How much RAM do you actually need in 2026? Home and enterprise

Sixteen gigabytes still runs office work, browsing and nearly every current game, but in 2026 it is the floor rather than the recommendation for anything beyond office work and esports: 32GB is the target for a new gaming or working desktop, 64GB for serious creative work, and local AI and servers are sized by arithmetic rather than by tier. The floors are now published by the platform owners themselves: Microsoft will not call a laptop a Copilot+ PC with less than 16GB. And the direction of travel reversed this year, as 8GB machines came back into the shops. Apple’s MacBook Neo has no other memory option, and Microsoft sells 8GB Surface models.

The tier lists that fill the search results - 8, 16, 32, 64 - are not so much wrong as incomplete, and the four things they leave out are the four that decide the real number. The graphics card: a card with 8GB of video memory pushes game data into system RAM, and in Valve’s August 2026 Steam survey about 69% of machines in the Windows video-memory table had 8GB or less (the all-platform table, which Valve counts separately, gives about 46%). The configuration: AMD rates its Ryzen 9000 desktop processors at DDR5-5600 with two modules and DDR5-3600 with four, so capacity has to come from the size of the modules rather than their number. What “used” means: PC Gamer measured a machine running Star Citizen with 29.3GB in use when 32GB was installed and 13.7GB when 16GB was. And the price of each step: at Tom’s Hardware’s US price index of 14 September 2026 the 32GB DDR5 kit was the cheapest DDR5 gigabyte on the list, cheaper per gigabyte than the DDR5 kits below and above it. None of the tier lists covers the other half of the market either, where memory is sized per session, per core and per memory channel, and where a single 256GB server module was reported by Tom’s Hardware in September 2026 to retail for around $19,000.

This article covers both halves: how to measure your own need; every home workload, with its published requirement beside its measured result; the arithmetic of local language models; the population rules that decide which capacities a desktop reaches at full speed; what each step cost in September 2026, and why a gigabyte you need is worth more than a speed grade you do not; laptops whose memory is fixed at purchase; five home buyers, from a student to a home server, worked from the workload to the kit; then fleets and workstation seats, virtual desktops, servers, databases, AI inference hosts and capacity planning at 2026 prices. It is long because the answer changes with the workload. Readers who want the numbers first 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, an analyst’s or a test’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 check it and disagree with the inputs. Tom’s Hardware’s price index, which this article leans on, lists the single cheapest US listing it found for each row, which is a floor rather than a median. Prices on this site move daily, so where the site is the evidence the article names the page to read rather than freezing a figure that will be wrong next week. Everything here is as of late September 2026.

Three neighbouring questions have their own articles. Why memory became expensive is the 2026 DRAM shortage. When to buy is buy RAM now or wait. DDR4 as a cheaper platform, used memory and laptops with slots are DDR4 as an escape route. This one is about how much.

Sixteen gigabytes is now a published floor, and 8GB has slipped below it

Microsoft’s Windows 11 specifications page gives 4GB as the minimum, which is the amount needed to install and boot rather than a working amount. Application vendors publish their floors in the same spirit, each as though nothing else were running:

Windows 11 minimum          4 GB   Microsoft, Windows 11 specifications page
Teams desktop client        4 GB   Microsoft Learn, updated 16 July 2026
Microsoft 365 Apps          4 GB   Microsoft Support, requirements page
                           -----
stated minimums, stacked   12 GB   before a single browser tab is open

**Each of those figures assumes it is the only thing on the machine, and an ordinary working day is all three at once plus a browser.

The one 2026 measurement of an 8GB Windows machine is not encouraging. The Verge tested one of Microsoft’s new 8GB Surface models, and TechSpot reported on 20 July 2026 that with about ten Chrome tabs, Slack and Signal open, Task Manager showed around 6.7GB of the 7.6GB available in use “most of the time”: 6.7 / 7.6 = 88%, and the laptop froze for several seconds whenever someone joined a Teams call. After a clean restart Windows used about 4.2GB, and half a dozen Chrome tabs took it to about 5.5GB: 5.5 / 7.6 = 72% with only a browser open. Microsoft’s store copy, quoted in the same report, calls 8GB “great for everyday use”, and TechSpot notes that the store page says a further 8GB is required for Copilot+ AI features.

The 8GB Mac has been measured too, and it fails more gently. Macworld’s Roman Loyola spent a week pushing a MacBook Neo in March 2026: 41 Safari tabs left “over 4GB of swap”, 41 Chrome tabs over four hours “over 5GB”, and at 59 tabs “The swap grew to nearly 8GB”, the size of the installed memory. A 67-minute 1080p podcast export took 31 minutes, “about 10 minutes longer” than on an M5 Max MacBook Pro. He noticed no stalls and called his results “all anecdotal”. An 8GB Mac keeps working by writing its overflow to the SSD, so the cost shows up as swap and slower long jobs, and the memory can never be added to.

The Copilot+ requirement is the only memory floor above 8GB found from a PC platform owner. Microsoft’s specifications page requires 16GB of DDR5 or LPDDR5 for a Copilot+ PC and says devices with less “are not Copilot+ PCs” even when the processor qualifies. The other platforms draw their lines lower, or leave them to the configurator:

Until 2025 the trend ran one way: Apple made 16GB standard on the M2 and M3 MacBook Air on 30 October 2024. In 2026 it reversed. Apple announced the MacBook Neo on 4 March 2026 at $599 with 8GB of unified memory and no other option, and raised it to $699 on 25 June 2026 (Tom’s Hardware, citing Bloomberg). Microsoft introduced 8GB base Surface models, from $849 in June 2026 according to TechSpot. Tom’s Hardware reported on 1 August 2026 that Microsoft plans to optimise Windows 11 for 8GB systems; that is a promise about future software, and the Surface measurement is the software shipping now. A 2026 base configuration is where the maker’s costs stopped, not a recommendation about what you need. Section 13 puts numbers on why the costs stop there.

The installed base moves more slowly. In Valve’s Steam Hardware Survey for August 2026 (the all-platform table), 16GB was still the most common amount at 41.20% and 32GB second at 37.45%. Tom’s Hardware had forecast on 2 September 2025 that 32GB would become the most common configuration “before the end of 2025”. It did not. Valve’s own August 2025 table, as the Internet Archive captured it that day, had 16GB at 41.88% and 32GB at 36.46% (Tom’s Hardware quoted 41.67% and 35.42%), so the gap narrowed from 5.42 points to 3.75 in a year and did not close. The survey is opt-in and covers Steam users only.

So 16GB holds office work, browsing, video calls and, on the evidence of sections 3 to 5, most current games, but not all of them at once with one heavier application, which is the working day the stacked minimums describe. On a desktop, 16GB is two modules from 16GB DDR5 kits, and the step to 32GB kits is priced in section 11; on a soldered laptop, 16GB is the least this article would put in a Windows machine meant to last (section 13).

“Per cent in use” is the wrong gauge, and the right ones are counters most people never open

Memory holds what the machine is working on now, and the drive holds everything else, which is why a laptop’s “512GB” is its storage. When programs ask for more than the memory holds, the operating system moves the least-used pages to the drive (the page file on Windows, swap on a Mac or Linux) and reads them back far more slowly. The question is how to tell that moment from memory that is merely busy.

Two retail guides published in 2026 give opposite rules for reading Task Manager. HP’s, last modified on 23 July 2026, says: “If usage consistently sits above 80%, your system needs more memory.” Newegg’s, published on 11 September 2026, says that sitting at 50 to 60% means you have the right amount, and that if use never crosses 40%, buying more “will change literally nothing you can feel”. Both read the same number, and it is one the operating system inflates on purpose.

PC Gamer’s April 2026 test shows by how much. On the same RTX 5070 system, Star Citizen averaged 29.3GB of system memory in use with 32GB installed and 13.7GB with 16GB; Microsoft Flight Simulator 2024 27.3GB against 14.9GB; Hogwarts Legacy 22.4GB against 12.8GB at near-identical frame rates. Games, browsers and the operating system spread into memory that is present and hand it back when asked. Memory in use on a large machine measures what was available, not what was needed. The signals that do mean a shortage record the machine being forced to give something up.

On Windows, read commit charge against commit limit, and then hard faults. Microsoft Learn’s page-file documentation (12 February 2026) says the commit charge cannot exceed the commit limit, “the sum of physical memory (RAM) and all page files combined”, and that a system-managed page file grows, up to three times physical memory or 4GB, when the charge reaches 90% of the limit. Task Manager’s Performance tab shows both under “Committed”. Hard faults are pages read back from disk, shown per process in Resource Monitor; sustained hard faults under your normal workload are the signal that more memory would help. Windows also compresses memory before paging, one more reason “in use” can run high with nothing wrong.

On a Mac, the gauge is the memory pressure graph, not “Memory Used”. Apple’s Activity Monitor guide says green means the Mac “is using all of its RAM efficiently”, yellow that it “might eventually need more RAM”, and red that it “needs more RAM”. macOS compresses inactive apps as the Mac approaches capacity, so a nearly full Mac with a green graph is working as designed.

On Linux, the pair is MemAvailable and pressure-stall information. The kernel documentation defines MemAvailable as “an estimate of how much memory is available for starting new applications, without swapping”, which is the number “free” is usually mistaken for. /proc/pressure/memory reports the share of time some tasks (“some”) or all non-idle tasks (“full”) were stalled on memory, over 10, 60 and 300 seconds; a persistent non-zero “full” line is the state the kernel documentation calls thrashing.

Windows: Performance Monitor counters
  \Memory\Committed Bytes      commit charge: memory promised to processes
  \Memory\Commit Limit         physical RAM plus every page file
  \Memory\Available MBytes     memory free for new work
  \Memory\Page Reads/sec       hard faults: pages read back from disk

Linux
  $ grep MemAvailable /proc/meminfo
  $ cat /proc/pressure/memory
  some avg10=0.00 avg60=0.00 avg300=0.00 total=0
  full avg10=0.00 avg60=0.00 avg300=0.00 total=0

Record peaks across a real working week, not an idle minute. Two machines, with illustrative numbers, show the reading on both sides of the decision:

Machine A: 16 GB RAM, 8 GB page file
  commit limit              16 + 8                 = 24.0 GB
  peak commit over a week   22.4 GB
  peak as share of limit    22.4 / 24.0            = 93%   past the 90% growth point
  Page Reads/sec            sustained during the peak
  verdict                   short of memory; 32 GB changes this machine

Machine B: 32 GB RAM, 8 GB page file
  commit limit              32 + 8                 = 40.0 GB
  peak commit over a week   29.0 GB
  peak as share of limit    29.0 / 40.0            = 73%
  Task Manager "in use"     85% while gaming
  Page Reads/sec            near zero
  verdict                   no signal; 64 GB changes nothing

Machine B is the one the 80% rule would send shopping, and it needs nothing. RAM caching goes further into working-set measurement, including the standby list that makes a busy Windows machine look full, and how to check what RAM fits covers reading what is installed and the memory a machine reserves for hardware.

Every page written to the page file or swap is also a write to the drive. No source measured what sustained paging costs in endurance, but it is a reason to add memory, never a reason to treat an SSD as spare RAM. Servers have their own misleading gauge, vSphere’s “consumed” memory (section 17).

Buy memory for sustained paging under your own workload, not for a high “in use” figure.

Game requirements have settled at 16GB, and the tests side with the minimum

The Steam listings of seventeen major releases from 2025 and 2026 were checked for this article on 26 September 2026: a selection of the big titles, not a random sample. Fourteen ask for 16GB as the minimum, two for 8GB and one for 12GB. Five recommend 32GB, all from 2025: Kingdom Come: Deliverance II, The Elder Scrolls IV: Oblivion Remastered, DOOM: The Dark Ages, Mafia: The Old Country and Borderlands 4. None of the five 2026 releases checked - Resident Evil Requiem, Marathon, Crimson Desert, PRAGMATA and 007 First Light - recommends more than 16GB. Escape from Tarkov, on Steam since 15 November 2025, is the only title found that recommends 64GB. Microsoft Flight Simulator 2024’s Steam listing stops at 32GB, although a Tom’s Hardware article of 3 May 2026 said it lists 64GB as “the ideal”.

Two independent tests with disclosed rigs measured what the games need. PC Gamer’s, published on 10 April 2026, ran a Ryzen 9 9900X with DDR5-6000 CL30 and nothing else running, on the eight most memory-hungry of 27 games examined. With an RTX 5090, Cyberpunk 2077, Kingdom Come: Deliverance II, Black Myth: Wukong and S.T.A.L.K.E.R. 2 (Stalker 2 from here on) “all run just as well with 16 GB of system memory as they do with 32 GB”; Cyberpunk 2077 averaged 83fps against 84fps. TechSpot’s, published on 8 January 2026, was deliberately messier: a Ryzen 7 9800X3D and RTX 5090 at 4K, on a three-year-old Windows installation with Chrome (six tabs) and Discord open, and capacity limited on one 64GB kit so timings and ranks stayed identical. Thirteen of its fourteen tests ran about as well on 16GB as on 32GB or 64GB: “if you have sufficient VRAM or can avoid running out, then most games will still run perfectly well with 16GB of system memory”. Tom’s Hardware’s best-RAM guide, updated 28 July 2026, agrees that 16GB “is far more affordable and sufficient for gaming and mainstream productivity tasks”. That is two independent tests and a buying guide on one conclusion.

Requirements are Steam’s. The measured column is memory in use, mostly with more than 16GB installed, which is why a figure above 16GB does not by itself mean that 16GB failed (section 2); Mafia: The Old Country is the one title where it did (section 4). Escape from Tarkov was measured by PC Gamer, not TechSpot.

Microsoft’s own advice has wavered. A Windows Learning Center page of November 2025, “How to optimize your gaming PC setup”, said “16 GB is plenty for most games. 32 GB is ideal for serious players who run the most demanding titles or use heavy mods”; TechSpot reported its removal on 5 August 2026, and the Internet Archive keeps the wording. Another, reported by Tom’s Hardware on 3 May 2026, called 16GB the “practical starting point” and was deleted a day later.

“Recommended” on a store page is not “required”, and the tests show it. A publisher writes the recommended figure for its recommended machine, and the game then expands into whatever memory is present. With a graphics card of 16GB or more and other programs closed, 16GB DDR5 kits run nearly everything. The next two sections are about everyone else, which is why this article’s target for a new gaming desktop is 32GB, or 32GB of DDR4 on a platform that takes it.

Where 16GB runs out in games: one heavy title, a small graphics card or a second program

Average frame rates barely move between 16GB and 32GB. The 1% lows - the slowest frames, which a player feels as stutter - move in specific, named cases.

The first is a heavy title. TechSpot called Mafia: The Old Country the first game it had tested where 16GB was not enough: “around 20GB of system memory is required for a smooth experience, meaning a 24GB to 32GB kit is recommended. The experience with 16GB was genuinely bad.”

The second is a smaller graphics card. On a 12GB RTX 5070 the extra memory showed in PC Gamer’s results: Escape from Tarkov’s 1% lows rose from 47fps to 63fps with 32GB while its average rose only from 110fps to 120fps, Microsoft Flight Simulator 2024 thrashed badly on 16GB, and PC Gamer noted that live Tarkov raids use more memory than its offline test.

1% lows, 32 GB against 16 GB, RTX 5070 (PC Gamer, April 2026)
  Escape from Tarkov    63 / 47   = 1.34    34% higher
  Stalker 2             51 / 46   = 1.11    11% higher
  MSFS 2024              6 /  2   = 3.00    PC Gamer's "200% increase"
  Star Citizen          28 / 15   = 1.87    87% higher

The third is a second program, and it matters most because it describes how most people play. PC Gamer measured that on 16GB its heavy games, with nothing else running, “leave between 3 and 5 GB of memory available for anything else”, so the game and Windows occupy 16 - 5 = 11GB to 16 - 3 = 13GB, and the encoder, voice client, browser and overlays must fit in what is left. PC Gamer warned that for anyone who likes “to stream your gameplay, 16 GB of system memory will almost certainly be a problem if you like to use maximum quality settings in some games”. That contradicts the line in system-builders’ guides, CyberPowerPC’s of February 2026 among them, that 16GB allows gaming while streaming “without compromising performance”.

TechSpot’s test kept Chrome and Discord open and still found 16GB adequate, but with a 32GB graphics card at 4K. A voice client and a few tabs fit; a streaming encoder, a browser playing video and a heavy title on a mid-range card do not.

At the other end, competitive titles stay light. Counter-Strike 2’s only stated figure on Steam is 8GB. Activision lists 8GB minimum and 12GB recommended for Call of Duty: Black Ops 7, with 16GB for its competitive and Ultra 4K tiers. Riot lists 4GB for VALORANT, and 2GB minimum and 4GB recommended for League of Legends. Minecraft moved the other way: Mojang raised Java Edition’s requirements on 21 July 2026 to a minimum of 8GB with a discrete graphics card or 12GB with integrated graphics, and 16GB recommended, noting that “Compared with many modern AAA games, these updated requirements are still relatively low.” TechSpot found 8GB poor or unplayable in most current titles, and saw Space Marine 2 silently swap in low-resolution textures at 8GB. TechRadar’s suggestion in December 2025 that, at current prices, gamers “may be able to skate by with half that” (8GB, half of 16GB) holds for esports titles and not beyond them.

The same tests settle the 64GB question for a machine that only games. TechSpot also tested every title with 64GB, and its one failure at 16GB, Mafia, needed what TechSpot called “a 24GB to 32GB kit”; only Escape from Tarkov’s publisher recommends 64GB. For gaming alone 64GB buys nothing the tests could find, and Steam’s August 2026 survey puts 64GB machines at 3.97%. It earns its place when the same machine edits, compiles or runs a model (sections 6 to 8).

16GB fails first when something else is running, and something else is usually running. The step from 16GB to 32GB buys headroom for the second program and nothing in the averages; whether to take that step now or later is the question of buy RAM now or wait.

An 8GB graphics card borrows system memory, and integrated graphics has none of its own

When a game’s textures and buffers do not fit in video memory, the overflow lands in system RAM. TechSpot measured it in January 2026 with two versions of one card. With the 8GB Radeon RX 9060 XT the system consumed 23GB of memory, “just over 30% more than the 16GB graphics card”, and read 26GB/s from system memory against 15GB/s. Going from 16GB to 32GB of system memory lifted the 8GB card’s 1% lows in Spider-Man 2 by more than 60%, while the 16GB card was “largely unchanged”. In Marvel Rivals the 8GB card on 16GB pushed data to the page file, “leading to terrible frame time performance”.

8 GB card, Spider-Man 2 (TechSpot, January 2026)
  memory in use with 32 GB installed                 23 GB
  memory in the smaller configuration                16 GB
  shortfall to be paged or dropped                   up to 23 - 16 = 7 GB
  1% lows recovered by installing 32 GB              more than 60%

The 60% is what the shortfall costs, and small cards are the common case. In Steam’s August 2026 survey of Windows machines, 8GB was the most common amount of video memory at 34.03%, and adding every smaller size gives the share at 8GB or below (Valve’s all-platform table, counted separately, gives 46.35%):

128 MB 6.80 + 512 MB 2.41 + 1 GB 0.70 + 2 GB 3.52 + 3 GB 0.67
  + 4 GB 9.89 + 6 GB 10.87 + 8 GB 34.03                  = 68.89%

The strongest measured case for 32GB in gaming is a graphics card with 8GB or less of its own memory, and the other fix for the same problem is a card with more video memory.

Integrated graphics has no memory of its own, so the system memory is the graphics memory, and both its size and its channel count decide what the graphics can do. Intel’s specification pages, read through a text-only proxy, say its Arc graphics branding on H-series Core Ultra systems requires “at least 16GB of system memory in a dual-channel configuration”, and that Iris Xe branding requires dual channel, “Otherwise, use the Intel UHD brand”. Mojang’s July 2026 minimum for Minecraft puts a number on the same effect: 4GB more system memory for a machine whose graphics borrow it. Tom’s Hardware put it bluntly on 31 July 2026: “Dual-channel memory is vital if you’re using an iGPU.”

Whatever is assigned to graphics is gone from the operating system, which is the hardware-reserved memory how to check what RAM fits explains. On a laptop with one slot filled, a second DDR5 laptop module is worth more to the graphics than a bigger first one; on a desktop running integrated graphics, two modules from 16GB DDR5 kits are the least that meets the 16GB dual-channel condition Intel sets for Arc graphics on its H-series laptop chips. On an integrated-graphics machine, capacity and channel count are the graphics card.

Vendor, benchmark and workstation builder disagree on creative work by a factor of three

Creative applications have three kinds of published evidence, and they answer differently for the same program.

The vendors’ figures are floors for one application on a short project. Adobe’s Premiere requirements page, updated on 16 September 2026, asks for 8GB and recommends 16GB “for HD media” and “32 GB or more for 4K and higher”. The DaVinci Resolve 21 figures come from ClipVerdict’s requirements guide (August 2026, checked against the 21.0.4 readme; Blackmagic’s own page was not read).

The benchmark isolates capacity and measures what a shortfall costs. Puget Systems’ test of 1 June 2026 ran a Ryzen 9 9950X3D2 and RTX 5080 with two 32GB DDR5-5600 modules, limiting Windows to 32GB and then 16GB with bcdedit so that nothing but capacity changed. 16GB lost ground in every application, by the margins in the table below. 32GB was level with 64GB in most suites, Lightroom’s AI tools being the exception at 15% slower. Puget concluded that 32GB is a cost-effective option for students and hobbyists, and that 64GB “was the sweet spot in our testing” for professionals.

The workstation builder sizes for long projects and several applications at once. Puget’s own recommendation page for Premiere, modified on 12 August 2026, gives a minimum of 64GB for 1080p footage, 96GB for 4K, 128GB for 6K, 192GB for 8K and 256GB for 12K, and its Resolve page carries the same table.

For 4K editing in Premiere the range runs from Adobe’s 32GB to Puget’s 96GB: 96 / 32 = 3, the factor in this section’s heading. Each end is right about something. Adobe’s figure is its recommendation for 4K work in the application. Puget’s benchmark is what a shortfall costs on test projects that Puget itself says do not reflect real edits, which vary in footage and length; that is why it still calls 16GB “not recommended” for Premiere. And Puget’s workstation figure is what a long professional timeline, with other applications open, grows into. Tom’s Guide’s video editor (8 March 2026) calls 32GB the sweet spot for most editors and recommends proxy media, lower-resolution copies of the footage that cost disk rather than memory, as the free alternative.

The configuration caps what a desktop can reach. With two modules at full rated speed the ceiling is 2 x 64GB = 128GB (section 10); 192GB and 256GB need four modules at the derated speed, and Puget’s Premiere page notes that going beyond four modules usually means a Xeon W or Threadripper PRO workstation on registered DDR5. On a desktop the working steps are 64GB, 96GB and 128GB kits of two modules each.

The benchmark sets the floor for professional work at 32GB, and the builder’s figure is what long projects consume: 64GB for most professionals, more only for 4K timelines and heavy compositing.

Developers pay for each virtual machine twice: once in the guest and once in the default

A development machine runs an editor, a browser, containers, often a Linux virtual machine and a phone emulator, and the largest default claim on its memory is one most people never see. Microsoft Learn’s WSL configuration page (updated 16 September 2026) says WSL 2 assigns “50% of total memory on Windows” to its Linux virtual machine unless told otherwise, so a 16GB laptop hands up to 8GB to Linux before the editor starts. Docker Desktop for Windows asks for 8GB of system RAM with either its WSL 2 or its Hyper-V backend (Docker Docs, modified 7 September 2026). Google’s Android Studio page, updated on 24 September 2026, lists 8GB for the IDE alone and 16GB with the Android Emulator, recommends 32GB, and says larger projects and several virtual devices “require higher RAM”. Microsoft’s Visual Studio 2026 requirements (updated 1 April 2026) give a 4GB minimum, recommend 16GB “for typical professional solutions” and say it “works best with 64 GB RAM”. Epic’s specifications for Unreal Engine 5.8 recommend 32GB, while the typical development workstation Epic describes for itself carries 256GB of DDR5-4800 ECC RDIMM.

Microsoft’s own remote-desktop sizing draws the same line, putting software engineers at 32GB and data work at 64GB (section 15).

32 GB Windows laptop, all defaults
  WSL 2 virtual machine        32 x 0.5          = 16 GB
  left for Windows             32 - 16           = 16 GB
  Android Studio with emulator                     16 GB system minimum (Google)
  plus the browser, the chat client, Docker Desktop and the IDE's indexer
  result                       the host pages as soon as the emulator
                               and a container run together

Fix 1: cap WSL 2 at 8 GB
  WSL 2 virtual machine                             8 GB
  left for Windows             32 - 8            = 24 GB

Fix 2: a 64 GB desktop, defaults left alone
  WSL 2 virtual machine        64 x 0.5          = 32 GB
  left for Windows             64 - 32           = 32 GB

The cap is one setting in the .wslconfig file in the Windows user profile, and it costs nothing:

[wsl2]
memory=8GB

For development, 32GB is the working floor and 64GB is where the defaults stop fighting each other; the cheapest fix before buying anything is capping WSL’s share. On a desktop both fixes exist, and the second is two modules from 64GB DDR5 kits rather than four from 32GB kits. On a laptop with soldered memory only the first exists after purchase (section 13).

Local AI is sized by arithmetic: parameters times bits, plus a cache that grows with context

Running a language model locally is the one home workload whose memory requirement can be calculated before buying anything, and the calculation has two parts.

The weights take parameters times bits per weight, divided by eight. Nvidia’s inference guide (Nvidia Technical Blog, published 17 November 2023, modified 27 December 2025) gives the base case: a 7-billion-parameter model at 16 bits would take roughly 7B * sizeof(FP16) ~= 14 GB in memory. Quantisation shrinks the bits. The llama.cpp project’s quantisation README lists Llama 3.1 8B at 14.96GiB in F16, 7.95GiB in Q8_0, 6.14GiB in Q6_K and 4.58GiB in Q4_K_M, at 4.8944 bits per weight, and says “memory and disk requirements are the same”. The same table gives the other formats’ bits per weight, and the sizes check against them:

bits per weight, README        check: size / F16 size x 16   (llama.cpp's 8B row)
  Q8_0      8.5008             7.95 / 14.96 x 16   = 8.50
  Q6_K      6.5633             6.14 / 14.96 x 16   = 6.57
  Q4_K_M    4.8944             4.58 / 14.96 x 16   = 4.90

70B-class model, 70.6e9 parameters (an assumption for a "70B")
  F16       70.6e9 x 16     / 8  = 141.2 GB
  Q8_0      70.6e9 x 8.50   / 8  =  75.0 GB
  Q4_K_M    70.6e9 x 4.8944 / 8  =  43.2 GB   (40.2 GiB)

Llama 4 Scout, 109e9 parameters in total
  Q4_K_M    109e9 x 4.8944 / 8   =  66.7 GB
  2 bits    109e9 x 2      / 8   =  27.3 GB

The README’s own example puts Llama 3.1 70B at 43.1GB in Q4_K_M, and Corsair’s April 2026 guide puts Llama 3.3 70B at Q4 at “about 42 GB just for the model weights”; both agree with the 70B line.

Qwen’s model card gives Qwen3-32B 32.8 billion parameters, so its weights at Q4_K_M are 32.8e9 x 4.8944 / 8 = 20.1GB. The last column adds the cache at 4,096 tokens from the per-model figures below and about 8GB for the operating system and tools, an assumption: 4.9 + 0.5 + 8 = 13.4GB for the 8B model, 20.1 + 1.1 + 8 = 29.2GB for Qwen3-32B and 43.2 + 1.3 + 8 = 52.5GB for the 70B. Scout’s layer count was not sourced, so its row keeps a 2GB placeholder, 66.7 + 2 + 8 = 76.7GB, and AMD’s own July 2025 figure for Scout at Q4_K_M, “roughly 96GB of VRAM”, lands on the same row. At 32,768 tokens the sums grow to 17.2GB for the 8B model, past a 16GB machine, 36.7GB for Qwen3-32B and 61.9GB for the 70B, which fills 64GB. The 32-billion-parameter class is the largest that fits a 24GB graphics card at four bits with a short context.

The cache is the part people forget, and it grows with every token of context. Nvidia’s guide gives the key-value cache as batch size x tokens x 2 x layers x hidden size x bytes per value, and works Llama 2 7B at 16 bits, 4,096 tokens and a batch of one to about 2GB:

per token      2 x 32 layers x 4,096 x 2 bytes   = 524,288 bytes = 0.5 MiB
4,096 tokens   0.5 MiB x 4,096                   =  2 GiB
32,768 tokens  0.5 MiB x 32,768                  = 16 GiB   more than the weights
cache at 8 bits (q8_0)                           =  1 GiB and 8 GiB

Current models need far less per token, because grouped-query attention keeps 8 key-value heads instead of one per attention head. Meta’s paper “The Llama 3 Herd of Models” (23 July 2024) gives Llama 3’s 8B model 32 layers and its 70B model 80, each with 8 key-value heads of 128 dimensions, and Qwen’s model card gives Qwen3-32B 64 layers and 8 key-value heads. In Nvidia’s formula the hidden size becomes the key-value heads times their 128 dimensions:

per token, 16 bits = 2 x layers x key-value heads x 128 x 2 bytes
  Llama 3.1 8B      2 x 32 x 8 x 128 x 2   = 131,072 bytes = 128 KiB
  Qwen3-32B         2 x 64 x 8 x 128 x 2   = 262,144 bytes = 256 KiB
  Llama 3.x 70B     2 x 80 x 8 x 128 x 2   = 327,680 bytes = 320 KiB

                    4,096 tokens    32,768 tokens    131,072 tokens
  Llama 3.1 8B        0.5 GiB          4 GiB           16 GiB
  Qwen3-32B           1 GiB            8 GiB           32 GiB
  Llama 3.x 70B       1.25 GiB        10 GiB           40 GiB

PromptQuorum’s local-LLM hardware guide (16 September 2026) prints the same nine figures. At long context a large model’s cache is as big as its weights: 40GiB, about 43GB, beside 43.2GB for a 70B model at four bits and 131,072 tokens.

Ollama’s FAQ still gives a 4,096-token default, but its context-length page (September 2026) says the default now follows video memory: 4k tokens under 24GiB, 32k from 24 to 48GiB and 256k at 48GiB or more, and it advises at least 64,000 tokens for agents and coding tools. On a machine with a large memory pool, set the context deliberately. The FAQ adds that a q8_0 cache “uses approximately 1/2 the memory of f16”, q4_0 about a quarter.

Mixture-of-experts models save compute, not capacity. Hugging Face’s explainer says “all parameters need to be loaded in RAM”, so Mixtral 8x7B needs the memory of a dense 47-billion-parameter model, and AMD’s July 2025 post says that for Llama 4 Scout, with 17 billion parameters active per token, “all 109 billion parameters need to be held in memory - so the footprint is the same as a dense 109 billion parameter model”. OpenAI’s model card says gpt-oss-20b (21B parameters, 3.6B active) runs “within 16GB of memory” and gpt-oss-120b (117B, 5.1B active) on one 80GB GPU.

Three errors circulate in 2026 guides. There is no “Llama 3.2 8B”: Meta’s Llama 3.2 text models come in 1B and 3B sizes, and the 8B is Llama 3.1. Scout cannot run in “about 10GB”, because every expert must be resident and even at two bits its weights are 27.3GB. And mixture-of-experts models are not memory-efficient; they are compute-efficient.

The built-in end of local AI fits the platform floor: Microsoft’s on-device model, Phi Silica, runs on the neural processors of 16GB Copilot+ PCs (Microsoft Learn, 15 July 2026, which says a model called Aion Instruct is scheduled to replace it on retail devices in November 2026), and LM Studio recommends 16GB or more, telling 8GB Mac owners to “stick to smaller models and modest context sizes”. The large-model end is a different machine: a 70B model at four bits wants two 48GB DDR5 modules from the 96GB kits, fits tightly in 64GB at short context, and at eight bits (75.0 + 1.3 + 8 = 84.3GB) fits tightly in 96GB and comfortably in 128GB. Capacity follows total parameters; speed follows active parameters.

Capacity decides whether a model runs, and bandwidth decides how fast

Generating each token means reading every active weight once, so peak memory bandwidth divided by the bytes read per token is a hard ceiling on tokens per second. Real systems land below it.

A dual-channel desktop’s bandwidth is the transfer rate times eight bytes per 64-bit channel times two channels: at DDR5-6000, 6,000 x 8 x 2 = 96GB/s. One 2026 local-AI guide quotes 32 to 70.4GB/s for DDR5, which is the range for a single channel from DDR5-4000 to DDR5-8800, not the bandwidth of a desktop. With four modules at AMD’s rated DDR5-3600 the figure falls to 57.6GB/s. The speed guide derives these numbers three ways.

The machines built for local AI break the two-channel limit, all of them with memory soldered next to the processor. AMD’s Ryzen AI Max+ 395 has a 256-bit LPDDR5x-8000 interface and up to 128GB, which is 8,000 x 32 bytes = 256GB/s. Nvidia lists its DGX Spark with 128GB at 273GB/s. Apple’s M5 Max supports up to 128GB at up to 614GB/s (Apple Newsroom, 3 March 2026). Apple’s M5 Ultra Mac Studio offers up to 512GB at 1.2TB/s, the 512GB option arriving in late October (Apple Newsroom, 22 September 2026). A graphics card is the other route, and its limit is capacity: Nvidia’s RTX 5090 has 32GB, so a 43GB model at four bits does not fit and the remainder runs at system-memory speed.

Every ceiling is bandwidth divided by bytes per token (96 / 43.2 = 2.2), and Scout’s 10.4GB is its 17 billion active parameters at Q4_K_M: 17e9 x 4.8944 / 8. These are upper bounds, not measurements. They ignore compute, cache reads and prompt processing, and no independent test comparing a DDR5 desktop with these machines was found. Two figures check the method. Tom’s Hardware’s review of the M5 Ultra Mac Studio (21 September 2026) measured almost four times the DGX Spark’s tokens per second on a dense 27-billion-parameter model at Q4_K_M, against a bandwidth ratio of 1,200 / 273 = 4.4. And AMD says its Ryzen AI Max+ 395 runs Scout at up to 15 tokens per second without stating the quantisation; at the Q4_K_M its own FAQ uses for Scout, 15 / 24.6 = 61% of the ceiling, as a real system should land below it.

Beside a graphics card, system memory holds whatever the card cannot, and the card’s share sets the speed. When the model and its cache fit in video memory, system memory only loads the file and runs everything else: PromptQuorum gives “16 GB minimum (with GPU)” and says system memory affects load time and CPU fallback speed “but not which model fits on the GPU”. Workstation builders size more generously. Puget Systems’ machine-learning and AI recommendations (modified 30 June 2026) give as “the first rule of thumb” at least double the system memory of the total video memory: 32GB beside a 16GB card, 48GB beside a 24GB card and 64GB beside a 32GB card, each a two-module configuration (section 10). When the model does not fit, the layers left over are read from system memory for every token, and that part alone caps the rate:

70B class at Q4_K_M, 43.2 GB of weights, DDR5-6000 system memory at 96 GB/s
  24 GB card, weights kept on the card (assumption)          22 GB
    left in system memory              43.2 - 22            = 21.2 GB
    ceiling from system memory alone   96 / 21.2            =  4.5 tokens/s
    system memory needed               21.2 + 8 (OS, tools) = 29.2 GB: 32 GB, barely
  32 GB card, weights kept on the card (assumption)          30 GB
    left in system memory              43.2 - 30            = 13.2 GB
    ceiling from system memory alone   96 / 13.2            =  7.3 tokens/s
    system memory needed               13.2 + 8             = 21.2 GB: 32 GB
  no card, the whole model in system memory  96 / 43.2      =  2.2 tokens/s

The 2GB left free on each card for the cache and the runtime is an assumption, and the card’s own read time comes on top, so these are upper bounds like the table’s. Offloading a 70B model to a 24GB card roughly doubles a desktop’s ceiling, and a 32GB card more than triples it, which puts 7.3 tokens per second just above DGX Spark’s 6.3 and at about half the M5 Max’s 14.2.

A desktop can hold a 70B model in 96GB, but it can read it only about twice a second. For a model that size the choice is capacity on DDR5-6000 sticks, as two modules from the 96GB or 128GB kits (256GB takes four modules and runs at the derated DDR5-3600), a graphics card with the rest offloaded, or bandwidth in a unified-memory machine whose memory is fixed the day it is bought, which DDR4 as an escape route weighs against socketed memory. For an 8B model the desktop is fine: 96 / 4.9 = about 20 tokens per second as a ceiling.

Two modules, not four: the population rules that set a desktop’s capacity steps

A desktop processor has two memory channels and most boards four slots, and filling all four costs speed. AMD’s product pages rate the Ryzen 7 9800X3D and Ryzen 5 9600X at DDR5-5600 with two modules of either rank and DDR5-3600 with four, with 256GB as the maximum. For Intel’s Core Ultra 200S, Tom’s Hardware’s reading of Intel’s datasheet (15 October 2024 and 26 November 2025) gives DDR5-5600 for standard modules at one per channel, and at two per channel DDR5-4800 for single-rank modules and 4400 for dual-rank ones; Intel’s product page gives the same 256GB ceiling.

Bandwidth is the rated speed x 8 bytes x 2 channels: 3,600 x 8 x 2 = 57.6GB/s, and 57.6 / 89.6 = 0.64, a cut of 36%. Two consequences follow.

Capacity comes from module size, not module count. DDR5 has 24GB and 48GB modules, so 48GB and 96GB are two-slot configurations at full speed; Steam’s August 2026 survey already shows 24GB and 48GB machines at 2.22% and 1.20%. The die behind those sizes, and the four-stick derating generation by generation, are in DDR4 vs DDR5. The only way to 256GB on AM5 is four 64GB modules at the derated speed, and Puget notes that filling every slot with high-capacity modules “often results in RAM being downclocked”.

A 2 x 8GB kit is an upgrade dead end. Newegg’s September 2026 guide says so: “8GB modules are what you will end up replacing rather than adding to.” A second pair means four modules and the lower rating, and Tom’s Hardware warned on 31 July 2026 that modules bought separately are not validated together. The same Tom’s Hardware test measured the other shortcut: across 13 games a single 16GB module cost around 8 to 10% against two on most processors (10.9% on a Ryzen 5 7600X) and 2.8% on a Ryzen 7 9800X3D, with capacity differing too.

Two more rules. DDR5-6000 on AM5 is an EXPO profile above AMD’s rated DDR5-5600, and AMD’s AM5 page says operating outside its published specifications “will void any applicable AMD product warranty, even when enabled via AMD hardware and/or software”. And Tom’s Hardware reported on 4 September 2026 that V-Color sells “1+1” packs pairing one real module with an RGB filler, which counts as nothing towards capacity.

Pick the total, then the module size that reaches it in two slots. The CPU index lists what each controller supports, how to check what RAM fits covers training and mixing, ranks and 3DS covers rank, and the 24GB, 48GB and 64GB DDR5 module pages list the two-slot routes to 48, 96 and 128GB.

32GB was September 2026’s cheapest DDR5 gigabyte, and the step up to it cheaper still

Tom’s Hardware’s US RAM price index, updated on 14 September 2026, lists the best price it found for each kit; its “lowest-ever” column is internally inconsistent, so only the best-price column is used.

The 16GB kit costs 14.94 / 12.78 = 17% more per gigabyte than the 32GB kit, and the 96GB and 128GB kits 38% and 43% more. The sixteen gigabytes that take a machine from 16GB to 32GB cost $170 / 16 = $10.63 each, less than the average gigabyte of either kit. The 128GB row is the least stable: the same index showed $3,399 for it on 17 August 2026, a figure Tom’s Hardware repeated on 24 August, $1,060 more than on 14 September.

The shape is not universal. PC Games Hardware found in August 2026 that the usual discount for doubling capacity had disappeared in Germany. And “the price of 32GB” depends on which price is read: from the $409.99 DDR5-5600 kit in a Tom’s Hardware deals article of 17 September to the $609.77 average for DDR5-6000 CL30 kits in RAM Price History’s tracker on 20 September (its cheapest that day was $472.99). This site’s pages state both the cheapest and the median price per gigabyte, with bid-only auctions excluded by default: read the median on 32GB DDR5 kits against 16GB, 48GB, 64GB and 96GB, and against every generation. Each row also states its condition, so a used kit that undercuts the median sits beside the new ones; the checks that make one safe are in buying used RAM on eBay.

The common question is whether to buy 16GB now and add more later. On AM5, at the index’s prices:

Route A: 32 GB now, as 2 x 16 GB                     $409
Route B: 2 x 8 GB now, a second 2 x 8 GB later       $239 + $239 = $478
         $69 more than A, and four modules rated at DDR5-3600 instead of 5600
Route C: 2 x 8 GB now, sold later, 2 x 16 GB later   $239 + later price - resale
         cheaper than A only if  later price - resale  is under $170

Route B loses on both counts. Route C wins only if prices fall far enough before the second purchase, and whether they will is the subject of buy RAM now or wait, with the causes in the 2026 DRAM shortage; retailer bundles and prebuilt machines, which have sometimes carried memory below the kit price in 2026, are covered there too. At September 2026 prices, capacity you will need within two years costs less as one kit now than as two kits later.

A missing gigabyte costs more than a missing megatransfer

At the same September index a 32GB DDR5-5600 kit cost the same $409 as a DDR5-6000 kit, while a 32GB DDR5-6600 kit cost $655. The speed premium from 6000 to 6600 was 655 - 409 = $246, within $6 of the 661 - 409 = $252 step from 32GB to 48GB.

A speed shortfall costs a few per cent, mostly on processors without the large cache of AMD’s X3D parts. A capacity shortfall costs paging: 43% in After Effects, 34% to more than 60% in the 1% lows of games that spill, and “genuinely bad” in Mafia. Buy the speed your platform is rated for, spend the rest on capacity, and never pay a speed premium while you are short of a gigabyte you need. At 5600 to 6000 the premium was nothing; at 6600 it was a whole capacity step. On servers the same trade appears as MRDIMM, which Micron says brings 39% more bandwidth than RDIMM (a vendor claim): it buys bandwidth, not capacity.

Whether 32GB of DDR4 beats 16GB of DDR5 is the same trade: the index listed 32GB of DDR4-3200 at $190, $49 less than 16GB of DDR5-6000, against 11 to 14% of gaming performance that Tom’s Hardware measured DDR4-3200 giving up to DDR5-7200 on Intel’s LGA 1700 processors (28 July 2026). On capacity alone DDR4 wins; whether the platform is worth buying into is the subject of DDR4 as an escape route. CAS latency and the speed guide cover the speed side, and DDR5-5600, DDR5-6000 and 32GB DDR4 kits list the options.

A soldered laptop is sized once, at the till, for its whole life

On a machine with soldered or on-package memory there is no second chance, and the makers price the configure-to-order step accordingly:

Apple, 14-inch MacBook Pro, 24 GB to 48 GB       $600 / 24 GB    = $25.00 per GB
  ($400 in March 2026 per 9to5Mac; after Apple's 25 June rise, $2,499 at
   24 GB per Thurrott and $3,099 at 48 GB per AppleInsider, 23 July 2026)
Apple, M6 Mac mini, 16 GB to 24 GB               $200 / 8 GB     = $25.00 per GB
  (Tom's Hardware, 25 August 2026)
Framework, socketed DDR5 SO-DIMMs                                  $13 to $18 per GB
  (Framework, March 2026)
desktop 32 GB DDR5-6000 kit, best price          $409 / 32 GB    = $12.78 per GB
  (Tom's Hardware index, 14 September 2026)

Apple's upgrade against the desktop kit          25 / 12.78      = 1.96 times
a $600 step spread over a five-year life         600 / 60 months = $10.00 a month

The ceilings are fixed too. Intel lists its Core Ultra 7 258V with a 32GB maximum, so no laptop built on it will ever hold more. The 8GB MacBook Neo sits below LM Studio’s 16GB recommendation for local models.

The makers are under the same pressure: TrendForce’s September 2026 estimate put the CPU, DRAM and SSD at 68% of a $900 notebook’s bill of materials by the third quarter of 2026, up from about 45% in the first quarter of 2025, and it said in July 2026 that consumers were beginning to keep notebooks longer. A longer life is the case for buying more at the start. On a soldered machine the capacity chosen at purchase is the capacity for life, so size it for year four or five rather than the first week: 16GB for office work and study, 32GB for development, creative work or local AI, and not 8GB on Windows, on the evidence of section 1. Laptops that keep DDR5 laptop modules or DDR4 laptop modules in slots, and the newer socketed formats, are compared in DDR4 as an escape route, used memory, and upgradeability.

Five home buyers, worked from the workload to the kit

Each case shows its sources and arithmetic, and ends in a capacity and, where the memory is socketed, a module count.

A student’s Windows laptop, meant to last five years, running Office, Teams, a browser and whatever the course adds:

8 GB Surface, measured (The Verge, via TechSpot)   88% in use with ten tabs and two apps
stated minimums, Windows + Teams + Microsoft 365   12 GB
Copilot+ PC requirement (Microsoft)                16 GB
buy                                                16 GB; 32 GB if the course
                                                   involves development or video
the extra 16 GB at $13 to $25 per GB               16 x 13 to 16 x 25 = $208 to $400
  (Framework's modules; Apple's upgrades)

A gamer with an 8GB graphics card who streams:

game with an 8 GB card (TechSpot)                  23 GB occupied when there is room
encoder, voice client, browser                     not measured by any source; above zero
16 GB                                              cannot hold 23 GB
32 GB as 2 x 16 GB                                 32 - 23 = 9 GB left for the encoder,
                                                   voice client and browser
step cost at the 14 September index                $409 - $239 = $170

Two modules from 32GB DDR5 kits. A graphics card with more video memory would remove the spill instead; 32GB covers the rest.

A 4K editor in DaVinci Resolve who uses Fusion:

Resolve 21 with Fusion (ClipVerdict)               32 GB
Puget benchmark, Fusion score against 64 GB        16 GB 31% lower; 32 GB 5% lower,
                                                   within Puget's normal variation
Puget workstation recommendation for 4K            96 GB
64 GB as 2 x 32 GB at the index                    $889
96 GB as 2 x 48 GB at the index                    $1,699
the extra 32 GB                                    $810, or 810 / 32 = $25.31 per GB

Buy 64GB, which Puget names as the threshold for Fusion; 32GB came within Puget’s normal variation on its test projects, which is why it is the budget choice rather than a mistake. The step to 96GB is the most expensive gigabyte on the ladder and buys long Fusion timelines; proxy media, which costs disk rather than memory, comes first.

A hobbyist who wants a 70B model at four bits:

weights, 70B class at Q4_K_M                       43.2 GB
cache at 4,096 tokens (80 layers, 8 KV heads)      1.25 GiB = 1.3 GB
cache at 32,768 tokens                             10 GiB = 10.7 GB; 5.4 GB at q8_0
operating system and tools (assumption)            about 8 GB
total at 4,096 tokens                              52.5 GB: 64 GB
total at 32,768 tokens                             61.9 GB: 64 GB full, 96 GB comfortable

ceiling on 96 GB of DDR5-6000                      96 / 43.2   =  2.2 tokens per second
ceiling, 24 GB card plus DDR5-6000                 96 / 21.2   =  4.5
ceiling, 32 GB card plus DDR5-6000                 96 / 13.2   =  7.3
ceiling on DGX Spark                              273 / 43.2  =  6.3
ceiling on MacBook Pro M5 Max                      614 / 43.2  = 14.2

Three routes, each with its own cost: 96GB of desktop memory that holds the model and reads it slowly, a 24GB or 32GB graphics card with the rest offloaded to 32GB of system memory, at ceilings of 4.5 and 7.3 tokens per second (section 9), or a 128GB unified-memory machine whose memory can never be changed. For an 8B model none of this applies, and 32GB is plenty.

A home server running Proxmox with a ZFS pool and a few virtual machines, a machine often built from ex-server parts:

Proxmox VE, OS and services (admin guide, 18 Sep 2026)       2 GB
ZFS, about 1 GB per TB of used storage: 16 TB (assumption)  16 GB
guests: a Windows VM and a Linux VM for containers           24 GB   assumption
total                                                        42 GB: 64 GB
64 GB of DDR4 ECC RDIMM at eRacks' $5.56 per GB (section 20) 64 x 5.56  = $356
64 GB of DDR5 RDIMM at eRacks' $31.05 per GB                 64 x 31.05 = $1,987
64 GB as a desktop DDR5-6000 kit (index, 14 September)       $889

A plain file server needs far less. TrueNAS’s hardware guide (modified 11 September 2026) asks for “at least 8 GB of RAM for basic TrueNAS operations with up to eight drives” and 1GB for each drive after that, at least 16GB and preferably 32GB when the pool serves virtual machines over iSCSI, and “the suggested 5 GB per TB of storage for deduplication”: deduplicating 10TB adds 10 x 5 = 50GB. Spare memory becomes read cache (RAM caching covers the ZFS ARC), and TrueNAS notes that most users “strongly recommend ECC RAM” (ECC vs non-ECC weighs that). A home server is sized from its pool and its guests, and registered DDR4 is usually the cheapest gigabyte it can take. The modules are on registered DDR4, which ones a given server accepts is on memory by server model, and testing them on arrival is in buying used RAM on eBay.

A company fleet needs memory tiers by role, and Microsoft has already written them down

No analyst’s fleet-memory standard could be sourced. Microsoft publishes the closest thing. Its Azure Virtual Desktop sizing guidance says that “VM sizing for single-session session hosts usually align with physical device guidelines”, and gives 2 vCPUs and 8GB for light work (basic data entry), 4 vCPUs and 16GB for medium work (consultants and market researchers) and 8 vCPUs and 32GB for heavy work, meaning software engineers and content creators. Windows 365’s size guidance, updated 2 June 2026, extends the ladder: 4GB for frontline and call-centre staff, 8GB for home and consultant users, 16GB for finance and healthcare, 32GB for developers, engineers and content creators, 64GB for software development and data analysis, and 128GB for AI model development that does not lean heavily on a GPU.

Two 2026 facts move the bottom rung: a Copilot+ PC needs 16GB, and Microsoft’s own 8GB Surface froze on Teams calls (section 1). A seat bought now also has to last: HP told investors in February 2026 that memory had become about 35% of the cost of building a PC (The Register, 27 February 2026), and Gartner forecast on 26 February 2026 that the sub-$500 entry-level PC segment would disappear by 2028.

An illustrative 1,000-seat fleet shows what tiering is worth. The mix of 200 task workers, 650 knowledge workers and 150 developers or creators is an assumption, and so is the price: no fleet price for memory is published, so Framework’s March 2026 retail range for DDR5 SO-DIMMs, $13 to $18 per gigabyte, stands in for it.

This article’s tiers come to 200 x 16 + 650 x 16 + 150 x 32 = 18,400GB. Putting everyone on 32GB costs 13,600GB more, $176,800 to $244,800; putting everyone on 16GB saves 2,400GB, $31,200 to $43,200, and leaves 150 developers below Microsoft’s heavy tier. Buy 16GB as the floor for every Windows seat bought in 2026 and 32GB for the named heavy personas, and put the difference in writing so the next refresh can be measured against it. When to place the order is the subject of buy RAM now or wait, and business laptops that still take DDR5 laptop modules in slots are covered in DDR4 as an escape route.

Workstation seats are sized by the size of the work, and two publishers give the rule. Dassault Systèmes’ SOLIDWORKS system-requirements page (read on 26 September 2026) lists 16GB, with 32GB recommended, for SOLIDWORKS 2025, 2026 and 2027. Puget Systems’ SOLIDWORKS recommendations (modified 29 April 2026) turn that into arithmetic: “about 5GB of RAM for Solidworks itself, then at least 20 times the largest assembly size you work with”, rounded to 16GB for assemblies under 500MB, 32GB up to 1.25GB and 64GB up to 3GB. For data analysis the rule comes from the author of pandas: Wes McKinney wrote in September 2017 that “you should have 5 to 10 times as much RAM as the size of your dataset”, before pandas gained Arrow-backed types, so treat it as an upper bound. Machine-learning seats with a graphics card take Puget’s rule from section 9, at least twice the card’s memory.

CAD seat, largest assembly 1.1 GB (Puget's rule)   5 + 20 x 1.1     = 27 GB: 32 GB
CAD seat, largest assembly 2.5 GB                  5 + 20 x 2.5     = 55 GB: 64 GB
analyst, 8 GB dataset in pandas (McKinney, 2017)   8 x 5 to 8 x 10  = 40 to 80 GB
ML seat with two 16 GB cards (Puget)               2 x (2 x 16)     = 64 GB

The results sit on Windows 365’s rungs: 32GB for engineers, 64GB for data analysis and 128GB for AI model development. Up to 128GB a seat is two modules on a desktop platform (section 10); beyond that it is a Xeon W or Threadripper PRO workstation on registered DDR5 (section 6). Size a workstation seat from its largest file, not from its job title.

Virtual desktops are sized per session, then rounded up to fill the memory channels

Virtual desktops come in two models whose memory needs differ by about an order of magnitude.

Multi-session hosts share one operating system among many users. Microsoft’s session-host guidance suggests at most 6 users per vCPU for light work, 4 for medium, 2 for heavy and 1 for power users, on a minimum virtual machine of 8 vCPUs and 16GB for the first three and 6 vCPUs with 56GB for power users. Its first example instance, Azure’s D8s_v5, has 32GiB, twice that minimum. The same guidance says “increasing memory from 8 GB to 16 GB can more than double the number of users you can support” where memory is the limit, and that virtual machines “can incur a 15-20% capacity cost compared to bare metal”.

Single-session desktops, persistent or not, give each user a whole virtual machine and follow the device tiers above. Servermall’s VDI sizing guide of 20 May 2026 gives 4 to 6GB per desktop for office work, 6 to 8GB with heavy browsing, 8 to 12GB with video calls, 12 to 16GB for light graphics and 16 to 64GB for CAD. Its formula is desktop memory plus service-machine memory plus hypervisor memory plus a reserve of 15 to 30%, it warns that “relying on aggressive memory overcommit is dangerous”, and it puts 10 office users at 128 to 256GB and 50 mixed users at 512GB to 1TB. No source sizes the service machines or the hypervisor, so those inputs below are labelled assumptions.

10 office users, single-session desktops (Servermall's profile)
  desktops              10 x 4 to 6 GB                   = 40 to 60 GB
  service machines      16 GB   assumption
  hypervisor             8 GB   assumption
  low                   (40 + 16 + 8) x 1.15             =  73.6 GB
  high                  (60 + 16 + 8) x 1.30             = 109.2 GB
  Servermall's figure   128 to 256 GB: 128 fits; 256 is more than twice the need

50 medium users
  multi-session         50 / 4 users per vCPU            = 12.5 vCPUs: two 8-vCPU hosts
                        2 x 16 GB, Microsoft's minimum   = 32 GB
                        2 x 32 GiB, its example VM       = 64 GiB
  single-session        50 x 16 GB, Microsoft's tier     = 800 GB
                        50 x 8 to 12 GB, Servermall      = 400 to 600 GB
                        plus 24 GB of service machines and hypervisor
                        (assumption), 15 to 30% reserve  = 488 to 811 GB

The factor of ten between the two models is the decision; the rest is rounding, which is not free, because a server socket is fastest with every channel filled by identical modules. On a 12-channel socket at one module per channel the balanced totals are 384GB, 768GB, 1,152GB and 1,536GB (12 x 32, 64, 96 or 128GB). The single-session 50-user range of 488 to 811GB lands on 768GB, as 12 x 64GB on one socket or 24 x 32GB on two, for all but its top 43GB, where the next step is 1,152GB or a 23% reserve instead of 30% (768 / 624 = 1.23). Servermall’s 512GB is not balanced on 12 channels. It would be on 16, and Lenovo reportedly said at ISC 2026, as Tom’s Hardware relayed from ComputerBase on 28 June 2026, that dual-socket servers due in 2027 with 16 channels per processor could need about 1TB just to use their full bandwidth: 32 x 32GB = 1,024GB.

A cluster must also be sized for the host it loses. Servermall’s guide lists “one large server without a fault tolerance plan” among the typical mistakes, and says that for 100 users “the failure scenario of one node should be calculated immediately”. Memory is where that decision is priced, because the surviving hosts must hold every desktop that restarts on them. For the 50 single-session users above, at the 768GB balanced total:

two hosts, each able to carry everyone      2 x 768       = 1,536 GB   100% spare
three hosts, any two able to carry everyone
  each host                                 768 / 2       =   384 GB   12 x 32 GB
  cluster                                   3 x 384       = 1,152 GB    50% spare
four hosts, any three able to carry everyone
  each host                                 768 / 3       =   256 GB, not balanced
                                                            on 12 channels: 384 GB
  cluster                                   4 x 384       = 1,536 GB
the 384 GB of spare in the three-host design
  at eRacks' $31.05 per GB (section 20)     384 x 31.05   = $11,923
  at the 256 GB module's $74.22 per GB      384 x 74.22   = $28,500

More, smaller hosts cut the spare fraction until channel rounding stops them; here three hosts is the cheapest design that survives a failure. Size for the busiest hour’s concurrent sessions: Servermall warns that a farm that copes at noon can “struggle through the first 15 minutes of the workday” as every profile loads, and advises a pilot with real users before buying every host. Size virtual desktops by concurrent sessions, not headcount, then round up to a channel-balanced total. Server memory capacity limits covers population and interleaving, RDIMM, UDIMM and LRDIMM the module types, memory by server model what each machine takes, and registered DDR5, 32GB and 64GB modules the parts.

Servers are sized per core and per channel, and the cloud has already published the ratios

The public clouds publish their memory-to-processor ratios in their instance tables, the nearest thing the market has to a sizing benchmark. AWS’s eighth-generation AMD instances step at 2, 4 and 8GiB per vCPU for compute-optimised, general-purpose and memory-optimised work (c8a, m8a and r8a), and its Graviton4 x8g carries 16GiB; Azure’s Dsv6 runs at 4GiB per vCPU and its Esv6 at 8GiB.

A vCPU is usually a hardware thread, and that matters. AWS’s m8i.large presents 2 vCPUs as one core with two threads, so its 8GiB is 8GiB per core; the m8a.large presents 2 vCPUs as two cores, so its 8GiB is 4GiB per core. “GB per vCPU” and “GB per core” must never be mixed in one plan.

On a 2026 socket the per-core figure depends on how many cores share the channels. AMD lists the 192-core EPYC 9965 and the 64-core EPYC 9575F with the same 12 DDR5 channels and 614GB/s per socket: 614 / 192 = 3.2GB/s per core against 614 / 64 = 9.6GB/s. AMD’s reference system for a two-socket EPYC 9965, in a benchmark footnote, carries 1.5TB as 24 x 64GB DDR5-6400, which is 1,536 / 384 = 4GB per core. Intel lists its 128-core Xeon 6980P with 12 channels, DDR5-6400 or MRDIMM-8800 and up to 3TB per socket (read through a text-only proxy), a ceiling of 24GB per core.

The cloud’s general-purpose 4GiB per vCPU is 8GiB per core where a vCPU is a thread, as on the m8i, twice what AMD’s reference system carries, and 4GiB where it is a whole core, as on the m8a, the same as the reference. The reference is a benchmark configuration rather than a sizing rule: the same AMD footnote gives a 256-core two-socket EPYC 9755 system the same 1.5TB, which is 1,536 / 256 = 6GB per core.

The measured evidence says allocated memory overstates use. Microsoft Azure’s Pond study (arXiv, 2022, so dated) found that “∼50% of all VMs touch less than 50% of their rented memory”, that up to 25% of DRAM becomes stranded as cores are allocated, and that DRAM can be half of a server’s cost. vSphere’s “consumed” memory is what was allocated after sharing savings, not what the guest actively uses, and Kubernetes places pods by their memory requests, so inflated requests strand memory and deflated ones get containers killed. On the other side, Tom’s Hardware reported on 24 August 2026, citing The Register’s reading of a Meta paper, that about 40% of Meta’s fleet is memory-capacity bound.

Size by measured active memory per core, then round up to a total that fills every channel. The ceilings each machine imposes are in server memory capacity limits, the rank rules in ranks and 3DS, and the parts in registered DDR5, MRDIMM and 256GB DDR5 modules; the CPU index and memory by server model give each platform’s channel count.

A database wants its working set in memory, and not a gigabyte more than its edition can use

Two vendors publish database memory rules precise enough to turn into host sizes. Microsoft’s guidance for SQL Server (Microsoft Learn, updated 10 September 2026) is to set max server memory to “75% of available system memory not consumed by other processes, including other instances”, after subtracting thread-stack memory, because the default of 2,147,483,647MB is effectively unlimited. Its editions page for SQL Server 2025 (updated 20 July 2026) caps the buffer pool at 256GB per instance on Standard and 1,410MB on Express, with Enterprise limited only by the operating system. The Standard cap covers the buffer pool only: the same table gives Standard separate caps of 32GB for the columnstore segment cache and 32GB of memory-optimised data per database. PostgreSQL’s documentation gives “25% of the memory in your system” as a starting value for shared_buffers on a dedicated server, says more than 40% is unlikely to work better because PostgreSQL also relies on the operating system’s cache, and warns that total work_mem use “could be many times the value of work_mem”.

SQL Server 2025 Standard, one instance
  buffer pool cap                          256 GB
  host memory at the 75% rule              256 / 0.75                  = 341 GB
  balanced single-socket host              12 x 32 GB                  = 384 GB
  12 x 64 GB = 768 GB                      adds nothing to the buffer pool past 341 GB

SQL Server Enterprise on a 768 GB host
  max server memory, before thread stacks  768 x 0.75                  = 576 GB

PostgreSQL on a 256 GB dedicated server
  shared_buffers to start                  256 x 0.25                  =  64 GB
  little gain past                         256 x 0.40                  = 102.4 GB
  work_mem at concurrency (illustrative)   200 sessions x 2 sorts x 64 MB = 25.6 GB

The 200 sessions with two sorts each are assumptions chosen to show that concurrency, not the buffer setting, is what runs a database out of memory. Microsoft adds that on a virtual machine “setting a min server memory (MB) value is essential”, so host pressure cannot take buffer-pool memory from the guest, which rules out overcommitting memory beneath a database guest. VMware, for its part, counts databases with moderate memory activity among the workloads that suit its NVMe tiering (section 20), so it is a memory-active database that belongs off it. SAP HANA’s and Oracle’s sizing rules could not be sourced.

The working set is measured, not guessed. On SQL Server the Buffer Manager counters do it: Microsoft Learn (updated 10 September 2026) defines Page life expectancy as “the number of seconds a page will stay in the buffer pool without references” and Page reads/sec as physical reads, so page life falling while reads rise under the real workload means a buffer pool smaller than its working set. On PostgreSQL the pg_buffercache module “provides a means for examining what’s happening in the shared buffer cache in real time”, table by table. A database is sized from its working set and its concurrency, not the size of its data, and its edition may cap what extra memory can do. RAM caching covers buffer pools and the page cache, ECC vs non-ECC why a database host wants ECC, and registered DDR5 and registered DDR4 the modules for new and existing hosts.

AI inference servers carry more host memory than GPU memory, and the cache is what grows

Nvidia’s reference AI server shows the proportions. Its DGX B200 product page pairs 1,440GB of HBM3e on the GPUs with 2TB of system memory, configurable to 4TB, on two Xeon Platinum 8570 processors with 112 cores between them; its GB200 NVL72 rack carries 13.4TB of HBM3E and 17TB of LPDDR5X.

DGX B200
  host memory against GPU memory       2,048 / 1,440     = 1.42 times
  at the 4 TB option                   4,096 / 1,440     = 2.84 times
  host memory per CPU core             2,048 / 112       = 18.3 GB
GB200 NVL72 rack, LPDDR5X against HBM3E  17 / 13.4       = 1.27 times

A published rule sits above Nvidia’s own configuration. Puget Systems’ recommendations for large-language-model servers (modified 16 July 2026) say “NVIDIA (and us) recommends at least 2 x the amount of CPU system memory as there is total GPU VRAM”, so that the buffers the GPUs read from can be pinned in system memory. DGX B200’s standard 2TB, at 1.42 times its GPU memory, sits under that rule; the 4TB option clears it. On a 12-channel socket the two often buy the same host: for 192GB of GPU memory, 2 x 192 = 384GB is 12 x 32GB, and 192 x 1.42 = 273GB rounds up to the same 384GB.

TrendForce said on 29 May 2026 that agentic inference is moving CPU-to-GPU ratios from the traditional 1:8 towards 1:4 or higher, with Nvidia’s NVL72 at 1:2, and that key-value cache capacity scales with context windows. The cache formula of section 8 is linear in the number of sequences served at once, which is what separates a serving host from a desktop:

Llama 3.x 70B, section 8's per-token figures, cache at 16 bits
  one sequence, 4,096 tokens          1.25 GiB
  64 sequences at once                64 x 1.25 GiB     =  80 GiB
  one sequence, 32,768 tokens         10 GiB
  16 sequences at 32,768 tokens       16 x 10 GiB       = 160 GiB
  cache stored at 8 bits              half: 40 GiB and 80 GiB
  the model's weights at four bits    about 43 GB

The Register noted on 1 April 2026 that “it’s quite common for inference engines to store KV caches at FP8”, which is the halving in the second-last line. The shortage is already cutting these configurations: TrendForce reported on 4 August 2026 that cloud providers and server makers reduced RDIMM capacities in the first half of 2026 and that Nvidia halved the SOCAMM memory on its Vera Rubin superchip modules. Why AI servers drive the whole market, and what that has done to their prices, is the subject of the 2026 DRAM shortage.

In inference, memory is sized from concurrency times context as well as from the weights, and at high concurrency the cache can outgrow them. The host side is built from registered DDR5, 256GB DDR5 modules and, where bandwidth matters more than capacity, MRDIMM.

When a server gigabyte costs $30 to $75, measurement is the cheapest capacity there is

What a DDR5 server gigabyte costs in September 2026 depends on who measures it, and the four published measures span more than a factor of two:

The same eRacks tracking put DDR4 ECC RDIMMs at $5.56 per gigabyte, down 25% since March while DDR5 RDIMMs rose 27%, which makes DDR4 a separate price layer for servers that take it. TrendForce said on 25 August 2026 that server DRAM contract prices rose a cumulative 64% in the second half of 2025; its forecasts for 2026 belong to buy RAM now or wait. Buyers have already adjusted: TrendForce reported on 9 July 2026 that cloud providers and server makers had moved some configurations from 96GB and 128GB modules to 32GB and 64GB ones.

VMware’s April 2026 guidance for Advanced Memory Tiering is to “keep active memory at 50% or less of your DRAM capacity”, with one gigabyte of NVMe per gigabyte of DRAM by default on drives of endurance Class D or higher (at least 7,300TBW); its May 2026 “up to 4x more available memory per host” is a vendor maximum, and the default doubles it. CXL pooling appliances, The Register reported on 10 May 2026, let memory be partitioned between servers but not worked on by two at once. Meta, per Tom’s Hardware on 30 June 2026, puts 256GB of reused DDR4-2400 behind its own CXL controller beside 768GB of DDR5-6400 in each server; DDR4 as an escape route covers that reuse.

two-socket host, 24 channels, one module per channel
  24 x 64 GB                                    = 1,536 GB
  24 x 32 GB                                    =   768 GB
  the difference                                =   768 GB
  at eRacks' median, $31.05 per GB              768 x 31.05 = $23,846
  at the 256 GB module's $74.22 per GB          768 x 74.22 = $57,001

measured, illustrative: 350 GB active across VMs allocated 1.2 TB
  VMware's rule, active at most half of DRAM    768 x 0.5 = 384 GB; 350 is under it
  NVMe tier at the default 1:1                  at least 768 GB of Class D NVMe
  memory presented to the VMs                   768 + 768 = 1,536 GB

Without the measurement, the 1,536GB configuration is bought on the allocated figure; with it, the smaller host is safe for everything except a memory-active database, where section 18’s rule holds, and the VM types VMware keeps out of tiering: latency-sensitive, fault-tolerant and encrypted-memory security VMs. Every gigabyte not bought because it was measured idle is worth $30 to $75 in 2026. Existing hosts that take DDR4 can be filled from registered DDR4 at that separate price layer, with the testing in buying used RAM on eBay; new ones from registered DDR5, checked against memory by server model. Drives for a memory tier are on the sister site’s NVMe pages.

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 “in use” overstates need at home and “allocated” overstates it in a data centre (sections 2 and 17). Reach a capacity at full speed: two modules on a desktop, every channel filled on a server (sections 10, 16 and 17). And at 2026 prices, spend on capacity before speed (sections 11 and 12).

What to do with this on a listing page

1. Measure first: a week of peak commit against the commit limit and hard faults on Windows, the memory pressure graph on a Mac, MemAvailable and the “full” pressure line on Linux, or active memory on a hypervisor.
2. Choose the total from the tables above, then the module count and size that reach it: two modules on a desktop, every channel on a server.
3. Open the total page for that capacity, such as 32GB DDR5 kits. It sorts cheapest per gigabyte first and states both the cheapest and the median price per gigabyte; judge a row against the median, not the cheapest row.
4. Read the module split in the title. 2 x 16GB is not 4 x 8GB, and 96GB of desktop DDR5 should be 2 x 48GB; on DDR5 a 96GB title can also be a single server module. “96GB DDR4” is always a kit total or an error, for the reason DDR4 vs DDR5 gives.
5. Start from the default view. Bid-only auction rows are excluded because a bid is not a price; ?bids=show brings them back and ?auctions=no removes auctions entirely.
6. Check the condition on each row, and match a pair by part number on the part pages rather than by title.
7. For a laptop, confirm there are slots at all before comparing SO-DIMM prices.
8. For a server, confirm module type and population against memory by server model and the CPU index.
9. If buying used, follow buying used RAM on eBay for testing and returns and DDR4 as an escape route for provenance. How this site reads titles is on the methodology page, and every generation is ranked together on the all-generations page.

The number is yours to measure; no tier list has seen your machine. This article supplies the floors the platform owners publish, the tests that show how far the recommendations overshoot them, the arithmetic for the workloads that can be calculated, and the configuration that reaches a capacity without paying twice: once for modules you later replace, and again in speed for slots you should have left empty.