Marvell Just Handed Google $12.2 Billion of Its Own Stock. Google Hasn’t Promised to Buy a Single Chip.

On Wednesday, the chipmaker Marvell Technology disclosed an expanded commercial agreement to build custom silicon for Google. Attached to it was something less ordinary: a warrant giving Google the right to buy 58.97 million Marvell shares at $206.58 apiece. Exercised in full, that is $12.18 billion of stock and roughly 7% of the company — enough to make Google one of the largest shareholders in its own supplier.

Google does not get the shares for signing. It earns them by shopping. About 1,360,867 shares vest quarterly over the first year on a simple time schedule. The other 57.6 million vest in 240 equal tranches of roughly 240,042 shares, and one tranche unlocks for every $500 million of custom-product revenue Google generates for Marvell through fiscal 2033.

Multiply it out and the ceiling on the deal is $120 billion of chip purchases.

Marvell’s most recent quarter was $2.4 billion in total revenue — $1.8 billion of it from the data center segment — across every customer the company has.

The Number That Does Not Fit

Annualise that quarter and Marvell is a roughly $9.6 billion-a-year business. Full vesting of the warrant requires $120 billion from Google alone, over about six and a half years. That is around $18.5 billion a year from one customer — nearly double what every Marvell customer combined delivered over the last four reported quarters.

For scale in the other direction: Marvell’s own stated target is to exceed $10 billion in custom revenue in fiscal 2029, across its entire customer base. The Google warrant is written against a number an order of magnitude past the company’s public ambition.

Which is the point. The final tranches are not a forecast. They are a ceiling deliberately set where nobody expects to arrive, so that every realistic outcome sits somewhere on the curve rather than at the end of it.

What Marvell Is Actually Selling

The deal is not about the TPU itself. Google’s Tensor Processing Units are the accelerators that run its AI models, and Broadcom has designed and supplied them for years — an arrangement that runs through 2031, with MediaTek cleared to supply cost-optimised variants. That did not change on Wednesday.

What Marvell won is everything bolted around the accelerator, the unglamorous silicon that decides whether a TPU actually runs at speed:

• Memory interface controllers, including custom high-bandwidth memory integration across all three major memory suppliers.
• Storage controllers — PCIe Gen 6 SSD parts such as the Bravera SC6 line.
• Network interface controllers plus PCIe Gen 6 and CXL controllers, which move data between chips and racks.
• Electro-optics, including the Photonic Fabric technology for memory shared across machines.
• Inference accelerators and near-memory compute aimed at the decode-heavy work that agentic AI generates.

Crucially, most of that portfolio already exists at Marvell — qualified, shipping, in production. Google is not commissioning a research programme. It is buying a second, independent source for the parts of its AI stack that a single supplier currently controls.

The Rebate Hiding Inside the Structure

The strike price is the detail worth slowing down on. At $206.58, Google is not being given shares — it is being given the right to buy them at roughly what they were worth the day the deal was announced. The stock traded near $230 afterwards. On that price, a single 240,042-share tranche carries about $5.6 million of intrinsic value.

Against the $500 million of purchases that unlocked it, that is an effective discount of about 1.1%. Modest — until you notice the loop it creates. The purchases that vest the tranches are the same purchases that grow Marvell’s revenue, which lifts Marvell’s share price, which widens the gap between the market price and Google’s fixed $206.58 strike. Google’s discount gets deeper the more successful its supplier becomes, and Google is the one making it successful.

The Precedent, and How This One Differs

This is the second time in under a year that a chipmaker has paid a giant AI customer in equity. In October 2025, AMD issued OpenAI a warrant for up to 160 million shares — about 10% of AMD — tied to deployments of up to 6 gigawatts of Instinct GPUs. AMD stock jumped nearly 24% the day it was announced.

The two deals look alike and are not. AMD’s warrant carries an exercise price of one cent per share. That is not a discount; it is a transfer. Marvell’s is priced at $206.58, which means Google has to write a cheque of roughly $12.2 billion to collect the full stake, on top of the $120 billion of chips that would earn it.

The difference tracks the balance of power. AMD needed OpenAI to validate a GPU line competing with NVIDIA, and paid accordingly. Marvell already ships qualified silicon into hyperscale data centers; it is buying share of wallet, not credibility. Either way, performance-vesting equity has gone from a novelty to something close to standard practice in how the largest AI buyers negotiate with silicon vendors.

Who Actually Lost

The market read this as a transfer, and it was fairly specific about from whom. Marvell rose roughly 10%. Alphabet barely moved. Broadcom fell about 4.6%.

Broadcom did not lose the TPU. Its design agreement runs to 2031 and Wednesday’s announcement did not touch it. What Broadcom lost is the assumption — the one baked into its valuation — that when Google needs custom silicon, there is exactly one phone number to call. Marvell’s attach across memory, storage, networking and optics establishes a second qualified vendor inside Google’s stack, and qualified vendors have a way of moving up the stack over time.

That is why a deal Google is not contractually obliged to spend a dollar on still knocked several percent off a rival’s market value. The drop was not about revenue lost this quarter. It was about a monopoly that stopped being one.

What To Watch

• The first performance tranche. Time-vested shares tell you nothing. The first 240,042 shares that vest on revenue are the first hard evidence Google is actually spending, and Marvell will have to disclose the custom-product revenue that triggers them.
• Marvell’s data center line. $1.8 billion a quarter today. Google reaching even a fraction of the implied run rate would be visible there within a year, or it will not be happening at all.
• Whether Broadcom’s 2031 agreement gets renegotiated. An incumbent facing a qualified second source usually responds on price before it responds on roadmap.
• Who copies the structure next. Two deals in eleven months is a pattern. Watch for the next hyperscaler that takes equity from a supplier instead of asking for a volume discount.
• Dilution arithmetic. Full exercise adds about 7% to Marvell’s share count. Existing shareholders are funding the customer’s discount, and they will only be happy about it if the revenue arrives on schedule.