Google discusses custom AI inference chips with Marvell to reduce Broadcom dependence

Google is in discussions with Marvell Technology to develop two new types of AI chips. One is a memory processing unit that works alongside existing TPUs, and the other is an inference TPU optimized for serving models rather than training them. However, no signed contract has emerged yet.

This move is not a signal of cutting ties with Broadcom, but rather a strategy to diversify the supply chain. Google already has Broadcom as its partner for high-performance TPUs and networking, MediaTek handling the design of the cost-reduced e variant that lowers costs by 20-30%, and TSMC handling production. Marvell could become a third design partner added to this structure.

Behind this is the reality that the center of computing demand has shifted from training to inference. Google's 7th-generation TPU Ironwood, which the company calls "the first Google TPU built for the age of inference," delivers 10x higher peak performance compared to TPU v5p, and produces 42.5 FP8 exaflops in a superpod connecting 9,216 chips. Google plans to build Ironwood in the millions of units this year.

The core of the discussion is cost structure. Training is a single large-scale task, but inference must continuously process every user request, so costs accumulate as scale grows. For a company like Google that operates AI search, Gemini, and Cloud AI API at massive scale, even a small reduction in per-user or per-request cost can save billions of dollars annually, making purpose-built custom silicon more valuable.

This relationship isn't a new story. In 2023, there were reports that Google was pursuing a chip codenamed Granite Redux with Marvell instead of Broadcom, aiming to save billions of dollars annually. But now, instead of excluding Broadcom, the direction has shifted toward a multi-supplier architecture where Broadcom, MediaTek, and Marvell each handle different segments.

Marvell is also growing its presence. The company's data center revenue reached $6.1 billion in fiscal year 2026, with total revenue of $8.2 billion, up 42% year-over-year. With Nvidia's $2 billion investment, the Celestial AI acquisition, and now the Google discussions overlapping, the company is expanding its position in the custom AI chip market.

The market is growing rapidly. TrendForce projects custom chip sales will grow 45% in 2026, while Counterpoint Research forecasts that by 2027, Broadcom will hold about 60% of the market and Marvell about 25%. Ultimately, Google's goal isn't simply cost reduction, but building an independent supply chain capable of sustaining the largest AI inference workloads.