Marvell shares rose 11–12% premarket on August 19 on an expanded custom-silicon agreement with Google. Broadcom fell more than 3% on the same news. The number attached to it everywhere — MarketWatch, CNBC, Reuters, a 1,046-post news cycle — is that Google received a warrant "worth about $12.2 billion," roughly a full year of Marvell's revenue.
The warrant is real and the 8-K is public. Read it and two things in that sentence stop being true — the table above is the whole correction, line by line.
The $12.18 billion is the exercise price
Marvell issued Google a warrant to purchase up to 58,970,907 shares at an exercise price of $206.58 per share. Multiply those together and you get $12.18 billion.
That is the aggregate exercise price — the cash Google would hand Marvell to exercise in full. It is not value transferred to Google, and a warrant is not a stake. What Google receives is the right to buy at $206.58, which is worth something only to the extent Marvell trades above $206.58. Marvell's price on this site is $195.22 as of August 19, a build-time snapshot rather than a live quote — the strike sits 5.8% above it. Struck out of the money against that snapshot, the warrant's value to Google at issuance is option value, not $12.18 billion.
Reporting a strike notional as a stake is a category error, and it is the number the entire cycle ran on.
And 97.69% of it has to be earned
The vesting terms are where the story actually is, and they are unusually explicit.
Of the 58,970,907 shares, 1,360,867 — 2.31% of the warrant, and 0.15% of Marvell's 898 million diluted shares — vest on time, in equal quarterly instalments over the first year.
Everything else vests on revenue. In the 8-K's words, the remaining shares vest "in 240 equal tranches, with one tranche vesting for each $500 million in Custom Products revenue", on discretionary purchases by Google and its affiliates from fiscal Q3 2027 through the end of fiscal 2033. The warrant expires August 18, 2033.
Do that arithmetic and the headline inverts:
- 240 tranches × $500 million = $120 billion of Custom Products revenue for the warrant to vest in full.
- Each tranche is 240,042 shares — so Google earns one share for roughly every $2,083 of custom product revenue it places.
- Marvell's full-year FY2027 revenue guide, across all products, is about $12.7 billion. Full vesting requires 9.4× that — in custom products alone.
- Spread evenly across the 25 quarters in the window, that is $4.80 billion a quarter of custom product revenue, against $2.418 billion of total company revenue in Q1 FY2027 and $1.833 billion of data-centre revenue. Custom products alone would have to run at about twice the whole company's current revenue, immediately and for six years.
Vesting need not be even, so $4.80 billion a quarter is a required average rather than a schedule. It does not need to be a schedule to make the point.
What the warrant actually is: a disclosed six-year revenue target
Strip the headline away and something more useful is left. Marvell has guided publicly to about $12.7 billion for FY2027 and approaching $15 billion for FY2028. It has never published a number for FY2033.
It just did, in vesting mechanics. The 8-K contains a $120 billion custom-product revenue scoreboard running to fiscal 2033, with a share count attached to every $500 million step. That is a far more informative disclosure than the press release, and it is the opposite of a giveaway: Google paid nothing up front, took no shares, and will own more of Marvell only in proportion to how much business it hands Marvell — capped at 6.57% of today's diluted shares, and only if it places $120 billion of orders.
Whether that is generous to Google or to Marvell depends on a number nobody has: how much of the $120 billion Google actually intends to spend. The structure tells you Marvell was willing to pay 6.57% of itself to find out, and that Google was unwilling to promise the revenue any other way.
The Broadcom leg is a category error too
Broadcom fell more than 3% on the read that Google is diversifying away from it. The figures we already hold say that is not the trade.
| Revenue | Period | |
|---|---|---|
| Broadcom AI semiconductor revenue | $10.8B | Q2 FY2026, +143% |
| Broadcom AI, guided | $16.0B | Q3 FY2026 |
| Marvell total revenue, all products | $2.418B | Q1 FY2027 |
| Marvell full-year guide, all products | $12.7B | FY2027 |
Marvell's entire FY2027 revenue guide is about 20% of Broadcom's annualised Q3 AI run rate alone. Broadcom's own Google agreement, signed in April, runs to 2031. Our Marvell coverage called this a duopoly with Marvell as the other half; the arithmetic says it is a lopsided one, and a 3% move in Broadcom on a Marvell announcement is pricing a share shift that these two revenue lines do not support.
What it means for Nvidia, and the limit of what can be said
The product list in the 8-K is specific, and it is the primary document rather than a press summary: "AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute."
Our Nvidia model names the first of those as the bear case, in as many words — "custom ASICs take the inference workload, and merchant Ethernet claws back the fabric." The spread it sits inside is the widest of any model here:
| Nvidia case | Fair value | Against the $225.16 stored price |
|---|---|---|
| Bear — ASICs take inference | $63.15 | −72.0% |
| Base | $245.07 | +8.8% |
| China licensed back | $350.31 | +55.6% |
| Bull — the ceiling is wrong | $625.45 | +177.8% |
So a hyperscaler funding an inference-accelerator and networking roadmap at a merchant supplier is directionally an input to that bear case. It does not move the model, and the reason is the same arithmetic as above: the warrant vests against custom revenue that has not happened, on a schedule running to 2033, and Marvell's entire current data-centre business is $1.833 billion a quarter against Nvidia's data-centre segment. This is evidence about a direction, not a quantity.
One thing this piece will not do is read the words "near-memory compute" as evidence about high-bandwidth memory demand. That inference needs the architecture, not the noun — whether near-memory compute displaces HBM content, complements it or is orthogonal is not established by a product name in an 8-K, and we have criticised exactly that kind of reading elsewhere. It goes in the watch list, not the argument.
What to watch
- Marvell's accounting treatment of the warrant, in the next 10-Q. If the revenue-vested portion is recorded as contra-revenue as the associated sales are booked, it reduces reported revenue against the guided $12.7B directly, and the economics of every tranche change.
- The first vesting disclosure. One tranche is $500 million of custom product revenue. The first time Marvell reports how many tranches have vested, the market gets a clean, audited read on Google's actual custom-silicon spend — a number no hyperscaler discloses voluntarily.
- Whether Broadcom says anything. Its Google agreement runs to 2031 and was signed in April. Silence is not confirmation that nothing changed.
- Whether this structure spreads. Amazon, Microsoft and Meta all buy custom silicon. If warrants struck near market and vesting on purchase volume become the standard way hyperscalers contract with suppliers, that is a story about how the AI buildout is financed, and it is much larger than these four tickers.
- The architecture behind "near-memory compute" — enough technical detail to say whether it changes HBM attach rates. Until then it is a product name.
Warrant terms — 58,970,907 shares at $206.58, 1,360,867 time-based shares vesting in equal quarterly instalments over the first year, the remainder vesting in 240 equal tranches of one per $500 million of Custom Products revenue from fiscal Q3 2027 through fiscal 2033, expiry August 18, 2033, and the July 29, 2026 commercial agreement — are quoted from Marvell's Form 8-K filed August 19, 2026. The product list is the 8-K's own wording. The $12.18B aggregate exercise price, the $120B full-vesting figure, the 240,042 shares per tranche, the $4.80B quarterly pace and all percentages are our arithmetic on those filed figures. Marvell's price of $195.22 and diluted share count of 898,000,000 are as stored on this site as of August 19; the price is a build-time snapshot and will differ from the live quote, and the diluted count is used because it is the one the site renders. Marvell Q1 FY2027 revenue of $2.418B, data-centre revenue of $1.833B, the $2.70B Q2 guide and the ~$12.7B FY2027 and ~$15B FY2028 outlook are from our Q1 FY2027 analysis. Broadcom's $10.8B Q2 AI semiconductor revenue and $16.0B Q3 guide are from our Q2 FY2026 analysis; its April agreement with Google running to 2031, and the 11–12% and 3% share moves, are press-reported and not independently verified. Nvidia scenario fair values are computed from our model as committed, against a stored price of $225.16. We have not read the full Warrant Agreement filed as Exhibit 4.1, and certain of its exhibits and schedules were omitted from the filing under Regulation S-K Item 601(a)(5).
