Google just did something a chip buyer rarely does: instead of simply writing a check for hardware, it negotiated the right to profit from its own supplier’s stock. In a securities filing this week, Marvell Technology disclosed that it has issued Google warrants to purchase nearly 59 million of its shares, a stake that could be worth as much as $12.2 billion if fully exercised. It’s an unusual structure for a chip supply agreement, and it says a lot about how central custom silicon has become to the AI infrastructure race.
The Mechanics Behind a $12.2 Billion Warrant
According to the filing, Google can buy up to 58,970,907 Marvell shares at a fixed price of $206.58 apiece, with the warrant remaining exercisable until August 2033. But this isn’t free money handed over on day one. Roughly 1.4 million shares vest in equal quarterly installments over the agreement’s first year, regardless of purchases. The remaining bulk of the warrant, more than 57 million shares, unlocks only as Google actually buys chips from Marvell, with one tranche vesting for every $500 million in qualifying revenue Marvell books from Google’s orders.
In other words, Google’s ownership position grows in direct proportion to how much custom silicon it purchases. If Google hits every spending threshold, the deal could translate into roughly $120 billion in cumulative chip purchases for Marvell running through fiscal 2033. That would also make Google the chipmaker’s fifth-largest shareholder, a striking outcome for what began as a components agreement.
What the Deal Actually Covers
The partnership doesn’t touch Google’s core Tensor Processing Units themselves. Instead, it covers what Marvell describes as products that “attach to the TPU ecosystem” — AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute hardware. These are the supporting components that let TPUs move data efficiently, store it, and communicate across a data center, which have become just as critical as the processors at the center of it all.
This expanded scope matters because it shows Google isn’t just chasing more raw computing power. It’s investing in the entire hardware stack that surrounds its custom silicon, treating data-center networking and storage as strategic assets rather than commodity purchases.
Markets React Fast
Investors didn’t waste time digesting what the deal meant. Marvell shares jumped as much as 14% intraday before settling to a close of roughly 8% higher, as the market read the agreement as a strong vote of confidence from one of the world’s largest cloud providers. Rival chipmaker Broadcom, which has been Google’s primary custom-silicon partner for years, slid more than 5% on concerns that it now faces real competition for Google’s business.
That reaction may be somewhat overblown. Morningstar analyst William Kerwin described the news as reflecting a growing opportunity at Google for new suppliers, rather than Marvell displacing Broadcom outright. Google has reportedly split its eighth-generation TPU program into separate training and inference workloads, and at the scale it’s now operating, one design partner may simply not be enough to keep up with demand.
A Pattern Reshaping the AI Chip Industry
This isn’t happening in isolation. Big Tech companies have increasingly turned to equity-linked arrangements to lock in chip supply while sharing in the upside they’re creating for suppliers. In October 2025, AMD struck a similar deal with OpenAI, agreeing to supply AI chips worth tens of billions annually while giving OpenAI the option to acquire roughly 10% of AMD. Nvidia, for its part, invested $2 billion directly into Marvell earlier this year through its NVLink Fusion partnership, and has also backstopped tens of billions in AI infrastructure spending for other partners.
For chipmakers like Marvell, these warrant-based deals offer a predictable, long-term revenue pathway and the market credibility that comes from being tied to a major hyperscaler. For buyers like Google, the structure works as a hedge: it locks in supply capacity for critical data-center hardware, all without requiring cash upfront or a fixed purchase commitment, since the shares only vest as spending actually happens.
Why This Matters for the Broader AI Buildout
The scale of this deal underscores just how much money is now flowing into the infrastructure layer beneath AI models, not just the software running on top of it. Demand for custom silicon like TPUs has surged as major tech companies search for cheaper, more efficient alternatives to Nvidia’s graphics processors, particularly for the inference workloads involved in running trained AI models day to day. The change has also caused a huge increase in capital spending by Google, which is spending tens of billions of dollars a quarter on its computing infrastructure.
What’s notable about the Marvell deal is how it blurs the line between customer and investor. Google is no longer just a buyer of chips, it is structurally aligned with the success of its supplier, and the two companies share a common incentive to continue to grow the relationship. As custom silicon becomes central to how the biggest cloud providers compete, expect more of these hybrid supply-and-equity arrangements to define the next phase of the AI hardware race, one where chip purchases and shareholder value increasingly move in lockstep.
