SSD vs HDD: speed, endurance, and how drives actually fail

Buy an SSD for the machine you use and a CMR hard drive for the data you only keep. Write endurance almost never decides it: a current TLC NVMe drive is rated for 600 terabytes of host writes per terabyte of capacity, far past what a desktop consumes in a warranty period. What differs is how the two die. Hard drives usually degrade through physical wear that leaves a trail in SMART counters. SSDs more often stop answering all at once, a controller or firmware event, flash intact underneath and nothing readable on top.

What a TBW rating actually promises

Consumer SSD spec sheets publish endurance as TBW (terabytes written) or DWPD (drive writes per day). TBW is the total host writes the vendor will stand behind, a tested figure and not a raw count of NAND program and erase cycles. Samsung footnotes both datasheets cited below: "All documented endurance test results are in compliance with JESD218 Standards."

Those two drives are the cleanest same-vendor comparison available. The 990 PRO, TLC on PCIe 4.0, is rated 600 TBW at 1TB and 1,200 TBW at 2TB, with a 5-year limited warranty or the TBW figure, whichever comes first. The 870 QVO, QLC on SATA, is rated 360 TBW at 1TB rising to 2,880 TBW at 8TB, with 3 years. QLC gives you 60 percent of TLC's per-terabyte endurance and two fewer years of coverage. Both print MTBF as 1.5 million hours, which shows how little MTBF separates two drives with a 40 percent endurance gap. And the TLC ratio is no 2022 artifact: Samsung's PCIe 5.0 9100 PRO rates 2,400 TBW at 4TB on the same 5-year term.

Convert it. 600 TBW on a 1TB drive is 328 GB written every day for five straight years. If your daily writes are a browser cache, a few document saves and the occasional game install, endurance is not your failure mode.

The program/erase cycle tables that circulate for SLC, MLC, TLC and QLC live on review sites, not on any NAND maker's product page. Only TBW and warranty years carry a vendor's signature.

Hard drive specs measure something different

Hard drive datasheets publish no TBW. They give a workload rating in terabytes per year, an MTBF figure, an unrecoverable read error rate, sometimes a rated annualized failure rate.

The Exos M sheet adds a rated AFR of 0.35 percent at full 24x7 operation, 7200 RPM on SATA 6Gb/s and 9.5W maximum operating power. Western Digital is unusually candid about its MTBF. The Red Plus figure is "estimated by statistical measurements and acceleration algorithms under typical operating conditions: workload of 90TB/year and drive temperature of 40C," and the footnote adds that it "does not predict an individual drive's reliability." Same physics in both drives; the enterprise model buys 2.5 times the MTBF hours, a ten times better error rate and two more warranty years.

What the field data shows

Backblaze published its 2025 report on 12 February 2026, covering 344,196 drives across 30 models. Fleet annualized failure rate improved to 1.36 percent for the year from 1.55 percent in 2024, with lifetime AFR at 1.30 percent. The worst Q4 outlier was an 8TB HGST HUH728080ALE600 at 10.29 percent. Backblaze ruled out temperature, floated vibration sensitivity as a working theory, then flagged those units for migration on age. Boot drives sit outside that dataset: hard drives only.

SSD field data is thinner and older. Backblaze's most recent SSD-specific report is the 2023 mid-year edition: lifetime AFR of 0.90 percent across its SSD boot fleet as of 30 June 2023, from 3,144 drives. Its own bar for trusting an AFR is 100 drives and 10,000 drive days per model, and much of that table failed it. No 2024, 2025 or 2026 SSD edition exists, so treat 0.90 percent as a dated data point.

The most useful SSD field study is still Meza, Wu, Kumar and Mutlu at ACM SIGMETRICS 2015, covering the majority of Facebook's flash fleet over nearly four years. Its central finding contradicts what most buyers assume: "SSD failure rates do not increase monotonically with flash chip wear; instead they go through several distinct periods." It also found that host-reported writes "do not always accurately indicate the amount of wear induced on flash cells."

Why SSD failure gives you less warning

Start with the hard drive baseline, because "hard drives always warn you" is half a myth. Backblaze's 2016 SMART study tracked five attributes. 76.7 percent of failed drives had at least one go non-zero before dying. The other 23.3 percent gave nothing, so roughly one hard drive failure in four is silent too. The signal is specific rather than sensitive: only 4.2 percent of operational drives showed a non-zero reading.

SSDs make the monitoring problem structurally worse. Of the 44 attributes Backblaze tabulated across three SSD models, only five were common to all three, about 11 percent. Across three comparable hard drive models the figure was 42 percent. A health script tuned on one vendor's SSD may not see the equivalent signal on another.

Then the category no monitoring habit catches. In November 2019, HPE warned that 20 enterprise SAS SSD models would fail completely and unrecoverably at exactly 32,768 power-on hours, three years 270 days and 8 hours, from a signed 16-bit counter overflow in firmware. Fixed firmware, HPD8, existed for 8 of the 20 when the advisory landed. Drives deployed together accumulate power-on hours together, so a whole RAID set reaches the number at once; reported cases include six drives dying inside 15 minutes. One logic defect kills a batch, data present and permanently unreachable.

SMART values worth watching

For hard drives, read the raw values of Backblaze's five rather than the drive's own pass/fail summary. Any of them off zero is a reason to start copying data away.

197 and 198 correlate closely enough that Backblaze calls them usable "as one indicator versus two," but not every manufacturer reports both, so keep both.

For SSDs there is no equivalent five. Watch these instead:

• Percentage Used on NVMe. Solidigm's guidance is blunt: "a value between 1 and 99 is acceptable, whereas 100 is not."
• The vendor-specific wear field on SATA SSDs: SMART 169 on Western Digital, 202 on Crucial, 231 on Seagate, 233 as Media Wearout Indicator.
• SMART 241 and 242, LBAs Written and LBAs Read, to cross-check consumption against the TBW rating.
• Unsafe Shutdowns on NVMe, after any power event.
• The NVMe Critical Warning field, which Solidigm calls "actually the most important one, as it captures all other important indicators": available spare below threshold, temperature, degraded reliability, read-only media, volatile memory backup failure.

Backblaze's caveat applies. Those counters "alone provide only a partial picture as they can miss a rapid acceleration of failures over a short period of time." Monitoring shrinks the no-warning risk on flash. It does not remove it.

Cold storage is where SSDs are simply wrong

An SSD on a shelf has a retention clock that shortens as it wears. Dell's support documentation states the JEDEC position plainly: "The JEDEC SSD Standards JESD218 & JESD219 provide industry-wide requirements around data retention for SSD media," and the requirement is retention "for a minimum duration of three months at the maximum level of rated endurance," with powered-off drives kept below 40 degrees C ambient. A lightly used drive stored cooler holds data far longer. The floor is what shrinks with every terabyte you write, and nothing about that suits a drive you fill once and put in a drawer.

Magnetic media has no wear-driven retention countdown, but no hard drive datasheet publishes an unpowered retention duration either. The Red Plus and Exos M sheets specify non-operating temperature, shock and vibration, not years in a cupboard. So: an HDD powered on and verified periodically, plus a copy elsewhere.

Purpose-built archival storage means tape. LTO-10 is current, with 30TB and 40TB media that LTO-10 drives read interchangeably, roughly 75TB and 100TB at the consortium's 2.5:1 compression figure, specified by HPE, IBM and Quantum. The consortium publishes no headline archival-life number. Its FAQ says the temperature, humidity and magnetic-field limits "are printed on the information sleeve that comes with the cartridge." Treat the 30-year numbers in archival guides as unsourced until you read that sleeve.

Skip M-DISC, which older guides still push on an "up to 1000 years" marketing claim. A 2012 accelerated-aging test at France's LNE, run at 90 degrees C and 85 percent relative humidity, put M-DISC under 250 hours, behind some discs tested alongside it. Millenniata went bankrupt in December 2016 and two licensees remain, Ritek and Verbatim.

Buying in 2026, during an actual shortage

Every older article assuming storage gets cheaper each year is now wrong. AI datacenter demand has absorbed essentially all 2026 hard drive and NAND output. Western Digital CEO Irving Tan described the company as "pretty much sold out for calendar 2026," with 89 percent of WD revenue now coming from cloud customers and 5 percent from consumer retail. Seagate's William Mosley said nearline capacity is "fully allocated through calendar year 2026." Consumer hard drive prices reportedly rose close to 50 percent in five months, a 4TB WD Blue going from $67 to $85 up to $99 on Amazon. NAND contract prices were reported up 50 to 60 percent quarter over quarter in Q1 2026, with 70 to 75 percent forecast for Q2, and relief is not expected before late 2027. Those last figures are trade press and analyst estimates, not vendor filings.

That changes the buying advice more than any endurance spec does. Buy TLC NVMe for the drive you boot and work from. Buy QLC only for read-mostly bulk where 360 TBW per terabyte and three years of coverage are enough. Buy CMR for anything in a NAS and confirm it against the datasheet for your exact model number: WD lists CMR for all eleven Red Plus models from 2TB to 12TB. Buy enterprise nearline when the drive runs 24x7 and the better error rate matters for surviving a rebuild.

Check the live vendor page before committing. Samsung's US store listed the 2TB 990 PRO at $669.99 in July 2026, and showed it unavailable. The 36TB Exos M that made headlines on 21 January 2025 went to select customers as samples, and the mid-2025 datasheet still tops out at 30TB, so size any capacity plan around the 30TB and 28TB models the vendor specifies as shipping.