, and that the variable actually driving it is not the drive: it is what replacing it costs you.
The authors also say when it does matter, and it is the opposite of what the headline implies: at scale. With Backblaze's 300,000-plus active drives, picking the less reliable one at the same price is over $3 million in extra total cost, and going from a 1% to a 2% annual rate means 3,000 more drives to replace per year — the paper labels that 1% as Toshiba and that 2% as Seagate, read off its Figure 2; those are raw levels and the order comes out the opposite way round to the adjusted Table I, which is exactly the confusion we flag below. And they add the detail that gives away people who have been in a machine room: the hardware may be under warranty, but the technician hours to swap it are covered by no warranty at all.
The number that actually changes your week: 2.4%
There is one sentence in the paper worth more than the whole brand table. It says that up to 50 months of age, no manufacturer's failure rate exceeds 2 in 1,000 drives per month, or 2.4% annualised. None of them. Not even the worst. After that it opens up — Toshiba spikes to between 5 and 6 per 1,000 a month, a spike no other manufacturer shows at any age, while the rest stay at 2.5 or below — but for the first four years everybody is in the same order of magnitude.
Translated to a cluster the size of the ones an SME argues about — say 36 drives under four years old — that 2.4% ceiling gives at most 0.86 failures a year. A ceiling, not an average: at that probability the most common outcome is that none of them dies. And with two conditions that come from the sample itself, enterprise drives and datacenter conditions: if your fleet lives in a hot room, that ceiling is not your ceiling, which is exactly what the 1.021 above is saying. Switching manufacturer moves that number by fractions of a drive. And do not confuse it with another figure we have used on this blog: real replacements in datacenters run at 2 to 4% a year, higher, because drives get swapped before they fail. Failure and replacement are not the same thing, and your storeroom knows it before your dashboard does. You are going to replace drives either way. The question that decides whether it is an anecdote or a bad week is not which one you bought: it is what the system does in the hours that follow. If the data is replicated and re-replication starts on its own, it is paperwork. If the copy lives in the same chassis, or the spare takes six weeks to arrive — a real problem we wrote about in your redundancy assumes the spare arrives tomorrow — then a failing component becomes an outage, and an outage is a design decision, not bad luck.
What this study does NOT say
- It does not measure workload. The authors put it in the limitations: each manufacturer implements SMART attributes differently, so reads and writes cannot be compared across brands. They assume workloads do not differ drastically, and lean on the fact that controlling for position does not move the ranking.
- And the 7% we have leaned on is the weakest of the four numbers. They write it themselves: the Toshiba-Seagate difference is significant at the 5% level but "not as clear-cut as the differences between other manufacturer pairs". On top of that, Toshiba drops below Seagate in two of the four tables (0.977 with no significance in one, 0.863 in another). If you got this far thinking Toshiba is the villain, the paper itself asks for restraint.
- It does not talk about your desktop drives. The paper says Backblaze largely uses enterprise units, in a climate-controlled datacenter spinning 24/7. Among consumer drives, they say, the differences could look different.
- It does not judge this year's models. The dataset includes models from over ten years ago, and the newest ones "did not yet have time to prove themselves". Their words.
- And be careful mixing two of their numbers. Table I gives adjusted ratios — drive against comparable drive — while Figure 2 gives raw observed levels. They are not the same thing, and the text itself uses both. Telling them apart is literally the reason the paper exists.
What you can move yourself, and in what order
Temperature comes first, because it is the one with a coefficient and the one nobody looks at. The attribute is SMART 194 and any agent you already run can read it; what changes the conversation is not the room reading, it is the list of drives sorted by their own average temperature. And sort by the 90th percentile, not the mean: what matters is the worst ten, not the number that reassures you.
Second is knowing where each drive sits and since when. Serial number, installation date and slot: three columns in the inventory, and without them the age lever — the biggest in the study — is unusable, because you do not know how many months anything has been spinning. With those columns you can cross last year's failures against chassis and rack unit and see whether they repeat in the same place. One warning, so as not to oversell it: that is not what the 2.63 measures — a Backblaze cluster is several vaults of twenty servers, not a chassis — but it is the same mechanism brought down to your scale, and it is the only way to tell a bad batch from a bad slot.
Third is what you buy, and the advice here is unglamorous: new before expensive. The installation-year coefficient belongs to Seagate drives, so the five-year sum — 39% fewer failures — is theirs and not everybody's; for WD the paper estimates about 17% a year, and warns in writing against extrapolating beyond the sample period. Even with that caveat the order of magnitude holds: refreshing old hardware pays about as well as switching manufacturer, and unlike the manufacturer, refreshing is within your reach.
And the fourth is not in the study: time a real rebuild. Pull a drive during office hours and measure how long the cluster takes to be whole again. That number, not the brand, is what describes your actual risk. While you are at it you will see that Ceph does not perform like the datasheet, which is another story from the same rack.
Where this fits into what we do
We are not resellers of any one platform: we recommend based on the case, not on the commission. What we do is run the hardware: colocation with your own hardware in a datacenter, with distributed storage with Ceph on top — the piece that turns a dead drive into a log line instead of a three-in-the-morning phone call. Each drive's average temperature and its installation date are, literally, two graphs in our Zabbix; and managed 24/7 support is who watches the panel when it turns orange.
If your fleet lives in a room that used to be an office and you suspect temperature is your actual problem, write to us: pulling SMART 194 from everything and sorting it by position is an afternoon's work and commits you to nothing. If what comes out is ugly, we talk about our colocation. If it comes out fine, we will tell you that too and you save yourself the move.
And if you keep one sentence from the paper, do not make it the ranking. Make it this one: the largest public study on manufacturer differences in drive reliability crowned a manufacturer, put a price on that crown — twelve dollars — and, in the same table, wrote down that the room you put the drive in weighs almost as much. The difference is that the room is actually yours.
Sources. Every figure, quote and coefficient in this piece comes from Are There Manufacturer Differences in Hard-Drive Reliability?, by Christoph Siemroth (University of Essex) and Yeomyung Park (Sungkyunkwan University), published in IEEE Transactions on Cloud Computing, vol. 14, no. 2, pp. 1015-1024 (DOI 10.1109/TCC.2026.3679404). We read the open-access version in the University of Essex repository end to end: repository.essex.ac.uk/43022. From it: manufacturer hazard ratios and the temperature coefficient (Table I, 442,992 observations), datacenter position, the 1.287/0.489 ≈ 2.63 ratio and the standard errors of both coefficients (Table II and section V-D), the capacity effect of 0.966 per TB (Table III), the installation-year effect of 0.905–0.884 (Table IV), the ceiling of 2 per 1,000 per month / 2.4% annualised up to 50 months and the Toshiba spike (section V-A), the acceptable-premium formula and the $11.80 worked example (section VI), and the limitations plus the 300,000-drive figure (section VII). Our own arithmetic, computed from their coefficients and flagged as such in the text: 1.021⁴ = +8.7% and 1.021¹⁰ = +23%; 0.966¹² = −34%; 0.905⁵ = −39%; the ceiling of 0.86 annual failures across 36 drives applying the 2.4%; and the acceptable premium of .72 for the Seagate-to-WD jump and at a replacement cost, using the section VI formula. Cover photo: "3.5 inch hard disk drive, side view" by Gunnar Wolf (Wikimedia Commons, CC BY-SA 4.0).
Microsoft 365
ES 
CAT 
EN