Google Expands AI Chip Partnership with Marvell, Diversifying Custom Silicon Supply Chain

Google is in talks with Marvell Technology to develop two new AI chips, a memory processing unit (MPU) and an inference-optimized Tensor Processing Unit (TPU). These discussions come days after Broadcom secured a long-term agreement to supply TPUs and networking components through 2031. The move underscores Google’s strategic shift toward inference as the dominant compute cost in its AI operations.

Why it Matters

Google's decision to engage Marvell Technology reflects a broader industry trend where inference is becoming the primary driver of compute costs. Training large AI models, while resource-intensive, is a one-time event that can take weeks or months. In contrast, inference involves continuously serving user queries and scales with demand, making it a more significant ongoing expense.

According to Google, its seventh-generation TPU, Ironwood, launched this month as "the first Google TPU for the age of inference." It offers ten times the peak performance of the TPU v5p and can scale to 9,216 liquid-cooled chips in a superpod, producing 42.5 FP8 exaflops. The company plans to manufacture millions of Ironwood units this year, highlighting the growing importance of inference in its AI infrastructure.

Key Risks

While diversifying its supply chain can mitigate risks associated with single-vendor dependency, Google faces several challenges:

1. Complexity in Integration: Adding a third design partner increases the complexity of managing and integrating multiple chip designs.
2. Market Competition: The custom Application-Specific Integrated Circuit (ASIC) market is projected to grow by 45% in 2026 and reach $118 billion by 2033. Competitors like NVIDIA, Intel, and AMD are also investing heavily in AI-specific chips, intensifying competition.
3. Cost Management: Balancing the cost of developing and deploying multiple chip designs while maintaining efficiency is crucial. Marvell’s role as a design-services provider for cost-optimized variants could help manage these costs.

The Opportunity

The partnership with Marvell presents several opportunities for Google:

1. Enhanced Flexibility: By working with Marvell, Google can develop chips tailored to specific workload profiles or cost points, complementing its existing Ironwood TPU.
2. Competitive Advantage: Purpose-built inference silicon offers better cost and efficiency compared to general-purpose GPUs, providing a competitive edge in the AI market.
3. Scalability: As AI products reach hundreds of millions of users, having multiple design partners ensures Google can scale its operations efficiently.

Strategic Context

Google’s current custom silicon supply chain includes:

• Broadcom: Primary partner for high-performance chip variants and networking components through 2031.
• MediaTek: Design services for cost-optimized "e" variants at 20 to 30% lower cost.
• TSMC: Fabrication of the chips.

The addition of Marvell as a third design partner is part of Google’s diversification strategy, not a replacement for existing partners. This approach allows Google to leverage the strengths of each partner while maintaining flexibility and reducing dependency on any single vendor.

Conclusion

Google's discussions with Marvell Technology highlight its commitment to staying at the forefront of AI innovation by optimizing for inference. As the custom ASIC market continues to grow, the company’s diversified supply chain strategy positions it well to meet the evolving demands of AI compute.