How much storage do you really need in 2026? SSD, games, photos, NAS

A 1 TB SSD holds a working system and about six big new games, 2 TB suits a gaming or creative PC, and each terabyte you cannot afford to lose needs about five terabytes of drives to be kept properly. That last figure is the one the search results leave out. Four terabytes of photos and documents on a six-drive home server with two drives’ worth of parity, a month of old versions and a fifth of the space left free, plus the two backup copies that the 3-2-1 rule asks for, comes to between 19 and 21 TB of drive labels, and each step of that sum is shown in section 3 so that you can change it.

The tier lists that fill the first page of results - 256GB for light use, 512GB to 1TB for most people, 2TB for gamers - are roughly right about the computer and silent about everything around it. Twenty of the 27 top-ranking articles read for this guide answer with tiers of that kind, and none of those twenty shows how to get from your own files to a number. They leave out the four things that decide it. What you already use, which every operating system will tell you in two clicks. What each workload costs per hour, per shot or per install, which the platform owners publish: Apple’s ProRes white paper puts an hour of 4K ProRes 422 at 265 GB, and the median storage requirement of the 2025 PC releases checked here is 100 GB. What the drives give back after the units, the filesystem, the parity and the free space are paid, which is never the number on the box. And how many copies the data needs, which for anything irreplaceable is at least three.

2026 made the question sharper in both directions. The floors rose: Apple’s whole iPhone 17 line starts at 256GB, double the iPhone 16’s entry tier; Microsoft will not call a laptop a Copilot+ PC with less than a 256GB SSD, and its Recall feature needs 50 GB free before it will start; a single local AI model runs to 404 GB. And storage became expensive on both media at once, so buying the largest drive “to be safe” costs more than it did a year ago, while buying too little still costs a migration later. The price side has its own two guides: why drives became expensive is the 2026 storage price hike, and when to buy is buy storage now or wait. This one is about how much.

It covers both halves of the market. For the home: how to measure what you use; the one formula that turns data into drives; the operating system, games and consoles, photos, video, music and films, and local AI, each with its published sizes; backups and versions; the gap between the label and the usable space; redundancy in a NAS; how full a drive can safely run; surveillance and the homelab. For organisations: the small office, growth rates, data reduction, tiering, erasure coding, retention law and lead times. Then where flash and platters each belong, how many drives of which size, and how to buy at 2026 prices. Readers who want only the numbers will find every workload in two tables in the second-last section, each row naming the section that argues for it.

One note on the numbers, because it changes how to read them. Where a figure is a vendor’s, a publisher’s or an analyst’s, the document is named in the text with its date. Where a figure is worked out here, the arithmetic is shown so that you can change the inputs and disagree with them. No price or listing count from this site appears in the prose: prices on the listing pages move every few hours, so where the site is the evidence the guide names the page to read rather than freezing a figure that will be wrong by the evening. Everything here is as of 2 October 2026, and anything dated after a source’s publication is labelled as that source’s forecast. The memory twin of this article, with the same method applied to RAM, is how much RAM you need.

How much storage you need, by profile: modest on the computer, five times the data around it

The table is the conclusion; the sections after it are the working. “Computer drive” is the SSD inside the machine you sit at. “Data” is what you keep. “Copies” is what keeping it properly adds, and it is zero for anything you could download again.

Three rules come out of the working and hold in every row.

The computer’s own drive is sized by what you work on now, and the library is sized by what you keep forever. These are different drives in different places for different reasons, and the tier lists blur them because a laptop’s drive used to be the whole library.

Replaceable data needs capacity; irreplaceable data needs capacity times copies. A game library is the largest thing on many home machines and needs no backup at all, because it downloads again. Ten years of family photos may be a tenth of its size and need five times their size in drives.

Plan for the three years you can see, and buy only what the next twelve to eighteen months need wherever the rest can be added later. Needs grow (section 22), but at 2026 prices a terabyte bought for year three is paid for at today’s price and sits idle until then. The plan fixes the layout and the bay count, and the purchases follow the growth. Buy the whole plan on day one only where nothing can be added later: a laptop with a soldered drive, a full chassis, a RAIDZ group at its final width.

The profile with the most room for error is the one with irreplaceable data and a single drive, because it looks like the cheapest row and is the only one that can lose everything at once.

Measure before you buy: three numbers, and where every system hides them

Every useful answer starts from what you already use, and every operating system will show it in under a minute. The work is in reading it the right way.

On Windows, Settings, System, Storage shows used space by category, and Microsoft’s Free up drive space page documents the two tools beside it: Storage Sense, which “can automatically free up drive space” by clearing temporary files and the Recycle Bin, and Cleanup recommendations, which list large or unused files, synced files and unused apps. For a per-folder view, WizTree reads the NTFS Master File Table directly and claims to be “46x faster than WinDirStat”; it is slower on network and non-NTFS drives. In PowerShell, Get-Volume lists each volume’s size and remaining space. Windows reports binary units under decimal names, so its “931 GB” on a 1 TB drive is 931 GiB, which section 12 explains.

On a Mac, System Settings, General, Storage (Ventura and later) shows the categories. Apple’s Storage support page (6 July 2026) defines the one that confuses people: “System Data” is “the storage space used by all Apple and third-party files that don’t belong to any more specific category”. The space Time Machine’s local snapshots take is shown differently again: Apple says the Mac counts it as available storage (section 11). macOS has counted in decimal units since 2009, the same units as the label on the drive, so a Mac and a PC show different numbers for the same files.

On Linux, df -h gives each filesystem in binary units and df -H in decimal, du -sh ~/* totals each folder in your home directory, and ncdu, “a disk usage analyzer with a text-mode user interface”, lets you walk the tree and see where the space went. On a ZFS pool, zfs list -o space splits each dataset’s use into the live data, its snapshots and its reservations, and zpool list reports the raw size with parity included, so the two never agree on a RAIDZ pool. Drive capacity explained lists what every Linux tool means by “G”.

On a phone, Settings shows the per-app breakdown (iPhone Storage on iOS, Storage on Android), and the cloud accounts show the rest: Google counts “up to 15 GB” per account across Gmail, Drive and Photos, according to Google Photos Help, and iCloud’s free tier is 5 GB. A family’s photo library often lives in four phones and two clouds at once, and the total is the sum, not the largest.

Whatever the system, the output to write down is three numbers:

1. used now           split into irreplaceable (photos, documents, projects,
                      mail archives) and replaceable (games, apps, downloads,
                      ripped media you still own on disc)
2. growth per year    the irreplaceable part's size a year ago, from an old
                      backup, a photo library sorted by year, or folders by year
3. working set       the biggest thing that must be on fast storage at once:
                      the games you play this month, the project you are editing

The first number sizes the library, the second sizes it for the years you are buying for, and the third sizes the computer’s own drive. Growth is the one most people guess, and it is the one easiest to measure from the past: a photographer who sorts a library by year can read off growth to the gigabyte. One working nature photographer, Chrissy Donadi, did exactly that in an April 2024 article: a 5 TB library, growth of 1.3, 1.1 and 1.3 TB in 2020, 2021 and 2022, and a Lightroom catalogue under 2 GB for 177,000 images, from which she bought a 12 TB drive for a five-year horizon. That is the method this guide generalises.

Measure the split between replaceable and irreplaceable before anything else, because it decides how many copies each terabyte needs. A 2 TB disk that is three-quarters games is a 500 GB backup problem.

One formula turns data into drives, and every factor in it has a source

Every sizing in this guide is the same multiplication, applied once per copy of the data and then added up:

labels for one copy  =  data  x  versions  /  redundancy efficiency
                              /  (1 - free space kept)  /  0.9095

total labels         =  the sum over every copy, then  x  growth

Each factor has a section and a source:

Worked for the home case in the first paragraph: 4 TB of irreplaceable data as the operating system reports it, a month of versions at 10% monthly change, a six-drive RAIDZ2 pool kept under 80% full, and two further copies on single drives, one of them off-site:

primary, six-drive RAIDZ2      4 x 1.1 / 0.667 / 0.80 / 0.9095  =  9.1 TB of labels
                               (six drives of about 1.5 TB each)
backup 1, one drive            4 x 1.1 / 0.9095                  =  4.8 TB
                               with 20% kept free                =  6.0 TB
backup 2, one drive off-site   the same                          =  4.8 to 6.0 TB
                                                                   ---------------
total                          9.1 + 4.8 + 4.8 to 9.1 + 6.0 + 6.0 = 18.7 to 21.2 TB
ratio to the data              18.7 / 4 to 21.2 / 4               =  4.7 to 5.3

About five terabytes of drive labels for each terabyte you cannot lose, when every copy is on your own drives. Nobody buys six 1.5 TB drives, so the real purchase rounds up, and the rounding is where growth lives. A RAIDZ2 group widened later keeps its old parity ratio for everything already written (section 13), so it is the case where the whole plan is bought at once, and six 4 TB drives give the pool room for more than four years at 25% a year (section 22).

The ratio falls when part of the job goes elsewhere. A laptop with a cloud backup and one external drive needs only the external drive in labels, and Apple’s Time Machine guidance (6 July 2026) is a drive “with at least twice the storage capacity of your Mac”. A cloud copy costs money each month rather than drive labels, and section 11 puts a figure on it. The ratio rises with longer version history, wider parity and a second on-site copy. Replaceable data skips the backup lines entirely, which is why a 4 TB game library needs 4 TB and not 20.

The formula has one trap. Redundancy, versions and copies multiply only the data they protect, and a pool that holds both is sized as two sums: 4 TB of photos at about five times, plus 8 TB of ripped films that are still on their discs at about twice, for parity, free space and units alone. A single multiplier across both would ask for 60 TB where about 37 is enough.

Laptop storage: Windows asks for 64GB, Recall and hibernation take more, and 512GB is the least that lasts

Microsoft’s Windows 11 specifications page asks for a “64 GB or larger storage device” and adds that the size of the operating system and the free space needed for updates “are highly variable”. That is the floor for installing it, in the same spirit as the 4GB memory floor on the same page, and it is a device size rather than a footprint: the 64 GB has to hold the system, the space Windows reserves for updates and the room to install them. Microsoft’s other documents add what a working Copilot+ laptop carries beside it:

Windows 11 floor: system, reserved storage   64 GB   Microsoft, Windows 11
and room for updates (a device size)                 specifications
Recall snapshots, default on a 256 GB device 25 GB   Microsoft Learn, 10 Dec 2025
hibernation file, 40% of 16 GB of RAM       6.4 GB   Microsoft Learn, 14 Jul 2025
                                            ------
Microsoft's figures, added                   95 GB
a 256 GB SSD as Windows shows it            256 x 10^9 / 2^30   = 238 GB
left for programs, games and files          238 - 95            = 143 GB

Counting the whole 64 GB floor as the system’s footprint is generous, since a fresh installation uses less, and the generosity is deliberate: updates, drivers and caches fill it over a laptop’s life. Inside it sits reserved storage, which Microsoft introduced in 2019 and said in its January 2019 announcement would start at about 7 GB, so that updates always have room to install.

The hibernation file is the line where Microsoft’s own pages disagree. The Win32 system power states documentation (14 July 2025) gives a full hibernation file a “Default size” of “40% of physical memory” and a reduced one, used only for fast startup, 20%. Microsoft’s hibernation troubleshooting article (12 February 2026) says instead that hiberfil.sys is “approximately equal to how much random access memory (RAM) is installed”. This guide uses the documented default, 40%, so 6.4 GB on a 16GB machine; on the other reading the file is 16 GB and the space left falls to about 133 GB. Either way it scales with memory, and powercfg /hibernate /type reduced shrinks it to a file that, in the words of Microsoft’s powercfg reference, “only supports hiberboot”. The page file scales with memory too, and is not in the ledger because it grows only under pressure: Microsoft Learn’s page-file documentation (12 February 2026) says a system-managed page file grows “up to three times the physical memory or 4 GB (whichever is larger, but no more than one-eighth of the volume size)” when the commit charge reaches 90% of its limit, which on a 256 GB drive caps it at about 30 GB.

Recall is the newest line in the stack, and the one that scales with the drive. Microsoft’s Recall support page says “To enable Recall, you’ll need at least 50 GB of storage space free”, that snapshot saving “automatically pauses once the device has less than 25 GB”, and that it needs a Copilot+ PC with at least 256 GB. The 50 GB is the room it asks for at the moment of switching on, not what it keeps: Microsoft’s Recall management page (10 December 2025) says that unless an administrator sets a limit, “25 GB is allocated when the device storage capacity is 256 GB. 75 GB is allocated when the device storage capacity is 512 GB. 150 GB is allocated when the device storage capacity is 1 TB or higher.” Microsoft’s Copilot+ comparison page (19 December 2025) sets that machine’s floor at a “256GB SSD” and 16GB of memory. On a 256 GB Copilot+ laptop with Recall switched on, Microsoft’s own figures leave about 140 GB for everything else: room for an office suite and some years of documents, or for one big game and little else. That is a thin client’s size, not a laptop meant to last five years with games or creative software on it.

The other platforms publish their floors in the same spirit:

The macOS row is the one where sources disagree. OSXDaily’s June 2025 Tahoe guide says at least 50 GB free with an installer of about 17 GB; other guides quote 20, 34 to 60 and 45 GB, because the real need depends on the state of the Mac and its snapshots. This guide uses 50 GB, the cautious figure. On a server the boot drive is small and the data is the point: Microsoft adds about 4 GB to Server Core’s 32 GB for the Desktop Experience, and warns that servers with 16GB of memory or more need extra room for paging, hibernation and dump files.

The computer’s own drive in 2026: 512 GB is the least for a Windows machine meant to last, 1 TB the comfortable size for anyone who installs games or creative software, and a 256 GB machine is a thin client whose files live elsewhere. Every current MacBook and some thin Windows laptops solder the drive to the board, and Apple sells no storage upgrade for a Mac after the sale, so on those machines the number chosen at the till is the number for the machine’s whole life. Check the specification sheet for a replaceable M.2 drive before counting on an upgrade. Desktops and laptops with an M.2 slot can grow later from the NVMe listings, and drive interfaces covers which slot takes which drive.

1TB or 2TB for gaming: a big new PC game asks for about 100GB, and the store page gives only the launch size

A PCWorld buying guide, first published in October 2024 and updated on 15 April 2026, says “Plenty of modern PC games take up between 20GB to 30GB each”, with Baldur’s Gate 3 and God of War Ragnarök as the exceptions above 150 GB. For big new releases the store listings tell a different story. The storage requirements of 32 current PC titles were read from Steam’s store data and the publishers’ own support pages for this guide on 2 October 2026, a selection of large and popular games rather than a random sample:

The two Black Ops 7 figures are the sources disagreeing: Steam’s listing says 161 GB and Activision’s own support page 116 GB, which looks like the same digits transposed. The sample was read from Steam, so the 161 GB is the one counted.

Across all 32 titles the median requirement is 102.5 GB and the mean about 108 GB; 19 of the 32 ask for 100 GB or more, and the 15 released in 2025 (the 14 in the table plus Dune: Awakening) have a median of exactly 100 GB. The 13 not in the table, from Steam on 2 October 2026: Assassin’s Creed Valhalla 160 GB, Red Dead Redemption 2 150, Microsoft Flight Simulator (2020) 150, Starfield 125, Indiana Jones and the Great Circle 120, Counter-Strike 2 85, Hogwarts Legacy 85, Space Marine 2 75, Ghost of Tsushima 75, Dune: Awakening 60 to 75, Cyberpunk 2077 70, God of War (2018) 70 and Persona 3 Reload 30, so the median and mean can be checked from the two lists together. A big new PC game needs about 100 GB, three to five times the figure that buying guide gives for a typical one.

Three things make the listed figure a floor rather than the size.

Patches, packs and modes. A listing states the size at launch. Monster Hunter Wilds launched asking 140 GB; Capcom then split off the High Resolution Texture Pack, which needs 16GB of video memory and adds more than 70 GB as free DLC, and the base listing now reads “75 GB available space”. The total is the same for a player with a large graphics card. A size that grows or splits after purchase is a normal feature of a live game, which is why the launch figure is the least a drive must hold rather than the most.

Streaming and caches. Microsoft Flight Simulator 2024 lists 50 GB on Steam with the warning that it “Requires download(s)”: the world streams from the cloud and caches locally, which is why survey articles quote 130 GB and “200 GB+ with add-ons” for the same game. Sources disagree because the size depends on how much of the planet a player has flown over.

Unreleased games have no size. Grand Theft Auto VI, a console-only release delayed for a second time to 19 November 2026, as Variety reported on 6 November 2025, was still scheduled for that date on Rockstar’s own page on 2 October 2026, with pre-orders open, and Rockstar has published no install size. Articles that print one are guessing.

Most of these titles now require an SSD, and several say so in the listing itself: Black Ops 7, Battlefield 6, Baldur’s Gate 3 and Monster Hunter Wilds. A hard drive still holds a backlog, at the cost of copying a game back to the SSD before playing it.

The library arithmetic follows. Windows shows a 1 TB drive as 931 GB; take off a 150 GB system and application footprint (section 4) and 15% kept free (the low end of section 14’s range, because a games drive is written once per install and then read), and divide by 100 GB a game:

drive    as Windows shows it    minus 150 GB and 15% free    games at 100 GB
1 TB          931 GB                   642 GB                    6
2 TB        1,863 GB                 1,433 GB                   14
4 TB        3,725 GB                 3,016 GB                   30

A 1 TB SSD holds about six big new releases, 2 TB about fourteen and 4 TB about thirty. Valve’s Steam Hardware and Software Survey for September 2026 shows where players stand: 53.30% of machines have more than 1 TB of total storage, up 3.79 points in a single month, 22.71% have 750 to 999 GB, and 15.86% have 250 to 499 GB. On free space, 14.37% have 10 to 99 GB free and 0.76% less than 10 GB, so about 15% of surveyed machines are within one large installation of full. The survey is opt-in and covers Steam users only.

Games are the clearest case of replaceable data. A library re-downloads, saves sync to the publisher’s cloud on most platforms, and the only backup a game needs is of the save files and any mods. For a gaming PC, buy a 2 TB NVMe drive for the games you play and a cheaper, larger drive for the backlog, and spend nothing on backing up what the store will give you again. The size step is on 2 TB NVMe drives and the step above on 4 TB NVMe drives.

PS5 and Xbox storage: usable space is 5 to 29 per cent below the number on the box

Every console reserves part of its drive for the system software, and the share varies more than the box suggests:

IGN’s hands-on, as reported by Windows Central, found 198 GB of the Series X reserved for the operating system and system files. On the Series S, the rest of the drive goes to the system and Quick Resume, and Microsoft told IGN at launch that it expected Series S versions of games to be about 30% smaller than their Series X counterparts. At 100 GB a game a launch PS5 holds six large games at once, and a Series S three, or five if Microsoft’s 30% holds.

Expansion is where the console owner’s real decision sits, and each console expands differently. Sony’s M.2 installation page accepts PCIe Gen4 x4 M.2 NVMe drives from “250 GB - 8 TB”, recommends “5,500MB/s or faster”, and limits the drive and its heatsink to 25 mm wide and 11.25 mm thick, so a large heatsink can stop a drive fitting. Nintendo’s Switch 2 offers “256 GB of internal storage - 8x the capacity of Nintendo Switch” and takes only microSD Express cards for expansion, so the microSD cards from the first Switch do not work in it. The Steam Deck OLED has a UHS-I microSD slot beside its NVMe drive.

Valve’s own pricing shows what a console maker charges for capacity in 2026. Game Developer reported on 22 June 2026 that the Steam Machine would sell at $1,049 with 512 GB and $1,349 with 2 TB, and that Valve tied the higher pricing to the memory and component shortage driven by generative-AI demand. The 1.5 TB step costs $300, which is $200 for each added terabyte; that is Valve’s price, not this site’s, and the 2 TB NVMe listings are the comparison to make before choosing the larger model.

For a PS5, a 2 TB M.2 drive more than doubles a launch console and 4 TB makes storage a non-question for most libraries. Check the drive’s heatsink against Sony’s 11.25 mm limit before buying, and check the interface on the listing: the PS5 needs NVMe, and an M.2 SATA drive in the same shape will not work, which drive interfaces explains.

Phone and photo storage: shots times format, and RAW multiplies the answer by ten

A photo library is sized by two numbers: how many pictures are taken in a year and how large each one is. The second depends almost entirely on the format.

Apple’s ProRAW support page (19 April 2024) gives the anchor figures: a 12 MP ProRAW file is “approximately 25 MB”, a 48 MP ProRAW file “approximately 75 MB”, and ProRAW files are “10 to 12 times larger than HEIF or JPEG files”. From those, a 12 MP HEIF is about 2 to 2.5 MB; that derivation is mine, since Apple publishes no HEIF size, and Apple’s HEIF page says only that HEIF and HEVC “offer better compression than JPEG and H.264, so they use less storage space”. Standard HEIF on current iPhones goes up to 24 MP and HEIF Max to 48 MP, so a modern phone photo is a few megabytes. A dedicated camera is larger by an order of magnitude: enthusiast measurements published by Alpha Shooters on 29 November 2022 put a 61 MP Sony a7R V uncompressed RAW at “around 125-135 MB”, its full-size lossless-compressed RAW at about 68 to 83 MB depending on the scene, and its compressed RAW at 64 to 73 MB.

Switching a phone from HEIF to ProRAW, or a camera from JPEG to uncompressed RAW, multiplies a library’s growth by ten or more. In a phone library the video usually outweighs the stills, as the family example below shows, and video is the next section’s subject.

Phones set the size of the family library, and their floors rose in 2025. MacRumors reported on 9 September 2025 that the whole iPhone 17 line starts at 256GB, double the iPhone 16’s 128GB entry tier, and Apple’s announcement offers the 17 Pro Max with “256GB, 512GB, 1TB, and for the first time, 2TB”. A phone that holds 256 GB holds a library that has to live somewhere else as well.

A phone is sized by what has to stay on it, not by the library. With Optimize Storage switched on, Apple’s iCloud Photos page (14 September 2026) says “your device keeps space-saving versions of your photos and videos, while iCloud Photos stores your photos and videos in their original, high-resolution version”. So once the cloud plan is paid for, a library far larger than the phone fits behind it, and the same logic holds for any phone whose photos back up to a cloud account. What has to stay on the phone is the system and its updates, the apps, offline music and maps, and whatever is shot before it uploads: an hour of 4K60 is 24 GB (section 8), and Apple’s ProRes page limits 128 GB models to 1080p ProRes. A 256 GB phone, the floor across the iPhone 17 line, covers that for most people; 512 GB is for a phone that regularly shoots 4K60 or ProRes or keeps its library offline; and a 128 GB phone works only with the library in the cloud and offline media kept small.

The cloud plans are where most phone libraries end up, and both large providers publish their tiers. Apple’s iCloud page lists, in the US, 50 GB for $0.99 a month, 200 GB for $2.99, 2 TB for $9.99, 6 TB for $29.99 and 12 TB for $59.99, after a free 5 GB. Google gives each account “up to 15 GB”, and its Storage saver setting resizes photos above 16 MP to 16 MP and video above 1080p to 1080p, which stretches the quota by throwing away resolution. A photo library in a single cloud account is one copy, not a backup, for reasons section 11 sets out.

For a library kept at home, the software adds its own share. The self-hosted photo server Immich says in its requirements that “generation of thumbnails and transcoded video can increase the size of the photo library by 10-20% on average”, and that its database, 1 to 3 GB, belongs on local SSD. A Lightroom catalogue is small by comparison: Donadi’s covers 177,000 images in under 2 GB.

Two worked libraries, with the assumptions labelled:

family of four, phones only (assumptions: 3,000 photos and 2 hours of 4K30
video per person per year)
  photos   4 x 3,000 x 2.5 to 5 MB                 =  30 to  60 GB a year
  video    4 x 2 h x 10.2 GB an hour (section 8)   =        82 GB a year
  total                                            = 112 to 142 GB a year
  five years, self-hosted, +20% for thumbnails     = about 0.67 to 0.85 TB

working photographer (Donadi's measured history, April 2024)
  growth                                           = 1.1 to 1.3 TB a year
  three more years at 1.2 TB, from 5 TB            = about 8.6 TB
  her purchase for a five-year horizon             = one 12 TB drive

The family library fits a 2 TB cloud plan for a decade; the photographer’s needs a NAS, and at about five times for proper copies (section 3), the 8.6 TB library becomes a household that owns several tens of terabytes of drives. A photographer is the one home user for whom the formula’s multiplier is the whole answer, because every frame is irreplaceable and every frame is large.

Video storage is sized by bitrate alone, and “an hour of 4K” runs from 7GB to 1.8TB

Video storage is one multiplication: bitrate times time. The Post Flow’s video storage calculator (Alex Bartz, updated 28 July 2026) gives the rule of thumb, and it is exact:

GB per hour  =  Mbps  x  3,600 seconds  /  8 bits per byte  /  1,000
             =  Mbps  x  0.45

The trouble is that “4K” names a resolution and says nothing about bitrate, and the bitrates span three orders of magnitude:

The often-quoted “an hour of 4K takes 7 GB” is Netflix’s streaming figure, and 4K recorded by a camera is between about one and a half and 250 times larger. The confusion is in print: a consumer storage guide published by Netpoint Solutions in April 2025 says “A 1-hour 4K video can take up 7GB or more”. An iPhone recording 4K60 fills about 24 GB an hour, an action camera at its top bitrate 54 GB, a mirrorless camera in XAVC S 45 GB and in its All-Intra mode 270 GB.

Apple’s ProRes on iPhone page (10 March 2026) adds the phone’s own ceiling: ProRes files are “up to 30 times larger than HEVC files”, 128 GB models are limited to 1080p ProRes, an external drive “must write at speeds of at least 220 MB per second for 4K60 ProRes and at least 440 MB per second for 4K120 ProRes”, and it must be formatted exFAT. 220 MB/s held for an hour is 792 GB, which matches the white paper’s 795 GB an hour for ProRes 422 HQ at UHD 60p, so Apple’s minimum write speed is the recording rate with a sliver of margin. The iPhone 17 Pro adds ProRes RAW and Apple Log 2 to that list.

Sources disagree on one row worth knowing. StreamerSize puts ProRes 422 at 320 GB an hour at 30 fps; Apple’s white paper says 265 GB for UHD at 30p. Apple wrote the codec, so this guide uses Apple’s figure.

Editing multiplies the camera figure. A project holds the originals, proxies for smooth playback, a render cache and the exports, and the proxies alone are not small: ProRes Proxy is 20 GB an hour at 1080p 30p and 82 GB at UHD in Apple’s table. YouTube’s recommended upload bitrates of 35 to 45 Mbps for 2160p SDR (53 to 68 at high frame rates) put a finished 4K upload at 16 to 20 GB an hour, which is the smallest file in the chain.

a creator shooting 3 hours a week of 4K (assumption)
  XAVC S at 100 Mbps    3 x 45  = 135 GB a week   = 7.0 TB a year of originals
  1080p ProRes Proxy    3 x 20  =  60 GB a week   = 3.1 TB a year, deleted
                                                    when the project closes
  the same 3 hours in ProRes 422 HQ, UHD 30p
                        3 x 398 = 1.19 TB a week  = 62 TB a year

A codec choice is a storage decision of about nine times, made in the camera menu. Professional post-production lives further up the same scale: Studio Network Solutions put an hour of ARRIRAW 4K at 1.6 TB and of 16-bit 4K OpenEXR at 4.3 TB in a June 2019 article, and Avatar (2009) at about a petabyte.

For a home or small studio, the working rule is two tiers. The project in progress goes on a fast SSD sized to the largest active project, originals plus proxies plus cache, and the finished archive goes on a parity array of 3.5-inch hard drives, then gets backed up like any irreplaceable data. A 4 TB NVMe drive holds about ten hours of ProRes 422 HQ at UHD 30p, which is why footage moves to the archive when a project closes; the price gap between the two tiers is a live number that the price-per-terabyte rankings show better than a guide can.

Music hardly registers; a Plex or Jellyfin film library is sized by the disc it came from

Music is the one library that stopped mattering. Apple’s figures for Apple Music, quoted in 9to5Mac’s Apple Music guide (Apple’s own page, updated 28 September 2026, no longer lists them), say 10 GB holds about 3,000 songs in high-quality AAC, 1,000 in Lossless and 200 in Hi-Res Lossless, and a three-minute song is about 6 MB at 256 kbps AAC, 36 MB in Lossless at 24-bit/48 kHz and 145 MB in Hi-Res Lossless at 24-bit/192 kHz.

a 10,000-song library
  AAC, 256 kbps               10,000 x   6 MB  =    60 GB
  Lossless, 24-bit/48 kHz     10,000 x  36 MB  =   360 GB
  Hi-Res, 24-bit/192 kHz      10,000 x 145 MB  =  1.45 TB

A lossless library fits on any computer’s drive; only a large hi-res collection needs planning.

Films are the opposite, and they are sized by the disc they came from. The Blu-ray Disc Association’s press release of 12 May 2015, completing the Ultra HD Blu-ray specification, gives discs of 66 GB (dual layer) and 100 GB (triple layer), against 25 and 50 GB for Blu-ray. Audioholics’ summary of the specification gives maximum video rates of 108 Mbps on 66 GB discs and 128 Mbps on 100 GB discs. At 128 Mbps an hour is 57.6 GB, so a full copy of a two-hour 4K film, a remux, lands near the size of the disc. A re-encoded copy is a choice of bitrate, and the arithmetic is the same as for video:

500 films
  4K remux, about 55 to 66 GB each           27.5 to 33 TB
  re-encoded at 20 Mbps (assumption)         500 x 2 h x 9 GB  =  9 TB

A media library of 500 films is a 9 TB or a 33 TB problem depending on one decision about format, which is why the profile table puts it at 9 to 33 TB. Media server software such as Plex or Jellyfin adds thumbnails, artwork and transcode caches on top; Immich’s 10 to 20% is the only published figure found, and it is for photos.

A film library is also the clearest case for the formula’s trap in section 3. If the discs are kept, the library is replaceable, and it needs parity for convenience but not three copies. A six-drive RAIDZ2 at 80% full holds 33 TB of remuxes on about 62 TB of drive labels (33 / 0.667 / 0.8, with no unit conversion because disc sizes are already decimal), six 12 TB drives with room to spare, and that is the whole cost. The same pool full of irreplaceable footage would need its backups added. Media is written once and read many times, which is the workload where shingled drives do least harm, but rebuilding a parity array of them is another matter, and CMR vs SMR explains why to keep them out of one.

Local AI models run from 1GB to 404GB each, and collections only grow

Running language models at home added a new kind of file to the home drive, and the sizes are published. Ollama’s library lists DeepSeek-R1 in seven sizes in its default quantisation:

Llama 3.3 70B is a 43 GB download in the same library. One copy of every DeepSeek-R1 size comes to 487 GB, and people who run models rarely keep only one: a collection grows by family, by size and by quantisation, and a hobbyist comparing three quantisations of two 70B-class models holds about a quarter of a terabyte before any data.

Two features make model files unusual. They are replaceable, since a model downloads again, so they need capacity but not backup copies, unlike the fine-tunes, embeddings and datasets made from them. And they are read whole, into memory, every time a model starts. A 43 GB model read at the 5,500 MB/s Sony recommends for a PS5 drive loads in about 8 seconds; from a hard drive at 250 MB/s it takes nearly 3 minutes (43 GB at 5.5 GB/s and at 0.25 GB/s). Keep the models you use on NVMe and the archive of models you might use on anything. How much memory each model needs once loaded is the memory site’s subject, in how much RAM you need.

The enterprise version of the same pressure is faster. Epoch AI’s dataset-size trend, read on 2 October 2026, says “The size of datasets used to train language models doubles approximately every six months”, and puts the growth at “3.7x per year” (an insight of 19 June 2024, its data updated on 24 November 2025). Solidigm’s product brief for the D5-P5336, a data-centre SSD of up to 122.88 TB, quotes an earlier version of the same page, at every eight months. Either way it is far faster than the 22 to 43 per cent a year the drive makers report for their own shipments (section 18): 3.7 times a year is growth of about 270%.

Backups are the multiplier the tier lists leave out: 3-2-1 means three copies before versions

Several of the articles that rank for this question mention the 3-2-1 rule, and none of them multiplies it in. It is the largest factor in the formula for anything irreplaceable.

The rule’s federal statement is a 2012 paper written for US-CERT by Paul Ruggiero and Matthew A. Heckathorn of Carnegie Mellon University, Data Backup Options:

Keep 3 copies of any important file: 1 primary and 2 backups.

The paper adds two media types and one copy off-site. Backblaze attributes the rule to the photographer Peter Krogh’s The DAM Book and dates it to 2009; other sources say the 2005 first edition, and this guide cannot settle which. The same Backblaze article (21 July 2021) defines two stricter variants: 3-2-1-1-0 adds one copy offline or air-gapped and zero errors when restores are tested, and 4-3-2 keeps four copies in three locations, two of them off-site. Many articles say that CISA “endorses” 3-2-1; its #StopRansomware Guide (September 2023) says instead to “Maintain offline, encrypted backups of critical data, and regularly test the availability and integrity of backups”, which is the same idea without the numbers.

Three copies means three times the data in capacity before a single old version is kept, and the primary copy usually costs more than one times, because it sits on redundant storage (section 13) with free space (section 14).

Two things that look like backups are not. Redundancy is not a backup: a mirror or a parity array copies a deletion, a corrupted file or a ransomware encryption to every drive within milliseconds. Sync is not a backup either: a folder synced to a cloud account replicates the same mistakes, unless the service keeps versions long enough to step back past them. Microsoft’s OneDrive service description (23 July 2026) adds, for its business accounts, that storing “system back-ups” on OneDrive “is not supported”. The confusion is common: in the Harris Poll for Backblaze of June 2020, 41% of US computer owners did not know the difference between cloud backup and cloud storage, 10% backed up all their data daily or more often, and 19% had never backed up. Backblaze’s May 2024 explainer cites 11% backing up daily in its later survey.

Versions are what make a backup useful, and they take space

A backup that keeps only the latest copy restores yesterday’s mistake. Version history is the protection against deletion, corruption and ransomware, and its cost is set by how much of the data changes and for how long old versions are kept. A rough multiplier, assuming every changed file is stored again whole:

versions multiplier  =  1  +  share of data changed per month  x  months kept

documents, 10% a month, kept 1 month        1 + 0.10 x  1   =  1.1
documents,  5% a month, kept 3 months       1 + 0.05 x  3   =  1.15
active projects, 10% a month, kept a year   1 + 0.10 x 12   =  2.2
photo library, new files only               about 1.0; new photos are growth,
                                            not versions

The multiplier is a rough one: it overstates block-level backup tools that store only the changed parts of a file, and it is right for anything that copies whole files.

Apple’s Time Machine page shows why its rule is twice the Mac. Time Machine keeps hourly backups for the past 24 hours, daily backups for the past month and weekly backups for all previous months, and deletes the oldest only when the disk is full, so its disk runs full by design and the size of the disk sets how far back you can go. Its local snapshots, about one an hour kept for 24 hours on the Mac’s own drive, plus the snapshot of the last good backup “until space is needed”, cost nothing in planning terms: macOS counts the space they use as available and deletes them as space is needed. Backblaze’s Computer Backup keeps 30 days of versions by default, a year if you switch that on at no charge, and forever for $0.006 per GB a month on versions older than a year.

Business systems have the same choice in their settings. Microsoft’s version history limits (updated 25 June 2026) let SharePoint and OneDrive keep versions by automatic rules or by count and age, for example 500 major versions with a 365-day expiry, and state that under a retention policy or eDiscovery hold “the versioning limits for the document library are ignored”, so a legal hold lets version storage grow without a cap. Virtual machine snapshots are the fastest-growing version of all. Broadcom’s snapshot guidance supports a maximum of 32 snapshots in a chain and recommends using “only 2 to 3”, says “Do not retain a single snapshot for more than 72 hours”, and notes that the snapshot file “continues to grow in size the longer it is kept”. A snapshot is an undo button with a meter running, not a backup.

What the off-site copy costs, in other people’s prices

The third copy is often cheaper in a cloud than on a drive in someone else’s house, and the providers publish their prices. Backblaze B2 “Starts at $6.95 / TB / mo”, with free download up to three times the stored data each month and $0.01 per GB after that. Amazon’s S3 price list for US East (N. Virginia), read on 2 October 2026, puts Glacier Deep Archive at $0.00099 per GB-month, with retrieval within 12 hours and a 180-day minimum. iCloud’s 6 TB tier is $29.99 a month (section 7).

5 TB off-site
  Backblaze B2            5 x $6.95            =  $34.75 a month,  $417 a year
  S3 Glacier Deep Archive 5,000 GB x $0.00099  =   $4.95 a month,  plus retrieval
                                                   and a 12-hour wait

Those are the providers’ list prices, read on 2 October 2026, not this site’s. Drive prices for the local copies are on the listing pages, and the arithmetic for comparing them is in section 21. For a household, a cloud copy is usually the off-site copy, a single external drive the local one, and the primary sits on the computer or a NAS: three copies, two places, and only one of them a drive you had to buy twice.

A “4TB” drive shows 3.64TB, and the filesystem takes its share first

Drive makers count in powers of ten and most operating systems count in powers of two while printing the decimal names. The binary prefixes exist: the IEC approved kibi, mebi, gibi and tebi in December 1998, and NIST’s page on binary prefixes notes they are not part of the SI, where kilo, mega, giga and tera stay powers of ten. Windows, df -h and most Linux tools use the binary units with the decimal letters; macOS uses decimal throughout.

The ratio is 10^12 / 2^40 = 0.9095 at every size, so 9% of every terabyte you buy disappears into the units before anything else touches it. Nothing is missing; the same bytes are counted with a larger yardstick.

The filesystem takes the next share, and each one documents its own:

• ext4 reserves 5% for root by default. The mke2fs manual says “The default percentage is 5%”, which on a 20 TB data volume is a full terabyte that ordinary users cannot write. tune2fs -m 1 reduces it on a data-only volume.
• ZFS holds back “slop space”. OpenZFS reserves 1/32 of a pool, 3.2%, from ordinary writes, but the source code caps the reserve at 128 GiB with a 128 MiB floor, so on any pool above about 4 TiB it is a flat 128 GiB, about 0.14 TB.
• A Synology NAS charges twice. Synology’s RAID calculator page says “Each drive in the RAID must reserve approximately 10 GB of system space”, and that Btrfs volumes reserve 4% for metadata and ext4 volumes 2%, and warns that real usable space is “less than the estimated available capacity”.
• A console keeps its system share (section 6), which ranges from 5.5% on a PS5 Pro to 29% on a Series S.

Two ledgers show how the shares stack:

one 12 TB drive, ext4, default settings
  label                    12 TB
  as Linux reports it      12 x 0.9095              = 10.91 TiB
  after the 5% reserve     10.91 x 0.95             = 10.37 TiB for users

four 12 TB drives, Synology SHR (one drive of parity), Btrfs
  raw                      4 x 12 TB                = 48.00 TB
  system space             4 x 10 GB                = -0.04 TB
  one drive of parity      (48 - 0.04) x 3/4        = 35.97 TB
  Btrfs metadata reserve   -4%                      = 34.53 TB
  as the NAS reports it    34.53 x 0.9095           = 31.41 TiB, 65% of the label

Four 12 TB drives make about 31 “TB” of space on the screen, before the free space that section 14 says to keep. Drive capacity explained takes the units and the reserves down to the sector, including the array metadata that Linux md and LVM take.

How much NAS storage you get: RAID returns half to four-fifths of the raw capacity

Any layout that survives a drive failure stores the data more than once, or stores parity from which it can be rebuilt, and pays for it in raw capacity. The share that remains is exact for mirrors and classic RAID, and slightly worse for RAIDZ. For 12 TB drives, with usable space in the binary units the system reports:

The RAIDZ figures are for 128 KiB records on 4 KiB sectors and come from the table in drive capacity explained, which works them from Matt Ahrens’s allocation rule. In his analysis of RAIDZ space, Ahrens shows that RAIDZ pads every allocation to a multiple of parity plus one sectors, and that “RAIDZ-p is no better than p-way mirrors for recordsize=4K or 8K”. A pool of virtual machine disks or databases with small blocks gets mirror efficiency from a RAIDZ layout, and should usually be built as mirrors instead, which also rebuild faster. A pool of photos, video and media lands close to the table.

The other platforms publish their own figures:

• Windows Storage Spaces Direct. Microsoft’s fault tolerance page (updated 22 August 2025) gives a two-way mirror 50% efficiency and a three-way mirror 33.3% (“to write 1 TB of data, you need at least 3 TB”), dual parity from 50% at four servers to 66.7% at seven, 72.7% at twelve for hybrid storage and up to 80% at sixteen all-flash, and says Microsoft “discourage[s] using single parity”.
• Btrfs. The Btrfs status page marks its RAID5 and RAID6 profiles “unstable” and its three- and four-copy RAID1C3 and RAID1C4 profiles “OK”, so Btrfs capacity planning means mirrors.
• Synology SHR. Synology recommends SHR or SHR-2 for drives of mixed sizes, which uses the larger drives’ extra space where classic RAID would waste it.

ZFS can now grow a RAIDZ group a drive at a time. OpenZFS 2.3.0, released on 14 January 2025, added RAIDZ expansion, and its manual page for zpool attach warns that after expanding “old blocks retain their old data-to-parity ratio”, so a five-wide RAIDZ2 grown to six keeps three data and two parity sectors in every old block, and that tools report “slightly less space than is expected”. Expansion adds capacity at the old efficiency for everything already written; a pool planned at its final width from the start keeps the better ratio for all of it. The TrueNAS capacity calculator shows usable capacity in both TiB and TB, with ZFS overhead and the reservation listed separately, and is the quickest way to check a layout before buying.

Choose the layout from the bay count and the drive size, then size the drives. Two bays make a mirror. Four bays make two mirrors or a RAIDZ2 with large drives, and a RAID 5 only with small ones. Six or more bays make a RAIDZ2, the layout this guide recommends for most home servers. Why double parity becomes the default as drives grow is a question of rebuild time (section 21), and drives for a NAS works through it, as NVMe arrays does for flash.

How full is too full for an SSD, a NAS pool or a hard drive, and why the answers differ

The advice in print runs from 10 to 25 per cent free, usually with no reason given and the same figure applied to every kind of drive. The mechanisms are different for flash, for copy-on-write filesystems and for plain hard drives, and so are the answers:

Flash. An SSD cannot overwrite in place, so it needs empty blocks to write into, and consumer drives size their fast single-bit cache from whatever is free. HDD or SSD works the arithmetic for a 2 TB TLC drive: at 90% full about 67 GB of cache remains, and at 98% essentially none, after which writes go straight to the slow native cells, and a nearly full QLC drive past its cache can write more slowly than a hard drive. Drive makers set aside spare flash at the factory for the same reason: an Ontrack guest article by Kingston (8 November 2018) gives about 7% for read-intensive drives and 28% for write-intensive ones, and says over-provisioning improves performance and “often increases the life of an SSD”. Leaving 15 to 20% of a consumer drive empty adds to that reserve; SSD endurance explains why it also slows wear.

ZFS and other copy-on-write filesystems. TrueNAS’s August 2026 article on the 80% rule traces it to fragmentation, metaslab pressure and, on hard drives, the slower inner tracks, says that “at approximately 96% full, ZFS will shift to a ‘best-fit’ behavior when allocating space”, and advises tracking the growth trend rather than today’s percentage. A pool that crosses 80% is not broken; a pool that will cross 90% before its next expansion is a pool that was sized too small.

Plain hard drives. A hard drive with ext4 or NTFS has no cache to lose and no copy-on-write to feed. It slows as it fills because data written last lands on the inner tracks, which pass under the head more slowly, and HDD or SSD puts a drive at 95% full at roughly half its headline write rate. The ext4 reserve already keeps 5% back; a little more is comfort, not necessity.

Recorders and backup disks run full on purpose. Frigate’s recording documentation says that “When less than roughly one hour of recording space remains”, it deletes the oldest recordings, and warns that “anything filling the drive, including non-Frigate files, can trigger deletion” of them. A camera disk and a Time Machine disk are sized by how far back they reach, not by how much free space they keep, and putting anything else on them shortens that reach.

Sources disagree on the consumer figure, and the disagreement is mostly about which medium the writer had in mind. Consumer articles read for this guide say 10 to 15%, 15 to 20% and 15 to 25%; TrueNAS uses 80% as its planning line; ext4 and ZFS hold back their own reserves regardless. This guide uses 20% free for SSDs that take steady writes and for ZFS pools, 15% for an SSD that mostly holds installed games, 10% for a plain hard drive, and zero for a disk whose software deletes the oldest data by design, and the formula in section 3 takes whichever applies.

CCTV and NVR storage is cameras times bitrate times days, and the law may cap the days

A camera recorder is the purest storage problem in the home or the office: a fixed stream written continuously, kept for a set time, then deleted. The formula is the video formula with days in it:

storage  =  bitrate (Mbps)  x  0.45 GB per hour per Mbps  x  24  x  days  x  cameras

one camera, 4 Mbps, 30 days         4 x 0.45 x 24 x 30           =  1.30 TB
eight cameras, 4 Mbps, 30 days      x 8                          = 10.37 TB
                                    as the recorder reports it   =  9.43 TiB

The bitrate is the input to get right. GuardSource HQ’s NVR storage calculator (last updated 2 July 2026) uses H.265 baselines of 2, 4, 5 and 8 Mbps for increasing resolutions and assumes about eight hours of motion a day when recording on motion, which cuts continuous storage to about a third. Axis Communications says in its July 2026 Zipstream white paper that the technology “lowers bandwidth and storage requirements by an average of 50% or more when compared to standard compression”, with the savings from its dynamic region-of-interest, frame-rate and group-of-pictures features each ranging up to 50% depending on how much the scene moves. A quiet car park compresses far better than a busy shop. Frigate, the open-source recorder, sets retention in days for each recording mode, with motion-only as its default.

eight cameras, 4 Mbps, 30 days, three ways
  continuous                                    10.37 TB
  continuous, with Zipstream's average 50%       5.18 TB
  motion only, about 8 hours a day               3.46 TB

The drive’s workload rating is the second limit, and no camera calculator found checks it. A rating counts data read and written per year, and a recorder writes all day. Western Digital’s WD Red Plus datasheet (September 2025) rates those NAS drives at 180 TB a year, and Seagate called 550 TB a year “three times that of standard surveillance hard drives” when it announced the SkyHawk AI 24TB in December 2023, rated for “up to 64 HD video camera streams and 32 additional AI streams”:

180 TB a year   = 180 x 10^12 / 31,536,000 s  =  5.7 MB/s  = 45.7 Mbps
                  at 4 Mbps a camera          =  11 cameras of continuous writing,
                                                 before any playback is read
550 TB a year   = 17.4 MB/s                   = 139.5 Mbps
                  across 64 cameras           =  2.2 Mbps a camera
                  64 x 2.2 Mbps fills 24 TB   in about 16 days

Eight cameras at 4 Mbps write about 126 TB a year, inside a 180 TB rating but not by much once someone reviews footage. A large recorder has to be checked on both counts: enough terabytes for the retention period, and enough rating for the streams. The ratings and what they assume are taken apart in drives for a NAS, and drives sold for the job are on surveillance-rated hard drives.

The retention period may be capped by law, which caps the storage. In the European Union, the European Data Protection Board’s Guidelines 3/2019 on video devices, adopted on 29 January 2020, say footage “should in most cases (e.g. for the purpose of detecting vandalism) be erased, ideally automatically, after a few days”. They go on: “The longer the storage period set (especially when beyond 72 hours), the more argumentation for the legitimacy of the purpose and the necessity of storage has to be provided.” Seventy-two hours of eight 4 Mbps cameras is about 1 TB. A business camera system in the EU that keeps 30 days needs a reason it can defend for keeping ten times what the guidelines describe, and a household camera that films a public street raises the same question.

Whether to add redundancy depends on what the footage is for. A recorder that exists to catch the night of a break-in loses its point if the one drive dies that night, so a two-drive mirror is the minimum for footage that matters, and a parity array suits a large system. Footage past its retention period is deleted rather than backed up, so surveillance is the one irreplaceable-looking workload with no 3-2-1 multiplier: the exported clip of an incident is the part that needs copies.

A homelab is sized by its virtual disks, its snapshots and the copies of both

A home server running virtual machines has three storage pools whether or not it names them: fast storage for the guests, bulk storage for data and media, and somewhere for the backups of both. Each is sized differently.

The guests are sized by what they use, but the pool must survive what they were promised. Virtual disks are usually thin-provisioned, so a 128 GB virtual disk holding 40 GB takes 40 GB of the pool. The promise is the risk: a pool whose guests have been promised more than it holds fails for all of them at once when the promises come due, and a ZFS or Btrfs pool already slows past 80% (section 14). Snapshots add to it, and Broadcom’s guidance from section 11 (two or three per chain, none older than 72 hours) is the right habit on any hypervisor, VMware’s or another.

Templates and installation images deduplicate extremely well, and little else does. Microsoft’s Data Deduplication overview (20 June 2025) lists typical savings by data type:

Ten Windows virtual machines built from one image share most of their blocks, and a deduplicating backup tool stores each unchanged block once, which is why a week of nightly guest backups can cost little more than one copy plus the changes. Photos, video and anything compressed or encrypted reduce close to 1:1.

Small blocks argue for mirrors. A guest’s virtual disk issues small writes, and RAIDZ spends a full parity sector on each small block (section 13), so the guests’ pool is usually a mirror of two SSDs. NVMe arrays covers building that mirror, and NVMe caching covers putting flash in front of a hard drive pool instead. Proxmox’s requirements page adds a memory cost to the storage: about 1 GB of RAM for every terabyte of used ZFS or Ceph storage.

A worked homelab, with every guest size an assumption:

guests (thin-provisioned, sizes as used)
  Windows desktop VM                          80 GB
  five Linux VMs and containers           5 x 30 = 150 GB
  ISO and template library                   200 GB
  snapshots, about 20% of guest disks         46 GB
                                            ------
  total on the guests' pool                  476 GB
  at 80% full, on a two-SSD mirror           476 / 0.8 = 595 GB  ->  two 1 TB NVMe drives

data pool (media, files)                     measured, sized as sections 3, 9 and 13
backups of the guests (deduplicated)         about one copy of 230 GB plus changes,
                                             on the data pool or a separate drive,
                                             with one copy off the machine

The ISO library is the line not to back up: every image in it downloads again. A homelab’s backup job covers the guests’ disks and their configuration, which is a few hundred gigabytes, not the multi-terabyte pool beside them, unless that pool holds irreplaceable data, in which case section 3 applies to it on its own terms. The boot drive is the smallest decision: Broadcom’s ESXi 8.0 asks for a 32 GB boot device and 128 GB for its data volume, and Proxmox sets no minimum.

A small office should measure its shares before trusting “per employee” figures

The figure most often printed for small business storage is a guess. Burgi Tech’s 2026 guide to cloud servers for small businesses says “Plan for 100-200 GB per employee minimum, but budget for 3-5 years of growth”, which for 25 people is 2.5 to 5 TB before the growth, and gives no derivation. Backblaze’s NAS buyer’s guide (May 2024) offers the only formula among the ranking articles: users times hard drive size, plus shared storage, times a growth factor, with two to four suggested. It sizes from drive capacities rather than measured data, and leaves out versions, units and free space; parity enters only as two extra bays.

An office already has its data on a server, a NAS or in a cloud tenant, and the measured size of it beats any per-head figure.

The office version of section 2’s three numbers:

1. The size of each shared folder or library, from the server’s own report, the NAS’s storage manager, or the cloud admin console.
2. Its size a year ago, from the oldest backup still kept.
3. The split between working files, which need versions, and the archive, which needs retention.

The cloud tenant may already be smaller than the share. Microsoft 365 Business Basic, Standard and Premium give each user 1 TB of OneDrive, according to Microsoft’s OneDrive service description (23 July 2026); E3 and E5 start at 1 TB and can be raised to 5 TB, and F3 users get 2 GB. Shared sites draw on a separate pool: SharePoint’s tenant storage is 1 TB plus 10 GB for each licensed user, as CIAOPS summarised Microsoft’s limits in May 2025, so 25 users share 1.25 TB for every team site, and a 1.5 TB file share moved to SharePoint overflows on day one. Google pools differently: Workspace’s edition comparison gives 30 GB per user on Business Starter, 2 TB on Business Standard and 5 TB on Business Plus, pooled across the organisation, so 25 Business Standard users have 50 TB between them.

A worked office of 25 people, with the measured size and growth rate as labelled assumptions:

measured shared data today (assumption)       1.5 TB, as the server reports it
growth, from last year's backup (assumption)  20% a year
in three years                                1.5 x 1.2^3       = 2.59 TB
with a month of versions at 10% change        2.59 x 1.1        = 2.85 TB
usable needed at 80% full                     2.85 / 0.8        = 3.56 TB
as drive labels                               3.56 / 0.9095     = 3.92 TB usable
four-bay NAS, two mirrors (RAID 10)           3.92 / 0.5        = 7.8 TB raw
                                              -> four 4 TB drives, with room past
                                                 year three
on-site backup, 90 days of versions           a NAS or drive of about the same
                                              usable size
off-site copy, Backblaze B2 list price        2.85 / 0.9095 x $6.95 = about $22 a month
"100-200 GB per employee"                     2.5 to 5 TB, with no versions,
                                              parity or backup in it

Deduplication helps a Windows file server: Microsoft’s typical 30 to 50% on user documents and 50 to 60% on general file shares would roughly halve the primary copy, and should be measured with the server’s own evaluation tool before it is counted on.

Retention law sets a floor, and in Europe a ceiling

How long records must be kept sets how much an office stores, and the rules are published:

Retention is a floor and, under the GDPR, a ceiling too. Records that must be kept for six years add six years of growth to the archive, and in the SEC’s case on storage that cannot be overwritten. Personal data kept past its purpose is a liability rather than a safety margin, and a deletion schedule is also the cheapest capacity an office will ever add. Surveillance footage is the sharpest case (section 15).

The seats themselves follow section 4: 512 GB is the least for a Windows laptop meant to last, 1 TB for anyone whose work is media, and their contents belong in the shared store or the tenant, which is what gets backed up.

Enterprise capacity starts from a growth rate, and the published ones run from 22 to 43 per cent

Enterprise planning reverses the home question. The footprint is known to the byte; the uncertainty is in how fast it grows, how much it shrinks under data reduction, and where each part of it should live. The growth evidence comes in three kinds, and they disagree in useful ways.

The analysts’ forecasts of data created. IDC’s forecast of data created, as quoted by Western Digital’s Brad Warbiany in an Embedded Computing Design article of February 2025, puts it at 132.4 ZB in 2023 rising to 393.9 ZB in 2028, a forecast that works out at 24.4% a year compounded. The same article quotes IDC’s Ed Burns on “around 13.6 ZB stored in 2028”, without saying whether that is the total stored or the year’s addition, and on “approximately 85% of enterprise data” being “still stored on HDDs”. IDC’s abstract for its 2025 to 2029 StorageSphere forecast (June 2025) quotes its research manager Adam Wright with a qualifier that 2026 has borne out: “Global data generation continues to accelerate, but the expansion of storage capacity is defined by uncertainty and restraint”.

The drive makers’ shipments. Seagate’s fiscal fourth-quarter results for the quarter to 3 July 2026, as Investing.com summarised the results slides on 28 July 2026, report 218 EB of hard drives shipped, up 34% on a year earlier, of which 195 EB was nearline, up 43%, with about 90% of exabytes going into data centres. Western Digital told its results call of 5 August 2026 that it shipped 231 EB in the quarter, up 22%, with exabytes up 25% for the fiscal year, and that it sees exabyte demand “growing at above 25%” from here. Those figures measure what the industry could make in an allocated market, not what customers wanted, which makes them a lower bound on demand.

Survey forecasts of enterprise data. Seagate’s Rethink Data report, an IDC survey of 1,500 organisations sponsored by Seagate in 2020, forecast enterprise data to grow 42.2% a year from 2020 to 2022, and found that “Only 32% of data available to enterprises is put to work”. It is vendor-sponsored and six years old, and it is the only published per-enterprise rate found.

published growth rates, per year
  Western Digital, exabytes shipped, fiscal Q4 2026           22%
  IDC, data created, forecast 2023-2028                       24.4%
  Western Digital, expected exabyte demand, Aug 2026          above 25%
  Seagate, all hard drive exabytes, quarter to July 2026      34%
  IDC for Seagate, enterprise data, forecast 2020-2022        42.2%
  Seagate, nearline exabytes, quarter to July 2026            43%

Plan on your own measured rate, and use 25% a year where there is no history, near the low end of the published range and the floor Western Digital gave for exabyte demand. At 25% a footprint doubles in a little over three years (1.25^3 = 1.95) and triples in five (1.25^5 = 3.05). AI workloads sit above the range: Epoch AI’s 3.7 times a year (section 10) is a growth rate of about 270% a year for the training datasets themselves.

Data reduction shrinks the footprint, and the vendors’ ratios do not compare

Every enterprise array advertises a reduction ratio, and the ratios measure different things:

Pure Storage, which renamed itself Everpure in February 2026, counts thin provisioning in its figure, so space promised but never written counts as saved; Dell’s applies only to data it considers reducible; NetApp’s guarantee is for SAN storage, and trade summaries give a lower figure for file storage that could not be confirmed at the source. None of these numbers can be multiplied into a plan until it has been measured on your data, and media, encrypted and already-compressed data reduce close to 1:1 on every array. Treat a vendor’s guarantee as a floor for the part of the estate it covers, and plan the rest at 1:1.

Most stored data is cold, and AWS prices cold at a twenty-third of hot

The largest saving in enterprise storage is not reduction but placement, because most of what is stored is rarely read. CTERA, which sells file-tiering systems, said on 15 September 2026 that across 16 PB of live production data in 856 file-share scans “just 4.4% of stored capacity is actively used” and “only 9.9% of stored data was accessed in any given 90-day period”. It has an interest in the finding, and the widely repeated claim that “60 to 80% of data is cold” has no primary source that could be found. Seagate’s 68% of data that “goes unleveraged” is a different measure, data not used for analysis, rather than data not opened.

Amazon’s price list shows what the cold side is worth. Its storage classes page and price list for US East (N. Virginia), read on 2 October 2026:

$0.023 / $0.00099 is a spread of 23 times between the hottest and coldest class, all quoted at the same eleven-nines durability. S3 Intelligent-Tiering moves an object to infrequent access after 30 days without access, to archive instant access after 90 and, if you opt in, to deep archive after 180 days or more. On premises the same split is flash for hot data, hard drives for warm and cold, and tape for the archive: the LTO Program’s roadmap gives LTO-10 a native 40 TB per cartridge, “up to 100 TB” at an assumed 2.5:1 compression, at up to 1,200 MB/s.

Protection is the next multiplier, and object stores publish theirs. Ceph’s erasure-code documentation gives the overhead as (k+m)/k: three data and two coding chunks store a terabyte in 1.67 TB of raw space against 3 TB for three-copy replication, and the default profile of two and two needs 2 TB. MinIO’s AIStor documentation says a 16-drive erasure set at its default EC:4 is “75% usable - tolerates 4 drive failures”, and that “Production deployments must use a parity of EC:3 or higher”.

A worked estate, with every input either sourced above or labelled as an assumption. It is in decimal terabytes throughout, as array vendors and AWS quote them, so section 3’s 0.9095 does not appear:

today, logical, decimal as billed (assumption) 600 TB
three years at 25% a year                     600 x 1.953         = 1,172 TB
hot share (assumption; CTERA measured 9.9%)   20%                 =   234 TB
cold share                                    80%                 =   938 TB

hot tier, flash
  reduction 2:1 (Microsoft's file-share range)   234 / 2          =   117 TB
  erasure-coded at 75% usable (MinIO EC:4)       117 / 0.75       =   156 TB
  kept 80% full                                  156 / 0.8        =   195 TB raw flash

cold tier, hard drives
  reduction 1.2:1 (assumption; media-heavy)      938 / 1.2        =   781 TB
  erasure-coded at 75% usable                    781 / 0.75       = 1,042 TB
  kept 80% full                                  1,042 / 0.8      = 1,302 TB raw
  in 24 TB drives                                1,302 / 24       = 55, built as
                                                 four 16-drive sets = 64 drives

second copy, off-site, AWS list prices for 1,172 TB
  S3 Standard, tiered                                             = $25,160 a month
  Glacier Deep Archive                           1,172,000 x 0.00099 = $1,160 a month

The hot-cold split decides what the estate costs, more than any reduction ratio: it puts 195 TB on flash and 1,302 TB on hard drives instead of the whole estate on flash, and the off-site copy of everything costs a twenty-second as much in deep archive as it would in standard storage, provided it is never needed in a hurry.

The last input is time. TrendForce reported in September 2025 that lead times for nearline hard drives had “ballooned from just a few weeks to over 52 weeks”, Western Digital said in August 2026 that it was negotiating long-term agreements for calendar 2029, 2030 and 2031, and TrendForce said on 21 September 2026 that it expects enterprise SSD orders in the fourth quarter to keep “a strong upward trajectory”, supporting “continued increases in enterprise SSD prices”, a forecast. An enterprise that needs capacity in twelve months has to order it now, which is why the plan above runs three years rather than one. Enterprise drives covers the used data-centre drives that fill some of the gap, and the SAS listings and enterprise SSDs are where they appear.

Flash for what you wait on, platters for what you keep

The split this site’s HDD or SSD guide settles on applies to capacity planning directly: flash for anything a person waits on, platters for anything measured in terabytes that a person does not wait on. In a home it divides the drives cleanly.

Endurance is rarely what limits a home SSD. SSD endurance works out that 30 GB written a day takes 55 years to reach a 600 TBW rating, and the workloads that come close are rare at home: a camera recorder, a download cache, a database. Retention is what limits an SSD as an archive. The JEDEC standard that HDD or SSD quotes specifies a client SSD’s powered-off retention at one year at 30 °C, so an SSD in a drawer is not a backup plan. A hard drive in the same drawer keeps its magnetic data for decades; its risks there are mechanical, a stuck head or a motor that will not spin up, so it needs checking too, as SSD endurance explains.

At rack scale the line moves, because power and floor space start to cost more than the drives. Micron began shipping a 245 TB data-centre SSD on 5 May 2026 and claims “82% fewer racks” than a hard drive deployment of equal raw capacity; TrendForce said in September 2025 that QLC SSDs draw “about 30% lower power” than nearline hard drives; and Solidigm’s D5-P5336 product brief rates that QLC family from 7.68 to 122.88 TB at 0.42 to 0.60 drive writes per day over five years, 5.9 PB written for the smallest and 134.3 PB for the largest. Those are the makers’ claims. The installed base still runs the other way: IDC’s Ed Burns, quoted in a Western Digital article of February 2025, puts about 85% of enterprise data on hard drives.

The price per terabyte between the two media is a live ratio, not a fact about them, and in 2026 both sides of it moved at once, which the 2026 storage price hike traces. Read today’s ratio from the hard drive listings and the SSD listings, both sorted by price per terabyte, and put each part of the plan where its access pattern says, not where this month’s price says.

How many drives, and which sizes: rebuild hours, bays and cost per usable terabyte

The formula gives a usable capacity; the layout from section 13 turns it into a count of drives; the drive size is what is left to choose, and three things decide it.

Rebuild time grows with the drive. When a drive in a redundant array fails, the replacement has to be filled from the others, and the floor on that is the drive’s capacity divided by its sustained transfer rate. Western Digital’s datasheets give the top rates: up to 196 MB/s and up to 260 MB/s for the two 7,200 rpm WD Red Plus 12 TB models in the September 2025 datasheet and up to 287 MB/s for the 26 TB WD Red Pro in the January 2025 datasheet:

12 TB at 260 and 196 MB/s   12 x 10^12 / 260 or 196 x 10^6  =  12.8 to 17 hours
26 TB at 287 MB/s                                             =  25 hours
44 TB at 287 MB/s                                             =  43 hours
24 TB at 200 MB/s, a whole-surface average on an idle array   =  33 hours
24 TB at 100 MB/s, an array still in service                  =  67 hours

Those are floors: “up to” rates are outer-track figures, and drives for a NAS tabulates the realistic range, from 200 MB/s on an idle array down to 50 MB/s on a busy one. The 44 TB row is not hypothetical for long. Seagate said in July 2026 that its Mozaic 4+ products, “which support up to 44 terabytes per drive”, were ramping with the two largest cloud providers, and Western Digital said on its August 2026 call that its 44 TB HAMR drive is “on track for the first half of calendar 2027”, with 50 TB products in the second half, both forecasts.

A single-parity array has no protection left while it rebuilds, and a larger drive keeps it unprotected for longer. The case against single parity on large drives is usually made with the unrecoverable read error rate: in February 2010, StorageMojo’s Robin Harris relayed Adam Leventhal’s finding that RAID 6 “will be as good as RAID 5 was until 2019”, and Harris’s own sum gave “a 62% chance of data loss due to an uncorrectable read error on a 7 drive (2 TB each) RAID 5 with one failed disk” at a specification of one error in 10^14 bits (Leventhal had assumed the enterprise figure of one in 10^16). Sources disagree on how literally to take it: the datasheet figure is a worst-case bound (the WD Red Plus quotes fewer than 1 in 10^14, the Red Pro fewer than 1 in 10^15), and rebuilds succeed far more often in practice than the literal arithmetic predicts. The direction holds regardless: more capacity per drive means more hours exposed. This guide uses double parity for any array of drives of 12 TB or more, and drives for a NAS does the probability properly.

More drives means more failures to rebuild from. Backblaze’s Drive Stats for the second quarter of 2026 (29 September 2026), across 354,415 drives, put the quarter’s annualised failure rate at 1.73%, “the highest it’s been in quite a while”, and the lifetime rate at 1.41%; drives of 20 TB and more are now over a quarter of its fleet, and a model of roughly seven-year-old HGST 12 TB drives averaged 7.63%. Blocks & Files reported from the first-quarter figures that 92% of the 10,220 drives Backblaze deployed that quarter were 20 TB or larger, and that Toshiba’s MG 24 TB had the quarter’s lowest failure rate, 0.42%, against 3.13% for a 7.5-year-old Seagate 10 TB model. At the 1.41% lifetime rate, the chance that at least one drive in an array fails in a year is:

Fewer, larger drives fail less often and take longer to rebuild each time; more, smaller drives fail more often, rebuild faster and spend less on parity. Drives for a NAS shows the two roughly cancel in exposure time, so the tiebreakers are the ones that scale with drive count: bays, noise, power, and the cost of the chassis.

Bays are the scarce resource, not terabytes. A chassis that is full can grow only by replacing every drive, one rebuild at a time. Plan the array at the width it will end at, leave bays free where the layout allows a second group of drives, and keep in mind that RAIDZ expansion keeps the old blocks at the old parity ratio (section 13). Mixed sizes waste the difference in classic RAID and ZFS, where every member counts as the smallest, and are what Synology’s SHR exists for. One cold spare on the shelf is worth more in 2026 than in most years, because a replacement ordered on the day of a failure may not arrive that week.

The price per usable terabyte is the price per raw terabyte times a multiplier

The listing pages on this site rank drives by price per raw terabyte, the number on the label. What a layout delivers per usable terabyte is that price times a fixed multiplier, which the formula gives:

multiplier  =  1 / (share of raw that holds data  x  0.9095  x  share kept full)

Multiply the price per terabyte a listing shows by the multiplier for your layout, and that is what each usable terabyte costs. Two copies more for irreplaceable data add their own: a six-wide RAIDZ2 primary plus two single-drive backups at 90% full comes to 2.06 + 1.22 + 1.22 = 4.5 times the price per raw terabyte for each TiB of data, or 4.95 with a month of versions, which is section 3’s “about five” again from the other end.

Three practical points follow from the arithmetic. A cheaper price per raw terabyte on a larger drive is worth having only if the layout can use the capacity: two large drives in a mirror pay 2.75 times over, while six smaller ones in RAIDZ2 pay 2.06. A drive from the cheapest row has to be compared delivered, which how this site works out price per terabyte explains. And a used enterprise drive is often the cheapest terabyte on the page, with its own checks: buying used drives on eBay for the SMART history, enterprise drives for workload ratings, sector formats and locked drives, and CMR vs SMR before any of them goes into an array.

Plan three years ahead, buy eighteen months at a time: 2026 prices make the rule stricter

Growth is the number most often underestimated, and the evidence of 2025 and 2026 is that it ran fast at the bottom of the market as well as the top. Valve’s survey shows the share of Steam machines with more than 1 TB rising 3.79 points in September 2026 alone (one month of an opt-in survey whose sample shifts, so a direction rather than a rate); Apple doubled the iPhone’s entry storage to 256GB in September 2025; a big new PC game asks for about 100 GB; a working photographer adds more than a terabyte a year. At the same time the opposite pull appeared: TrendForce said on 30 September 2026 that PC brands were cutting SSD capacities in mainstream models as flash costs rose. A 2026 base configuration is where the maker’s costs stopped, not a statement about what the buyer needs, the same conclusion the memory site reaches about RAM.

Compound growth is the arithmetic to do before buying:

Size the plan for three years of your measured growth, and buy the next twelve to eighteen months of it now wherever the layout lets you add the rest. Three years is long enough to decide a pool’s width, bays and parity once rather than every spring. Twelve to eighteen months is as far ahead as anyone should buy at 2026 prices, which is where buy storage now or wait lands after working through the forecasts. Order the expansion when the pool crosses 70% full, so that it arrives before 80% even with a long delivery, and track the growth trend, as TrueNAS advises, rather than today’s percentage.

2026 is the year that rule became stricter, because both media became expensive at once. TrendForce’s quarterly forecasts for NAND flash, the chips in every SSD, called for rises of 85 to 90% in overall NAND prices in the first quarter of 2026 (a revised forecast of 3 March 2026, with the quarter already under way), and for contract prices 70 to 75% in the second (31 March), 10 to 15% in the third (3 July) and 15 to 20% in the fourth (30 September). If all four forecasts held, contract prices would end 2026 about four times where they began: 1.85 x 1.70 x 1.10 x 1.15 = 3.98 at the low end and 4.59 at the high. That compounding is mine, and it multiplies forecasts, not settled prices. In the shops, hard drives rose as fast as SSDs: ComputerBase’s price check of 17 September 2026 found the most popular hard drives in German shops 134% dearer than a year earlier, and SSDs 129%. A maker’s average across all its sales moves more slowly: Western Digital said on its August 2026 call that the year-on-year rise in its price per terabyte had moved “from high single digits last quarter to high teens this quarter”. Single drives show the extremes: Tom’s Hardware found a Samsung 990 Evo Plus 2 TB that it had recommended at a $113 deal price on Prime Day 2025 on sale for $363 on 29 July 2026, “triple the price”, a figure the PC builder MetaPCs repeated on 28 September 2026.

What that changes in practice, in order of weight:

1. Do not buy year three at today’s price where expansion is possible. A NAS with free bays, a mirror that can take a second pair, a desktop with a free M.2 slot and a cloud plan can all grow later; buy the next twelve to eighteen months.
2. Buy the final width where expansion is not possible, and pick the drive size on price per usable terabyte. A full chassis and a RAIDZ group that would keep its old parity ratio are the cases where the layout must be right on day one. In ComputerBase’s September 2026 check, large drives were still about a third cheaper per raw terabyte, as buy storage now or wait shows, while a wider layout of smaller drives pays a smaller multiplier (section 21). Multiply both out on the day’s listings.
3. Buy a soldered laptop’s storage for its whole life at the till, because no one will sell you more later, and put the overflow on an external drive.
4. Count measurement and deletion as capacity. The duplicate photo library, the ISO folder that downloads again, the camera footage past its legal retention and the old VM snapshot are terabytes that cost nothing to recover.

Why the prices rose and how long the analysts expect them to stay high is the 2026 storage price hike; what the forecasts mean for the timing of a purchase is buy storage now or wait.

The whole answer, on one page, by workload

No new facts appear here; the number in brackets is the section, counted from the top of the article, that argues for the row.

Three rules hold across both halves. Measure before buying, because a per-head or per-persona figure has never seen your data (sections 2 and 17). Multiply only what needs it: copies for irreplaceable data, parity for convenience, nothing for what downloads again (sections 3 and 11). And plan three years of measured growth at the final layout, but buy only the next twelve to eighteen months where the rest can be added later, because at 2026 prices idle capacity is the most expensive kind (sections 21 and 22).

What to do with this on a listing page

1. Measure first: used space split into replaceable and irreplaceable, its growth over the last year, and the largest working set, from the tools in section 2.
2. Run the formula once per copy (section 3), with the versions, layout and free-space factors that apply, and convert to drive labels with 0.9095.
3. Choose the layout from the bay count and the drive size together, then check the rebuild hours for the size you are considering (section 21).
4. Open the section for the medium: NVMe SSDs for the system, games and active projects; 3.5-inch hard drives for libraries, backups and recorders; every drive together to compare across both.
5. Pick a capacity point, not a range. The cap filter selects one exact capacity in decimal gigabytes, so compare 8 TB drives with 12 TB and 16 TB side by side; min and max bound price, not size.
6. Convert the price per terabyte the page shows into a price per usable terabyte with the multiplier for your layout from section 21, and compare capacities on that, not on the raw figure.
7. For an array, keep shingled drives out: hard drives with SMR hidden, checked against CMR vs SMR, and NAS-rated drives for their error-recovery behaviour.
8. For a recorder, check the workload rating against the streams as well as the capacity against the retention, on surveillance-rated drives.
9. Start from the default view. Bid-only auction rows and for-parts listings are hidden because neither is a price for a working drive; buy=now removes auctions altogether.
10. If buying used, follow buying used drives on eBay for the checks on arrival, and for SAS pulls from the SAS listings, confirm the controller first with drive interfaces.

The number is yours to measure; no tier list has seen your files. This guide supplies the sizes the platform owners publish, the multipliers that every copy, layout and filesystem applies, and the arithmetic that turns a measured library into a count of drives, so that the terabytes bought are the ones that will be used, and the ones that must not be lost are kept more than once.