Google Strikes $12 Billion Custom AI Silicon and Optical Interconnect Deal with Marvell

On August 19, 2026, Alphabet's Google finalized a landmark $12 billion multi-year custom semiconductor and optical networking agreement with Marvell Technology. The partnership aims to co-develop next-generation custom AI application-specific integrated circuits (ASICs) and high-throughput optical interconnects, strengthening Google's in-house TPU and Axion datacenter infrastructure to support massive cluster scaling for frontier Gemini model training.

Scaling Custom Silicon and Optical Networking Infrastructure

As foundational AI models scale past trillions of parameters, the primary barrier in datacenter performance has shifted from raw compute density to inter-rack networking bandwidth. By partnering with Marvell, Google secures dedicated silicon design capacity and proprietary electro-optics technology to connect tens of thousands of custom TPU accelerators across hyperscale data campuses with near-zero latency and reduced power dissipation.

Key Agreement Objectives & Technical Scope

• Custom ASIC Accelerator Co-Design: Marvell will provide specialized IP and silicon design services for upcoming generations of Google's custom AI accelerators.
• High-Speed Optical Interconnects: Deploys next-generation optical DSPs and silicon photonics interconnects delivering multi-terabit-per-second rack-to-rack bandwidth.
• Optimized for Trillion-Parameter Training: Built specifically to eliminate networking bottlenecks during distributed frontier model pre-training and agent inference.
• Diversifying Datacenter Silicon Supply: Reduces long-term reliance on commercial merchant GPU silicon by accelerating Google's vertically integrated infrastructure roadmap.

Strategic Reshaping of the Hyperscale Silicon Ecosystem

The $12 billion deal represents one of the largest custom semiconductor commitments to date, underscoring how major cloud providers are aggressively investing in custom silicon architectures and advanced networking to maintain competitive cost and performance moats in generative AI.

Deal Overview