10GbE network switch and cabling for a local AI homelab

10GbE for Local AI: 6 Proven Facts and 1 Costly Mistake

10GbE for local AI: when it actually helps, the $50 point-to-point build, and why buying 10GBASE-T copper is the costly mistake.

This guide is built from published pricing, hardware specifications and third-party homelab testing rather than a specific network we built. Component prices move frequently, so verify before purchasing.

Most people running local AI do not need 10GbE. That is the honest answer, and it contradicts nearly every homelab guide you will read.

The exception is specific and worth knowing: if your models live on a NAS rather than local disk, the difference between 2.5GbE and 10GbE is roughly two and a half minutes versus thirty-five seconds every time you switch models. That is felt. Everything else about local AI happens after the model is loaded, where the network sits idle.

The costly mistake, if you do upgrade, is buying the wrong kind of 10GbE.

In this guide:

10GbE SFP+ DAC cabling between local AI machines
Inference never touches the network. Loading a model from a NAS is the only place this spending shows up.

What the Network Actually Does in Local AI

Three moments, and only one of them is network-bound.

Loading a model. Weights move from wherever they are stored into memory. If that storage is a NAS, this is entirely a network operation and the only place 10GbE changes your day.

Running inference. The model sits in VRAM or unified memory. The network does nothing. A 1GbE connection and a 10GbE connection produce identical tokens per second.

Serving to other machines. Request and response text is tiny. A conversation is kilobytes. Any network from 1GbE upward handles this without noticing.

That distribution is why the answer is usually “no.” Inference is where the time goes, and inference does not touch the network. Our storage for local AI guide makes the same point about disks: sequential read matters for load time only, and load time is the one thing people optimise last.

The Model Loading Numbers

The arithmetic that decides this, using a 40GB model, which is roughly a 70B at 4-bit quantisation.

Link Real throughput 40GB model load
1GbE ~110 MB/s ~6 minutes
2.5GbE ~280 MB/s ~2.4 minutes
10GbE ~1,100 MB/s ~36 seconds

The jump from 1GbE to 2.5GbE is the large one in proportional terms and costs almost nothing, since most hardware shipped since 2023 includes 2.5GbE natively. Mini PCs, Synology and QNAP NAS units all ship with it, which often makes the upgrade cost just the switch.

The jump from 2.5GbE to 10GbE saves roughly two minutes per model load. Whether that is worth $150 depends entirely on how often you switch models. Loading one model each morning, it is not. Switching between five models an afternoon while testing, it is.

Local storage sidesteps the question entirely. An NVMe drive loads that same 40GB model in around eight seconds, which beats 10GbE by a factor of four. If your models fit on local disk, network speed is irrelevant to loading and you should stop reading here.

The 10GBASE-T Mistake

This is the costly mistake, and it is easy to make because the cheap-looking option is the expensive one.

There are two ways to do it. 10GBASE-T runs over familiar RJ45 copper and looks like a straightforward upgrade. SFP+ uses a different port with DAC cables or optical transceivers and looks more intimidating.

SFP+ wins on every axis that matters:

  • Power. SFP+ NICs draw roughly 2 to 4W per port. 10GBASE-T draws 10 to 15W per port. On an always-on machine that is a permanent difference in heat and electricity.
  • Cost. SFP+ switches are cheaper. A four-port SFP+ switch runs around $95 to $140, where comparable 10GBASE-T copper switches sit near $160 and up.
  • Latency. SFP+ with DAC is lower latency than copper 10GBASE-T.
  • Cabling. 10GBASE-T needs Cat6A for full-distance runs. A passive DAC cable under five metres costs $12 to $20 and needs no cabling decisions at all.

The published guidance is blunt: avoid 10GBASE-T for home labs. If you are connecting machines in the same rack or the same room, SFP+ with DAC is cheaper, cooler and simpler.

The exception is if your existing hardware only has RJ45 10GbE built in, in which case you use what you have rather than buying adapters to avoid adapters.

What 10GbE Costs in 2026

Prices dropped enough that the gap is smaller than most people assume.

Point to point, two machines, roughly $50. Two used Mellanox ConnectX-3 cards at $20 to $30 each plus a one-metre DAC. No switch at all. If you are connecting one AI box to one NAS, this is the highest return upgrade available and it is startlingly cheap.

Switched, four nodes, roughly $200. Add a MikroTik CRS305-1G-4S+IN at around $95 to $140 for four SFP+ ports at 9W idle. This has been the default homelab 10GbE recommendation since 2022 and remains it.

2.5GbE for comparison. An eight-port unmanaged 2.5GbE switch such as the TP-Link TL-SG108-M2 costs around $65, or the Netgear MS308G at $80. Managed 2.5GbE with 10GbE uplinks, like the TP-Link TL-SG3210XHP-M2, runs about $150.

The overall gap between a basic 2.5G setup at $70 to $180 and a basic 10G setup at $200 to $280 is roughly $100 to $150. That is real money and it is not the several-hundred-dollar barrier the older guides describe.

When 2.5GbE Is Genuinely Enough

Being direct, because the upgrade is easy to talk yourself into.

Models on local NVMe. Local storage already beats 10GbE by four times on load speed. The network is not in the path.

Spinning-disk NAS. A mirrored HDD array will not exceed roughly 200 MB/s. 2.5GbE at 280 MB/s is already faster than the disks behind it, so 10GbE would connect to nothing.

One machine and one NAS, occasional transfers. If you load a model in the morning and use it all day, two minutes once is not worth rewiring for.

Serving text to other devices. Chat traffic is negligible at any speed.

The inverse also holds, and it is the case our GMKtec EVO-X2 review flags: a machine with a single 2.5GbE port that serves models to a network hits that limitation first, before anything about the GPU matters. The Beelink GTR9 Pro exists in the same category with dual 10GbE precisely because that is the differentiator worth paying for in a serving box.

Clustering Needs Different Thinking

If your plan is linking several machines to run a model too large for any one of them, the network stops being a convenience and becomes part of the compute path.

Distributed inference with Exo or llama.cpp RPC passes activations between nodes for every token generated. That is latency-sensitive in a way that file transfer is not, and it is why clustering guidance leans toward the fastest interconnect available rather than the cheapest adequate one.

Two honest caveats.

Clustering helps capacity, not speed. Splitting a model across nodes lets you run something that does not otherwise fit. It does not make a model that already fits run faster, and inter-node latency usually makes it slower.

Thunderbolt beats Ethernet for two machines. Where both machines support it, Thunderbolt 5 offers higher bandwidth and lower latency than 10GbE. Our Mac mini review notes that Apple reserves Thunderbolt 5 clustering for the M5 Pro model, which matters if you were planning a fleet of the cheaper ones.

Cabling and Power

Two practical details that catch people out.

Cabling is less demanding than expected. 2.5GbE runs on ordinary Cat5e up to 100 metres, so no rewiring. 10GBASE-T also runs on Cat5e up to about 30 metres and Cat6 to 100 metres, though signal integrity requirements are tighter. For SFP+, a passive DAC under five metres avoids the question entirely; use fibre for longer runs or through walls.

Power adds up on always-on gear. Budget roughly 6 to 9W idle per 10G NIC, and the MikroTik CRS305 draws about 9W idle. On a machine running continuously, that is worth including in the running-cost arithmetic alongside the box itself.

What Actually Bottlenecks a Model Server

Worth placing networking in context, because it is rarely the first constraint people hit.

Memory capacity comes first. A model either fits or it does not, and no network fixes that. This is the binary wall covered in our how much VRAM you need guide, and it decides what you can run at all.

Memory bandwidth comes second. Once a model fits, generation speed tracks bandwidth almost directly. This is why a Mac Studio outruns a Strix Halo box holding the same model, and why the Framework Desktop review spends more time on 256 GB/s than on anything else.

Storage speed comes third, affecting load time only.

Network speed comes fourth, and only when storage is remote.

That ordering is worth keeping in mind when a budget is finite. Spending on networking while a model spills out of memory is optimising the wrong end of the pipeline, and the symptom of the real problem, throughput collapsing when a model does not fit, looks nothing like a network issue.

The one case that reorders this list is clustering, where the interconnect moves into the compute path and becomes latency-critical rather than throughput-critical. That is a different problem from “my NAS is slow,” and it deserves different hardware.

Who Should Buy What

If you are Buy Why
Running models from local NVMe Nothing Local storage already beats 10GbE
One AI box and one NAS, frequent model switching Two used ConnectX-3 plus a DAC, ~$50 Point to point, no switch, highest return
Three or four 10G-capable devices MikroTik CRS305, ~$95 to $140 Four SFP+ ports at 9W idle
Backed by spinning disks 2.5GbE switch, ~$65 Disks cap out below 2.5GbE anyway
Clustering for capacity Fastest interconnect available Activations cross the network per token
Choosing between copper and SFP+ SFP+ with DAC Cheaper, cooler, lower latency
Serving models to a network A box with 10GbE built in Saves adding a NIC later

Frequently Asked Questions About 10GbE for Local AI

Does 10GbE make local AI inference faster?
No. Once a model is loaded into memory the network is idle, and tokens per second are identical on 1GbE and 10GbE. It only affects how long loading takes from network storage.

How much faster does a model load over 10GbE?
A 40GB model takes roughly six minutes over 1GbE, two and a half minutes over 2.5GbE, and about 36 seconds over 10GbE. Local NVMe does the same job in around eight seconds.

Should I buy 10GBASE-T or SFP+?
SFP+ with DAC cables for anything in the same room or rack. It costs less, draws 2 to 4W per port against 10 to 15W for 10GBASE-T, and has lower latency. Use copper only if your hardware already has it built in.

What is the cheapest useful 10GbE setup?
Two used Mellanox ConnectX-3 cards at $20 to $30 each plus a $12 to $20 DAC cable, connecting two machines directly with no switch. Around $50 total.

Do I need to rewire my house?
No for 2.5GbE, which runs on existing Cat5e to 100 metres. Probably not for 10GbE either, since 10GBASE-T works on Cat5e up to roughly 30 metres, and SFP+ DAC needs no structured cabling at all.

Is 10GbE worth it for a single NAS and desktop?
Usually not, unless you move large files often. If your NAS is spinning disks, definitely not, because the drives cap below what 2.5GbE already delivers.

A Sensible Upgrade Order

If you are optimising a local AI setup and wondering where networking sits in the queue.

First, put active models on local NVMe. This removes the network from the path entirely and is faster than any network upgrade.

Second, make sure you are on 2.5GbE. Most modern hardware already has it, so this is often just a $65 switch. The jump from 1GbE is the biggest proportional gain available.

Third, go 10GbE only if you have a specific reason. Frequent model switching from network storage, an all-flash NAS, or a clustering plan. Not because the number is bigger.

Fourth, if you do go 10GbE, use SFP+. The power and cost difference compounds on hardware that runs continuously.

Most people should stop after the second step. That is the genuinely useful conclusion, and it is the one that saves money rather than spends it.

Verdict

Networking is the least important upgrade in most local AI setups, and 10GbE is worth it for a narrower group than homelab enthusiasm suggests. Inference does not touch the network, chat traffic is negligible, and local NVMe loads models four times faster than 10GbE anyway. For most people, 2.5GbE is already present in their hardware and already sufficient.

The exception is real though, and if you fall into it the upgrade is cheap. Models on a NAS plus frequent switching means 10GbE turns a two-and-a-half-minute wait into thirty-five seconds, several times a day. Two used ConnectX-3 cards and a DAC cable cost around $50 for a point-to-point link, which is the best value networking upgrade available anywhere in this hobby. The mistake to avoid is buying 10GBASE-T copper because the ports look familiar, and paying more for a hotter, slower, more cable-dependent version of the same speed.

Concrete next step: time how long a model actually takes to load in your current setup, and count how many times you do it in a working day. Multiply those together. If the answer is under five minutes a day, spend the money on storage or memory instead, because they affect what you can run rather than how long you wait to start.

Sources and Further Reading

For related coverage on this site, see storage for local AI for the local-disk alternative, the Beelink GTR9 Pro review for a machine with dual 10GbE built in, the GMKtec EVO-X2 review for where single 2.5GbE becomes a limitation, and the home AI server build for the machine this network connects to.