Google could partner with MediaTek for next gen AI chips: Report

Summary

• Google may switch TPU partners from Broadcom to MediaTek for cost optimization.
• MediaTek might handle I/O modules in partnership, with Google focusing on more in-house TPU design.
• Changing partners could reduce Google's Broadcom dependency in the AI market.

AI models are computation-intensive, requiring powerful compute units which Google calls Tensor Processing Units (TPUs). Its Gemini chatbot and other AI tools are available to most users for free, and only a small subset of AI features are paywalled, limiting access to Workspace customers. So, it is understandable if the company wishes to optimize the costs associated with acquiring the processing horsepower. With AI and demand for the tech showing no signs of letting up, a new report based on insider reports states that Google may consider a new TPU production partner in chipmaker MediaTek, after favoring Broadcom for the past nine years.

Related

What are Google Cloud TPUs?

Google's aiming for the cloud-based AI creation sector: Google Cloud TPUs are helping it win

Posts

Since 2015, Google has used Broadcom TPUs for AI. While the first generation hardware has been retired, TPUs from 2017 through 2021 are currently serving cloud customers. Subsequent generations of the chip have enabled Google engineering efforts while also serving cloud customers. 2025's will soon be available and according to a report from The Information, it could be the last generation exclusively handled by Broadcom. The tech partner co-designs Google's AI chips based on blueprints the latter supplies, and Taiwanese foundry TSMC handles production, a Broadcom staffer reportedly said.

An unnamed source at TSMC added that MediaTek charges Google less per chip and also has strong ties with the same foundry. Retaining a cost advantage would help Google stay competitive in a market where rivals like OpenAI have been buying NVIDIA's chips to power AI applications like ChatGPT. That doesn't mean Google isn't an NVIDIA customer. It placed an order for over $10 billion worth of Blackwell chips. Separately, the company spent an estimated $6-9 billion on TPUs last year alone.

MediaTek will only handle I/O modules

Nothing set in stone yet

Source: Google Blog

An important detail outlined in the report is that MediaTek will only handle the design of the I/O module designs in the planned partnership, while Google will design most of the TPU in-house. The new partner will also liaise with TSMC and be responsible for quality control. This is a shift for the Search titan that had depended on Broadcom until now for all the heavy lifting. Moving more operations in-house while using the same manufacturer in TSMC could prove beneficial for Google in the long term.

Google seems to be doubling down on in-house TPU design since it is recruiting chip design engineers in Taiwan, including former TSMC staffers. However, the company reportedly produces two different TPUs — one for training new models and another to power services like Search, YouTube, and Gemini. It's still unclear if the company's internal strategy will shift to making a single chip if MediaTek is onboarded as a partner for the 2026 production.

While the impact on consumer-facing AI applications will be none to minimal, we suspect Google could reap the benefits of reducing its Broadcom dependency in the coming years as demand for AI compute modules continues to soar. That said, Broadcom has other big-ticket clients in Meta Platforms, Arm Holdings, Apple, and OpenAI, so it isn't going under anytime, even though Google is the company's biggest revenue source from the sale of AI chips.

Embed Gemini here