Create and destroy population inversions to make lasing gain positive or negative, thus providing high-gain amplification. Pipeline bits through optical fiber with length comparable to fiber diameter (≈8μm or 30fs, thus enabling switching speeds in the tens of THz). Doped fiber sections pumped from the side by another laser produce an AND gate in the obvious interpretation, but you can do better still with Manchester encoding: the stimulated emission from one pulse leaves behind a lack of population inversion which will attenuate following pulses for a while. Wideband-excitable materials like trivalent-erbium-doped fiber amplifiers may permit wavelength-division multiplexing in computation; the inhomogeneous-broadening-induced effect known as “spectral hole burning”, usually considered a nuisance, thus provides a way for pulses at one wavelength to suppress pulses at nearby wavelengths. Dynamic memory can also be constructed by using a metastable population inversion to store each bit, as long as the spontaneous-emission half-life is long enough to permit a reasonably low-frequency refresh cycle.
Taking full advantage of these effects will require micron-precision optical-path-length matching.
Increasing characteristic operating frequencies with these approaches past the tens of THz requires stimulated emission or at least induced emission at shorter wavelengths. Induced or stimulated emission from nuclear isomers provides a plausible route to six to nine orders of magnitude faster operation; note that these timescales are so short that even deep-sub-nanosecond-half-life nuclear isomers could be useful, dramatically broadening the possible range of possibly useful substrate materials, which should reduce potential conflicts between this computational technology and proliferation concerns. Aside from the present difficulties of inducing metastable nuclear state decay, this approach has another serious difficulty: controlling the flow of high-energy gamma rays is more difficult than simply using optical fibers.
Aside from the possibility of induced emission from majority-metastable nucleus populations, neutron-induced fission has been known since the 1930s and a practical energy source since the 1940s, and produces prompt neutrons within 10fs of the fission, and they can be channeled to some extent by neutron reflectors. However, it is not entirely clear to me how to use this effect for computation, since I don’t see how to combine two different neutron signals in anything other than a sort of logical OR. Moreover, aside from proliferation concerns, computational devices built using this approach would need the fissile nuclei physically replaced after firing, a process that will surely take at least microseconds if not entire seconds; so although a nuclear prompt-fission chain-reaction logic device could compute with propagation delays in the femtosecond range (if we figure out how to combine signals usefully), its clock rate would be very low.