Google in Talks with Marvell for New Custom AI Inference Chips 2026 | TPU Expansion

Google in Talks with Marvell to Develop New AI Inference Chips: A Major Step in Cost Reduction and Efficiency

April 19, 2026 — In a significant development in the AI hardware landscape, Alphabet’s Google is in advanced negotiations with Marvell Technology to co-develop two new custom chips specifically optimized for AI inference — the process of running trained AI models to generate responses for real-world applications.

This potential partnership highlights Google’s continued push to lower the massive operational costs of AI while reducing dependency on any single supplier.

What Chips Are Google and Marvell Planning to Build?

According to reports, the discussions focus on two key chips:

1. A Memory Processing Unit (MPU) designed to complement Google’s existing Tensor Processing Units (TPUs) by improving memory bandwidth and overall efficiency.
2. A new inference-optimized TPU built from the ground up for running large AI models more efficiently and at lower cost.

Marvell would serve as a design-services partner, playing a similar role to MediaTek in Google’s recent Ironwood TPU project. The companies are aiming to finalize the design of the Memory Processing Unit as early as 2027, followed by test production.

Why AI Inference Is the New Battleground

While training large language models gets most of the attention, inference now accounts for the majority of AI compute costs in production environments. Every search query, Gemini response, recommendation, or chatbot interaction relies on inference.

Google’s move to develop more specialized inference hardware aims to deliver:

1. Significantly lower cost per query
2. Better power efficiency (critical amid growing energy constraints in data centers)
3. Higher throughput to serve billions of users
4. Greater independence from third-party GPU suppliers

Google’s TPU Strategy and Supply Chain Diversification

Google has been a leader in custom AI silicon since introducing its first TPU in 2015. The company has long partnered with Broadcom for TPU development and continues to strengthen that relationship.

Bringing Marvell into the mix represents a clear strategy of supply chain diversification. By working with multiple design partners (Broadcom, MediaTek, and now potentially Marvell), Google can accelerate innovation, reduce risk, and negotiate better terms across its AI infrastructure buildout.

Marvell’s Rising Role in the AI Chip Market

Marvell Technology has emerged as a major player in custom AI and data center semiconductors. The company already powers high-performance networking, storage, and custom silicon solutions for several hyperscalers. This potential collaboration with Google further strengthens Marvell’s position in the booming custom AI chip sector.

Strategic Impact on the AI Industry

This development is part of a broader trend among big tech companies:

1. Reducing heavy reliance on Nvidia GPUs
2. Investing heavily in custom silicon for better performance-per-dollar
3. Preparing for the explosive growth of AI inference workloads as agentic AI and real-time applications scale

If successful, the Google-Marvell chips could help make advanced AI more affordable to run at scale, benefiting Google Cloud customers and the entire ecosystem.

What Happens Next?

The talks are still ongoing and no final agreement has been signed. However, the momentum suggests progress could be made relatively quickly, with initial design work potentially starting later in 2026.

Conclusion

Google’s negotiations with Marvell to build new AI inference chips signal the company’s determination to control costs and maintain technological leadership in the AI era. As inference becomes the dominant driver of AI spending, custom silicon optimized for efficiency will be a decisive competitive advantage.

This partnership — if finalized — could mark another important milestone in the shift toward specialized AI hardware tailored for real-world deployment.

What’s your take? Will more custom inference chips from Google and partners like Marvell challenge Nvidia’s dominance even further? Share your thoughts in the comments.

Subscribe for weekly AI hardware and industry updates delivered straight to your inbox.

FAQs

What is Google developing with Marvell?

Google is in talks with Marvell to create two new chips: a Memory Processing Unit (MPU) to enhance TPUs and a new TPU optimized specifically for AI inference workloads.

When could these new chips be ready?

Design of the Memory Processing Unit could be finalized as early as 2027, with test production to follow. Full commercial timelines are not yet confirmed.

Why is Google focusing on inference chips?

Inference (running AI models) represents the largest portion of ongoing AI operational costs. More efficient chips will dramatically reduce expenses and power consumption.

Does this replace Google’s partnership with Broadcom?

No. This appears to be an expansion and diversification strategy. Google continues to work closely with Broadcom on TPU development.

How does this affect Marvell?

The potential deal strengthens Marvell’s growing reputation as a key player in custom AI silicon design for hyperscalers.