Google is negotiating with Marvell Technology to develop two custom AI chips—a memory processing unit and an inference-optimised TPU—as part of a broader effort to diversify its silicon supply beyond Broadcom and MediaTek.

The proposed TPU would be tailored for AI inference, optimized to run trained models in real-world applications like chatbots, search, and recommendations.

Designs for the memory processing unit are targeted to be finalized and handed off to test production as soon as next year.

Industry sources say the two-chip collaboration aims to improve efficiency and reduce latency for large-scale AI workloads.

Google’s move comes as part of a broader shift from training to inference in AI, increasing demand for purpose-built inference silicon.

Adding Marvell would broaden Google’s supplier base to improve pricing and risk resilience while keeping Broadcom as a core partner under an extended deal through 2031.

Ironwood, Google’s seventh-generation TPU, already targets inference with substantial performance gains and scalability for large deployments.

Marvell brings a strong data-center footprint and growing custom silicon business, with design wins at major cloud giants, positioning it as a meaningful partner in Google's AI hardware strategy.

The collaboration aims to make TPUs a stronger alternative to Nvidia GPUs and to boost Google Cloud’s AI revenue by expanding its cloud-based inference offerings.

Google’s aggressive 2026 capex, potentially up to $185 billion, and rapid growth in Google Cloud revenues underscore the scale of AI-infrastructure investments driving hardware collaborations.

The information on talks comes from secondary sources, with no official confirmation from Google or Marvell at this time.

Despite exploring partnerships, Broadcom remains a core supplier with continued collaboration through 2031, indicating a gradual diversification rather than a full break.