Expert Home Server Setup: A $7,890 Eight-Bay Build

This is the ceiling of a sensible home build: eight bays, 10 gigabit networking, double parity and a real backup strategy, for about $7,890. It is also the tier where the economics of home storage become impossible to ignore, because the eight drives cost $4,640 and the enclosure costs $1,370. The box is 17 percent of the bill. Everything the internet argues about is 17 percent of the bill.

Read this before the parts list. Most people who want this build should buy the four-bay build instead. The extra $4,351 buys four times the usable capacity, and unused capacity earns nothing while still drawing power, ageing, and eventually needing replacement. If you cannot name the data that will fill 96 TB within three years, you are buying a specification rather than a solution. There is a section at the bottom of this page on exactly who should not buy it.

What capacity do you actually end up with?

Eight 16 TB IronWolf Pro drives give 128 TB raw. Parity takes whole drives, so what survives depends on the layout, and at eight members the choice matters far more than it does at four.

All figures are decimal terabytes, the number printed on the drive. Your system reports binary tebibytes and labels them "TB": one TB is 1,000,000,000,000 bytes, one TiB is 1,099,511,627,776, a fixed ratio of 0.909495 that makes every capacity look 9.05 percent smaller. On a 96 TB volume that is an apparent difference of nearly 9 TB, which is enough to alarm people who have not met it before. Nothing is missing. See the TB against TiB chart.

If you run ZFS rather than a vendor operating system, plan on keeping the pool below about 80 percent full for write performance, and note that ZFS separately reserves 1/64 of the pool, about 1.56 percent, as slop space. Between the two, treat 76.8 TB as the practical working ceiling of a RAIDZ2 pool at this size rather than the full 96 TB.

Why single parity is unacceptable at eight members

The capacity argument says RAID 5 gives you 112 TB against RAID 6's 96 TB, a 16 TB gain for free. The rebuild argument says do not do it, and the rebuild argument wins.

When a member fails under single parity, reconstructing it requires reading every sector of every surviving drive: about 112 TB on this build. At a realistic 80 to 150 MB/s, with the array still serving files, that is roughly 9 to 16 days of continuous reading. For the whole of that period there is no redundancy at all, and the seven drives being read flat out are the same age as the one that just died, in the same thermal environment, often from the same manufacturing run.

Applying the manufacturer's rated unrecoverable read error floor of one per 10^15 bits to that read volume gives roughly 59 percent. Under RAID 6 the equivalent figure is 54 percent, and more importantly a second failure or a bad sector no longer costs you the array.

Treat those percentages as a ranking, not a forecast. The URE rating is a warranty floor rather than a measurement, observed error rates across large drive fleets run substantially better than specification, and modern systems survive a read error rather than aborting: ZFS reports which file was affected and continues, and mdadm can be told to carry on. What the numbers legitimately tell you is the ordering, and the ordering is unambiguous at eight members with large drives.

Run the same comparison for your own drive set in the cross-vendor RAID capacity calculator, and put annualised failure rates against eight drives over five years in the drive failure probability calculator. At a large-fleet annualised failure rate of 1.4 percent, eight drives over five years is not a remote possibility, it is close to a coin flip that at least one dies.

Can you actually back up 96 terabytes?

No, not affordably, and this is where most eight-bay build guides quietly stop talking. Two 20 TB external drives cover 40 TB for about $1,400. Covering 96 TB the same way would cost roughly $3,400 in external drives, and covering it in cloud storage at a typical $6 per TB per month would be about $6,900 a year, every year.

The correct answer is not to back up the array. It is to back up the data. Split what you store into two categories and treat them completely differently.

Two 20 TB units cover the first category with several years of headroom, split by data set and rotated so one is always somewhere other than your house. Keep a plain text inventory of the second category, because the expensive part of losing replaceable media is not re-acquiring it, it is not remembering what you had. Size all of it properly with the 3-2-1 backup sizing calculator, and read the 3-2-1 guide for the version-history and restore testing that make a backup real rather than theoretical.

None of this is optional because you have double parity. RAID 6 survives two drives failing. It does nothing about a deletion you do not notice for a month, ransomware that encrypts every share it can reach, a controller writing garbage to all eight members, theft, fire or flood. Snapshots, covered in NAS security hardening, are the defence that actually works against the first two.

Does 10 gigabit networking pay off here?

This is one of the few home configurations where it genuinely does, and the reason is arithmetic rather than enthusiasm. A single 7200 rpm drive sustains roughly 150 to 260 MB/s sequentially. Six data drives striped under RAID 6 sustain roughly 900 to 1,500 MB/s. That is comfortably past 2.5 gigabit, which tops out around 280 MB/s, and into the range 10 gigabit was built for at about 1,100 MB/s.

Two cautions. First, the link is only as fast as its slower end: if your workstation has no 10 gigabit port, buy the network card before the switch. Second, 10 gigabit gear runs hot and idles higher than 2.5 gigabit, typically 15 to 40 W for a small switch against 3 to 5 W, which is a real addition to an always-on build. Cabling needs to be Cat6 or better for reliable 10 gigabit over household run lengths. The 2.5GbE against 10GbE comparison has the full argument, and the network throughput calculator will tell you whether your array or your link is the limit.

What can you skip?

The NVMe cache, in its usual form. A read cache in front of a 96 TB media array has almost no hit rate, because sequential streaming does not benefit from caching. What the four M.2 slots are genuinely good for at this tier is a separate fast pool: put active projects, virtual machine disks or a photo catalogue on two NVMe drives mirrored, and leave the spinning array for bulk. Same hardware, completely different and much better use.

The 10 gigabit switch, if your workstation is gigabit. The enclosure already has the ports. Buy the card first, prove you need the throughput, then buy the switch.

Buying all eight drives at once, if cash flow says otherwise. Six drives in RAID 6 is a perfectly good array, and buying the last two later spreads about $1,160. Note the trade honestly: expanding a ZFS vdev is far more constrained than expanding an mdadm array, so check what your chosen operating system supports before planning to grow into empty bays.

What you cannot skip is double parity and the UPS. Eight members on single parity with 16 TB drives is the configuration that loses people arrays, and eight drives spinning up together draw a surge that a small battery will not carry. The 1500 VA unit is sized for the surge, not the steady load.

What should you upgrade first?

• Memory, before anything else. Eight gigabytes is thin for an eight-bay box running containers, and it is genuinely inadequate if you put ZFS underneath. This is the cheapest upgrade on the list and the one with the most obvious effect.
• A third offsite copy of the irreplaceable tier. Cloud storage priced per terabyte is affordable for 10 to 20 TB and absurd for 96 TB, which is exactly why the tiered approach above matters.
• A flash pool rather than a cache. See above. It is the same money and a much better outcome.
• Not more bays. If eight 16 TB drives are not enough, the next move is larger drives, not a second chassis. Replacing drives one at a time with 24 TB models raises the array to 144 TB usable under RAID 6 without adding a second point of failure. Plan the sequence with how to migrate to a bigger NAS, and note that under standard RAID you gain nothing until every member is replaced.

Who should not buy this build?

Four kinds of buyer, stated plainly, because talking people out of the expensive tier is the only reason to trust anything else on this page.

• Anyone who cannot name the data that will fill 96 TB within three years. Unused capacity earns nothing, draws power, and ages toward replacement regardless. The four-bay build is $4,351 cheaper and its 24 TB is more storage than most households ever fill.
• Anyone buying this primarily for Plex. Transcoding is a processor feature, not a bay-count feature. A four-bay with good Quick Sync silicon serves more simultaneous streams per dollar than eight bays ever will.
• Anyone who needs the machine to be quiet or efficient. Eight spinning drives are audible in a living space and idle around 80 to 100 W, which is roughly $120 to $150 a year at 17 cents per kWh, or $600 to $750 over five years. Check the numbers in the power draw calculator before committing.
• Anyone who has not yet solved backup. An unbacked 96 TB array is not eight times better than an unbacked 12 TB one, it is eight times more data to lose in a single event. Solve copies two and three first, at any capacity, and then scale.

Alternatives worth considering at this tier, if the specification rather than the brand is what you are after: the QNAP TS-873A for its 64 GB memory ceiling and two PCIe slots, the Synology DS1825+ if you want ECC memory and SHR-2 with the longest software track record, or the TerraMaster F8 SSD Plus if silence matters more than cost per terabyte. All three are covered in the eight-bay roundup.

What does it cost to leave running?

Eight drives at 6 to 9 W idle plus a Core i5 board at 25 to 35 W plus a 10 gigabit switch at 15 to 40 W puts this build around 90 to 145 W idle. At 8,760 hours and 17 cents per kWh, using W / 1000 x 8760 x tariff, that is roughly $134 to $216 a year, or $670 to $1,080 over five years. That is most of the price of the enclosure again, spent invisibly, and it is the strongest practical argument against buying capacity before you need it. The power consumption chart has the per-component figures.

Where this build sits against the others

Related reading

• Best eight-bay NAS, the enclosure alternatives
• Best NAS hard drives, where 59 percent of this budget goes
• RAID 5 against RAID 6, the rebuild arithmetic in full
• Cross-vendor RAID capacity calculator
• NAS security hardening, because parity does not stop ransomware
• All three complete builds compared