Homelab Regrets: Five Purchases I'd Take Back

The used rack server is the purchase I would undo first. Cheap to buy, expensive to run, and far louder than the listings suggest — “quiet” in that market means quiet for a datacenter.

Every regret below has that shape. The purchase price was the number that looked good, and the cost that mattered turned up somewhere else: in what the box draws around the clock, in the noise floor, in a chassis outgrown almost immediately.

Five purchases I would take back, worst first.

1. Used rack servers — the one I would undo first

The noise part is measurable, and Dell publishes the measurements. Its 13G PowerEdge acoustical document covers the 13G line — the R630, R730 and R730xd among them, the generation that fills used listings — and it reports per configuration rather than per model, which turns out to be the whole story:

Sound power is the upper-limit LwA per ISO 9296; the dBA figure is bystander sound pressure, taken in a 24U rack at 23°C per ISO 7779.

Two things fall out of that table. The minimum configuration is one processor, one 4GB stick, one SSD and no expansion cards, and nobody sells that machine used. And the feature-rich R630 prints the same number for idle and operating — 6.4 bels either way. It does not quieten down, because there is nothing to quieten down to.

The dependency notes in the same document are where the buying advice actually lives:

• Low-power Xeons make it louder. Dell states that configurations with 65W L-series parts such as the E5-2650L v3 run “about twice as loud as typical configurations” under moderate or heavy use. Those are the exact chips people buy to cut the power bill.
• A GPU doubles it again. Any GPGPU card in an R730 is “about twice as loud” as the typical build.
• Drive count raises the floor. An R730xd 3.5-inch chassis with 16 drives idles about 50% louder than one with four — 6.2 bels against 5.6.
• PCIe SSDs raise the ceiling. Under highly stressed conditions the R730 goes to 7.0 bels.

Dell’s own reference table settles what “quiet” means to the people who designed these. It maps 75 dBA and 39 sones to “Data center, vacuum cleaner, voice must be elevated to be heard,” and 45 dBA and 4 sones to “Whispering, open office layout, normal living room.” A machine engineered to be acceptable in the first room is not being engineered for your second bedroom.

The running cost points the same direction. The idle power costs page puts a single-CPU Dell R720 at 80–100W against 15–22W for a current mini PC, on the lightly-threaded workloads most homelabs actually run. What the rack server still wins on is RAM capacity, drive bays and expansion slots — not compute per watt.

Replacing the rack gear with small nodes is the best change I have made to the setup. If you are still weighing it, used enterprise versus mini PC puts the two side by side, and servers for a homelab covers the classes in between.

2. Over-building generally

I spec’d for a workload that never showed up.

Every watt of headroom you buy ahead of the work is a watt you pay for twenty-four hours a day while the work fails to arrive, and the hardware depreciates the entire time it waits. The power cost calculator will price the headroom you are considering before you commit to it.

Buy for the workload you are running this quarter. The homelab that grows into more capacity is a cheaper machine than the one that was already sitting there.

3. Two-bay NAS units

Two bays get outgrown almost immediately. Four bays or build.

The arithmetic is the whole argument. Two bays mirrored leave you one drive of usable capacity, and the only way to grow is to replace both drives and rebuild the pool. Four bays under single parity give you three drives of capacity and let you swap disks one at a time as they fail or as prices fall. The NAS capacity calculator shows the gap at whatever drive size you are shopping.

Start at four bays or build your own. What you save on the two-bay enclosure you spend again on the replacement, plus a migration you did not plan for.

4. Cheap 2.5GbE switches

They are a coin flip on thermals.

The warranty is the filter I would use now, because it is the one line in a listing that costs the manufacturer something to write. TP-Link’s US warranty schedule gives switches with a plastic chassis two years, non-Omada switches with a steel chassis three, and reserves limited lifetime cover for the Omada S4500, S5500, S6500 and S7500 series. TRENDnet publishes a lifetime warranty on the TEG-S350, its 5-port unmanaged 2.5G box.

If the switch in front of you offers a single year, you have left the tier that stands behind these. Buy from a brand whose warranty terms you can find without emailing anyone.

5. Exos bought on price and capacity

I bought Exos on price and capacity, then regretted them for noise.

The published acoustics say where the gap actually is. Here are three 24TB enterprise drives, sound power in bels, each from its manufacturer’s own document:

Under seek all three print the same figure. The entire published gap sits at idle — 0.8 bels, eight decibels of sound power — and idle is where a NAS spends its life. The drives are not working during most of the hours you can hear them, which is why the seek number is the wrong one to shop on and the idle number is the right one. The gap also does not go away when there are six of them in one box. One caveat from Seagate’s own footnote: offline SMART activity during idle “may increase acoustic and power to operational levels,” so the idle figure is a floor rather than a promise.

None of that makes the Exos a bad drive. It is the cheapest 24TB drive in its class by enough to matter across six bays — the six-bay rebuild math priced it at $39.50/TB against $44.17 for the WD Red Pro in August 2026. If the array lives in a basement or a garage, buy on price and buy the Exos. If it shares a room with you, shop the published idle figure and pay for the Ultrastar or the MG.

What I would buy instead

• Small nodes, not a rack. This is the change that improved the setup most.
• Four bays minimum, or build the box yourself.
• The workload you have, not the one you are planning for.
• A switch with a warranty you can read, over the cheapest 2.5GbE port count.
• Idle acoustics over price on drives, but only when the array is in earshot.

Each of these looked good on one number. The rack server was cheap. The switch was cheap. The Exos are still the cheapest per terabyte. The number that decided the outcome was always the second one.