FS2 is developing three-dimensional AI chips out of MIT, integrating logic and memory into a single seamless stack rather than placing them side by side on a board. The constraint this addresses is the defining one in modern AI hardware: moving data between memory and compute costs far more time and energy than the arithmetic itself, so accelerators spend much of their power budget shuttling weights rather than multiplying them. Stacking memory directly against logic shortens that path dramatically.

Monolithic three-dimensional integration is difficult for well-understood reasons, involving thermal management, yield, and process compatibility between layers that were historically fabricated separately. If those are solved, the gain is not incremental but architectural, applying to every workload that is memory-bandwidth bound. FS2 is an AI infrastructure position at the deepest layer of the stack, with a long horizon and a payoff measured against the physical limits that currently cap accelerator efficiency. Research of this kind takes years to reach production silicon, but the ceiling it addresses is one every accelerator vendor eventually confronts.