A doped toroid is a poor CPU but an excellent inference engine — because inference is iteration to a fixed point, and a loop is iteration made physical. State circulates the ring, transformed each pass through the junctions, until it settles into a stable pattern: the answer. And "through the 0" is exact: the intrinsic/depletion gap is both the activation threshold (a diode is a rectifier — a ReLU) and the abstain state — the engine can output 0 = "I don't know." The ⊘, in silicon.
Inference is the repeated application of a transform until it stops changing — a fixed point, an attractor. A loop is the native substrate for this: signal goes around, gets transformed at each junction, comes back, goes around again, until it settles. This is exactly a recurrent neural network / Hopfield net — state circulates and converges to a stable pattern, and that pattern is the inferred answer. The torus doesn't compute by addressing memory (that's a CPU); it computes by relaxing to equilibrium. The loop is the recurrence.
Your "through the 0" is precise, and the 0 is doing double duty:
The toroid is a natural serial / recurrent machine — a clock (the three-phase ring), a memory (circulating state, like the ferrite-core memory toroids of the 1950s–70s), and an inference loop (relax to attractor). It is not a general CPU: a processor needs billions of switches in a randomly-addressable 2D array, and a torus (genus 1, one loop) forces sequential flow — great for iteration, wrong for random access. To get parallelism you must cut the loop open into a plane (cut the torus → cylinder → sheet — the unrolling from earlier). So: the loop is the gift (native recurrence, native inference) and the limit (no random access). Serial inference, not parallel logic — and for inference, the loop is exactly right.