INT8 Hits 58x, Voltage Underscaling Saves 36% — Semiconductor Physics Limits Are Being Bypassed by Software in 2026

Why This Matters Now Last week, Tesla started recruiting engineers in Taiwan for "Terafab" - Elon Musk's vision for an in-house AI semiconductor fab. Around the same time, IBM Japan announced development of a 2nm neuromorphic accelerator led from Japan. Read these headlines individually and they're just semiconductor noise. But overlay them with three ArXi papers published this month, and a very different picture emerges. 2026's semiconductor industry is quietly shifting from "push physics harder" to "bypass physics with software." This is my reading, but the evidence isn't thin.