Marvell hands Google a $12.2B stock option to build its custom chips

Marvell has granted Google a warrant to buy up to $12.2 billion of its stock as part of an expanded deal for Google's custom AI silicon, the companies disclosed on August 19, 2026, according to Reuters and The Next Web. The structure is the notable part: Google does not pay cash for the option - it earns the right to buy the shares by buying chips, tying the size of its potential stake directly to how much custom silicon it purchases through Marvell's fiscal 2033.

For developers and teams that run workloads on Google Cloud, the story under the finance headline is supply: Google is locking in a second major partner to build the accelerators behind its Tensor Processing Units, and a bigger, more secure TPU supply chain is what determines whether the cheaper-than-Nvidia capacity you rent actually exists when you need it.

How the warrant works

The warrant covers 58,970,907 Marvell shares at $206.58 apiece - $12.2 billion if fully exercised - which would make Google the fifth-largest investor in Marvell, per both Reuters and The Next Web. It vests in tranches: The Next Web reports each $500 million in chip purchases triggers an additional block of exercisable shares, with roughly 1.4 million shares vesting in the first year. Reuters puts the potential revenue to Marvell at about $120 billion through fiscal 2033 if Google hits its purchasing targets.

The chips are not a single product. Per The Next Web, they integrate with Google's TPU ecosystem across inference accelerators and the storage and networking controllers around them - Reuters describes the scope as processors that run AI models, manage data storage, and move information across networks.

Why Google is widening its supplier base

Google's custom chips have until now leaned heavily on Broadcom. Analysts framed this deal not as a switch but as an expansion: Morningstar's William Kerwin called it "a big win for Marvell" and described it as "a growing pie at Google for new sources, rather than a competitive displacement of Broadcom," per both sources.

The market read the split accordingly. The Next Web reports Marvell shares rose about 8% on the news while Broadcom fell more than 5%. The Next Web also cites the scale driving the urgency: Google's TPU revenue projected at roughly $3 billion in 2026 and about $25 billion in 2027 - the kind of ramp that makes a single-supplier dependency a risk worth paying to reduce.

What this means for developers

The direct lever here is TPU availability and pricing. Reuters frames the custom-silicon push as giving Google "cheaper alternatives to Nvidia's graphics processors"; a second builder scaling that supply is what turns a roadmap promise into rentable capacity. If your inference or training budget depends on TPU instances - or on the price pressure they put on Nvidia GPU rates - a more resilient TPU supply chain is a tailwind worth factoring into 2027 capacity planning.

The pattern is the wider signal. Reuters notes this warrant structure echoes the Nvidia-OpenAI and AMD-OpenAI arrangements: buyers and chip makers binding themselves together with equity that vests on volume, rather than arm's-length purchase orders. For anyone choosing a cloud AI platform, the strategic frame is vertical integration - the hyperscalers are increasingly co-owning their silicon suppliers, which favors the provider's own accelerators over merchant GPUs over time. Betting a long-lived architecture on one accelerator family is the risk to watch; keeping model-serving code portable across TPU, Nvidia, and AMD backends is the hedge.