Google has entered preliminary discussions with Marvell Technology to co-develop custom artificial intelligence processors, according to a report by The Information published Monday.

The talks, as reported, remain exploratory. No deal value, timeline, or silicon specification has surfaced. Reuters carried the headline globally. Neither Google nor Marvell had issued a statement at the time of writing.

For a sector accustomed to leaks measured in teraflops and tape-out dates, a bare-bones report counts as something of a novelty.

What does the report actually say?

The Information's piece attributes the talks to people familiar with the matter and frames the discussions as early-stage. That is the entire substance of the public record. Everything else in circulation amounts to guesswork by reporters and analysts reading the same thin set of facts.

Why would Google want a second partner?

Google has co-designed its Tensor Processing Units with Broadcom since the TPU's debut in 2015. The chips now power Search ranking, the training runs behind Gemini, and a growing share of Alphabet's internal infrastructure workload. They have also become the lever on which some of the company's most consequential capital-expenditure decisions rest.

A second custom-AI relationship would give Google room to negotiate, hedge against supply constraints, and pursue designs that fall outside the Broadcom partnership's remit. Hyperscaler procurement teams have grown fond of optionality since the AI accelerator shortage of 2023.

Amazon has Graviton, Trainium, and Inferentia. Microsoft has worked with AMD and Marvell. Meta has leaned on Broadcom and AMD. The pattern is now the rule rather than the exception.

What does Marvell bring to the table?

Marvell has spent the past decade pivoting from a hard-disk-drive controller business into a custom-silicon franchise anchored by ASIC engagements with hyperscalers. Its data-center revenue, of which custom silicon forms the largest component, surpassed $1.5 billion in a single quarter earlier in 2025.

The company also retains a meaningful Ethernet networking business, with switch silicon that competes head-to-head with Broadcom's Tomahawk and Trident lines. A Google partnership would, in other words, hand Marvell a strategic beachhead against its larger rival on two fronts at once.

Marvell's custom-silicon unit sits inside a broader portfolio that spans data-center switching, storage controllers, and 5G infrastructure silicon. That breadth, combined with a track record of delivering working ASICs at advanced process nodes, makes the company a credible second partner in a way that a startup would not be.

What changes for Broadcom?

Broadcom remains the senior partner on TPU development. Its Google-related revenue has been estimated by sell-side analysts at several billion dollars annually, though Broadcom does not break the figure out in segment reporting. A Google program with Marvell would not, on its own, displace it.

It would, however, introduce the kind of competitive friction that senior procurement teams tend to cultivate on purpose. The semiconductor industry has learned, through a decade of supply shocks and consolidation, that a single-source dependency is a strategic liability.

What comes next?

For now, nothing visible. Custom-silicon programs typically run two to three years from kickoff to first silicon, so any product that emerges from these talks would not arrive before 2027 at the earliest, assuming the discussions mature at all.

The Information's sources describe the talks as preliminary. They may yet collapse, as preliminary talks often do. Should they hold, the next data points to watch are the usual ones: tape-out disclosures from TSMC, foundry capacity chatter, and the inevitable hiring posts from both companies in the relevant engineering centers.

In semiconductors, optionality is rarely free. It is, however, increasingly standard practice.