The mechanism
An ultrafast (femtosecond-scale) optical pulse dumps energy into a ferrimagnetic film and drives the two magnetic sublattices out of equilibrium at different rates. In the right material (canonically Co/Gd and GdFeCo), the magnetisation reverses deterministically from a single pulse. The switching is helicity-independent in these ferrimagnets, meaning the pulse doesn’t need circular polarisation of a particular handedness, which matters a lot for building a real device.
Speed is the headline: this is a picosecond-scale write, which is where the “~1,000x faster than HBM cell activation” class of claim comes from. It is non-volatile and low-energy, because you are not holding a charge, you are flipping a domain and walking away.
Where it actually is (Jul 2026)
The strongest integrated result is single-pulse AOS of Co/Gd on a SiN waveguide at 90% contrast (arXiv 2511.02440, Nov 2025, TU Eindhoven / LioniX). Two caveats define the frontier:
- Readout is electrical (Hall), not optical. The write is integrated; the read is not. This is the class-wide gap.
- Switching goes stochastic above ~500 nm device width. Deterministic behaviour lives below that, which is both the constraint and, awkwardly, exactly where the density argument needs it to work.
Why it isn’t sufficient on its own
AOS solves the write. It does not solve addressing (how one diffraction-limited port serves many nm-scale domains) or readout. A company that has AOS working has the least-hard of the three problems solved. Worth remembering when a founder leads with switching speed: the binding constraint is downstream of the thing being demoed.