Marvell Gave Google a Right to 7% of Its Stock; Google Promised Nothing

One Filing, Two Companies, $69 Billion

On the evening of August 18, Marvell filed an 8-K with the securities regulator. It disclosed that the company had issued Google LLC a warrant to purchase 58,970,907 of its shares at $206.58 each. The warrant expires on August 18, 2033.

The reaction arrived the next day, and most of it landed not on Marvell but on Broadcom. At the August 19 close, Broadcom fell 4.61% to $362.48, erasing roughly $87 billion of market value. Marvell rose 9.85% to $237.27, adding about $18 billion. Alphabet, the other party to the agreement, finished up 0.15%.

Together the two suppliers lost about $69 billion. A customer moving an order from one supplier to another is, by definition, a transaction in which one side's loss is the other side's gain. A shortfall that large means the market did not price this as a transfer of orders.

The "$12.2 billion deal" figure circulating in headlines is the product of the first two lines above: 58,970,907 × $206.58 ≈ $12.18 billion. That is not money Google paid Marvell. It is money Google would pay Marvell if it exercised the entire warrant. The headline number is an exercise cost, not revenue. Google has paid nothing so far.

A Ladder With 240 Rungs

None of the warrant is immediately exercisable in full. Under the 8-K, 1,360,867 shares vest in equal quarterly installments over the first year, on the passage of time alone. The remaining 57,610,040 shares vest on sales: they are split into 240 equal tranches, and each tranche unlocks when another $500 million of custom-silicon revenue from Google accumulates. The measurement period runs from August 1, 2026 to January 29, 2033.

240 tranches times $500 million is $120 billion. That is the figure the press has been calling a $120 billion deal. It is not an order. It is the top rung of a ladder.

The distinction has practical weight. Marvell's total revenue last quarter was $2.418 billion, of which $1.833 billion came from data center. The company's own target is $16.5 billion in total revenue in fiscal 2028, and more than $10 billion in custom silicon revenue across all customers by fiscal 2029. The top of the ladder requires an average of $18.5 billion of Google sales per year for six and a half years. Sales to Google alone would have to exceed what the company expects to earn from every customer combined.

The upper rungs are therefore unoccupied in practice. On a realistic path — say $2 billion a year of Google sales — 26 tranches vest over six and a half years, about 6.2 million shares, and dilution stays under 1%. The 7% dilution figure being discussed applies only if the ladder is climbed to the top.

What Was Said, and What Was Bound

Two questions have to be asked separately here: what did the parties say, and what is the binding instrument behind it?

The stated side is broad. Marvell enters Google's TPU architecture at several points: inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute. That is not a single socket. It is several sockets at once.

The binding side is narrow. Nothing in the 8-K requires Google to purchase any specified amount; purchases are discretionary. The warrant gives Google a right. It gives Marvell no guarantee. The only obligation that binds Marvell is to deliver shares once sales occur.

The gap between the two is the gap between a supply contract and an incentive program. In the first, the buyer commits. In the second, the seller rewards the buyer for buying, using its own equity.

What Each $500 Million Costs Marvell

The cost of the warrant is not fixed. It rises with the share price. The reason is simple: each tranche unlocks at a fixed $206.58, while the market value of those shares moves.

An ordinary volume discount is a fixed percentage: however much the customer buys, the per-unit discount stays the same. This discount is indexed to the seller's own share price. The more successful Marvell becomes, the more each dollar of Google revenue costs it. Marvell shares are up about 180% this year; the exercise price was set after that run, and the warrant already carries value.

Where this cost lands in the accounts also matters. When a company issues shares or warrants to its own customer, the US standard ASC 606 generally treats it as consideration payable to a customer and records it as a reduction of revenue — not as an expense, but as a direct deduction from the top line. Marvell did not disclose its chosen treatment in the filing. If the general rule applies, reported revenue from Google sales will run below the cash collected.

The chart above is live. It shows where the stock stands today, not the single session this article describes.

Why the Loss Was Bigger Than the Gain

The question that matters: why did Broadcom lose $87 billion against Marvell's $18 billion gain?

Because the market was pricing the loss of negotiating position, not the loss of an order. When a supplier is the sole source, scarcity sets the price. The moment a second source appears, far more revenue than the volume actually transferred goes back on the negotiating table.

Consider the arithmetic. The figures are illustrative and are not Broadcom's disclosed data. Take a supplier with $40 billion in annual sales to one customer at a 60% gross margin. A second source takes a fifth of the volume: $8 billion of revenue leaves, and $4.8 billion of gross profit goes with it. That is the visible loss.

The invisible loss sits in the remaining $32 billion. That volume is now negotiated against a live alternative. If pricing slips just 5%, another $1.6 billion disappears — and all of it comes out of margin, because nothing on the cost side moves with it. The volume loss happens once and is visible. The pricing effect repeats every year, across the entire book.

That second item is what the market priced on August 19. Market capitalization reflects the present value of future earnings, not one year's profit, and a permanent margin decline always costs more than a one-time order loss.

Broadcom Learned This in June

For Broadcom shareholders, none of this was entirely new. On June 4, Macquarie cut the stock from Outperform to Neutral, lowered its price target from $513 to $437, and reduced its 2028 earnings estimate by 21%. The reason was not Marvell. It was Google's work with MediaTek and the expansion of its in-house design team. Analyst Arthur Lai projects Broadcom's share of Google's AI chip spending falling from roughly 95% in 2026 to 80% in 2027 and 65% in 2028.

Google's structure had been visible for some time as well. The company split TPU supply across four partners: Broadcom holds the high-end training chip under an agreement running to 2031, MediaTek handles the cost-optimized inference variant, Marvell takes the memory processing unit and inference side, and Intel supplies server processors and networking infrastructure. Dependence on a single partner is a pricing risk for the buyer. Google chose to spread it.

The August 20 data carries its own message. The SOXX semiconductor ETF barely moved all session, closing at $517.27. The market read the event as a shift between two suppliers rather than a problem spreading across the sector.

The Other Side

Both readings hold together, and both come from the same contract. The difference is which end of it you are standing at.

What Is Left

Broadcom's business is intact: the company guides to more than $100 billion of AI semiconductor revenue in fiscal 2027 and reported over $30 billion of AI bookings last quarter. The Google agreement runs to 2031. What changed is not the business but the position that sets its price.

Marvell's picture cuts both ways too. The company gained a second large customer that reduces its concentration in Amazon, and in exchange agreed to pay for part of every sale to that customer in its own stock. The cost looks small now. If the stock rises, it grows.

This article draws on the warrant terms in Marvell's 8-K filing dated August 18, 2026, revenue figures from the company's own investor releases, estimates relayed from Macquarie and Bank of America client notes, and price data reported by CNBC, Reuters and Investing.com. Analyst estimates rest on each firm's own methodology and differ from one another. The margin calculation is illustrative and is not data disclosed by Broadcom. Marvell did not state its accounting treatment for the warrant in the filing; the assessment here rests on the general rule under ASC 606.