Mini PCs and NUC-style nodes keep showing up as the default homelab building block on this site, and most of them ship with Thunderbolt 4 or USB4 ports that never get used for anything but an external drive. Those ports are capable of carrying IP traffic directly between two machines, no switch required, at speeds a 2.5GbE or even 10GbE port can’t touch for the price. It’s a real option for a small Proxmox or TrueNAS cluster sitting on a desk, not a parlor trick, but it comes with sharp edges that the glossy “40Gbps!” marketing number doesn’t mention.

What Thunderbolt/USB4 networking actually is

Thunderbolt 3 and 4, and the USB4 spec that absorbed Thunderbolt’s networking capability, support a feature usually called Thunderbolt Networking or Thunderbolt IP: plug two Thunderbolt-capable machines together with a Thunderbolt or USB4 cable, and each end exposes a virtual Ethernet-like network interface to the OS. No switch, no separate NIC, no SFP+ module. The link shows up as a standard network interface (en* on macOS, thunderbolt0-style or a USB4-backed interface on Linux) and gets an IP address like any other adapter.

The headline bandwidth numbers (40Gbps for Thunderbolt 3/4, up to 80Gbps for some USB4 v2 implementations) are the raw link capacity, not what IP traffic actually achieves. Real-world throughput for Thunderbolt Networking commonly lands in the 10-20Gbps range once you account for protocol overhead, CPU-bound packet processing, and the fact that this is emulated networking over a bus designed primarily for PCIe tunneling and DisplayPort, not for sustained line-rate Ethernet frames. Still genuinely faster than 10GbE for the cost of a cable you may already own, and dramatically cheaper than anything approaching real 40GbE hardware.

The point-to-point limitation that shapes everything

This is not a networking technology that scales past two directly-connected hosts without real complication. A single Thunderbolt or USB4 port gives you one link to one other port. There’s no equivalent of a switch that fans a Thunderbolt connection out to five nodes the way an Ethernet switch does. Daisy-chaining through a hub device that has Thunderbolt networking support exists on paper for some docks, but it’s inconsistent across vendors and firmware versions, and most homelabbers who’ve tried it report it either doesn’t negotiate the link at all or drops to a fraction of direct point-to-point speed.

The practical pattern that actually works at scale past two nodes is a ring topology: node A connects to node B, node B connects to node C, node C connects back to node A, each machine using two of its Thunderbolt/USB4 ports. With IP forwarding enabled and a routing protocol like OpenFabric or plain static routes layered on top (Proxmox cluster guides built around this pattern typically reach for frr and OpenFabric specifically, since it handles a ring without needing a central route reflector), every node can reach every other node even though no single link touches all three directly. This is the setup most often seen in 3-node Proxmox HA clusters built from mini PCs, where it replaces a dedicated 10GbE switch for the cluster/corosync and storage-replication network while leaving the regular Ethernet port for the LAN-facing bridge.

Beyond three or four nodes in a ring, the hop count and the fact that intermediate nodes are now forwarding someone else’s traffic on top of their own workload starts to matter. This is a small-cluster trick, not a scale-out fabric.

Setting it up on Linux

The mechanics are the same whether you’re using it for a Proxmox cluster network or just a fast link between a workstation and a NAS:

  1. Confirm both ports are actually Thunderbolt, or USB4 with Thunderbolt 3 compatibility, rather than a USB4 port that only negotiates USB3.2 speeds to non-Thunderbolt peripherals. boltctl on Linux lists Thunderbolt domains and will show the device once connected.
  2. Connect the cable. A proper Thunderbolt or USB4-certified cable matters here, a generic USB-C charging cable will often negotiate down to USB2/3 speeds silently rather than failing outright, which is a confusing way to end up with a 480Mbps “40Gbps” link.
  3. Each end gets a thunderbolt-net or similarly-named kernel interface once the Thunderbolt security level allows the connection (set Thunderbolt security to at least “user” in firmware/BIOS if it’s locked to “secure” or “DisplayPort only,” which some vendors ship as the out-of-box default specifically because they don’t expect this use case).
  4. Assign static IPs on both ends in a dedicated subnet, separate from the main LAN, the same way you’d treat a dedicated storage-replication or corosync network on real Ethernet hardware.
  5. For a ring topology across three or more nodes, enable IP forwarding (net.ipv4.ip_forward=1) on each node and bring up a routing daemon, or hand-write static routes if the topology is small and fixed enough that you’re comfortable maintaining them manually.

Proxmox’s own documentation on low-budget, high-speed cluster networking covers this pattern directly for Proxmox clusters, specifically because it lets small mini-PC-based clusters get genuinely fast node-to-node bandwidth for Ceph or ZFS replication traffic without buying a 10/25GbE switch and matching NICs.

Where this fits and where it doesn’t

Worth doing:

  • A small (2-4 node) Proxmox cluster built from mini PCs that already have spare Thunderbolt/USB4 ports, where the goal is a fast private network for corosync, storage replication (pvesr), or Ceph traffic.
  • A direct link between a workstation and a NAS or TrueNAS box for occasional large transfers, when both already have the ports and you don’t want to buy 10GbE NICs and a switch for a link used a few times a week.
  • Anywhere the alternative is “no fast link at all” because the budget or rack space for real 10/25GbE gear isn’t there yet.

Not worth doing:

  • Anything needing more than a handful of nodes, or a topology that changes often. The point-to-point nature makes this brittle to grow and annoying to reconfigure compared to a switch everyone just plugs into.
  • Production traffic that needs to survive a cable getting bumped. A Thunderbolt/USB4 link renegotiating after a disconnect is generally less graceful than an Ethernet link coming back up, and ring topologies in particular can see a single cable pull disrupt routing across the whole ring until it reconverges.
  • Mixing vendors and firmware versions without testing first. Thunderbolt Networking interoperability between different laptop/mini-PC vendors has a real history of being flaky, especially across Thunderbolt 3-to-4 generation boundaries. Test the actual link speed and stability between your specific hardware before committing a cluster design to it, don’t trust the spec sheet.

The honest comparison to just buying real networking gear

A pair of basic 10GbE NICs and the DAC or SFP+ cabling to connect them directly (no switch needed for a two-node link) will usually cost less than people assume, and gives you a dedicated, well-understood Ethernet link with none of the ring-topology routing complexity. Thunderbolt/USB4 networking wins specifically when the ports are already sitting unused on hardware you own, and the setup is temporary, small-scale, or genuinely budget-constrained enough that free ports beat a cable-and-NIC purchase. Once a cluster is expected to live for years and grow past three or four nodes, migrating to real 10GbE or better is the more maintainable long-term answer, and the Thunderbolt setup becomes the bridge that got you there cheaply in the meantime.