Why RAM Prices Exploded in 2026: DDR5 Up 5x, NAS Drives Up 2x

The parts list in my Proxmox homelab build was fine in June. I re-priced every line of it last week against live listings and the memory row had gone from about $300 to $1,880. Same kit, same part number. That one line cost more than the CPU, board, case and power supply put together.

So this is the audit, not a forecast. I pulled current buy-box prices for every component in two of our own build guides, plus our firewall hardware picks, and compared them against the bands those guides shipped with three months ago. RAM prices in 2026 are the headline, but the interesting part is what did not move: every cooler is within a few dollars of its June band, the desktop CPUs in the table below have barely moved, and the cheapest board got cheaper. The shortage is narrow and deep rather than broad. Cases are the one quiet third category that did rise, by 4% to 50% depending on the model, which is not memory-scale but is enough to notice.

Ran the whole parts list past live listings again in September 2026. Same part numbers where those still make sense to buy; a current equivalent where they do not, called out below.

1. What actually changed, part by part

Every row here is a part we had already recommended and priced in June, re-checked in September. The June column is the band our own parts table carried at the time. The September column is a live new-condition price captured the week of 1 September. Treat each one as a point on a spread rather than the price: the same 20TB class runs $720 to $944 and the same 8TB model $355 to $480, all in new condition and all shipping now, so any single figure here is one listing out of several.

Read the bottom half of that table before the top half. Most of the silicon that is not memory did not move. A 16-core CPU costs what it cost, the micro-ATX board got cheaper, and the cooler and the 850W supply are within a few dollars of their June bands. Not everything below the memory rows is flat: the bigger X870E board and the smaller power supplies both went up, and all four cases in these builds rose. But if someone tells you PC parts got expensive in 2026, they mean memory and storage.

2. Memory is the line that broke the builds

DDR5 is roughly $15 a gigabyte now. In June it was closer to $3. That is the whole story of why a homelab build costs what it costs today, and it lands hardest on exactly the builds people come here to make, because virtualization and ZFS both want memory before they want anything else. When you do have to buy at these prices, our DDR5 kit picks for a homelab or workstation track which modules are still worth the money.

ECC is worse. A single 16GB DDR5-4800 ECC UDIMM is $484, so the 64GB an eight-bay NAS wants is about $1,936 of memory in a machine whose board, CPU, case and power supply come to $1,035 to $1,140 together. Registered ECC on the older DDR4 server platforms held up better at around $345 for a 32GB module, which is the first time in years that buying a used server platform has been the cheaper way to get a lot of ECC. If you are choosing between the two, our ECC RAM for a homelab breakdown covers which boards actually validate which type.

The gotcha here is capacity planning. Everyone’s instinct is to buy the RAM last, after the case and the drives, because RAM used to be the flexible line. Invert that this year.

3. Storage went up twice, for two different reasons

Flash and spinning disks both roughly doubled. I can measure that; I cannot tell you why with any authority, so take the rest of this as observation rather than explanation. The two categories moved differently in shape. Spinning disks rose remarkably evenly: every new-condition listing shipping now that I checked, at 4TB, 8TB and 20TB, lands between about 2x and 2.9x, which is about as uniform as a market gets. Flash was all over the place, from 2.1x on a 1TB consumer NVMe to 4.7x on an enterprise M.2 with power-loss protection. If you want one rule of thumb, spinning rust roughly doubled everywhere and flash punished whichever part you specifically needed.

Cost per raw terabyte still favors the big drive: $720 for 20TB is $36 per terabyte against $44 for an 8TB at $355. If you are buying capacity, buy it in large drives, same as always. What changed is that the absolute numbers are now large enough that the topology decision costs real money. Four 8TB drives in a RAIDZ1 give you 24TB usable at $59 to $80 per usable terabyte, depending which of the live 8TB listings you buy. Two 20TB drives in a mirror give you 20TB at $72 to $94. The RAIDZ1 route is the cheaper one at any matched shopping quality, by about 15%, but the ranges overlap enough that shopping badly on it costs you more than the layout saves. Once you have accepted that, pick on the things that are not price: the mirror resilvers as a straight sequential copy with no parity reconstruction and only one surviving drive to depend on, and it scales write IOPS as you add pairs, though a 20TB member takes longer to fill than an 8TB one. The four-drive RAIDZ1 gives you more space, more drives to fail, and the write IOPS of a single disk. RAIDZ expansion also only arrived in a recent OpenZFS release and does not re-stripe what is already written, so a RAIDZ vdev is a longer commitment than a mirror. Check the CMR and SMR difference before you go shopping on price alone though, because the cheapest drive per terabyte at any given capacity is usually the one you do not want in a ZFS pool.

Enterprise flash with power-loss protection took the worst of it. The 960GB M.2 drive our Proxmox guide uses as its datastore went from about $160 to $749. There is a real 960GB SATA equivalent at $399, so the fix is usually to drop from M.2 to SATA rather than to give up PLP, which you should not do on a datastore that takes synchronous writes.

4. Some parts left the shelf entirely

Price is the visible problem. Availability is the one that actually stops a build. Auditing our own guides turned up a long list of picks that are not buyable new any more, and two of them are server boards with no successor at a sane price. The firewall hardware picks had the same problem on a smaller scale.

The Supermicro H12SSL-i, the standard single-socket SP3 board for an EPYC 7003 homelab, has no new stock. Neither does the ASRock Rack ROMED8-2T; the only listings are compatible-part resellers at $1,580 and $2,577 against a real market well under $700. Both are used-market parts now, which is fine if you were already buying a used EPYC to put in them, and a problem if you were not. One Protectli four-port 2.5GbE box we had linked went out of stock, though the lesson there is the opposite of the one I expected: Protectli still sells four-port 2.5GbE appliances, and the FW4C is in stock at $359, so the dead link was a dead listing rather than a dead product. Worth remembering when a search suggests a category has vanished. Even small power supplies thinned out: the smallest Seasonic ATX 3.1 unit on the shelf is 650W, so a 15W idle NAS now carries more headroom than it needs.

5. What it does to a real build budget

Here is the same arithmetic applied to four builds we publish, re-priced line by line. Around a dozen parts had to be replaced along the way. Most are off the shelf entirely. A handful are still listed but at prices that stopped making sense, so the guides moved to a current equivalent, and that is worth being explicit about because it cuts in the reader’s favor here rather than against. Keeping every June part that is still listed at today’s asking price would put the ZFS four-bay at $1,991 to $2,251 instead of $1,840 to $2,100, and the rackmount’s new-parts portion at about $2,761 to $2,951 instead of $2,445 to $2,635. The figures below are the cheaper, current-equivalent route. One specification change was deliberate: the Proxmox mid tier dropped from 96GB of memory to 64GB, which is noted in that guide. The mid-tier June figure below is its parts-table sum rather than the total that guide printed, because the printed total was itself lower than the parts beneath it.

The mid-tier Proxmox build is the one worth staring at. It rose about four fifths, and it has caught up with the used-EPYC “beast” build in the same guide, whose new parts come to $2,445 to $2,635 before the used platform goes in. A ladder whose middle rung has reached its top rung no longer sorts by price. I ended up cutting that tier from 96GB to 64GB to keep it recognizable as the middle option, and that is the honest shape of buying advice this year: the tier you want is one memory step smaller than it was.

Notice also that the entry build rose about three quarters while its CPU, board and NIC all got cheaper or stayed flat. A 32GB kit went from being about 14% of that build to about 39% of it. Memory is not a line item on these builds any more, it is the build.

6. What to do about it right now

Buy memory and drives first, and price them the day you order rather than the day you plan. They are the volatile lines and everything else on the list is stable enough to leave until later in the week.

Buy fewer, larger modules than you think you need slots for. Every consumer board in our builds has four DIMM slots, so a 2×32 kit today leaves room for a 2×32 later, and both ZFS and Proxmox will take more memory whenever you add it. Half-populating deliberately is the cheapest hedge available. The exception is an eight-channel server board, where half-populating costs you memory bandwidth, so there you buy the full set of smaller modules: eight 16GB RDIMMs came out level on price with four 32GB ones when I checked, within ten dollars, and they fill all eight channels.

If the pool is the point, buy the capacity in large drives, then be aware that the four-drive RAIDZ1 route runs about 15% cheaper per usable terabyte than a two-drive mirror at comparable listing quality. Shop the listing harder than you shop the topology. Picking between live listings for the same drive swings about $21 per usable terabyte, while the choice between the two layouts above swings about $13. Which listing you click matters more than which layout you pick, by roughly half again, and that is before you have argued with anyone about parity. Our NAS drive picks list the current CMR models at each capacity, and the ZFS storage server build is the guide those drive rows come from. And if you are building for VMs rather than storage, the how much RAM Proxmox actually needs numbers are worth re-reading before you commit to a kit at these prices, because the difference between a comfortable 64GB and an aspirational 96GB is now about $900.

Used server gear is genuinely competitive again for the first time in a while. Registered DDR4 ECC went up about a quarter to a third as much as DDR5 did, and the SP3 and Rome platforms that host it are cheap because everyone has moved on. If you can live with the noise and the idle power, a used EPYC with 128GB of RDIMMs is now the cheapest route to a lot of ECC memory by a wide margin.

How to check a parts list before you buy

Any build list you find, including ours, is a snapshot. Three things are worth checking yourself before you trust one, because all three caught real problems in our own guides this month.

Check the condition of the buy box, not just whether the page says in stock. Two picks I audited were reporting availability normally while the actual offer was a used unit, including a warehouse motherboard and an enterprise SSD in a slot where write endurance was the entire reason for choosing it. The listing does not shout about it.

Check whether a quoted price is per unit or per set. A row that reads “buy two” against a $740 band is telling you $370 a drive, and a row that reads 128GB against a four-module band is telling you something different again. This trips people up on RAM more than anything else, and I am not above it: two rows of the table above were wrong in the first draft of this article because I read a pair price as a per-module price and made the increase look smaller than it was.

And check the date on the list. Any parts list without a month on it is unusable this year. A June price on memory is not stale by a few percent, it is wrong by a factor of five, and that is a big enough error to turn a build you can afford into one you cannot.