Google is in talks with Marvell Technology to develop two new AI chips, according to a report from The Information published Sunday. The deal would include a memory processing unit and an inference-optimized TPU, adding Marvell as a second major design partner alongside Broadcom for Google's custom silicon program. Marvell shares jumped nearly 6% on Monday. Broadcom fell 2%.

The move is significant because Google has relied almost exclusively on Broadcom for its Tensor Processing Unit (TPU) design work since the chip program began in 2015. Google was the first hyperscaler to build custom AI accelerators, and the TPU partnership with Broadcom has been one of the most valuable relationships in the semiconductor industry. Broadcom's AI revenue doubled to $8.4 billion in its most recent quarter, driven largely by that Google work plus deals with Meta, ByteDance, and OpenAI.

Why Marvell, Why Now

The Google-Broadcom partnership isn't dying. Earlier this month, Broadcom disclosed an expanded long-term agreement with Google to supply future generations of TPUs through 2031. That deal also revealed that Broadcom, Google, and Anthropic expanded their relationship for AI chip work. The Broadcom relationship is alive, well, and recently extended.

But Google is clearly hedging. When your AI infrastructure costs billions per quarter, single-supplier dependence is a strategic risk. Marvell brings expertise in a different area: memory processing and inference optimization. The two proposed chips target workloads where Google may see Broadcom's current offerings as insufficient, or where it simply wants competitive pressure on pricing.

Marvell has been on a tear. The company's stock has gained more than 50% since March after posting strong Q4 earnings driven by AI demand. Nvidia took a $2 billion stake in Marvell in March, making it easier for Nvidia customers to access the ASICs (application-specific integrated circuits) that hyperscalers like Google are building. Marvell expects its revenue to approach $15 billion in fiscal 2028.

The Custom Silicon Land Grab

Google is not alone in diversifying its chip supply chain. Every major AI company is now designing custom silicon: Amazon has Trainium, Meta has MTIA, Microsoft has Maia, and OpenAI has a 10-year deal with Broadcom for its own chips. The common thread is that Nvidia's GPUs, while still dominant, are too expensive and too scarce for companies running AI at hyperscale. Custom ASICs offer lower cost per inference and can be optimized for specific workloads.

Google's 7th generation "Ironwood" TPU launched in November, and the company may reveal its next generation at Google Cloud Next later this week. With both Broadcom and Marvell now working on Google's AI silicon, the company is building the deepest custom chip bench in the industry.

For Broadcom, the news is a mixed signal. Its $120 billion AI revenue forecast by 2027 (raised by Mizuho Securities last week) still looks achievable. Six XPU clients, long-term Google contract, Meta's 1GW MTIA deal. But losing its monopoly on Google's TPU design work means the negotiating leverage shifts. When your biggest customer has a second option, the conversation changes.

Marvell trades at 33.35 times forward earnings, compared to 27.84 for Broadcom. The market is pricing in the upside of diversification into Google's chip program. If the deal closes, Marvell would add more than $6 billion to its market capitalization of $122 billion.

Google and Marvell did not respond to requests for comment.

First reported by The Information. Additional reporting by CNBC, Reuters, The Next Web, and Sherwood News.