Google Signs Marvell for Custom AI Silicon, Ending Broadcom's Clean Run

Google has brought a second chip design partner into its AI silicon program, and Wall Street reacted within hours. An SEC filing this week disclosed that Marvell Technology will develop custom silicon for Google, and Broadcom shares closed roughly 4 percent lower on Wednesday.

The filing is short on specifics but broad in scope. Marvell described work on custom silicon programs that attach to Google's TPU ecosystem, spanning AI inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute.

An open secret, formalised

Hyperscalers rarely design accelerators entirely in-house. Large, undifferentiated blocks of the work are outsourced to IP houses — Broadcom, Marvell and Arm chief among them — while the cloud provider keeps the pieces that define its architecture. Broadcom has been Google's partner of choice on Tensor Processing Units, although that relationship was only publicly confirmed earlier this year.

Nothing in the filing suggests Broadcom is losing the TPU work. Every technology Marvell is now contracted to supply is something Broadcom could equally deliver. What has changed is that Google no longer has a single source for any of it.

A $12.2 billion sweetener

Marvell clearly expects the relationship to matter. It issued Google a warrant to acquire close to 59 million of its own shares, worth roughly $12.2 billion at current prices — an unusually large alignment mechanism to attach to a supply agreement.

Whether Google exercises the warrant is a separate question, and doing so would commit the company to a chip design collaboration measured in years rather than quarters. But the size of the incentive is a reasonable proxy for how much future revenue Marvell believes is attached to it.

What Google might actually want

Marvell has spent several years assembling hardware IP that maps closely onto the problems large TPU fleets encounter: high-speed SerDes for switches and network cards, multi-die XPU reference designs, CXL memory controllers and silicon photonics interconnect.

The most intriguing candidate is the photonic memory technology Marvell picked up with its acquisition of Celestial AI late last year. Its recently unveiled Photonic Fabric NIC and chiplet products allow racks as far as 50 metres apart to share a common memory pool.

That fits Google's architecture unusually well. TPUs already communicate over high-speed optical circuit-switched links, so pooled optical memory would let Google reassign capacity between TPU clusters according to workload, rather than wiring memory to compute in a fixed ratio. Memory stops being a specification to provision and becomes a resource to schedule.

Leverage, mostly

The simplest reading may also be the correct one. Alphabet has watched what concentrated supplier power does to pricing elsewhere in the industry, and a second qualified vendor is the cheapest insurance available against it. With Marvell in the mix, Google can put two IP houses in competition on both price and performance for every block it does not build itself.

For Marvell, the agreement is validation after a stretch in which Broadcom absorbed most of the attention around custom AI accelerators. Its IP portfolio now has a reference customer at the largest possible scale, which tends to attract others.

For Broadcom, the 4 percent slide is less about revenue already lost than about an assumption quietly expiring — that Google's custom silicon roadmap ran through one company. Going forward, it plainly does not.