For thirty years, the promise of optical computing has been the same: replace electrons with photons, and the heat problem disappears. For thirty years, the engineering has been the obstacle. Light is superb at carrying information and terrible at being told what to do.
That is finally changing. A new class of chip built from silicon photonics and thin-film lithium niobate can now switch light at speeds that were laboratory curiosities five years ago. Three startups and two university labs have demonstrated matrix multiplication — the core operation of modern AI — entirely in the optical domain.
What changed
The breakthrough was not a single invention but a manufacturing one. Photonic components can now be fabricated on standard CMOS lines, which means they can be made in the same factories that produce ordinary chips. The first commercial parts are small, power-hungry, and narrowly useful. They are also shipping.
The question is no longer whether light can compute. It is whether the software can be rewritten to care.
The hard part is now the stack above the hardware. Every machine-learning framework assumes that moving data costs almost nothing and that arithmetic costs a little. In a photonic system, the economics invert. The companies that win may not be the ones with the fastest chip, but the ones that teach developers to think in light.
