Best Used Server for a Homelab (2026 Guide)
By kciservers.net Technical Team · 2026-07-24
You want a real server in your homelab. Not a desktop pretending to be one, not a mini PC that thermal-throttles under a sustained backup job, and not a consumer NAS that locks you into a vendor's app store. You want ECC memory, hot-swappable drives, a proper out-of-band management controller, and a platform that was designed to run 24/7 in a data center.
The good news is that three- and four-year-old enterprise gear hits the refurbished market at a fraction of its original price, and it still has years of useful life left. The catch is that enterprise servers were not designed for your living room. They are loud, they draw more power than a desktop, and the storage controller that comes installed may not be the one you want for your workload.
This guide walks through what to buy, what to avoid, and how to match a platform to what you actually plan to run.
What matters in a homelab server
Before picking a model, the four constraints that determine whether a server works in a home environment:
Form factor and noise. A 1U rack server has small, high-RPM fans that move a lot of air and make a lot of noise doing it. A 2U server has larger fans that run slower and are noticeably quieter at idle. A tower server can sit on a desk or floor without a rack and is the quietest option. If the server lives in a basement or garage, any form factor works. If it lives in your living space, a tower or 2U is the practical choice.
Power draw. Enterprise servers are not subtle about electricity. A dual-socket server with a full load of memory and drives will draw more at idle than a consumer NAS, and significantly more under load. The platforms in this guide have minimum PSU ratings from 495 W to 750 W and maximum ratings up to 2800 W. You will not run them at maximum, but you should budget for higher idle power than any consumer device.
Memory capacity. This is where enterprise gear wins decisively. DDR4 ECC RDIMMs are inexpensive on the secondary market right now, and the platforms below have 24 or 32 DIMM slots. Filling those slots with 16 GB or 32 GB modules gives you hundreds of gigabytes of ECC memory at a price that is impossible to match with consumer hardware.
Storage flexibility. The storage controller and backplane determine what you can do. A server with a hardware RAID controller and SAS backplane is fine for Proxmox local storage. A server with an HBA in passthrough mode is what you want for TrueNAS and ZFS. Some backplanes support NVMe, some do not. This is the single most important thing to check before buying, and we cover it in detail below.
Three homelab personas
Most homelab builds fall into one of three profiles. Knowing which one you are narrows the field quickly.
```mermaid flowchart TD A["Start: What will you run?"] --> B{"ZFS storage\n(TrueNAS / Unraid)"} A --> C{"VMs and containers\n(Proxmox / ESXi)"} A --> D{"GPU compute\n(LLM inference / passthrough)"} B --> B1["Need: HBA controller\n+ drive bays + RAM for ARC"] B1 --> B2["Pick: DL380 Gen10 or R740\nwith HBA or HBA-mode controller"] C --> C1["Need: CPU cores + RAM\n+ hardware RAID with cache"] C1 --> C2["Pick: R740 with PERC H730P\nor DL380 Gen10 with P408i-a"] D --> D1["Need: GPU slots + PSU wattage\n+ PCIe bandwidth"] D1 --> D2["Pick: R750 for dual 300W GPUs\nor R740 for single-card setups"] ```
The storage labber. You want to run TrueNAS or Unraid, pool a lot of drives, and serve files, backups, or media to the rest of your network. Your priorities are drive bay count, an HBA controller for ZFS passthrough, and enough memory for ARC caching. CPU core count is secondary; you need enough to run the storage OS and a few lightweight containers, not to compile code.
The virtualization labber. You want to run Proxmox or ESXi with a cluster of VMs and containers: a Docker swarm, a home automation stack, a CI runner, maybe a Windows VM for testing. Your priorities are CPU cores, memory capacity, and a hardware RAID controller with cache for local VM storage. You may also want a second server for clustering.
The AI/ML tinkerer. You want to run local LLM inference, GPU passthrough to a VM, or experiment with model fine-tuning. Your priorities are GPU slot capacity, PSU wattage to feed the GPUs, and enough PCIe bandwidth to support them. Storage and CPU are secondary to having enough power and physical space for accelerator cards.
The picks
The all-rounder: Dell PowerEdge R740 (14G)
The R740 is the sweet spot for most homelabs. It is a 2U dual-socket server with 24 DIMM slots, up to 3072 GB of DDR4 memory at 2933 MHz, and six memory channels per CPU with two DIMMs per channel. The CPU platform is Intel Xeon Scalable on the LGA3647 socket. You can drop in anything from a low-wattage Silver to a 28-core Platinum.
A common homelab configuration uses two Xeon Gold 6138 processors (20 cores per CPU, 40 cores total, 125 W TDP per chip) with 128 to 256 GB of DDR4 ECC and a PERC H730 or H730P RAID controller. The H730P has 2 GB of flash-backed cache and supports RAID 0, 1, 5, 6, 10, 50, and 60. The R740's chassis options include 8x 2.5-inch, 8x 3.5-inch, and 16x 2.5-inch SAS/SATA front bays. The common refurb configs use SAS/SATA backplanes. If you need NVMe front bays, look at the R750 or R640 instead.
The R740 supports up to six single-width GPUs or three double-width GPUs at up to 300 W each, which is enough for a single inference card or a couple of low-power accelerators. PSU options range from 495 W to 2400 W, hot-swappable and redundant. For a homelab without GPUs, a pair of 495 W or 750 W units is sufficient. If you add a 300 W GPU, step up to 1100 W or higher.
Dell's 14G management is handled by iDRAC9. The Express license gives you monitoring (temperature, power, event log) but does not unlock remote console or virtual media. For those, you need iDRAC Enterprise. Many refurb units ship with Express; check before buying if remote management matters to you.
The HPE alternative: ProLiant DL380 Gen10
The DL380 Gen10 is HPE's direct competitor to the R740, and it is worth checking both before committing. It is a 2U dual-socket server on the same LGA3647 socket, with 24 DIMM slots and up to 3072 GB of DDR4 at 2933 MHz. Six memory channels per CPU, two DIMMs per channel, the same memory architecture as the R740.
The DL380 supports Intel Xeon Scalable processors up to 28 cores per CPU and 205 W TDP. A typical homelab config uses two Xeon Silver 4216 processors (16 cores per CPU, 32 cores total, 100 W TDP per chip) with 128 GB of DDR4 ECC. The Silver 4216 is a lower-wattage part (100 W versus the Gold 6138's 125 W), which means slightly lower power draw and heat, at the cost of fewer cores and lower base clock.
Where the DL380 differs from the R740 is storage and management. The DL380's SFF chassis can take NVMe by fitting a dedicated 8 NVMe cage, or a Premium 6 SAS/SATA + 2 NVMe cage, in place of a drive box — an option the R740's common configs do not offer at all. The 24 SFF chassis goes further, with up to 24 front bays plus 6 rear bays. HPE's Smart Array controllers (P408i-a, P408i-p) handle hardware RAID with 2 GB of cache, and they can also be set to HBA mode for ZFS passthrough, a useful flexibility that Dell's PERC controllers do not offer natively.
HPE's out-of-band management is iLO 5. The Standard license gives basic monitoring; iLO Advanced is required for remote console and virtual media. As with Dell, check the license level on refurb units.
GPU support on the DL380 matches the R740: up to six single-width or three double-width GPUs, with a 300 W maximum per double-width slot. PSU options range from 500 W to 1600 W, hot-swappable and redundant.
The quiet tower: HPE ProLiant ML350 Gen10
Not every homelab has a rack. The ML350 Gen10 is a 4U tower that can also be rack-mounted, and it is the quietest option in this guide. It uses the same LGA3647 socket and Intel Xeon Scalable processors (first and second generation) as the DL380, with 24 DIMM slots and up to 3072 GB of DDR4 at 2933 MHz.
The ML350's max CPU TDP is 165 W (lower than the DL380's 205 W), which limits you to mid-tier processors but also keeps heat and fan noise down. A Xeon Silver 4216 (16 cores, 100 W) is a comfortable fit. The chassis options are simpler: 8x 2.5-inch SFF with SAS/SATA/NVMe support, or 4x 3.5-inch LFF for larger drives. The SFF config includes a Smart Array P408i-a controller by default.
GPU support is more limited than the rack servers: up to three single-width or two double-width GPUs, with a 300 W maximum per double-width slot. PSU options are 500 W, 800 W, or 1600 W. For a storage-focused homelab in a shared living space, the ML350 is the platform that trades rack density for livability.
The compact 1U: Dell PowerEdge R640 (14G)
If you have a wall-mount rack or a shallow rack and 2U is too tall, the R640 is the 1U option. It shares the R740's platform: second-generation Intel Xeon Scalable, 24 DIMM slots, DDR4 at 2933 MHz. The maximum memory capacity is 3072 GB, the same as the R740 — the 1U chassis costs you expansion, not memory.
The R640's storage options are where it gets interesting for a compact homelab. The 10x 2.5-inch chassis supports SAS, SATA, and NVMe in the same bays, giving you NVMe boot drives alongside SAS storage drives. The 8x 2.5-inch and 4x 3.5-inch configs are SAS/SATA only. The R640 also supports up to two rear 2.5-inch bays for boot or cache drives.
GPU support is the main limitation: the 1U chassis only fits single-width, low-power GPUs at 75 W maximum, up to three of them. No double-width cards. If you need GPU compute, look at the R740 or R750 instead. PSU options match the R740: 495 W to 2400 W, hot-swappable and redundant.
The high-end: Dell PowerEdge R750 (15G)
The R750 is the 15G jump. It moves to the LGA4189 socket and third-generation Intel Xeon Scalable processors, which brings eight memory channels per CPU, up from six on 14G. The DIMM slot count increases to 32, and maximum memory capacity reaches 8192 GB of DDR4 at 3200 MHz.
A common high-end homelab config uses two Xeon Gold 6338 processors (32 cores per CPU, 64 cores total, 205 W TDP per chip) with 256 GB or more of DDR4 ECC. The R750's chassis options are extensive: 8x 2.5-inch NVMe, 8x 2.5-inch Universal (SAS/SATA/NVMe tri-mode), 12x 3.5-inch, 16x 2.5-inch, and 24x 2.5-inch configurations. If you want front-bay NVMe, the R750 is the platform in this guide that supports it most broadly.
GPU support is strong: up to eight single-width or two double-width GPUs, with 300 W per double-width slot. The R750 supports NVIDIA A100 and A40, which makes it a viable platform for serious AI inference work. PSU options range from 700 W to 2800 W, hot-swappable and redundant. If you are running dual 300 W GPUs, plan for at least 1800 W PSUs.
One correction worth noting if you have read older guides: the R750 uses DDR4, not DDR5. The DDR5 platform is the R760 (16G), which uses DDR5 at up to 5600 MHz and supports fourth and fifth-generation Xeon Scalable processors with up to 64 cores per CPU. The R760 is appearing on the refurb market, but pricing is still high for a homelab. The R750 is the practical high-end pick for 2026.
The Lenovo option: ThinkSystem SR650
Lenovo's SR650 is the 2U alternative if you want to look beyond Dell and HPE. It uses the same LGA3647 socket and second-generation Intel Xeon Scalable processors as the R740 and DL380, with 24 DIMM slots and up to 3072 GB of DDR4 at 2933 MHz. Max CPU TDP is 205 W, matching the DL380.
The SR650's chassis options are straightforward: 24x 2.5-inch with SAS/SATA/NVMe support, or 14x 3.5-inch SAS/SATA. The 24-bay config is attractive for storage-heavy homelabs: it gives you 24 front bays in a 2U chassis, as many as any server in this guide. Storage controllers are Lenovo's RAID 930-16i or RAID 530-16i.
GPU support is more limited than the Dell or HPE 2U options: up to three GPUs, with the catalog listing NVIDIA T4 and V100 as supported models. PSU options are 750 W, 1100 W, or 1600 W, hot-swappable and redundant. Lenovo's management controller is the XClarity Controller (XCC), with a similar licensing model to iDRAC and iLO.
Candidate comparison
| Server | Form factor | Socket | DIMM slots | Max RAM | RAM speed | Max GPU (double-width) | PSU range | NVMe front bays | |:---|:---|:---|:---|:---|:---|:---|:---|:---| | Dell PowerEdge R740 | 2U | LGA3647 | 24 | 3072 GB | 2933 MHz | 3 × 300 W | 495–2400 W | No (SAS/SATA configs) | | HPE ProLiant DL380 Gen10 | 2U | LGA3647 | 24 | 3072 GB | 2933 MHz | 3 × 300 W | 500–1600 W | Yes (SFF NVMe cage) | | HPE ProLiant ML350 Gen10 | 4U tower | LGA3647 | 24 | 3072 GB | 2933 MHz | 2 × 300 W | 500–1600 W | Yes (8 SFF config) | | Dell PowerEdge R640 | 1U | LGA3647 | 24 | 3072 GB | 2933 MHz | 0 (single-width only) | 495–2400 W | Yes (10x 2.5 config) | | Dell PowerEdge R750 | 2U | LGA4189 | 32 | 8192 GB | 3200 MHz | 2 × 300 W | 700–2800 W | Yes (multiple configs) | | Lenovo ThinkSystem SR650 | 2U | LGA3647 | 24 | 3072 GB | 2933 MHz | 3 × 300 W | 750–1600 W | Yes (24x 2.5 config) |
HBA vs RAID: the storage controller decision
This is the decision that catches the most homelab buyers off guard. The storage controller in a refurb server is usually a hardware RAID controller: a PERC H730 or H730P on Dell, a Smart Array P408i on HPE. These are fine if you are running Proxmox with local storage and want the controller to manage redundancy. They are the wrong choice if you are running TrueNAS or any ZFS-based storage OS.
ZFS wants direct access to each drive. It manages redundancy, checksums, and scrubbing at the filesystem layer, and it does not work well with a hardware RAID controller interposing its own abstraction. If you feed ZFS a hardware RAID virtual disk, you lose the ability to detect silent corruption, and you cannot use ZFS features like self-healing and device replacement at the pool level.
The solution is an HBA, a Host Bus Adapter that operates in passthrough (IT) mode, presenting each physical drive to the OS as an individual device. On Dell 14G servers, the HBA330 is the HBA option. On HPE Gen10 servers, the Smart Array controllers can be switched to HBA mode in the controller settings, which means you may not need to swap the card at all.
If you are buying a refurb server for TrueNAS, check what controller is installed. If it is a PERC RAID controller, you will need to either cross-flash it to IT mode (possible on some PERC models, not others) or replace it with an HBA330. If it is an HPE Smart Array, you can switch it to HBA mode without changing hardware.
If you are running Proxmox with local storage and do not plan to use ZFS for the boot pool, a hardware RAID controller with cache is a perfectly good choice. The PERC H730P has 2 GB of cache and supports RAID 0, 1, 5, 6, 10, 50, and 60. The HPE Smart Array P408i-a has 2 GB of cache and similar RAID levels. The cache improves small random write performance, which matters for VM disk I/O.
The rule: HBA for ZFS, RAID with cache for VM local storage. Know which one you need before you buy.
Firmware: check before you deploy
Enterprise servers that come off a three-year refresh cycle may not have been firmware-updated before decommissioning. A server that was running fine in a data center on a known-good firmware revision does not need updates to keep running, but if you are installing a fresh hypervisor or storage OS, you want the firmware current to avoid known bugs and security vulnerabilities.
The four things to update before deploying a refurb server:
1. Management controller (iDRAC, iLO, or XCC). This is the out-of-band controller that runs independently of the OS. It has its own firmware, its own network stack, and its own security surface. Update it first, before anything else, because it often coordinates the rest of the update process.
2. BIOS. The system BIOS controls CPU microcode, memory training, and boot behavior. Updates often fix stability issues with specific CPU steppings or memory configurations.
3. Storage controller. RAID and HBA firmware updates fix compatibility issues with specific drive models and can resolve intermittent link errors. If you are cross-flashing a PERC to IT mode, you need the HBA firmware image, not the RAID firmware.
4. NIC firmware. Network adapter firmware updates fix PXE boot issues, iSCSI offload bugs, and link negotiation problems with specific switches.
Dell servers can be updated in bulk using the Dell Update Package system or the iDRAC Lifecycle Controller, which can pull updates from Dell's repository over the network. HPE servers use the Smart Update Manager or iLO firmware update. Lenovo uses the XClarity Administrator. All three vendors make firmware available for free download. You do not need a support contract to access firmware for these generations.
Noise, power, and physical fit
A few practical notes that are not on a spec sheet:
Depth. A 2U rack server is deeper than a consumer NAS or a desktop tower. A standard 4-post rack needs to be deep enough to hold one with the bezel off. Wall-mount racks are often too shallow unless explicitly rated for enterprise servers. Check the server's physical depth against your rack before buying.
Rails. Refurb servers sometimes ship without rails. Dell Rapid Rails and HPE Gen10 rails are specific to the chassis; you cannot mix generations. If the listing does not mention rails, ask. Buying them separately costs more than you would expect.
Noise at idle. Enterprise 2U servers are not silent at idle, but they are not jet engines either. The fans ramp based on inlet temperature and component load. A server in a cool basement with a light load will be quieter than the same server in a warm closet running a compile job. Tower servers like the ML350 run the slowest fans and are the most livable in a shared space.
Power. A dual-socket server with 128 GB of RAM and a handful of drives will draw more at idle than a consumer NAS, and significantly more under load. The PSU ratings in this guide tell you the maximum the platform can deliver, not what it will draw in your config. Check the PSU options for your target platform and budget accordingly, and if you are adding GPUs, the PSU wattage needs to cover the GPU TDP on top of the base system draw.
What to verify before you buy
| Priority | What to check | Why it matters | |:---|:---|:---| | 1 | Storage controller model | Determines whether you can do ZFS passthrough or need to swap to an HBA | | 2 | Backplane type | SAS/SATA only vs NVMe-capable vs tri-mode. Sets your storage ceiling | | 3 | DIMM population | How many slots are filled and with what size. Affects memory bandwidth and capacity | | 4 | PSU wattage | Must match your power budget, especially if adding GPUs | | 5 | Management license | iDRAC Enterprise / iLO Advanced / XCC Enterprise needed for remote console | | 6 | Rails included | Missing rails are expensive to source separately | | 7 | CPU model and TDP | Confirms the processors match what the listing claims | | 8 | Firmware versions | Older firmware may have known bugs or security vulnerabilities |
Where to start
The best homelab server is the one that matches your workload, fits your space, and does not require you to fight the hardware to run your software. If you are running TrueNAS, look for a DL380 Gen10 or R740 with an HBA, or an HPE Smart Array you can switch to HBA mode. If you are running Proxmox with VMs, an R740 with a PERC H730P and 128 GB of RAM is a proven combination. If you want GPU compute, the R750's eight memory channels and 32 DIMM slots give you the most headroom.
You can configure any of these platforms at kciservers.net/build. The configurator verifies component compatibility and estimates power draw as you select parts. Contact us for current pricing and availability.
*Written by the kciservers.net Technical Team