Summary

  • Marvell issued Google a warrant for as many as 58,970,907 shares at US$206.58. Only 1,360,867 shares vest with time; the remaining 57,610,040 depend on qualifying product revenue.
  • Performance vesting uses 240 cumulative steps of US$500 million. Complete performance vesting therefore corresponds to US$120 billion, but Marvell calls Google's purchases discretionary and discloses no minimum order.
  • A sale counts only if the product meets a narrow contract test: written custom development plus Google's right to control whether Marvell may sell the whole product elsewhere, or a later mutual written designation.
  • Vesting is not issuance. Cash or cashless exercise, the strike price, a 4.999% initial beneficial-ownership limit and the future share count determine what dilution, if any, eventually appears.

Marvell has given the market an unusually precise scale and an unusually dangerous temptation.

The scale is 240. Each time cumulative qualifying revenue from Google reaches another US$500 million, one performance tranche of the warrant vests. Multiply the two and complete performance vesting requires US$120 billion of qualifying revenue during the measurement period.

The temptation is to call that number a contract value.

It is not. Marvell's Form 8-K says the relevant purchases are discretionary. It does not disclose a minimum volume, an irrevocable order, backlog, remaining performance obligations or a delivery calendar. US$120 billion describes the upper end of an equity incentive. It does not describe what Google has promised to buy.

That distinction is the centre of the instrument. The warrant is designed to make a very long commercial relationship measurable without making the whole relationship public. It reveals the revenue ruler, the equity reward and several control mechanisms. It withholds the commercial agreement, current qualifying-product schedule, unit economics and purchase plan.

The 240-step ledger

Marvell and Google entered their custom-semiconductor agreement on 29 July. Marvell issued the warrant on 18 August and filed it the next day. The public programme spans AI inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute attached to Google's TPU ecosystem.

The maximum warrant covers 58,970,907 Marvell common shares. Its two portions should never be blended.

The smaller portion, 1,360,867 shares, is time-based. It vests in four instalments during the first year: 340,216 shares after three months, followed by 340,217 shares at six, nine and twelve months. These rights do not require a disclosed revenue threshold.

The remaining 57,610,040 shares are performance-based. The first 239 US$500 million steps each vest 240,042 shares. The final step vests 240,002, completing the exact total. Vesting events must occur between 1 August 2026 and 29 January 2033.

This architecture gives observers two independent receipts. Time-based vesting shows that the partnership has survived specified dates. Performance vesting shows that Marvell has recognised cumulative revenue inside the contractual perimeter. Neither receipt identifies a product shipment, margin or Google cash payment by itself.

One US$500 million step is also not a quarterly forecast. Revenue can cross a cumulative threshold at any point after enough qualifying sales have been recognised. A quarter may contain less than one step, one step or several. The useful public sequence is cumulative qualifying revenue, steps vested and shares actually exercised—not a straight line obtained by dividing US$120 billion by years.

The product perimeter begins with control

The warrant's most consequential definition is not “AI chip”. It is “Qualifying Product”.

A custom semiconductor enters the revenue meter only when a written statement of work or product exhibit governs its development and Google or an affiliate holds what the contract calls a Sale Control Right. That right must let Google determine whether Marvell may sell the product to any customer outside Google's group.

Control over a feature, function or component is expressly insufficient. The right must apply to the product as a whole. The parties can also add products by mutual written designation.

This test ties the equity reward to a specific economic sacrifice. A product that Google controls at the whole-product level may be harder for Marvell to sell across a wider customer base. In return, revenue from that product can unlock equity for Google. A reusable component lacking that control may contribute to Marvell's ordinary revenue without advancing the warrant.

The distinction prevents the public programme list from becoming the contractual list. The 8-K names five custom-silicon areas, but the warrant says Exhibit A contained the only qualifying products at issue. The filed public exhibit ends at the signatures and does not expose that schedule. It is therefore possible to know the programme's breadth without knowing which current designs are inside the vesting meter.

The missing list is not a minor disclosure gap. It is the denominator of the commercial bargain: how much product exclusivity or sale control Marvell accepts in exchange for demand and how much of its custom-silicon platform remains reusable elsewhere.

Revenue is recognised, adjusted and contestable

The contract does not count purchase orders. It counts revenue recognised under US GAAP during the measurement period from or on behalf of Google and its affiliates for qualifying products.

Actual rebates, refunds, credits and returns adjust the total. The calculation excludes the contra-revenue effect created by allocating the warrant's own value. That exclusion matters: otherwise the customer incentive could reduce the same contractual meter that determines how much of the incentive vests.

Marvell also promises not to take an affirmative action, including a policy change, with the specific intent of prejudicing or reducing what qualifies as revenue in a quarter. That clause recognises the conflict embedded in the ledger. Marvell calculates the revenue, but a lower number delays Google's vesting.

The answer is a certification and dispute process. Within 30 days after each fiscal quarter, Marvell must notify Google when a vesting event occurred and provide quarterly qualifying revenue unless Google waives the latter. Google then has 20 business days to object or say that it cannot verify the calculation.

If the parties fail to agree during the next 30 business days, Google may send the dispute to an internationally recognised independent accounting firm. The firm acts as an expert rather than an arbitrator, yet its decision carries the force of an arbitral award and is final absent manifest mathematical error. Google normally pays. Marvell pays for the specific review if the firm finds an unreported vesting event.

This process makes qualifying revenue operationally real while keeping it mostly bilateral. A threshold can be crossed before public investors see a detailed customer ledger. Later Marvell filings may reveal warrant vesting or accounting, but the contract does not promise a public quarter-by-quarter Google revenue series.

The strike creates a second gate

Vesting gives Google an option. It does not make Google a shareholder for those shares.

The initial exercise price is US$206.58. Google may pay cash for vested shares or use a cashless formula. In a cashless exercise, the number delivered is reduced according to the difference between Marvell's 30-trading-day volume-weighted average price and the strike.

At expiry, vested and unexercised rights convert automatically only when that VWAP is above US$206.58. They convert cashlessly. Unvested rights become void.

This makes the warrant a two-gate instrument. Revenue decides whether performance rights vest. Marvell's future equity value decides whether exercise is economically useful and how many shares a cashless exercise delivers.

Even the US$12.182 billion obtained by multiplying the full share ceiling by the strike is not promised cash. Full vesting may never occur and Google can choose cashless exercise. The number is only the theoretical payment if every right vested and every share were bought for cash at the unadjusted strike.

The issue-date share denominator provides scale without predicting the result. Marvell had 876,926,613 common shares outstanding. The maximum warrant equals about 6.72% of that fixed count: approximately 0.16% for the time portion and 6.57% for the performance portion.

That is not a forecast of 6.72% dilution. Marvell's denominator will change through compensation, repurchases, acquisitions and other capital activity. Only vested shares can be exercised. Cashless settlement issues fewer shares than full cash exercise. The strike may be out of the money. Exercise may occur in blocks over years.

There is also an initial beneficial-ownership limit of 4.999% after an exercise. Google can amend or waive that percentage by notice effective on the 61st day, subject to the contract's transaction rules. The limit stages how much can be exercised at one moment; it does not erase the remainder of a vested warrant.

Ownership has a disposal clock

The warrant also separates exercise from sale.

The time-based portion cannot generally be transferred before the earlier of 10 November 2027 and the Kestrel Product Launch. Kestrel launches for this purpose when Google completes qualification of the final production version of that named custom product.

Once shares can be sold, Google and its affiliates generally cannot transfer more than 10% of Marvell's average daily trading volume on one trading day, measured over the preceding ten days. An underwritten offering or block trade is excepted. Marvell must cooperate with orderly sales designed to reduce market impact.

These clauses do not eliminate capital-market pressure. They make its path observable. Vesting, exercise, beneficial ownership, registration and sale are five different events, each with its own disclosure and timing possibilities.

The old warrant is evidence, not a template

Marvell already had a revenue-linked customer warrant before Google. Its fiscal-Q1 Form 10-Q describes a fiscal-2025 warrant for up to 4.2 million shares at US$87.77. Marvell valued that older award at US$227.6 million, recognised it as a revenue reduction as qualifying revenue was recognised and reported 0.9 million shares vested by 2 May 2026.

That history shows that customer equity is an operating economic cost for Marvell, not merely a footnote about share count. It does not supply the new warrant's missing accounting.

Marvell has not yet disclosed the Google warrant's grant-date fair value, the amount of any contra-revenue, the expense or revenue timing, or the first qualifying-revenue result. The old warrant differs in size, strike, customer, product perimeter and terms. Copying its fair-value ratio or accounting balance into the Google instrument would manufacture evidence.

The same discipline applies to customer concentration. Before the Google warrant, Marvell reported one unnamed direct customer at 16% of fiscal-Q1 revenue and three customers representing 75% of gross receivables. Those figures describe real concentration, but the filing does not identify Google. The quarter also predates the new instrument. It cannot be relabelled after the fact.

Marvell's Q1 results do establish the operating scale: US$2.418 billion of revenue, 28% annual growth and US$638.8 million of operating cash. The US$120 billion full-vesting threshold is therefore enormous relative to one quarter, but it is measured across more than six years and a confidential product set. That comparison explains duration; it does not forecast a growth rate.

What the warrant actually proves

The instrument proves that Marvell and Google were willing to write a long, granular alignment mechanism around custom silicon. It proves that the commercial programme reaches several parts of the TPU system. It proves that Marvell is prepared to reserve a large equity ceiling if qualifying revenue scales.

It does not prove US$120 billion of orders. It does not prove which product is shipping, how many units Google will buy, what margin Marvell will earn, whether development costs are recovered or how much equity will be issued.

The durable market signal is a chain of receipts:

  1. a product satisfies the whole-product control test;
  2. Marvell recognises adjusted GAAP revenue for it;
  3. cumulative revenue crosses a US$500 million boundary;
  4. a tranche vests and survives any dispute;
  5. Marvell's share value makes exercise rational;
  6. Google exercises for cash or cashlessly inside ownership constraints; and
  7. issued shares are held or sold under the transfer rules.

Every skipped step turns an observable contract into an invented forecast.

Sources