unlike huge moat for NVDA GPU/CUDA, without SW stack, the moat for ASIC is narrow, MRVL, QCOM, MTK can catch up easily