Google has struck a broad custom silicon partnership with Marvell Technology, rattling investors in Broadcom — the chipmaker that has co-designed Google's tensor processing units for nearly a decade.
Marvell disclosed the agreement in a securities filing Wednesday, describing a partnership that "spans a comprehensive range of custom silicon programs that attach to the TPU ecosystem," including components related to memory and storage. As part of the arrangement, Marvell issued stock warrants to Google valued at roughly $12.2 billion.
Broadcom shares fell 5% on the news, putting the stock on pace for its lowest close since early July. The decline came amid a broader selloff in semiconductors, with the iShares Semiconductor ETF down approximately 2% on the day. Marvell shares, by contrast, climbed about 7%.
Broadcom has co-designed TPUs with Google since the chip's early days — it is now in its eighth generation — serving as a key conduit to Taiwan Semiconductor Manufacturing Co. to have the chip blueprints fabricated into physical products. TPUs belong to the class of chips known as AI accelerators, tailored to perform the mathematical computations that underpin generative AI models.
The deal reinforces a longstanding concern among investors: that Google would increasingly diversify its custom AI compute supply chain, either by designing more components in-house or by enlisting additional partners such as Marvell and Taiwan's MediaTek. Those worries persisted even after Broadcom disclosed in April an agreement to develop and supply future generations of TPUs, along with networking and other components, through 2031.
Broadcom CEO Hock Tan addressed the diversification question directly on the company's most recent earnings call in early June. "It's a very, very strong agreement," Tan said of the 2031 deal. "It's a commitment that is very substantial in dollars — a very, very substantial amount of dollars. Now we also accept the fact that, while we like to win every design in that program, we also accept the fact that given the growth of development and consumption of AI compute even by our partner, Google, that we fully expect that there will be some diversity of sources for them. But our commitment from them is a very substantial dollar amount."
Broadcom's custom silicon business has also added new customers in recent periods, including OpenAI, Anthropic, Meta, and Apple, reducing the company's overall revenue dependence on Google. Broadcom expects to generate $56 billion in AI chip revenue in the current fiscal year and has projected more than $100 billion in fiscal 2027.
The warrant structure embedded in the Marvell-Google deal drew particular attention from analysts. Advanced Micro Devices has issued warrants to Meta and OpenAI as part of compute supply agreements, a tactic some chipmakers have used to secure business during the AI buildout. Tan, however, has a well-documented reluctance to dilute existing shareholders through such arrangements — a posture that may cost Broadcom business in the near term even as it protects long-term shareholder value.
Wednesday's broader pressure on AI-related equities came as The Wall Street Journal reported sluggish second-quarter revenue growth and widening losses at OpenAI, though OpenAI noted its growth rate has accelerated in July following the launch of new AI models. Growing political resistance to data center construction — Pennsylvania Gov. Josh Shapiro has moved to restrict new approvals in the state — has added to investor uncertainty around the AI infrastructure buildout.
The Marvell-Google partnership signals that hyperscalers are moving aggressively to spread custom silicon development across multiple vendors, a trend that is likely to intensify as AI compute demand scales. How Broadcom responds — and whether Tan's approach of competing on technical merit rather than warrant incentives holds — will be closely watched as the company heads into the second half of its fiscal year.
Disclaimer