Google assembles four-partner chip supply chain with Broadcom, MediaTek, Marvell to challenge Nvidia in inference

Google is building the AI industry's most diversified custom chip supply chain with four design partners: Broadcom (TPU v8 training chip 'Sunfish'), MediaTek (inference chip 'Zebrafish' at 20-30% lower cost), Marvell (in talks for a memory processing unit and inference TPU), and Intel (Xeon and infrastructure processors). All chips are fabricated at TSMC. The current Ironwood TPU (7th-gen, inference-focused) is shipping in the millions, with Anthropic committed to up to one million units. The strategy targets AI inference economics — where cost per query matters at Google's scale — as the primary battleground against Nvidia. Google projects 4.3 million TPU shipments in 2026, scaling to 35 million by 2028, while the broader custom ASIC market is projected to grow 45% in 2026 versus 16% for GPUs.