Why does Ethernet use magnetics, and can two PHYs on one board be linked without them?

The transformer's main job is isolation. Twisted-pair Ethernet ports are specified to withstand 1500 V RMS between the cable side and the equipment, because a cable can run up to 100 m between devices powered from different outlets, buildings or earthing systems, and their ground potentials may differ or carry fault and surge voltages. The transformer passes the differential signal and blocks DC and low-frequency common-mode voltage, so no ground current flows along the cable.

It also improves signal quality and emissions. The transformer and the common-mode choke that is usually in the same package let the differential signal through while attenuating common-mode noise in both directions, which helps with both immunity and radiated-emission tests.

Centre taps are where designs go wrong, because the correct connection depends on the PHY's driver type. PHYs with current-mode line drivers need the PHY-side centre tap tied to a supply rail, since the driver sinks current through the windings. PHYs with voltage-mode drivers want the centre tap connected only to a capacitor to ground, and tying it to a supply is wrong. The datasheet's reference schematic is the authority; do not copy the arrangement from a different PHY.

On the cable side, the usual arrangement is the so-called Bob Smith termination: each pair's centre tap goes through 75 Ω to a common point, which connects to chassis ground through a high-voltage capacitor, typically 1 nF rated at 2 kV. It reduces common-mode emissions from the cable while keeping the isolation barrier.

For two PHYs on the same board with a common ground, the isolation is not needed, and capacitive coupling is an accepted technique. Series capacitors in each line of each pair block the DC bias, which differs between PHYs, and each side gets the biasing or termination that its driver type requires. PHY vendors publish application notes with the exact networks for their parts, and these differ between current-mode and voltage-mode drivers, so follow the note for both of your devices and keep the traces short, routed as 100 Ω differential pairs.

This is only for links inside one enclosure. Anything that leaves the box on a cable needs the transformer, both for safety and to meet the standard.

Before doing that, check whether you need PHYs on that link at all. A microcontroller MAC and a switch chip can usually be connected directly through their MII or RMII ports, MAC to MAC, with no line-level signalling in between. Many embedded switch ICs have a port intended for exactly this, and it saves the power of two PHYs as well as the magnetics.

For RMII back to back, cross transmit and receive data, provide the shared 50 MHz reference to both, and configure both ends for a fixed 100 Mbit/s full duplex, because there is no auto-negotiation without PHYs. The switch's datasheet normally shows this configuration.

If you keep the magnetics for the external ports, two practical notes. Jacks with integrated magnetics are not interchangeable between manufacturers even when the footprint matches: pin assignment, centre-tap arrangement and LED polarity vary, so check the internal schematic in the jack's datasheet against your PHY's needs. And match the transformer to the speed: 10/100 parts have two transformer channels, while gigabit needs four, since 1000BASE-T uses all four pairs in both directions at once.

In the layout, keep the area under the transformer and between the transformer and the RJ45 free of ground and power planes, so the isolation gap is not bridged by plane capacitance, and keep the spacing that the 1500 V rating implies.