Google has deepened its push into custom artificial intelligence silicon by expanding its partnership with Marvell Technology, granting the chip designer a warrant that could give the search giant up to a $12.2 billion stake. 

 

The deal, disclosed in regulatory filings and reported on August 18–19, 2026, ties equity upside directly to Google’s future purchases of Marvell-designed chips for its Tensor Processing Unit ecosystem.

 

Under the agreement, Marvell issued Google a warrant to purchase as many as 58.97 million shares at $206.58 each. If fully exercised, the stake would make Google roughly the fifth-largest shareholder in Marvell and represent about 7 percent of the company. 

 

Nearly 1.4 million shares vest in the first year. The remainder unlocks in 240 equal tranches—one for every $500 million in custom-chip revenue Marvell records from Google—running through fiscal 2033.

 

If Google hits the targets, the commercial relationship could generate approximately $120 billion in cumulative purchases for Marvell. The products covered include AI inference accelerators, memory processors, storage and network interface controllers, and near-memory compute solutions that attach to Google’s TPU infrastructure.

 

The structure is notable because Google does not pay cash upfront for the equity. The warrant vests only as Google actually buys the silicon. This aligns incentives tightly: Marvell gains a multi-year volume commitment from one of the world’s largest buyers of AI hardware, while Google secures a preferred design partner and potential ownership upside without immediate dilution of its own capital.

 

Marvell shares rose sharply on the news, climbing as much as 8–14 percent intraday in some reports, while shares of rival Broadcom, Google’s primary long-term custom-chip partner, declined more than 5 percent. Alphabet stock was little changed. Analysts described the move as evidence that Google is diversifying its custom silicon supply chain rather than displacing Broadcom outright.

 

Google has relied heavily on Broadcom for high-performance TPU variants and related networking silicon under multi-year agreements extending into the 2030s. Adding Marvell expands options for inference-optimized designs and other components as the cost of running trained models continues to dominate AI compute budgets. Demand for specialized inference hardware has accelerated as enterprises and cloud providers seek alternatives to general-purpose GPUs.

 

The deal fits a broader pattern in which hyperscalers lock in chip suppliers with equity-linked structures. Similar arrangements have appeared elsewhere in the industry as companies race to secure capacity for the next generation of large language models and agentic systems.

 

Marvell, historically stronger in networking and storage, gains a clearer path into the high-growth AI accelerator market and a powerful endorsement from Google’s engineering and procurement teams.

 

For Google, the arrangement reduces single-supplier risk and gives it leverage over design roadmaps for chips optimized for its own software stack and data-center architecture. 

 

Custom silicon remains central to Alphabet’s strategy of lowering the cost per token of inference while improving performance on its internal models and Google Cloud offerings.

 

Market reaction underscored the competitive stakes. Broadcom has been a major beneficiary of the custom-chip boom, helping design and supply silicon for Google, Meta, and others. 

 

Any perception that Google is spreading work more widely weighed on Broadcom’s shares even though the Marvell agreement does not cancel existing Broadcom contracts.

 

The timing coincides with intense activity in AI infrastructure financing. Chip designers, foundries, and cloud providers continue to arrange large debt and equity packages to fund the build-out of data centers and specialized accelerators. 

 

Google’s warrant structure is one more example of how equity, volume commitments, and long-term roadmaps are being combined to underwrite the capital intensity of the AI hardware cycle.

 

Marvell’s filing and subsequent reporting indicate the commercial terms were signed in late July, with the public disclosure and market reaction arriving in mid-August. The company positioned the partnership as covering a broad range of components that sit alongside or within Google’s TPU systems rather than replacing the core TPU designs themselves.

 

Industry observers note that inference workloads are expected to account for a growing share of total AI compute spending. Models that can deliver higher throughput or lower power consumption at the edge of the data center or in specialized racks therefore carry strategic value.

 

Marvell’s expertise in high-speed interconnects and controllers is seen as complementary to pure accelerator design.

What happens next depends on execution. Google must place the purchase orders that unlock the warrant tranches. Marvell must deliver competitive designs on schedule and at the required power and performance targets. 

 

If both sides perform, the relationship could become one of the larger custom-silicon programs of the decade and further entrench Google’s dual-track approach of internal innovation plus selective external partnerships.

 

The transaction also highlights how the AI chip market is evolving beyond a pure Nvidia-versus-everyone contest. Hyperscalers are building multi-supplier ecosystems that mix proprietary architectures, external design partners, and foundry capacity. Equity warrants and volume-linked incentives are becoming tools to secure those relationships for the long term.

 

For investors and industry watchers, the Marvell-Google deal is a concrete signal that custom AI silicon demand remains robust and that Google is actively managing its supply chain to support continued rapid growth in AI services. 

 

The $12.2 billion warrant and the associated $120 billion revenue opportunity provide a measurable yardstick for how large these partnerships can become.