Google builds four partner chip supply chain to rival Nvidia

Tuesday 21 April 2026 - 11:40

By: Dakir Madiha

Google builds four partner chip supply chain to rival Nvidia

Alphabet is assembling a diversified custom chip supply chain involving four design partners as it seeks to reduce reliance on Nvidia and strengthen its position in artificial intelligence infrastructure. The effort spans current inference chips to future processors based on 2-nanometer manufacturing expected by late 2027, with production anchored at TSMC.

At the center of the strategy is Google’s next generation of tensor processing units, including Ironwood, a seventh-generation chip designed specifically for inference workloads. The company plans to produce millions of these units this year and is expected to present further details at its upcoming cloud event in Las Vegas. Demand for inference computing has surged as AI deployments scale, making it a primary cost driver in data center operations.

Broadcom plays a key role under a long-term agreement running through 2031 to design and supply high-performance TPU variants and networking components. It is also developing the next training-focused chip, known internally as Sunfish, targeting advanced 2-nanometer fabrication. MediaTek is working on a lower-cost inference-oriented version, codenamed Zebrafish, also planned for the same manufacturing node.

Additional partners are expanding the ecosystem. Marvell Technology is in discussions to develop memory processing units and new inference chips, while Intel has signed a multi-year agreement to supply Xeon processors and custom infrastructure chips for Google’s data center networking layer. All chips are manufactured by TSMC, allowing Google to control architecture while outsourcing fabrication.

The initiative reflects a broader shift toward specialized silicon. Google is prioritizing inference chips, which handle the execution of trained AI models and now account for a growing share of computing demand. Chief scientist Jeff Dean has indicated that separating training and inference workloads at the hardware level improves efficiency and cost control.

Commercial partnerships are reinforcing the strategy. Meta has agreed to rent Google’s TPU infrastructure under a multi-year deal, while Anthropic has secured large-scale compute capacity based on future TPU systems. Google executive Amin Vahdat said additional details on the chip roadmap are expected soon as the company expands its role in the AI hardware market.