Fi-Wi is building infrastructure: an inside fiber plant connects many small 802.11 radio heads throughout a building to one concentrator that makes the building-wide scheduling and forwarding decisions. Each head uses high-volume consumer Wi-Fi silicon and its own fiber. The concentrator runs one wireless forwarding plane, sees the radio environment as a system, and decides centrally which radio talks to which client and when.
Wired networks already have forwarding planes for routing and Ethernet access. Wireless networks have centralized management and sometimes tunneling, but the forwarding and scheduling entities remain autonomous 802.11 MACs in access points that contend independently. Fi-Wi adds the third: a scheduled wireless forwarding plane whose unit of work is the air interface itself.
In the vocabulary of wired networking, the concentrator is a shared-memory switch fabric whose egress ports are radios, and what makes it unlike a fabric card is that those ports are contended rather than deterministic. Proximity improves capacity and lowers required transmit power, so abundant radio heads put a radio near each client. Their density only works with coordination: one scheduler can make each decision once with knowledge of every link, queue, and deadline instead of allowing nearby APs to make overlapping decisions independently. The concentrator also consolidates telemetry from the whole building's RF environment into a single vantage point, so the network can be measured, diagnosed, and tuned as one system. Umber is measurement-first; our founder created and maintains iperf2, and Fi-Wi is built to expose its numbers.
The design-partner system is one workstation-class concentrator with six identical 16-port adapter cards serving 96 low-power radio heads with internal antennas. Data and timing travel over fiber; head power can travel alongside it in a hybrid cable, so the physical deployment is designed to fit established structured-cabling practices. Ninety-six is the target capacity set by slot, adapter, lane, and compute geometry and to be confirmed by aggregate-resource measurements. Smaller systems use the same concentrator and software with fewer fibers populated; the 8-head evaluation system runs the same forwarding plane, scheduler, and telemetry as the system it grows into.
Failure domains are explicit. A head that loses power or fiber, or whose firmware misbehaves, is required to cost that one head, isolated by the same class of hardware protection that confines device access in data-center virtualization, with the other 95 undisturbed and the head rejoining alone when restored; an adapter-card failure costs that card's sixteen heads, and the card is the field-replaceable unit; the concentrator is the 96-head failure domain. Concentrator availability is therefore a first-class design input. Fronthaul failover to a standby concentrator is under evaluation, and service-level redundancy can instead use two complete overlapping systems. The partner's tolerance for losing one head, sixteen heads, or a concentrator for some interval determines which model is appropriate. The wired uplink is not yet in that list. One connection carries all traffic between the building and the outside network, so losing it takes external connectivity from every client while service between clients inside the building continues. Whether that makes the uplink a failure domain in its own right, and what redundancy it warrants, follows from the availability cost measured against your requirement, the same way the concentrator question does. It is named here rather than left for you to discover.
Twenty-four heads have been brought up on the development system. The path to 96 is defined by roughly twenty measurements and qualification gates covering isolation, fault containment, timing and synchronization, scale, recovery, and availability. Platform-independent pass conditions are specified now; deployment-specific thresholds, including availability, come from the design partner. Some gates are already established, and the remaining results are deliverables rather than internal artifacts.
The proposition. You are not being asked to trust an architectural claim. Your building is instrumented, your requirements become acceptance conditions, and you see the numbers.
A companion architecture note provides the full interface, isolation, memory, timing, and verification details under NDA.
A design partnership has three stages: an instrumented 8-head evaluation in a real area of the building, measuring the environment before and after with the data shared openly; definition of traffic, availability, and operational requirements as pass conditions for the 96-head system; and the 96-head deployment with its acceptance measurements delivered at handover. The operator supplies the real requirements that turn the architecture into a product.