Fabric
moves a network-gear decision from vendor mythology to honest used-enterprise — and connects your wagons into one computer.
Get your wagon first, then connect it. Box by FI teaches the wagon.
Recommended tier
The three Fabric tiers
Three operational profiles. Pick by Box count, speed target, and where it lives — not by tier number.
Why this shape
About the vendor mythology
The standard advice for "what network gear should I buy" is either "get a Netgear from Best Buy" or "you need Cisco." Both are wrong. Consumer routers can't enforce VLANs, can't be configured off-band, and don't have the QoS to handle ZFS replication. Cisco is enterprise-priced because Cisco has a salesforce, not because switching is harder than what used Mellanox does. Used enterprise > new SMB at the same price point — same argument as Box and Bucket, applied layer-for-layer.
The transceiver tax
A Cisco SFP+ transceiver from Cisco: $500. The same physical part from FS.com, unbranded: $15. Vendor lock-in on transceivers is one of the most reliable cost extractors in enterprise gear. Mellanox switches accept FS.com unbranded modules without firmware games; this is a real product advantage and a structural reason to prefer Mellanox for homelab use. FS.com is the doctrine vendor for transceivers — not because of any affiliate relationship, but because they manufacture and sell the right part at the right price. If FS.com later operates a partner program, the link earns commission. If not, the recommendation stays. Doctrine first.
DAC cables vs. optical
Direct-attach copper (DAC) is a passive twinax cable with SFP transceivers pre-attached to both ends. Under 5m, DAC is cheaper, simpler, and more reliable than optical. A FS.com 10G SFP+ DAC at $10 does the work of two FS.com optical transceivers ($30) plus fiber. Fabric recommends DAC by default for short runs; optical wins at >5m or when future upgrades beyond the current speed are likely.
Layer 3 at the edge isn't optional
Once you expose services from your Box to the internet, or run multiple tenants on one Box, VLANs and routing are not optional. VLANs separate traffic classes in hardware — the same way ZFS separates datasets at the storage layer. Without them, a compromised service has a straight shot to every other service on the same L2 domain. Managed enterprise switches do VLANs at zero cost (configuration only); consumer routers cannot. This is why the Closet tier starts at a Mikrotik managed switch, not at a Netgear unmanaged.
10G is the new 1G; 25G is approachable
A 10G link is 1.25 GB/s — enough for ZFS replication at meaningful speed, NFS without bottleneck, most containerized services. In 2026, used 10G equipment on eBay is at the price 1G equipment was at in 2020. 10G is the doctrine default. 25G is approachable with used ConnectX-4/5 NICs and a used SN2010 switch; total BOM for two 25G-connected Boxes is under $1,500. 100G is available used if you have a rack.
The fabric IS the rail network
Drives are buckets, servers are wagons, and the network is the rail network connecting the yards. Two Boxes connected by a properly configured managed switch are not two computers — they're one computer with a hardware-enforced topology. The switch enforces VLAN segmentation, MAC allowlist, port-level QoS. The build decided the topology; the switch makes the topology real. Consumer routers can't do this; that's the structural reason they're the wrong answer for sovereign use.
Cold spares for network gear
Same availability-decay doctrine. Mellanox ConnectX-3 NICs are abundant on eBay today; supply will thin over 3-5 years. Buy your spare NIC when you buy your main NIC. For switches at production-scale: one spare unit buys you the ability to swap and continue while diagnosing. Fan replacement is a common failure mode in used enterprise switches; stock 2-3 fans per switch.
When the network breaks
Power-cycle first — resolves ~60% of transient issues. If the switch doesn't come back, swap in the cold spare (configured identically via USB config; prepare that ahead of time). The Box reconnects automatically once the switch is back. Downtime: minutes, not hours. This is only possible if the spare's USB config matches the production switch.
Fabric vs. the alternatives
| Option | Cost | Speed ceiling | Sovereignty | Where it fits |
|---|---|---|---|---|
| Consumer router + unmanaged switch | $100–300 | 1G (real: ~400Mbps) | None | Family with no VLANs |
| New SMB managed (Ubiquiti, Netgear M-series) | $300–1,000 | 1–10G | Medium | Prosumer; reasonable Closet alternative |
| Cisco Catalyst / Nexus (new) | $5,000–50,000+ | 10–400G | Full | Enterprise with SmartNet budget; wrong answer for homelab |
| Fabric by FI (used enterprise) | $300–1,800 | 10–100G | Full | The sovereign homelab operator |
| Cloud SD-WAN (Tailscale, Cloudflare Tunnels) | $0–$100/mo | Variable | None | If your workload tolerates vendor dependency |
| New prosumer (Mikrotik CRS326 new) | $200–300 | 1–10G | Full | Closet alternative; no used-market arbitrage |
A note on cloud SD-WAN: Tailscale and Cloudflare Tunnels are overlay networks that add latency and vendor dependency. They're not alternatives to physical networking for the Box use case; they're a different layer.