Marvell granted Google a warrant worth more than twelve billion dollars. Its structure is far more interesting than its size.
Errors or inaccuracies are possible. Spotted something that looks wrong? Write to me and I will correct it.
What Exactly Was Granted
The figures here come from the Form 8-K Marvell filed with the US Securities and Exchange Commission on 19 August 2026.
| Commercial agreement signed | 29 July 2026 |
| Warrant issued | 18 August 2026 |
| Number of shares | 58,970,907 |
| Exercise price | $206.58 per share |
| Value on full exercise | about $12.18 billion |
| Exercisable until | 18 August 2033 |
What the agreement covers, in the language of the filing: AI inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute - all of them products that attach to Google's TPU ecosystem.
And the Structure: 2.31% Against 97.69%
This is the line that decides everything, and it appeared in almost no headline.
| Shares | Share | |
|---|---|---|
| Vesting with time | 1,360,867 | 2.31% |
| Vesting on performance | 57,610,040 | 97.69% |
The first part vests in equal quarterly instalments over the first year. That is the part Google receives simply because time passes.
And all the rest vests entirely differently: in 240 equal tranches, with one tranche vesting for every $500 million of custom-product purchases - by Google or on its behalf, from Marvell's third quarter of fiscal 2027 through the end of fiscal 2033.
And this is an instrument worth knowing, because it is appearing more and more in AI infrastructure deals.
A warrant is a right to buy shares at a price fixed in advance. It is not a share, and it costs the recipient nothing - it is worth something only if the stock rises above the exercise price.
And when vesting depends on purchases, the instrument stops being compensation and becomes a mechanism for aligning interests. Google is not handed value; it earns it, half a billion dollars at a time.
The meaning for a Marvell shareholder: the dilution does not arrive alone. It arrives only together with the revenue that created it.
And How Much Google Has to Buy
240 tranches, $500 million each.
That is, full vesting of the warrant requires Google to buy $120 billion of custom products from Marvell.
And this needs proportion, because the number is hard to grasp.
| Required for full vesting | $120 billion |
| Marvell revenue, trailing twelve months | $8.72 billion |
| The ratio | 13.8 times |
The window runs about six and a half years. Spread over it, that is roughly $18.5 billion a year, from Google alone - 2.1 times everything Marvell sells today to all of its customers combined.
So it is worth saying plainly: the headline "$12.2 billion" describes a scenario that requires a change in the order of magnitude of the entire business, not a continuation of a trend. Partial vesting is far more plausible than full vesting.
And what I do not know is the probability of that. The filing contains no purchase commitment, only the mechanism - the purchases are described in it as "discretionary purchases." There is no guaranteed backlog here.
The Exercise Price, and What It Says
$206.58.
Marvell's close on the day preceding the announcement was $216.00 - meaning the exercise price was set 4.36% below the market price at that moment.
And on the day of the announcement itself the stock rose 9.85% and closed at $237.27.
| Exercise price | $206.58 |
| Close, 19 August | $237.27 |
| Difference per share | $30.69 |
| Across the full warrant | about $1.81 billion |
In other words, the warrant went into the money on the same day it became known. That does not mean Google can exercise it - the overwhelming majority has not vested - but it does mean the instrument carries economic value from day one.
And the Dilution
58,970,907 shares against 875.8 million existing shares.
On full exercise: 6.73% dilution.
And this is the part a shareholder should hold in both hands. Dilution and revenue are tied together here: the shares vest only as the revenue is generated. But they vest on the basis of revenue, not profit - and what Marvell earns on those $500 million does not appear in the formula at all.
And What Happened to Broadcom
On the same day Broadcom's shares fell 4.61% and closed at $362.48.
Broadcom had been Google's principal partner in developing the TPU, and a move like this widens the supplier circle.
But it is worth being precise about what the filing says and what it does not. It describes an expansion of Marvell's collaboration with Google, in products that attach to the TPU ecosystem. It does not mention Broadcom, does not speak of replacement, and names no shares of business.
So the move in Broadcom's stock is the market's pricing, not a fact from the filing. What I do not know - and did not find in a primary source - is Broadcom's share of the TPU business today. Without that figure, there is no way to quantify what was taken.
And Where Photonics Comes In
This is not stretching the story - it is written into the scope of the agreement itself.
Of the five product categories the filing names, three are interconnect components: network interface controllers, memory interface controllers, and near-memory compute.
And why that matters.
In a modern data center, the constraint has long since stopped being only how much compute can be packed onto a chip. It is how fast data can be moved between chips - and between racks, and between halls.
And there, information no longer travels in copper but in light. Between the optical fibre and the chip sits a layer of signal-processing chips and interface controllers, which convert between the two and correct errors.
That is the layer where Marvell has historically been strong, and it is also the most concentrated layer in the photonics supply chain - few manufacturers can supply it at the rates now required.
So this deal is not only about inference accelerators. A substantial part of it touches what connects the accelerators to each other - and that is exactly where the bottlenecks in AI infrastructure sit today.
הזווית שלי
דעה אישית של אילן אברמוב - לא ייעוץ ולא המלצה
What interests me in this deal is the structure, not the sum.
A company that grants its customer shares in exchange for purchases is doing something very specific: it is saying it is prepared to give up a slice of ownership in order to lock in a customer of a size that changes the business. That is neither generous nor naive - it is a trade.
And what I think is easy to miss here is that 97.69% of the instrument costs Marvell nothing if nothing happens. If Google buys little, few shares vest. The dilution is tied to the outcome, and that is a far fairer structure than a blanket grant.
What does trouble me, by contrast, is that vesting is tied to revenue rather than to profit. In custom silicon, the margin is negotiated with a single large customer, and a customer like that knows exactly how much leverage it has. It is entirely possible for revenue to grow, for the shares to vest, and for the margin to erode. The formula does not protect the second line.
So the number I will follow is not custom-product revenue but Marvell's gross margin over time. If it holds while sales to Google grow, the deal did its job. If it erodes, the company bought turnover and sold ownership.
And one thing worth remembering about the timing: the agreement was signed on 29 July and disclosed on 19 August, three weeks later - when the warrant was issued. Marvell reports its quarter at the end of August, and that is where the detail this filing omits will be.
(It is important to stress: this is my personal opinion only, and nothing herein constitutes a recommendation to take any action.)
