SANTA CLARA, 21 AUG 2026 — Marvell has granted Google a warrant to buy almost 59 million of its shares at US$206.58 each, a stake worth about US$12.2 billion. Most of it vests in tranches tied to every US$500 million of chips Google buys.
The customer is being paid to keep ordering, in the supplier's own stock.
The structure
Marvell will develop AI inference accelerators, storage, networking and memory interface controllers, and near-memory computing technologies for Google. The arrangement was disclosed on 19 August.
Marvell shares rose more than 10 per cent on the news. Broadcom, its larger competitor in custom silicon, fell more than 3 per cent.
A warrant is a discount that only exists if you keep buying
The warrant is a carefully designed instrument.
Google has the right, not the obligation, to buy shares at a fixed price. If Marvell's stock rises above that price the right is worth money; if it does not, the right expires and Google has lost nothing. The tranches vest against purchase volume, so the value accrues in proportion to how much Google spends.
This is functionally a volume rebate paid in equity, not cash. The structure has two properties a cash discount lacks: it costs Marvell nothing up front, and it gives Google a direct financial interest in Marvell's stock performance, not just in receiving its chips.
The widely reported US$120 billion figure is not a revenue forecast. It’s the implied spending by Google through fiscal 2033 that would be required for the full warrant to vest, which makes it a measure of the incentive, not a commitment.
There is a risk in the structure that belongs to Marvell's existing shareholders rather than to either party. Dilution at full vesting is substantial, and it is triggered precisely when the business is doing well, which is when existing holders would otherwise expect to benefit most. The trade they are being offered is a smaller share of a much larger company, and whether that is good depends entirely on whether the purchase volumes that trigger the vesting would have happened anyway.
Customer equity is becoming the standard instrument
This is the third arrangement of its kind we have covered in a month. We reported on Nvidia disclosing equity stakes in customers, and the same logic runs through the compute-supply agreements now common between labs and cloud providers.
This kind of deal is a response to a matching problem in the AI hardware market. Suppliers need multi-year demand visibility to build capacity; buyers need supply certainty to design products around a specific chip. Neither can get a normal contract to carry that risk, because the volumes are enormous, the technology moves annually and nobody wants to be locked to a price.
Equity bridges it. The supplier gets a customer with a reason to stay, the customer gets preferential attention and upside, and neither has to sign a fixed-price commitment for a decade. It works, and it also means the boundary between a customer and an investor is dissolving in a way that makes the sector's revenue figures harder to interpret from outside.
The strike price is worth a glance too. A warrant exercisable at a fixed level is only valuable if the shares trade above it, so the price at which it was set encodes a shared expectation about where the business is going. Setting it too low is a giveaway; setting it too high makes the incentive theoretical. That both parties agreed on the strike price makes it one of the few informative figures in the announcement.
One practical consequence for anyone reading Marvell's numbers from here on. Revenue recognised from a customer who is simultaneously an option holder is not straightforwardly comparable to revenue from an ordinary buyer, because part of the consideration flows back as equity value rather than staying as margin. That does not make the revenue less real. It does mean the gross margin line and the share count have to be read together, and most summaries will quote only the first.
What Google is actually buying
The component list is more interesting than the headline. AI inference accelerators, storage, networking, memory interface controllers and near-memory computing is not a chip order. It is most of a system.
Google already designs its own accelerators. What it does not build in the same depth is the surrounding silicon that moves data to and from them, and at current model sizes that surrounding silicon is frequently the constraint rather than the compute. An accelerator waiting on memory bandwidth is an expensive idle asset.
Near-memory computing points the same direction. Putting processing closer to where data sits attacks the movement cost directly, and a buyer commissioning it is telling you where its bottleneck is.
Broadcom's 3 per cent fall is the market reading it as share moving rather than as a market expanding, which may or may not be right — but it does confirm that custom silicon for hyperscalers is now understood as a small number of very large relationships rather than a competitive market with many buyers.
What it means for buyers outside that circle
While few in Southeast Asia will ever negotiate a chip supply agreement directly, the deal still matters to regional buyers.
Capacity committed to a hyperscaler under a decade-long arrangement is capacity not available to anybody else, and it is committed before it is built. The practical effect for a regional cloud operator or a national AI programme is that the queue is being allocated years ahead, in deals whose terms are not public, to buyers who can offer something a purchase order cannot.
We wrote yesterday that Samsung raised advanced foundry prices while losing market share, which is what a fully booked supplier does. This is the other end of the same condition: the bookings that fill it are being locked in with instruments that ordinary buyers have no access to.
What we could not establish
The chip pricing. A volume rebate paid in equity is only assessable against what the chips cost, and neither the unit economics nor the purchase commitments have been disclosed.
Also unestablished are the warrant's expiry date; any conditions beyond purchase volume; whether Google has committed to minimums; how the deal affects Google's own accelerator programme; if Marvell has similar deals with other customers; the accounting treatment; and whether the US$120 billion figure came from the companies or from analysts.
What to watch
Watch the vesting disclosures. Tranches tied to purchase volume make the warrant a public meter on how much Google is actually buying, and that will be visible in filings long before any revenue breakdown is.
Then watch whether Broadcom responds with a comparable instrument. If customer equity becomes the price of competing for hyperscaler business, the smaller suppliers who cannot offer it are effectively excluded from the largest segment of the market.
Finally, watch the accounting commentary. Warrants settled against future purchases are an awkward disclosure problem. The accounting presentation of this deal will likely set a precedent for how investors interpret others.
