What Proxmox actually asks for
Proxmox's recommended hardware list is short. It wants an Intel 64 or AMD64 CPU with the Intel VT or AMD-V flag, at least 2 GB of memory for the OS and Proxmox services plus whatever your guests need, fast storage (SSDs give the best results), and redundant gigabit NICs. Nearly any used business mini PC from the last several years clears that bar. What makes a box pleasant to live with is RAM first, then storage layout, then network ports and remote management.
Count your guests before you pick RAM
Add it up before you buy. The host wants about 2 GB. Home Assistant's own install guide lists a minimum of 2 GB of RAM and 2 vCPUs for its VM. Pi-hole needs 512 MB, so it lives happily in a small LXC container. Then add n8n, a reverse proxy, a dashboard or two, and any agent runtimes you plan to try. 16 GB covers a starter stack, and 32 GB is where you stop rationing. See Home Assistant, workflow automation and AI agent server for what each of those workloads needs on its own.
If you use ZFS, budget for it. The Proxmox ZFS page says to start with at least 8 GB, and by default Proxmox gives ZFS's ARC cache 10% of host memory, capped at 16 GiB. Its rule of thumb for sizing the ARC is 2 GiB base plus 1 GiB per TiB of storage. On a 32 GB node with a 1 TB pool that works out to about 3 GiB, which is comfortable. On a 16 GB node it starts to compete with your VMs.
Check the platform's ceiling as well as what's installed. Lenovo's PSREF lists 32 GB as the maximum for the M720q and M920q and notes that the figure comes from testing with Lenovo's own memory offerings. Intel lists the N100 at 16 GB on a single memory channel, which is why an N100 box makes a good light node or third cluster vote and a cramped main node. The N100 vs N150 vs N305 guide covers the rest of that family.
NVMe for guests, a second drive for safety
Put VM disks on NVMe. Most 1-liter business PCs pair one M.2 NVMe slot with a 2.5-inch SATA bay, as the M720q and M920q do. Use the SATA drive as a vzdump backup target or an ISO and template store, and keep a copy of anything you care about off the box. A ZFS mirror needs two similar drives. The ThinkCentre M920x is unusual in Lenovo's tiny line because it has two M.2 SSD slots plus a 2.5-inch bay (you give up the bay if you fit its PCIe card). Proxmox's ZFS documentation recommends enterprise-class SSDs for cache and log devices and high-quality ECC RAM. Lenovo lists non-ECC memory for tinies such as the M920x and M710q, so treat backups as mandatory.
Networking, and the Intel I219 hang
One gigabit port is fine for a single node. Most Intel business tinies use an Intel I219-V or I219-LM controller on Linux's e1000e driver, and a long-running Proxmox forum thread collects users seeing "e1000e … Detected Hardware Unit Hang" messages and dropped connections, including on I219-V hardware. The workaround posted there is to turn off TCP and generic segmentation offload with ethtool -K <interface> tso off gso off. If the message shows up in your kernel log, apply it and add it to the interface configuration so it survives a reboot.
Need more ports? The M720q and M920q take an optional 4-port Intel I350-T4 PCIe card in place of the 2.5-inch drive. Consumer ODM boxes such as the Beelink EQ12 ship with two 2.5 GbE ports instead. For firewall duty, see network services.
vPro and remote KVM
Intel AMT, sold under the vPro name, is a management engine that runs independently of the OS. It can power a machine on or off and give you a hardware-level remote screen, keyboard and mouse, which is exactly what you want when a headless node hangs at boot. The M920q has Intel vPro with AMT 12.0 on supported configurations, and the M720q has none. Only units built with a vPro-capable CPU have it, so look for the vPro sticker or an Intel AMT / MEBx entry in the BIOS.
For years the go-to free tool was MeshCommander, and its GitHub page now says Intel has discontinued support for it. MeshCentral, a free, open-source, self-hosted remote management server, documents full Intel AMT support, and it runs fine in a small LXC container on the cluster itself. AMT is a powerful network-facing interface, so set a strong password when you provision it and keep it on your LAN.
Clustering three tiny nodes
Proxmox's cluster documentation says you need at least three nodes for reliable quorum, a network with under 5 ms latency between nodes, and UDP ports 5405–5412 open for corosync. A two-node cluster can use a QDevice for the third vote. Three used tinies on one switch is a classic setup. You can migrate guests, patch one node while the others carry the load, and survive a dead power brick. If you plan to try Ceph, read the requirements first. Proxmox budgets roughly 1 GB of memory per TB of Ceph or ZFS storage and calls for additional NICs depending on your storage design, which is a lot to ask of single-port 1-liter boxes.
Where the other workloads fit
A Proxmox node can host the rest of this site's use cases. Put network services like Pi-hole and a reverse proxy in containers, Home Assistant in a VM, and workflow automation with n8n alongside. Frigate and a media server can run here too, but plan how their container or VM will reach the Intel iGPU. A local LLM is the workload most likely to outgrow a 32 GB tiny. Before buying, run through the used mini PC buying checklist and compare the big three business lines in ThinkCentre vs OptiPlex vs EliteDesk.




































