Marvell handed Google ~7% of itself in warrants (rights to buy shares at a set price), vesting on chip purchases Google isn't required to make, mirroring the AMD OpenAI October 2025 template.
Marvell just paid Google to keep buying its chips. The bill comes in Marvell stock.
A warrant, the right to buy shares at a fixed price later, is the mechanism. Under a chip-supply agreement signed July 29, 2026, Marvell issued Google a warrant for 58,970,907 Marvell shares, roughly 7% of the company, at an exercise price of $206.58 per share. Most of those shares vest as Google accumulates purchases, with a tranche unlocking for every $500 million of covered Marvell silicon it buys. A separate 1.36 million shares vest on a time schedule over the first year regardless of purchase activity.
The covered products are inference accelerators, storage controllers, and other chips that sit alongside Google's in-house TPUs. Google runs its own TPU stack; Marvell's role is to attach across it with networking, memory, and storage silicon.
The deal is a buyer-favorable inversion of the usual supplier relationship. Purchases under the agreement are fully discretionary. Google controls the pace and is not obligated to spend a dollar. The warrant is upside-only for Google: if the purchases never come, Google keeps the time-vested slice. If they do come, the share count grows by roughly 7%, and the cost of that growth lands on Marvell's existing shareholders through dilution.
At disclosure, the math was already tilted Google's way. With Marvell trading near $243, the warrant was roughly $2.1 billion in the money. The package costs Google about $12.2 billion to exercise in full, a hypothetical ceiling rather than a confirmed purchase commitment, and the Marvell 8-K does not specify aggregate purchase commitments. Marvell shares jumped about 10% on the news and closed near $237, reflecting investor relief that Marvell had landed a hyperscaler win, but the dilution math applies whether the stock pops or drops.
The Marvell arrangement follows the template set by AMD's October 2025 supply deal with OpenAI, in which AMD agreed to supply 6 gigawatts of GPUs and issued OpenAI a warrant for up to 160 million shares, roughly 10% of the company, at a $0.01 per share exercise price, vesting on deployment milestones and AMD stock-price targets. AMD extended the same equity-for-purchases template to Meta in February 2026 with its second mega chip-supply deal of the year. Three deals in under a year suggest the structure is now a recurring condition of large AI chip purchases, not an anomaly.
The constructive read for chip suppliers is real. Marvell gets a TPU-adjacent foothold inside Google's data center stack, where the alternative was no deal at all. Analysts framed the win as Marvell's "white whale", the customer it had been chasing for years. The deal may also be a strategic premium to win a position competitors would have paid for in cash, rather than a sign of supplier weakness.
The dilutive read is also real. The cost of winning the business falls on shareholders who did not approve the spend, and the pace at which those shareholders get diluted is set by the buyer's purchasing decisions. When a supplier hands a buyer equity as a condition of purchase, the supplier is paying for the privilege of selling, and the existing cap table is the one that funds it. The wire will frame the Marvell deal as a chip-supply win. The structural read is closer to a supplier-side concession in disguise.
The next data point is whether Marvell's competitors in TPU-adjacent silicon can land comparable hyperscaler business without issuing warrants of their own. If they can, the "paid to buy" framing is overstated. If they cannot, the template is the new floor for AI chip deals with the largest buyers.