Google Unveils Custom Chips to Challenge Nvidia's Dominance

Google unveiled a multi-partner chip supply chain strategy designed to challenge NVDA Nvidia's dominance in AI inference, pairing Broadcom's "Sunfish" training chip with MediaTek's "Zebrafish" inference chip at 20-30% lower cost than competing offerings. The Ironwood TPU, Google's seventh-generation custom silicon, delivers four times the performance of its predecessor, offering 192 gigabytes of HBM3E memory per chip and scaling to 42.5 FP8 exaflops across a 9,216-chip superpod.

Anthropic has committed to deploying up to one million Google TPUs, while GOOGL projects total TPU shipments of 4.3 million units in 2026, scaling to more than 35 million by 2028. This demand trajectory positions Google's silicon program as the most direct institutional challenge yet to Nvidia's stranglehold on AI training and inference infrastructure, with Amazon, Meta, Microsoft, and OpenAI all accelerating similar custom silicon programs to reduce per-token compute costs.

The move marks a structural inflection in hyperscaler chip strategy. Google's advantage lies in vertical integration — designing chips alongside its own AI workloads — giving it optimization advantages that third-party hardware buyers cannot replicate. By splitting the next-generation roadmap explicitly between Broadcom for training and MediaTek for inference, GOOGL is building a cost-optimized, redundant supply chain that hedges against single-vendor concentration risk while targeting Nvidia's highest-margin product lines.