We Build the Silicon Other People’s Models Run On

NeuralPeak was founded in 2021 by three engineers who had each spent a decade on opposite sides of the same problem — one in physical design, one in memory systems, one in compilers. The observation was not new: neural inference at the edge is memory-bound, and general-purpose silicon spends most of its die on the wrong thing.

What the three of them shared was a conviction that the right answer was a purpose-built part with an unusual amount of SRAM, paired with a compiler that treated the memory hierarchy as the resource to be scheduled — not an afterthought to the arithmetic.

We did not start with a pitch deck. We started with a floorplan.

Five years later that floorplan is NP-1 Apex. The company remains small, the toolchain remains ours, and the work remains the same: make on-device inference faster by moving less data, not by adding more transistors.