Japan's Aisin develops new type of AI chip that cuts energy use 98% - Nikkei Asia

Aisin's foray into AI silicon is not a diversification play—it's a paradigm shift in edge-inference energy economics. The 98% reduction almost certainly targets automotive perception-decision workloads via sparse ASIC or neuromorphic architectures, sidestepping the brute-force FLOPS arms race entirely. Technical ripple: If this clears ISO 26262 functional-safety certification, the pricing architecture of Mobileye EyeQ, Horizon J6, and Qualcomm SA8650 in the ADAS stack faces structural pressure. Upstream EDA toolchains (Synopsys, Cadence) must adapt to new instruction sets, and OEM BOM cost models get rewritten. Compliance exposure: Aisin has no foundry. Fabrication almost certainly routes through TSMC (Taiwan, China) or Rapidus. In a tightening geopolitical environment, IP provenance and capacity lock-in are the critical fragility points. Market dynamics: NVIDIA DRIVE's moat is software ecosystem, not silicon. But if "inference per joule" becomes a hard procurement KPI for OEMs, the 3-5x premium logic collapses. Qualcomm and Mobileye will likely ship counter-products within six months. Forward look: Within 18 months, "TOPS" yields to "Joules/inference" as the industry's core metric. Aisin's real threat isn't the chip—it's the proof that a non-semiconductor company can redefine AI compute boundaries, a structural challenge to the fabless business model itself.