What a 40-year MTBF actually promises

Forty years on the datasheet

The Model 4077's datasheet gives one reliability number: MTBF 40 years. In YS-EN-0004 we used it to size a spares kit for the SS Mariposa's 225 cores, assuming the rate stays constant for the whole trip. That note varied the ageing shape as a what-if. This one asks what the number means and checks the constant-rate assumption against real hardware.

We don't have 40 years of field data on any computer, and nobody else does either. What does exist is the largest public record of hardware failures: Backblaze, a cloud storage company, has published a snapshot of every hard drive in its data centres every day since 10 April 2013, including which drives failed.¹ Hard drives are not compute cores, and they wear mechanically in ways solid-state parts don't. But the question we're asking is about the statistics: whether a single rate, measured early, describes a whole service life. That question transfers, even if the numbers don't.

A rate, not a lifetime

With a constant failure rate λ, the MTBF is 1/λ and the share of units still running at time t is an exponential:⁸

So 63% of units fail before they reach the MTBF, and the median unit lasts only ln 2 ≈ 0.69 of it: 27.7 years for the 4077. For the Mariposa's 225 cores, with no repairs, a 40-year MTBF means about 50 failed by year 10 and 142 by year 40. With every failed core replaced, the fleet sees 225 failures in 40 years, one per core on average, though some cores fail twice and others never.

Vendors often quote the same rate as an annualized failure rate (AFR). Seagate's datasheet gives its MTBF and AFR as a pair, and they match the usual vendor practice of dividing 8,760 power-on hours a year by the MTBF: 2.5 million hours gives 0.35%.^3,6 Backblaze computes AFR from its fleet as failures divided by drive-years, which is the maximum-likelihood estimate of a constant rate.^1,8

For small rates the two are nearly equal. A 40-year MTBF is 350,400 hours: a rate of 2.5% a year, a 2.47% chance of failing in any given year. The same arithmetic turns an enterprise drive's 2.5 million hours into a mean life of 285 years. No drive lasts 285 years, and the datasheet doesn't claim one will. The number is a rate, and a rate is only meaningful over the period in which it was measured.

Where the number comes from

Seagate's knowledge base is unusually direct about this. An MTBF "is based on a large (statistically significant) number of drives running continuously at a test site, with data extrapolated according to various known statistical models"; "based on the observed error rate over a few weeks or months, the MTBF is estimated and not representative of how long your individual drive, or any individual product, is likely to last."⁵ The same article says that historically, "the field MTBF, which includes all returns regardless of cause, is typically 50-60% of projected MTBF."⁵

The conditions matter as much as the number. The Exos X24 product manual promises its 0.35% AFR "over a 5 year service life", at 8,760 power-on hours a year, a drive-reported temperature of 30 °C or less, a "typical workload" and less than 550 TB a year, and adds that the AFR "is a population statistic not relevant to individual units".⁴ A 2.5-million-hour MTBF is a promise about the first five years under those conditions, stated in units that make it sound like a lifetime.

Field data has disagreed with datasheets for as long as anyone has collected it. In 2007, Schroeder and Gibson compared replacement logs for about 100,000 disks with datasheet MTTFs of 1.0–1.5 million hours, an AFR of at most 0.88%: "annual disk replacement rates typically exceed 1%, with 2-4% common and up to 13% observed on some systems."⁶ They also found "replacement rates in our data grew constantly with age", rather than the flat middle that the bathtub curve assumes.⁶ Google's study of more than 100,000 drives that year saw AFR range "from 1.7%, for drives that were in their first year of operation, to over 8.6%, observed in the 3-year old population", while warning that model mix changed with age.⁷

The gap is still there. Seagate rates the Exos X24 at 0.35%.³ Backblaze runs about 10,000 of the 24 TB model, ST24000NM002H, and reports a lifetime AFR of 3.24%, more than nine times higher.² Our own count from the raw data, 302 failures in 10,032 drive-years at ages up to 1.5 years, gives 3.0%. That doesn't make the datasheet false. Backblaze's temperatures and workloads are not Seagate's test conditions, and Backblaze doesn't publish what it counts as a failure. It does mean the datasheet figure won't predict what happens in someone else's data centre.

Half a million drives

Backblaze publishes one CSV row per drive per day, with the date, serial number, model, capacity, a failure flag and the drive's SMART statistics. A failed drive is marked "1" on its last operational day and then disappears from the files.¹ We downloaded every quarterly archive from 2013 to June 2026, 35 GB compressed, and reduced it to one row per drive: when it was first and last seen, and whether and when it failed.

To estimate a failure rate by age, each drive's age must be known on every day it was watched. Drives that joined later are taken as new on the day they first appeared, which is Backblaze's own convention. That is close but not exact: a quarter of them already show more than 30 days of power-on hours when first seen (median about two months, perhaps testing before service), and 0.4% more than a year. The 21,195 drives already running on 10 April 2013 enter at the age given by SMART attribute 9, their power-on hours on that day.^11,12 Those drives are left-truncated: we only see the ones that survived until the records began, so each one only contributes from its entry age onwards. Drives still running at the end, or retired while healthy, are right-censored. The rate in any age band is failures divided by drive-years spent in that band, Backblaze's AFR computed per age instead of per calendar quarter.

We dropped 7,214 boot SSDs and small drives, and 825 drives whose SMART 9 was missing or didn't advance by about 24 hours a day. That leaves 504,571 hard drives of 112 models, with 31,901 failures in 2,022,902 drive-years, an overall rate of 1.58% a year. The raw files contain two traps worth knowing about. On 94,826 drives the capacity is logged as −1 on the first day they appear, and from late 2020 some Western Digital models are logged with a double space in the model name, which splits each of those drives in two unless names are normalised.

Failure rate by age

Pooling every model gives Figure 1. In the first month the rate is 2.0% a year. It falls to 0.9% by six to nine months, then climbs fairly steadily: 1.9% at four years, about 2.5% at six and a peak of 3.0% at six and a half. Beyond eight years it falls again, but by then only a few models are left, so the late drop reflects which models remain rather than ageing.

This is close to the shape Schroeder and Gibson described: modest infant mortality and an early, steady rise rather than a long flat floor. Backblaze's first look at its own fleet, in 2013, saw something similar: "We are not seeing that much “infant mortality”, but it does look like 3 years is the point where drives start wearing out."¹⁰ But a pooled curve mixes ageing with fleet history. Each age band is a different blend of models, and the models differ by a factor of ten. Google's study has the same warning: its 3- and 4-year figures were "more strongly influenced by the underlying reliability of the particular models in that vintage than by disk drive aging effects".⁷ To see ageing, look at one model at a time.

Two drives with a 40-year MTBF

Two of the largest models in the data averaged almost exactly what the 4077's datasheet promises. The Seagate ST4000DM000, a 4 TB desktop drive, logged 5,790 failures in 223,250 drive-years: a rate of 2.59% a year, or an MTBF of 38.6 years. The HGST HUH721212ALN604, a 12 TB enterprise drive, logged 1,856 failures in 76,281 drive-years: 2.43% a year, an MTBF of 41.1 years. Taken as averages over the years observed, either could carry the 4077's datasheet line. Figure 2 shows how differently they got there.

The Seagate is the textbook exponential. Its rate wanders between 1.8% and 3.6% but has no trend, its fitted Weibull shape is 1.01, and a constant rate taken from its first year predicts 15.4% failed by year 7 against 15.8% observed. For this drive the MTBF really did describe the fleet, over the nine years for which there is enough data.

The HGST ran at about 0.55% a year for its first three years. A test in that window would have measured an MTBF of about 1.6 million hours, the kind of number that goes on a datasheet. Between ages 4 and 8 the same drives failed at 4.0–5.8% a year, eight to ten times as often. The first-year rate predicts 4.0% failed by year 7. The observed figure is 16.1%, four times as many. The third model, the HGST HMS5C4040BLE640, goes the other way: 0.75% in year one, then 0.2–0.5% for the next eight years. If a constant rate measured early were applied to it, it would look about twice as unreliable as it turned out to be.

Is the rate constant?

The Weibull distribution is the usual way to ask the question in one number. Its hazard is

"If the Weibull shape parameter β > 1, the Weibull hazard function is increasing (corresponding to wearout); if β = 1, the hazard function is a constant; and if β < 1, the hazard function is decreasing."⁹ With β = 1 it is the exponential, and the MTBF is the whole story. We fit β and the scale η to each model by maximum likelihood, allowing for drives that entered at a nonzero age and drives that were still running when they left the data:⁹

Here e_i is the age at which drive i entered the data, x_i the age at which it left, and δ_i is 1 if it left by failing. Subtracting the e_i term is what corrects for left truncation. Fitting a plain Weibull to the 2013 drives would overweight the survivors. We checked the fitter on simulated fleets with known shapes: it recovers β, its 95% intervals cover the true value 95–98% of the time, and it wrongly rejects a constant rate 2% of the time when the truth is exponential.

Figure 3 shows all 28 models with at least 2,000 drives and 50 failures. Sixteen have a failure rate that rises with age, six have one that falls, and six are consistent with a constant. The rising shapes run from 1.2 to 2.3, plus one outlier: the Seagate ST3000DM001, a 3 TB drive whose fleet failed at 25% a year with β = 4.0. The Exos X24 already shows β = 1.43 in its first 18 months, though over that short a window early-life effects and batch differences can mimic ageing.

Extrapolating the first year

An MTBF from a short test amounts to measuring the rate early and assuming it holds. The data lets us run that experiment for every model followed long enough. We take each model's failure rate in its first year of life, treat it as constant, and predict the share of drives that will have failed by year 5. Then we compare with what actually happened.

Of 18 models followed to year 5, only seven landed within 25% of the prediction. Eight failed more than 1.25 times as often as predicted. The worst, a Western Digital 16 TB model, failed six times as often: 0.20% in its first year, 6.0% failed by year 5. Three failed less than 0.8 times as often. The median model ended 1.20 times above its forecast. At year 3 the errors are smaller (median 1.07×). The first year tells you about the first year. It doesn't tell you which kind of drive you have.

Back to the 4077

Suppose the 4077's 40-year MTBF is exactly right as an average, but the ageing follows one of the shapes we just measured. Equation 3 fixes η for each β so that the mean life stays at 40 years. The table follows the Mariposa's 225 cores with no repairs.

With the same datasheet line, the first decade costs anything from less than one core to 87. With β = 2.18 a core almost never fails in year one, but a 30-year-old core has a 3.0% chance of failing that year and about 20 years left on average. The exponential promises 40 more years at any age; that is what "memoryless" means. The decreasing shape is worst early and best late. The share of cores that outlive the MTBF itself runs from 30% to 51%, and is 37% only when the rate is constant.

For spares this cuts both ways. Under a rising hazard, a spares kit sized with a constant rate is heavy in the early years and short late, which is the case YS-EN-0004 flagged as wear-out the datasheet never saw. Under a falling hazard, failures come early, when the kit is fullest. Either way, the MTBF alone cannot tell you which you have.

What we do with this

The YS·CORE team works to four rules when it reads or writes a reliability number:

1. Read MTBF as a rate. Convert it to a failure rate a year (8,760 hours ÷ MTBF) and to a survival curve with Equation 1. Never read it as a lifetime.
2. Ask over what window, and under what conditions. A rate measured over weeks of testing or five years of service says nothing about year 30. Seagate's own manual limits its rating to five years at 30 °C or less.⁴
3. Track the rate by age, per model and lot. Count drive-years and failures in each band of age, correct for units that join or leave partway, and fit a shape. The pooled fleet curve mixes ageing with purchasing history.
4. Plan with the shape you have, not the exponential you hoped for. When the shape is unknown, plan for more than one, as YS-EN-0004 did for spares, and replace the guess with field data as it arrives.

Limits of this note

• Backblaze is one operator, with its own environment, workloads and purchasing. Its data is hard drives, most of them in a single kind of storage server. Backblaze doesn't publish what triggers a failure flag; 1,063 drives were flagged and later reappeared, and we count them as failed on the first flag.
• Drives retired while healthy are treated as censored, which assumes their retirement says nothing about their risk. Backblaze does take drives "that are still performing well" out of service,¹¹ so the oldest age bands come from the drives it chose to keep.
• Drives that first appear after April 2013 are assumed new. Re-running with SMART 9 as every drive's age drops 21,519 drives whose SMART 9 fails the checks (against 825), and gives an overall rate of 1.68% and 17 rising, 6 falling and 4 constant shapes out of 27 models: the same picture. Days missing from a drive's record are counted as exposure; 2.6% of drives are missing more than 10% of the days between their first and last appearance.
• A single Weibull per model ignores production batches, firmware, data-centre and calendar effects, and the rise-then-fall shapes in Figure 2. The shapes summarise; they don't explain. The intervals treat drives as independent, so batch effects make them too narrow, and with 28 models tested, the borderline cases (p between 0.01 and 0.05) could go either way.
• No drive here is older than about 12 years, and the 4077 table is arithmetic on the fitted shapes, not a prediction for any real core. The 40-year MTBF is the 4077 datasheet figure, not a measurement.

References

1. Backblaze. Hard Drive Data and Stats (Drive Stats data, schema and terms of use), and "How Backblaze computes failure rates" (February 2015, in docs_Q4_2019.zip). backblaze.com/cloud-storage/resources/hard-drive-test-data; docs_Q4_2019.zip. Accessed 2026-10-05.
2. S. Doyle. "Backblaze Drive Stats for Q2 2026." Backblaze Blog, 29 September 2026. backblaze.com/blog/backblaze-drive-stats-for-q2-2026.
3. Seagate Technology. Exos X24 Data Sheet, DS2080-2307US, 2023. seagate.com.
4. Seagate Technology. Exos X24 SATA Product Manual, publication 205194300, Rev. A (draft), September 2023, §2.9.1. Read in a copy hosted by a retailer, trovaprezzi.it; we could not find it on seagate.com.
5. Seagate Technology. "Hard disk drive reliability and MTBF / AFR." Knowledge base article 174791en. seagate.com/support/kb/hard-disk-drive-reliability-and-mtbf-afr-174791en. Accessed 2026-10-05.
6. B. Schroeder, G. A. Gibson. "Disk failures in the real world: What does an MTTF of 1,000,000 hours mean to you?" USENIX FAST '07, pp. 1–16. usenix.org.
7. E. Pinheiro, W.-D. Weber, L. A. Barroso. "Failure Trends in a Large Disk Drive Population." USENIX FAST '07, pp. 17–28. usenix.org.
8. NIST/SEMATECH. e-Handbook of Statistical Methods, ch. 8, "Assessing Product Reliability": §8.1.2 (hazard rate, bathtub curve), §8.1.6 (exponential, Weibull), §8.4.1.2 (maximum likelihood). itl.nist.gov/div898/handbook/apr.
9. Y. Hong, W. Q. Meeker, J. D. McCalley. "Prediction of remaining life of power transformers based on left truncated and right censored lifetime data." Annals of Applied Statistics 3(2):857–879, 2009. doi:10.1214/00-AOAS231; arXiv:0908.2901.
10. B. Beach. "How long do disk drives last?" Backblaze Blog, 12 November 2013. Original version: web.archive.org.
11. Backblaze Drive Stats Team. "Are Hard Drives Getting Better? Let's Revisit the Bathtub Curve." Backblaze Blog, 15 October 2025. backblaze.com/blog/are-hard-drives-getting-better-lets-revisit-the-bathtub-curve.
12. A. Klein. "10 Stories From 10 Years of Drive Stats Data." Backblaze Blog, 10 April 2023, §4. backblaze.com/blog/10-stories-from-10-years-of-drive-stats-data.

Reproduce

Data: Backblaze Drive Stats, used under Backblaze's terms.¹ The 45 quarterly and yearly archives total 35 GB. extract.py reduces each one to a drive table of a few megabytes, so the archives can be deleted as you go. Extraction takes a few hours of CPU time; the rest runs in under a minute. You need Python 3.10+, pandas, NumPy and SciPy.

python code/ys-en-0005/extract.py data_Q1_2024.zip out/Q1_2024.csv.gz   # once per archive
python code/ys-en-0005/merge.py out/ drives.csv.gz                       # one row per drive
python code/ys-en-0005/analyse.py drives.csv.gz                          # → results.json
python code/ys-en-0005/analyse.py drives.csv.gz smart9                   # sensitivity: SMART 9 ages
python code/ys-en-0005/check_weibull.py                                  # fitter on simulated fleets
python code/ys-en-0005/build_figures.py                                  # rebuild the figures in this page

• extract.py and merge.py: daily snapshots to one row per drive
• analyse.py: hazard by age, Weibull fits, extrapolation test, 4077 arithmetic
• results.json and results_smart9.json: every computed value
• check_weibull.py and its output: checks the truncated, censored Weibull fit
• build_figures.py and svgchart.py: figure generation
• LICENSE: the scripts are MIT-licensed. Use them freely, keep the notice; they come with no warranty

Changelog

• 2026-10-06Rev 1. First publication.