Best SSD and NVMe for Servers and a Homelab

Two consumer SSDs sat in the same server. One acknowledged 642 durable writes a second, the other 2,024, and the interface label predicted neither result. The slower drive was newer. The faster one was older. The spec that sells an SSD, the big sequential-read number on the box, had nothing to do with the gap.

This guide picks the best SSD and NVMe drives for a server or homelab by the numbers that actually move that gap: power-loss protection, write endurance, and sustained fsync rate. Capacity and the interface come last. Five drives are below, one per role, from a cheap boot disk up to an enterprise NVMe drive sized for a Ceph cluster, with the spec that earns each one its slot stated plainly.

Benchmarked on Proxmox VE 9; every price and listing on this page was re-checked in October 2026.

The picks at a glance

Each product name links to its current Amazon listing. Prices were read off those listings on 10 October 2026 and will move, because flash pricing has gone somewhere ugly this year. Four drives that held slots here in June are gone from this list entirely, and the section below the table explains why.

If only one row matters: the Micron 5300 PRO is the drive to put under anything you would be unhappy to lose, and it is SATA. Why a SATA drive wins that title is the rest of this guide.

How to read an SSD spec sheet for a server

A desktop reads big files in long streams, so the sequential-read number on the box describes its life well. A hypervisor does not. It commits small, scattered writes from a dozen guests and calls fsync() on each one to force it to stable storage before moving on. Databases do it per transaction, filesystems do it for journals, the qemu layer does it to keep guest disks consistent. The drive spends its day servicing a flood of tiny synchronous writes, not streams.

So read the spec sheet in this order, top to bottom:

• Power-loss protection (PLP). A bank of capacitors that lets the drive acknowledge a sync write the moment it lands in cache instead of waiting for a physical flush to NAND. It is the single largest lever on fsync rate and the only one that also protects data on a power cut. Enterprise drives have it. Consumer drives do not.
• Endurance, as DWPD or TBW. DWPD is how many times you can overwrite the whole drive per day for the warranty term; TBW is the total write budget. A busy node with ZFS or Ceph write amplification chews through a consumer budget in months.
• Sustained fsync rate. What pveperf reports as FSYNCS/SECOND. A drive that streams 7 GB/s but acknowledges only a few hundred durable writes a second makes guests feel sticky under load.
• Sequential throughput and the interface. Last. SATA tops out near 550 MB/s, NVMe runs from 3,500 to 14,000, and almost none of that reaches a datastore bottlenecked on small sync writes.

The gap from the intro is what that ordering looks like in practice. On the lab nodes a five-year-old SATA SSD on ZFS posted 642 fsync/s while a newer OEM NVMe drive managed 355, and forcing every write durable on a consumer NVMe drive collapsed it from 334,000 IOPS to 2,041, a 160-fold cliff that a PLP drive simply does not fall off. The full bench, drive by drive, is in the NVMe versus SATA for a Proxmox datastore breakdown. The buying lesson is shorter: pay for PLP and endurance, let the interface be whatever your slots accept.

The 2026 price reality: buy endurance, not headline speed

Timing matters this year more than usual. NAND contract prices rose 55 to 60 percent in the first quarter of 2026 and a further 70 to 75 percent in the second, with enterprise SSDs leading the climb, and analysts expect prices to stay elevated through 2027 because meaningful new fab capacity does not arrive until late 2027 at the earliest. The driver is AI: a single NVL72 rack consumes on the order of a petabyte of NAND, and Samsung, SK Hynix, and Micron are reallocating wafers toward high-bandwidth memory and datacenter QLC, which squeezes everything else.

For a homelab budget that changes the math. Spend on the endurance tier the workload needs, buy the capacity you will actually fill rather than aspirational headroom, and treat a drive bought today as one you keep for years rather than flip on the next sale.

It also changes which drives are worth naming at all, and that is the real difference between this list in June and this list now. Enterprise SATA has effectively left the ordinary Amazon storefront: the Kingston DC600M that held the top slot here is now sold only through an Amazon Business listing with no consumer buy box, and its 1.92TB model prices at over a thousand dollars. The WD Red SN700 1TB that was the value pick at about a hundred dollars now has a single overseas reseller in its buy box asking more than five times that. The Crucial MX500, the budget boot pick for years, has no capacity left with a new buy box: the 1TB is a single scalper listing at $400, and the 500GB and 2TB both sit on used stock. The Micron 7450 MAX, the mixed-use drive that answered the write-heavy slot, is out of stock outright. Three of those four slots have a new drive below. The write-heavy one does not, because nothing new under seven hundred dollars replaces a 3 DWPD M.2 drive today, so this list is five drives rather than six and a database node goes to the used market covered later on. The Micron 5400 PRO 960GB that took the top slot in September has since gone out of stock as well, so its predecessor fills that slot now.

The best SSDs and NVMe drives worth buying

One pick per role. Match the drive to the job rather than buying a single type for everything, because the boot disk and a Ceph OSD have nothing in common except the slot they plug into.

Micron 5300 PRO 960GB, the best overall pick

Start here for anything that holds data you would be unhappy to lose, and notice that it is SATA.

The 5300 PRO proves the whole spec-sheet argument. It is a SATA drive, so it loses every sequential-throughput race against any NVMe stick, yet it is the one to put in a node before any consumer NVMe drive because it carries hardware power-loss protection with capacitor monitoring. That single feature lets it acknowledge sync writes from cache, so it holds the high, flat fsync rate consumer drives cannot, and it survives a power cut without leaving a guest unbootable. Micron rates the 960 GB model at 2,628 TBW over a five-year warranty, which works out to 1.5 drive writes a day, better per gigabyte than the 1 DWPD enterprise SATA drives it sits beside. That is the same endurance rating as the newer 5400 PRO it replaces here; the 5300 uses 96-layer TLC instead of 176-layer and Micron no longer builds it, so what you are buying is new old stock. Most older rackmounts and many mini servers have 2.5-inch SATA bays sitting empty, and this is the cheapest honest way to fill one with real datastore storage. About $395 new, from a marketplace seller rather than Amazon itself, which is now true of every new enterprise SSD on this page.

Samsung 990 Pro 1TB, the homelab datastore value pick

The drive most single-node labs should actually run, where PLP is a nice-to-have rather than a requirement.

For a single-node lab that does not justify enterprise pricing, this is the consumer NVMe drive to get, and it is here because the WD Red SN700 that held the slot in June is no longer buyable at a sane price. The 990 Pro ships with onboard DRAM, so it dodges the slowdown and faster wear that plague cheap DRAM-less designs under sustained random writes, and its controller posts the strongest sustained random numbers of any consumer drive on this list. Samsung rates the 1 TB model at 600 TBW on a five-year warranty. It has no power-loss protection, so leave ZFS on sync=standard and keep the host on a UPS. About $240, and Amazon itself holds the buy box, which after the past quarter counts for something.

WD Green SN3000 500GB, the budget boot disk

The OS drive is the one place to spend almost nothing, because Proxmox writes very little after boot.

This is the cheapest new named-brand drive on the page and the only one whose weaknesses do not matter. It is DRAM-less, it carries a three-year warranty rather than five, and its 100 TBW budget is the smallest number in this guide by an order of magnitude. None of that reaches a boot disk, because the hypervisor barely touches the OS drive once it is up, and 100 TBW still covers more than a decade of the twenty or so gigabytes a day a Proxmox host writes to it. Buy it for that job and nothing else: put it under a database and the endurance budget is gone in months. Mirror two of them with ZFS and the host survives a dead boot drive, which matters more here than any spec on the label. About $96, and it is true whether you are building a full rack or a small mini PC for a homelab.

Samsung 9100 Pro 2TB, the prosumer NVMe pick

Reach for this when you want genuine speed for a workstation or a fast lab and accept the lack of PLP.

When the box is a workstation, a build server, or a lab where compile times and clone speed matter more than power-loss safety, this is the drive to buy. It reads at 14,700 MB/s and writes at 13,400 on PCIe Gen5, carries 2 GB of LPDDR4X cache, and Samsung rates the 2 TB model at 1,200 TBW over five years, double the per-terabyte budget of the 990 Pro above. It still has no power-loss protection, so it is not a database or Ceph drive. Treat it as fast scratch and fast VM storage for a single box you keep on a UPS, and be honest that the interface is the reason to buy it, because a datastore bottlenecked on small sync writes will never see the Gen5 numbers. About $400, roughly what the Gen4 990 Pro 2TB costs today, which is the only reason to prefer it.

Solidigm D5-P5430 7.68TB, the capacity pick

For a Ceph cluster or a node that needs a lot of fast, safe space rather than a lot of writes.

When the goal is capacity that is still safe, this is the enterprise NVMe drive that gives you PLP and bulk space. It is a U.2 PCIe Gen4 drive, so the host needs a U.2 backplane or an adapter. The 7.68 TB model carries a large 32,768 TBW sequential-write budget with a more modest 0.58 DWPD random rating, which is normal for the QLC NAND it uses and exactly right for mostly-read VM storage and Ceph capacity tiers. About $2,943, which is four to five times what it cost in June and no longer a casual purchase, though at about $385 a terabyte it is still the cheapest safe capacity on this page and the smaller 3.84 TB model is worse per terabyte, not better. Confirm your host has U.2 connectivity before buying.

Before you click buy on any of them, confirm the form factor the node accepts. M.2 slots, 2.5-inch SATA bays, and U.2 backplanes are not interchangeable, and the right drive in the wrong slot is an expensive paperweight. Check the seller as well as the price: three of the five drives above sit in someone else’s buy box rather than Amazon’s own, and a buy box that looks fine today can flip to a reseller or to used stock inside a week.

Match the drive to the role

The same server often wants three different drives. This is the short mapping from job to pick.

One row has no new drive in it, and that is the honest state of the market rather than an oversight. A busy database VM or a ZFS SLOG wants a mixed-use drive at 3 DWPD or better, and no new M.2 or SATA drive in that class is currently buyable for sensible money, so that row points at the used section below. The rest of the pattern holds: every drive that touches data worth keeping has power-loss protection, and the interface follows the form factor the host happens to offer. That is the same conclusion the ZFS RAID level choice and a three-node Ceph cluster push you toward from the software side, where a missing PLP capacitor turns into the sync-write cliff measured earlier.

The used enterprise SSD play

This section used to be an afterthought. It is now the main event, and for a write-heavy node it is the only sensible route: with new enterprise drives climbing every quarter, a decommissioned datacenter SSD is the cheapest way to get power-loss protection and mixed-use endurance at the same time. Drives like the Intel and Solidigm S4510 and D3-S4610, Samsung PM883, Micron 5300 (the same model as the top pick, bought used), and Kingston DC450 and DC500 sell for a fraction of new pricing and carry the same PLP and endurance class as the picks above. A used 1 DWPD enterprise SATA drive with 90 percent of its life left beats a new consumer drive for a datastore and costs less than one, and a used 3 DWPD mixed-use drive is currently the only affordable answer to a database VM or a SLOG device.

Three rules keep that from going wrong. Read the SMART wear before paying, covered next; a drive listed as “tested good” with 40 percent of its life already consumed is not a bargain. Confirm the exact model carries PLP rather than assuming the whole family does, because vendors split read-intensive and mixed-use lines under near-identical names. And check the connector and sector format: U.2 and the newer U.3 look alike but are not always cross-compatible with a given backplane, and an enterprise drive may arrive formatted 512e or 4Kn, which you can change with nvme format but should expect rather than discover. Buy from a seller who reports SMART data, and treat a listing without it as a gamble.

Three checks before you trust a drive

Specs on a box are a starting point. These three checks confirm what a drive actually is, and they are worth running on any used enterprise SSD before it carries real data.

First, read the wear counter. On an NVMe drive the line to watch is Percentage Used, which counts toward 100 as the endurance budget is consumed:

sudo smartctl -a /dev/nvme0 | grep -iE 'Percentage Used|Data Units Written'

A healthy drive reports a low percentage and a host-writes figure well under its rated budget:

Percentage Used:                    7%
Data Units Written:                 63,001,322 [32.2 TB]

On a SATA SSD the equivalent attributes are Wear_Leveling_Count and Total_LBAs_Written. A used drive with most of its life consumed is worth skipping no matter how cheap it is.

Second, confirm power-loss protection. The datasheet is the authority here: enterprise drives state “power loss protection” or “PLP” outright, consumer drives say nothing because they have none. The behavioral confirmation is the fsync test. Point pveperf at the mounted filesystem:

pveperf /var/lib/vz

A drive with real PLP holds tens of thousands of durable writes a second, a figure enterprise drives post routinely in this test; a consumer drive lands in the hundreds to low thousands and tells you the same thing the capacitor count does:

FSYNCS/SECOND:     2024.31

Third, match the form factor before the order ships, not after. An M.2 enterprise drive will not seat in a U.2 backplane, a U.2 drive needs the right cabling, and a 2.5-inch SATA drive needs a SATA bay and port. Confirm the slot, then pick the drive that fits it from the table above. With the host assembled and the drive verified, the Proxmox VE 9 install guide is where the storage actually goes to work.