Fifty-nine million shares at $206.58 apiece. That is the option Marvell Technology just handed Google, a warrant worth as much as $12.2 billion if fully exercised, in exchange for something Marvell wants more than cash: a guarantee that Google keeps buying its chips for years. Marvell’s stock jumped nearly 8 percent on the news.
The warrant, issued on August 18 and first reported by Bloomberg, is exercisable until August 2033. If Google uses all of it, the search giant becomes Marvell’s fifth-largest shareholder. Not through an acquisition, not through a venture round, but as a byproduct of a chip purchase agreement.
Chips first, equity later
The mechanics are the interesting part. Nearly 1.4 million of the warrant shares vest in the first year of the deal. The rest vest in tranches, with each tranche unlocking as Google buys another $500 million worth of Marvell silicon. Google does not get the stock for showing up. It gets the stock for spending, which turns the warrant into a loyalty program measured in billions.
The underlying commercial agreement was signed on July 29 and covers a wide range of custom silicon programs supporting Google’s tensor processing unit ecosystem: AI inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory computing solutions. Analysts estimate the arrangement could bring Marvell roughly $120 billion in revenue through fiscal 2033, if Google hits the purchase targets its stake depends on.
Why Google wants a second silicon partner
Google has designed its own TPUs for a decade, but designing a chip and getting it built are different businesses. Broadcom has long been the dominant partner on TPU development, a position that gives it pricing power over one of the most important product lines in Google’s future. Bringing Marvell deep into the TPU ecosystem changes that negotiation. Two capable custom-silicon partners bidding for the next generation is a much healthier position than one.
There is also the Nvidia question. Every hyperscaler is trying to shift more of its AI workloads onto chips it controls, because Nvidia’s margins are, from the buyer’s perspective, a tax. Custom accelerators for inference, where workloads are predictable and volume is enormous, are the most direct route to cutting that bill. It is the same logic driving Wall Street’s $500 billion pipeline for AI compute: whoever controls the silicon supply controls the margin.
Silicon deals now come with stock attached
A year ago, a warrant of this size attached to a supply agreement would have been exotic. Now it is close to standard practice at the top of the AI market. AMD handed OpenAI warrants for up to 160 million shares last October, tied to deployment milestones. Nvidia has taken direct positions in its largest customers. The pattern is consistent: the companies on each side of a strategic silicon relationship keep buying pieces of each other, binding supplier and customer together so tightly that neither can walk away cheaply.
Skeptics call this circular financing, and the concern is not frivolous. When suppliers fund customers and customers own suppliers, revenue can look stronger than the underlying demand justifies. But the Marvell structure is more conservative than most: Google gets nothing unless it actually spends, and Marvell books real chip revenue before any dilution lands.
The thing to watch is Broadcom’s response. Google just gave its longtime TPU partner a public reminder that exclusivity was never part of the deal. How Broadcom prices the next TPU generation, and whether it concedes ground on the inference accelerators Marvell now builds, will show whether Google’s $12.2 billion signal worked.
The deal also says something about where TPU demand is heading. Google is no longer building these chips only for itself. Anthropic agreed last year to run its Claude models on up to one million of Google’s TPUs, a commitment measured in gigawatts of capacity, and Google has been pitching the chips to other outside customers since. Serving external demand at that scale requires far more silicon than Broadcom alone was ever going to supply, and it turns the TPU program from an internal cost-saving project into a product line with its own revenue ambitions. Marvell is being hired to help build inventory for a business Google has only started to sell.
For more coverage of AI chips and infrastructure, visit Mylistingo.
