- Lead. Marvell Technology granted Google the right to purchase up to 58.97 million shares worth a potential $12.2 billion, in exchange for a multi-year agreement to develop custom AI chips built around Google’s tensor processing unit ecosystem.
- Fact. The warrant carries an exercise price of $206.58 per share and extends until August 18, 2033; if Google buys the maximum amount, it becomes Marvell’s fifth-largest shareholder.
- Stake. The deal could generate roughly $120 billion in revenue for Marvell through fiscal 2033, while signalling that hyperscalers are building equity-aligned custom-silicon partnerships to reduce their dependence on Nvidia’s GPU platforms.
The Warrant Mechanics
The commercial agreement between Marvell and Google was signed on July 29; the warrant was issued on August 19. Under the terms reported by Yahoo Finance and Bloomberg, nearly 1.4 million of the 58.97 million warrant shares vest in the first year of the partnership. The remainder come in tranches tied to commercial throughput: for every $500 million worth of chips that Google purchases from Marvell, another tranche vests. The structure aligns Marvell’s equity upside directly with Google’s consumption, creating an incentive for both companies to deepen the relationship rather than revisit its terms at contract renewal.
At the announced exercise price of $206.58, the full warrant would cost Google approximately $12.18 billion to exercise — a fraction of Alphabet’s capital expenditure in any recent year. Marvell shares jumped nearly 8 percent on the announcement. Broadcom, which holds a prior custom-chip agreement with Google, fell about 5 percent as investors processed what a second major silicon partnership means for its share of Google’s infrastructure spending.
What the Chips Do
The scope of the agreement covers processors built to complement Google’s TPUs, including AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute components. TPUs handle model training and much of the inference workload directly; the surrounding infrastructure — how data moves between memory, storage and the compute fabric — is where Marvell specialises. As the partnership terms make clear, the deal is less about replicating Nvidia’s GPU than about optimising the entire data-centre stack around Google’s proprietary silicon.
That distinction matters for the competitive landscape. Nvidia’s own infrastructure commitments to hyperscalers are growing, but the Google-Marvell alliance is predicated on a view that custom silicon, purpose-built for specific inference workloads, will ultimately prove more efficient than general-purpose accelerators — a bet that TSMC’s expanding advanced-packaging capacity is making technically feasible at scale.
Implications for the Custom-Chip Ecosystem
The deal reinforces a pattern that has been building for several years: large cloud providers are using equity instruments — warrants, convertible notes, strategic stakes — to lock in chip supply and align supplier incentives. Amazon has a comparable relationship with Annapurna Labs. The Google-Marvell warrant is unusual in its transparency and scale, but the underlying logic — hyperscaler capital expenditure translating into quasi-equity relationships with hardware partners — is becoming the industry standard.
For Marvell, the arrangement reduces customer concentration risk by tying revenue to a long-duration volume commitment rather than annual purchase orders, while converting a customer into a financial stakeholder who has reason to want Marvell’s product roadmap to succeed. The 2033 expiry on the warrant gives both companies a seven-year runway in which the economics of AI inference are likely to be reshaped multiple times by advances in model architecture, chip design, and data-centre infrastructure — all of which Marvell and Google now have shared incentives to navigate together.