The home-lab upgrade conversation usually starts with someone quoting 10 gigabits, then stalls while everybody tries to remember what category their landlord pulled through the walls. That is the wrong order. The useful first question is not "how fast can the network go" but "what is the slowest thing between the two ends of this link" — because that is the only number a file copy will ever see. In most home labs the slowest thing is not the cable, and often not even the switch.

Line rate is not file-copy speed

Gigabit Ethernet is 1,000 megabits per second, which is 125 megabytes per second once you divide by eight. Framing, inter-packet gaps and protocol overhead then take their cut, so a healthy 1GbE link moves real files at roughly 110–118 MB/s in practice. Scale it up and the numbers look like this: 2.5GbE is a 312.5 MB/s line rate and lands near 280–295 MB/s usable; 10GbE is a 1,250 MB/s line rate and lands around 1.1–1.18 GB/s. Useful — but only if something on both ends can actually feed that rate.

The number nobody budgets for: the drive

This is where most home-lab upgrades quietly fail. A modern 3.5-inch CMR hard drive reads sequentially at roughly 150–250 MB/s, and that figure is optimistic: rotational drives are fastest on the outer tracks, so the same drive degrades as the platter fills and the head moves inward. A SATA SSD is capped by its own interface at about 500–550 MB/s regardless of how good the network is. Only NVMe changes the arithmetic.

Put those two facts next to the line rates and the decision mostly makes itself. One spinning drive over 2.5GbE is already at or above its own ceiling — the network is not your bottleneck, the platter is. The same drive over 10GbE copies at exactly the same speed it did before, minus the money you spent. 10GbE only starts paying when both ends are flash, or when you are moving data across several drives at once (a rebuild, a multi-stream backup, a RAID array that aggregates). And if you are running a model server off your NAS, remember that the weights file is read once at load, not continuously — a fast network helps you start faster, not generate tokens faster.

Where 2.5GbE came from, and why it fits your walls

2.5GBASE-T and 5GBASE-T were standardised in IEEE 802.3bz in 2016 for a very specific reason. Wi-Fi access points had started outrunning a gigabit uplink, 10GBASE-T needed better copper than the world had installed, and nobody was going to rewire millions of buildings over it. So the industry built intermediate speeds designed to run on the installed base — 2.5GBASE-T over Cat5e or Cat6, 5GBASE-T over Cat6 — at the full 100 metre channel length.

The consequence for a home lab is blunt: 2.5GbE is a gear upgrade, not a cabling project. Buy the switch and the NICs, plug in what is already there, and let auto-negotiation settle on the highest speed the link can hold. If a degraded or badly terminated Cat5e run cannot sustain 2.5G, NBASE-T hardware negotiates down to 1G automatically — you will not break anything by trying. The honest caveat is that 5GBASE-T is not specified for Cat5e at the full 100 m; if 5G is the target, Cat6 becomes the floor.

10GbE is where cable becomes the spec

10GBASE-T runs at a much higher signalling frequency, and the specification reflects it. The channel is still 100 m (90 m of permanent link plus up to 10 m of patch cords combined), but what it takes to get there changes: Cat6a is rated to the full 100 m, plain Cat6 is quoted at roughly 37–55 m and Cat5e is not specified for 10G at all. The category frequencies are the underlying reason — Cat5e is characterised to 100 MHz, Cat6 to 250 MHz, Cat6a to 500 MHz.

Note that the Cat6 figure is a range rather than a number, and that is not sloppiness. The limiting factor at 10G is usually alien crosstalk — interference from the other 10G cables bundled against it. One Cat6 cable in free air does considerably better than the same cable in the middle of a tight bundle of twenty. That is the problem Cat6a was designed around, which is why it carries a tighter alien-crosstalk limit and often a separator or a shield. So the practical rule is: a 5 m Cat6 patch cord inside a rack will carry 10G without complaining, and an 80 m Cat6 run through a ceiling void next to nine other cables probably will not.

What actually breaks links is the termination, not the category

Given how much attention the category label gets, it is worth saying plainly that most home installations fail certification on workmanship, not on cable grade. Three rules carry most of the weight. Keep the pairs twisted to within about 13 mm of the termination — untwisting them at the patch panel or keystone is the single most common cause of a link that tests badly. Respect the bend radius, which for these cables is roughly four times the cable diameter; a sharp kink crushes the pair geometry and raises crosstalk permanently. And keep network cable away from mains cable — around 15 cm of separation, crossing at right angles when they must cross, and never sharing the same conduit.

If you are pulling new cable anywhere, the economics point one way: the labour is the expensive part and the headroom is cheap, so run Cat6a and stop thinking about it for the next twenty years. If you are not pulling cable, do not talk yourself into a rewiring project you do not need — measure the link speed first.

If the switch is feeding power, the power budget is the real constraint

Home labs increasingly have one switch doing networking and PoE duty for access points and cameras, and that is where the arithmetic gets less forgiving. The IEEE tiers are worth memorising because they are the most stable numbers in the trade: 802.3af (Type 1) reserves 15.4 W at the switch and guarantees 12.95 W at the device over two pairs; 802.3at (Type 2) is 30 W and 25.5 W; 802.3bt Type 3 is 60 W and 51 W over four pairs; 802.3bt Type 4 is 90 W and 71.3 W over four pairs.

Two things follow. First, the gap between those columns is not vendor padding — it is the allowance for power lost as heat in up to 100 m of copper. At Type 4 the current is 960 mA per pair, and across the standard's maximum loop resistance for Cat5e that dissipates almost 19 W in the cable itself. On Cat6a, whose lower DC resistance cuts that loss to somewhere around 13 W, the same device sees closer to 77 W rather than 71.3 W. At high power, the cable is part of the power supply. Second, PoE budgets are set per chassis, not per port: a 48-port switch with a 370 W budget can carry twenty-four 30 W PoE+ devices on paper, and fewer once anything on the list is Type 3 or 4. Budget in PSE watts, specify devices in PD watts, and do not assume the port count is the limit. Above 60 W per cable in a bundle, installation code takes over and imposes derating by conductor gauge and bundle size — one more reason the cheap thin patch cables are the wrong place to save money.

The decision, honestly

If you have Cat5e or better in the walls and one machine that would like faster access to a NAS, buy 2.5GbE gear and keep the cable. That covers the overwhelmingly common case: a workstation, a media box and a NAS on one switch. If you are pulling cable anyway, pull Cat6a regardless of what speed you think you need today, because the incremental cost is small against the labour. Reserve 10GbE for the specific case where both ends are flash and large files genuinely move between them — that is the only configuration where the line rate and the workload match. And before any of it, look at the drives, the switch's power budget and the terminations, because in a home lab those three decide your throughput long before the cable category does.