The wrong NVMe drive in a Proxmox host does not announce itself on day one. It benchmarks fast, and then the wear counter climbs faster than anyone expected. The cluster filesystem, the metrics database, the journal and every VM fsync land on the same SSD, and ZFS writes most sync writes twice. Sequential speed, the number every spec sheet leads with, says almost nothing about how long the best NVMe SSD for Proxmox will survive that load.
Two specs decide it instead: write endurance in terabytes written, and capacitor-backed power-loss protection. The NAND shortage that our SSD price tracker follows has made the third number, price per terabyte, move week to week. The cheapest TLC drive with a DRAM cache in this guide came to about $195 per TB (Samsung 990 PRO 2TB, Amazon price checked November 2026). The enterprise M.2 with power-loss protection cost four times that. We cover boot drives, VM pools, ZFS log devices and the mini PCs most of these drives end up in.
Best SSD for Proxmox: the five drives compared
For most single-node Proxmox hosts, the answer is a 2TB TLC NVMe drive with a DRAM cache, and the Samsung 990 PRO is the safe default. None of the consumer drives has power-loss protection. A cluster boot mirror or a ZFS SLOG wants an enterprise M.2 such as the Kingston DC2000B or the Micron 7450 PRO. Every drive in the table is TLC and M.2 2280. Check height before a mini PC, though. Kingston lists the DC2000B 960GB only as a heatsink part, 8.3mm tall. The 4TB SN850X is 3.88mm, and the 7450 PRO needs its own clearance check. Kingston and Micron list no DRAM figure, shown as n/s.
| Drive | Capacity | DRAM | PLP | Endurance | Price per TB (Nov 2026) | Best for |
|---|---|---|---|---|---|---|
| Samsung 990 PRO | 2 TB | Yes | No | 1,200 TBW | About $195 | Single-node VM host |
| Crucial T500 | 2 TB | Yes | No | 1,200 TBW | About $250 | VM host, when it is cheaper |
| Kingston DC2000B | 960 GB | n/s | Yes | 700 TBW (0.4 DWPD) | Business price only | Boot drive with PLP |
| Micron 7450 PRO | 960 GB | n/s | Yes | 1,700 TBW (1 DWPD) | About $780 | Write-heavy pool, SLOG |
| WD Black SN850X | 4 TB | Yes | No | 2,400 TBW | About $196 | High capacity |
Read the table by column. The two cheapest drives per terabyte, the 990 PRO and the SN850X, are both consumer drives within a dollar of each other, so capacity decides between them. Only the two enterprise drives have a Yes under PLP. That column, not the price, decides whether a drive can hold a SLOG or a cluster boot mirror.
How we picked: endurance, PLP, NAND and size figures come from the maker’s datasheet or spec page. Controller and DRAM details for the T500 and SN850X come from published reviews. The Proxmox advice comes from the Proxmox VE admin guide. Prices are Amazon listings we checked on 12 November 2026. For servers that do not run Proxmox, our SSD and NVMe guide for servers and a homelab is the wider roundup.
1. Samsung 990 PRO 2TB: the safe default for a VM host
For a single-node Proxmox host that runs a handful of VMs and containers, this is the drive to buy and stop thinking about.
The 990 PRO pairs Samsung V-NAND TLC with a 2GB LPDDR4 DRAM cache and Samsung’s own controller on a PCIe 4.0 x4 link. Samsung rates the 2TB at 1,200 terabytes written with a five-year warranty. For Proxmox, the DRAM cache matters most. It holds the flash translation table on the drive itself, so the small random writes of a VM host never borrow host memory to find their place on the flash. It is also 2.3mm thick at most, so it fits under the low lid of a typical mini PC. Expect roughly $370 to $400 at 2TB, which made it the cheapest drive with a DRAM cache per terabyte in this guide in November; check the live price.
Skip it if the host will run a busy database or a write-heavy log pipeline. It has no power-loss protection, so an ungraceful shutdown can lose the last in-flight writes. And 1,200 TBW is generous for a home host but modest next to an enterprise drive. For a boot-plus-VMs box that idles most of the day it is more than enough. For a drive that is hammered around the clock it is the wrong tool, and picks three and four exist for that.
2. Crucial T500 2TB: the same endurance, when the price is right
The T500 matches the Samsung on the two specs that matter, 1,200 TBW and a full DRAM cache, so it only wins on the weeks it costs less.
Crucial pairs Micron’s own TLC with an LPDDR4 DRAM cache on a Phison E25 controller, rated at 1,200 TBW at 2TB with a five-year warranty. On paper it matches the 990 PRO for a VM host: the same endurance rating, a DRAM cache, and a sequential rating no Proxmox host will saturate. In early September the two listings were within $5 of each other. By November the T500 listing was close to $500, about $250 per terabyte against $195 for the Samsung. Expect $480 to $520 for now; check the live price.
Its trade-off matches the Samsung’s: no power-loss protection, consumer endurance. No performance gap justifies paying 28% more per terabyte for it. Buy it only when the price falls back under the 990 PRO, which is the kind of swing a Black Friday sale can produce. On any other day, the Samsung is the better buy for the same job.
3. Kingston DC2000B 960GB: power-loss protection in an M.2 boot drive
Most homelab builders do not know this drive exists: an enterprise boot SSD with real power-loss protection on an M.2 2280 card that ships with a heatsink at 480GB and 960GB.

The DC2000B is Kingston’s data-center boot drive: 3D TLC behind a PCIe 4.0 x4 controller, with onboard capacitors that give it full hardware power-loss protection. Kingston rates it at 0.4 drive writes per day, which works out to 700 terabytes written at 960GB over a five-year warranty. Sequential write tops out around 1,300 MB/s against a 7,000 MB/s read. Those numbers look weak next to a consumer drive, and that is the point. Its job is the operating system and the cluster filesystem, where the value is the capacitors, not the write ceiling. With a pair of them, PLP keeps acknowledged pmxcfs writes intact through a power cut.
Expect an enterprise premium and a business sales channel. Amazon shows its price only to business accounts, and reseller listings we checked in September ran $700 to $850 at 960GB; check the live price, and consider the 480GB model, which is plenty for a boot device. Kingston lists both only as heatsink parts, 8.3mm tall. Confirm the clearance. Skip it as a VM store. At 0.4 DWPD it is not built to absorb heavy guest writes, and it has the lowest rated sequential write (1,300 MB/s) of any drive here. For heavy NFS or iSCSI sync traffic, the 1 DWPD 7450 PRO is the safer log device.
4. Micron 7450 PRO 960GB: enterprise endurance for a write-heavy pool
When a Proxmox node actually earns its endurance rating, a busy database or a log-heavy VM stack, this is the M.2 built for it.
The 7450 PRO is a full enterprise drive in M.2 2280 form: Micron 176-layer 3D TLC, an enterprise controller, capacitor-backed power-loss protection, and a 1 drive-per-day rating that the datasheet puts at 1,700 terabytes written at 960GB. That is about two and a half times the endurance of the DC2000B and enough to run a write-heavy VM pool for years without coming close to the wear budget. Its capacitors are what make it safe for synchronous database writes. Plan on roughly $730 to $780 at 960GB; check the live price on the MTFDKBA960TFR listing, because the older listing for the same drive was out of stock in November.
Price is the whole objection. At about four times the cost per terabyte of a 990 PRO, the 7450 PRO makes sense only once the workload threatens a 1,200 TBW consumer drive inside its warranty. Most home nodes never get there. Run the wear-rate check at the end of this guide before buying it. If a consumer drive would take fifty years to wear out on your workload, this one is money spent on endurance you will never use. On a node that chews through a consumer drive in two years, it pays for itself.
5. WD Black SN850X 4TB: capacity without dropping to QLC
Four terabytes is where consumer drives are tempted into QLC to hit a price. The SN850X stays on TLC, which is why it is the capacity pick here.
The SN850X uses TLC NAND with a DRAM cache on WD’s own controller, PCIe 4.0 x4, rated at 2,400 terabytes written at 4TB with a five-year warranty. At four terabytes it gives a single-node host room for a real VM library, ISOs and local backups on one drive without stepping down to QLC. Endurance scales with capacity, so the 4TB carries double the TBW of a 2TB drive, which makes it the most forgiving consumer option for a busy single-drive host. In November it cost about $196 per terabyte, level with the 990 PRO 2TB. Expect around $760 to $800, with stock running thin; check the live price.
Like picks one and two, it has the consumer limitation: no power-loss protection, so it is a VM store and a capacity drive, not a place for synchronous writes you cannot afford to lose. Its 4TB board is 3.88mm tall, so it wants airflow or a thermal pad in a sealed mini PC. Skip it if the host only needs a terabyte or two, or if the workload needs PLP, where the enterprise drives are the answer regardless of capacity.
Price per TB during the NAND shortage
Price per terabyte is the fairest way to compare these drives while flash prices keep moving, because it cancels out capacity. The table adds two 4TB alternatives to the five picks. All prices are the Amazon listing price we pulled on 12 November 2026, divided by the capacity on the label.
| Drive | Capacity | DRAM | Endurance | Amazon price, 12 Nov 2026 | Per TB |
|---|---|---|---|---|---|
| WD Black SN7100 | 4 TB | No (DRAM-less) | 2,400 TBW | About $610 | About $153 |
| Samsung 990 PRO | 2 TB | Yes | 1,200 TBW | About $390 | About $195 |
| WD Black SN850X | 4 TB | Yes | 2,400 TBW | About $784 | About $196 |
| Samsung 990 PRO | 4 TB | Yes | 2,400 TBW | About $800 | About $200 |
| Crucial T500 | 2 TB | Yes | 1,200 TBW | About $499 | About $250 |
| Micron 7450 PRO | 960 GB | n/s | 1,700 TBW | About $749 | About $780 |
| Kingston DC2000B | 960 GB | n/s | 700 TBW | Business accounts only | n/a |
The cheapest row is the one with a catch. The WD Black SN7100 4TB (about $610 in November) is TLC with the same 2,400 TBW as the SN850X, but the SanDisk datasheet calls it a DRAM-less drive. We would use it as a boot drive or backup target, and only as a compromise for the main VM store. The Samsung 990 PRO 4TB (about $800) carries a 4GB DRAM cache and costs about $200 per terabyte, close to its 2TB sibling. So the choice between two 2TB drives in a ZFS mirror and one 4TB drive comes down to how many M.2 slots the host has.
For Black Friday, ignore the list price. Amazon showed the 990 PRO 2TB at $639.99 list against a $389.99 sale price, a 39% discount on a number nobody pays. What counts is whether the per-TB figure drops under the $195 floor for a DRAM-equipped TLC drive. Our Black Friday homelab deals roundup tracks the drives that actually fall during the sale.
Why Proxmox writes so much to disk
A Proxmox host is a constant low-level write generator even when the VMs are idle. The cluster filesystem, pmxcfs, fsyncs its SQLite config database to disk on every change. The high-availability stack persists its LRM and CRM state, the built-in RRD metrics database updates every few seconds, and the system journal logs the lot. None of that is large, but it is relentless small random writes, which is the workload that ages an SSD fastest. A drive that looks bored on a throughput graph can still be burning through its wear budget.
ZFS makes it worse in the way that matters. Most synchronous writes are committed twice, once to the ZFS intent log and again to the pool. Copy-on-write and metadata updates add their own amplification on top. Without a dedicated log device that traffic lands on the same VM drive. So a consumer SSD that would last a decade in a desktop can wear out in a year or two under a busy ZFS Proxmox node. That is also why giving ZFS enough RAM for its cache is part of protecting the disk, not a separate concern.
Power-loss protection, and when you actually need it
Power-loss protection is a bank of capacitors on the drive that holds enough charge to flush its write cache to flash if power drops mid-write. The enterprise drives here have it; the consumer drives in this guide do not, no matter how fast they are. A drive with PLP can safely acknowledge a synchronous write the instant it reaches the cache. Enterprise drives are designed to keep their sync-write speed under load for that reason, while consumer drives typically fall off. PLP also protects the data already in flight when a node loses power. For the cluster filesystem or a database, that is the difference between a clean reboot and a corrupt config.
Proxmox itself is blunt about it. The Proxmox VE admin guide says: “SSDs with Power-Loss-Protection (PLP) are recommended for good performance. Using consumer SSDs is discouraged.” For a home node, that advice has to be weighed against a fourfold price gap per terabyte. A single-node host on a UPS, running VMs that are backed up, survives fine on a consumer drive. PLP earns its premium on a cluster that cannot afford a corrupt pmxcfs and on a ZFS SLOG that must honour every sync. Any node running a database, where a lost write is a lost transaction, belongs on the list too.
Does the Samsung 990 PRO have power-loss protection?
No. The 990 PRO datasheet lists no power-loss protection, and neither do the Crucial T500, WD Black SN850X or SN7100 sheets. The DRAM cache on these drives speeds up writes but is volatile, so any write not yet flushed to flash when power drops is lost. That is fine for a VM store on a UPS with backups; it is the wrong property for a SLOG or a cluster boot drive.
Consumer vs enterprise NVMe for Proxmox
Speed is not what splits the two classes. On paper the consumer drives are faster; under a sustained synchronous load the enterprise drives are the ones that hold their numbers. The price row comes from the November 2026 Amazon listings above.
| Spec | Consumer (990 PRO, T500, SN850X) | Enterprise M.2 (DC2000B, 7450 PRO) |
|---|---|---|
| Power-loss protection | None | Onboard capacitors |
| Endurance rating | 1,200 TBW at 2TB, 2,400 TBW at 4TB | 0.4 to 1 DWPD for five years |
| Sync writes (fsync, ZIL) | Slow once the cache must be flushed | Acknowledged from protected cache |
| Headline sequential write | 6,600 to 7,000 MB/s | 1,300 to 1,400 MB/s |
| Price per TB (Nov 2026) | About $195 to $250 | About $780 (7450 PRO) |
| Where it belongs | VM store on a single node | Cluster boot, SLOG, databases |
DWPD and TBW are roughly the same budget stated two ways. Multiply the drive-writes-per-day figure by the capacity in terabytes, by 365, and by the warranty years, and you get terabytes written. For the 960GB DC2000B, 0.4 DWPD over five years comes to about 700 TBW, exactly what Kingston rates it at. Use that conversion to compare an enterprise sheet against a consumer one on equal terms.
DRAM, HMB, and why QLC struggles in a VM store
Two spec-sheet details separate a drive that holds up as a VM store from one that does not. The first is the DRAM cache. A drive with onboard DRAM keeps its flash translation table on the drive; a DRAM-less drive borrows a slice of system RAM through the host memory buffer. That works well for a desktop and is weaker under the sustained small random writes a hypervisor produces. Every consumer pick here has a real DRAM cache for that reason. Kingston and Micron publish no DRAM figure. Treat DRAM-less drives such as the SN7100 as a boot or backup compromise, not the drive a busy VM store runs on.
NAND type is the second detail. TLC stores three bits per cell; QLC stores four, which raises capacity and lowers cost but cuts both write endurance and sustained write speed hard. A QLC drive writes fast into its SLC cache and then falls off once that cache fills, which under a VM workload happens routinely. Every drive in this guide is TLC on purpose. QLC has a place for cold bulk storage and backup targets, not for the drive the VMs run on. If a SATA drive is still on the table, our NVMe vs SATA datastore comparison settles the interface question.
Boot, VM pool, SLOG or special vdev: one drive per job
Most Proxmox builds give an SSD one of four jobs, and each job weights the specs differently. The table maps the picks onto them.
| Role | What it holds | Spec that decides | Pick |
|---|---|---|---|
| Boot mirror | Proxmox OS, pmxcfs, logs | PLP on a cluster, small capacity | DC2000B pair; 990 PRO pair on a single node |
| VM pool | Guest disks, containers | TLC, DRAM, TBW per dollar | 990 PRO 2TB, SN850X 4TB |
| ZFS SLOG | Synchronous writes only | PLP and low commit latency | 7450 PRO, DC2000B for light loads |
| Special vdev | Metadata, small blocks | Redundancy equal to the pool | Mirrored PLP drives |
Pick RAID1 in the Proxmox installer with two disks of the same size and it builds a ZFS mirror for the root filesystem. Then proxmox-boot-tool keeps the EFI partitions on both disks in sync. The mirror protects against a dead drive, not against both drives wearing out together, which is a real risk when two identical consumer drives absorb identical writes for years. On a single home node, two 1TB or 2TB consumer TLC drives are acceptable if you watch the wear counter on both.
A SLOG only ever holds synchronous writes, so it helps NFS, iSCSI and databases that fsync, and does nothing for async guest writes. The Proxmox guide asks for “fast SSDs with power-loss protection” and notes that “using more than half of your installed memory won’t provide you with any real advantage”. A 960GB enterprise drive is therefore mostly empty space as a SLOG, and the guide shows splitting such a drive into a small log partition and a cache partition. A consumer drive makes a poor SLOG: without capacitors it flushes its cache on every sync write and can end up barely faster than the pool behind it.
A special vdev can take the whole pool down, so read the admin guide’s rules before you buy for it:
- Redundancy: “The redundancy of the special device should match the one of the pool, since the special device is a point of failure for the whole pool.” A RAIDZ2 pool wants at least a three-way special mirror.
- Permanence: “Adding a special device to a pool cannot be undone!” Size it for the pool’s life, not for today’s data.
- special_small_blocks: at or above the dataset recordsize (128K by default), every block lands on the special device, which fills a small SSD fast.
For those two ZFS roles, the Optane and PLP drives in our SLOG and special vdev SSD guide go further than the M.2 picks here. The ZFS RAID levels guide covers where each device sits in the pool layout. Bulk cold storage belongs on none of these drives: that is the job of NAS hard drives priced per terabyte behind an HBA, at a fraction of the NVMe cost.
Fitting NVMe into a mini PC
Most of these drives go into a homelab mini PC running Proxmox, and the chassis narrows the choice before the spec sheet does. Check four things on the mini PC’s own spec page before buying a drive:
- Slot count: a ZFS boot mirror needs two M.2 slots. A one-slot box gets a single drive and good backups, so buy the larger 4TB TLC drive instead of two 2TB. The 4TB SN850X is 3.88mm tall. The 990 PRO 4TB stays at 2.3mm.
- Slot length: every pick here is 2280. A 2230 or 2242 slot rules all of them out, so check the datasheet of any short drive for DRAM and for TLC before trusting it with VMs.
- Lane width: a second slot wired at PCIe x1 or x2 caps sequential speed but not endurance or PLP, so it is fine for a boot drive and a poor home for the VM pool.
- Height and heat: the 990 PRO is 2.3mm and the SN7100 2.38mm at most, which clears a low lid; the 4TB SN850X is 3.88mm and the DC2000B 480GB and 960GB are heatsink parts at 8.3mm, which may not fit, and a 4TB drive under a sealed lid benefits from a thermal pad against the case.
Enterprise M.2 drives add one more problem, because many of them are 22110, not 2280. Micron sells the 7450 PRO in M.2 2280 only at 480GB and 960GB; the 1.92TB and 3.84TB models are 110mm long, and a mini PC has nowhere to put them. That is why the enterprise picks here top out at 960GB.
M.2 22110 and U.2 when the board allows it
A tower or a server board with a 22110 slot opens up the larger 7450 PRO models, but new stock is priced for data centers. We found the 1.92TB 22110 listing on Amazon at about $2,700 in November, roughly $1,400 per terabyte. That is close to double the 960GB model per terabyte and seven times the 990 PRO. At that price the 22110 slot is worth using only for a drive that is already on hand.
U.2 is the cheaper route to enterprise endurance with PLP. Used 2.5-inch data-center drives sell well below their new price, and they connect through a PCIe adapter card with airflow over them, which most mini PCs cannot offer. Our used U.2 NVMe homelab guide covers the drives, adapters and cooling, and the wear figures to check on a used drive before trusting it with a pool.
How long these drives actually last under Proxmox
The endurance rating and the real lifespan are not the same number, because write amplification sits between them. Take a worked example. A quiet single-node host that writes 20 gigabytes a day to its system drive uses 7.3 TB a year, so a 1,200 TBW rating covers about 164 years on paper. If ZFS sync writes and VM churn multiplied that load tenfold, a figure we pick only to show the math, the same drive would still last about 16 years. A node that writes 200 GB a day before any amplification lands in that same range, which is why the measured rate matters more than the rating. The rating tells you the ceiling; only measuring your own node tells you the rate.
Two counters in every NVMe drive’s SMART health log carry the real figures. Percentage Used is the drive’s own estimate of consumed endurance; Kingston’s DC2000B datasheet notes that a drive at its warranty limit reports a value greater than or equal to 100, so it can climb past 100. Data Units Written counts host writes in units of 1,000 sectors of 512 bytes, as the NVMe specification and smartmontools define it, so multiply by 512,000 to get total bytes written. That host figure is what a TBW rating counts too, so the two compare directly.
A single reading says little. The write rate is what turns the rating into years, and it takes two readings to get it:
- Baseline: note Data Units Written and Percentage Used today.
- One week later: read both again after a week of normal load, with backups and scrubs included.
- Projection: divide the TBW rating minus the terabytes already written by the weekly terabytes, then by 52, to get years left.
If that number is comfortably longer than the warranty on a consumer drive, the enterprise drives are endurance you will pay for and never use. A drive on course to wear out inside a couple of years is the signal to move the write-heavy guests onto a 7450 PRO and let the wear counter tell the truth.



