A 3×3×3 kernel holds 27 weights, and one application of it produces one number: nine products per channel, all twenty-seven summed together. Step through the nine window positions and read the arithmetic — the three channel subtotals never survive on their own. Then switch to two filters: the second filter has its own 27 weights and writes its own map, and the two maps stack into a two-channel output — the input shape of the next layer's kernels.