Google builds chip partners

- Google is assembling a four-partner chip supply chain to reduce reliance on Nvidia for AI inference workloads. - The partnership list includes Broadcom, MediaTek, Marvell and TSMC as key collaborators. - That approach turns supplier selection and dependency management into enterprise-strategy problems requiring scenario planning and operating-model design (thenextweb.com).

Google is widening the roster behind its artificial-intelligence chips as it tries to cut how much of its AI serving work depends on Nvidia. (thenextweb.com) The company now has Broadcom, MediaTek, Marvell and Taiwan Semiconductor Manufacturing Co. tied to different parts of its Tensor Processing Unit pipeline, according to recent reports and Google’s own product disclosures. Broadcom remains a long-term design and supply partner through 2031, while Google is also working with MediaTek on seventh-generation TPU work and is in talks with Marvell on two inference chips. (aol.com) (trendforce.com) (theinformation.com) One of the Marvell projects is a memory processing unit that would sit next to a TPU, and the other is a TPU built specifically for inference, the stage when a trained model answers prompts instead of learning from data. Google said in April 2025 that Ironwood, its seventh-generation TPU, was its first chip designed specifically for inference. (theinformation.com) (blog.google) Inference has become the expensive part of the AI business because companies must keep models running for every user query, image request or software call, not just train them once. Google’s push toward dedicated inference silicon follows a year in which it started positioning TPUs more directly for that workload inside Google Cloud. (blog.google) (docs.cloud.google.com) That changes the chip contest from a race for the fastest training hardware into a race for cost, power use and supply assurance across millions of repeated model responses. Broadcom’s filing last week showed Google is locking in not only future TPU generations but also networking components for next-generation AI racks through 2031. (datacenterdynamics.com) Google has built TPUs for more than a decade, but the supplier map has shifted as AI demand surged. Reuters reported in March 2025 that Google was preparing to partner with MediaTek on its next TPU, with TSMC expected to manufacture the chip. (finance.yahoo.com) TSMC’s role is the factory piece of the plan: it fabricates the chips that Google and its design partners specify. Reports on Google’s TPU roadmap have consistently pointed to TSMC as the manufacturer for recent and upcoming generations, including Ironwood. (trendforce.com) (blog.google) The Marvell talks also show Google is not replacing Broadcom so much as adding options around it. CNBC reported on April 20 that Marvell shares rose on the report even as Broadcom shares slipped, with investors reading the move as an expansion of Google’s custom-chip bench rather than a clean handoff. (cnbc.com) For Google, the next question is not whether it has a TPU program. It is whether this wider partner network can keep enough inference chips flowing, at low enough cost, to make its own silicon a real counterweight to Nvidia in 2026 and beyond. (thenextweb.com)

Get your own daily briefing

Scout delivers personalized news, insights, and conversations tailored to your role and industry.

Download on the App Store

Shared from Scout - Be the smartest in the room.