Model-specific inference silicon got its first fully open design-to-fabrication path. Ankhdjet is an open-source compiler that lowers a HuggingFace BitNet b1.58 checkpoint (1.58 bits per weight) to a via-mask program on the SKY130 open PDK, verified end-to-end with open tools: KLayout DRC zero, netgen LVS zero, clean timing. The first compute-in-ROM macro built with this flow is already submitted to the TinyTapeout/ChipFoundry SKY130 shuttle, with silicon expected in 2027.
The commercial baseline here is Taalas HC1, which validated hardwired-weight inference on an advanced node behind closed tooling. Ankhdjet does not close that performance gap -- SKY130 at 130nm is not 5nm -- but it removes the toolchain constraint entirely. Two weight matrices were pushed through an identical flow with only the mask differing. That reproducibility is the point: model-specific silicon becomes a repeatable flow, not a one-off.
The missing variable is the process node. Teams building inference hardware on commercial nodes can now study and fork a clean open reference with measured signoff data. The 2027 shuttle result will be the first production data point on whether compute-in-ROM is practical below the advanced-node tier, and on how much of the Ankhdjet flow survives contact with a non-SKY130 PDK. If it holds, the closed-tooling moat around model-specific silicon has a working alternative.