Broadcom’s $12K Inference Chips and Google-Marvell Tie-Up Threaten Nvidia
Google and Marvell plan to finalize next-year designs for a memory processing unit to complement Google’s TPU and a separate inference TPU, potentially cutting Google's GPU spending with Nvidia. Broadcom’s custom inference chips cost about $12,000 each with 67% higher energy efficiency versus Nvidia’s $30,000-$40,000 GPUs, posing a competitive threat.
1. Google-Marvell AI Chip Collaboration
Google and Marvell have initiated a joint effort to develop two specialized AI chips, including a memory processing unit designed to work alongside Google’s existing TPU and a separate inference TPU. The companies aim to finalize chip designs by next year, marking a strategic move to diversify Google’s AI hardware sources. Completion of these designs could reduce Google’s dependency on Nvidia GPUs for certain workloads.
2. Broadcom’s Custom Inference Chip Strategy
Broadcom partners directly with hyperscalers to produce tailored AI inference chips costing around $12,000 per unit, compared with $30,000–$40,000 for flagship GPUs. These chips deliver approximately 67% greater energy efficiency, supporting continuous inference workloads at scale. Broadcom’s AI semiconductor revenue grew 106% year-over-year to $8.4 billion in Q1 2026, with an AI backlog exceeding $73 billion, underscoring strong demand.
3. Competitive Implications for Nvidia
Emerging alternatives from Marvell and Broadcom threaten to erode Nvidia’s pricing power and market share in the AI inference segment. Cost and efficiency advantages could prompt hyperscalers to diversify beyond GPU-based solutions, pressuring Nvidia to innovate or adjust pricing strategies. Sustained competition may impact Nvidia’s margins and accelerate development of next-generation AI accelerators.