Google's $12.2B Marvell warrant: compute gets a second source

Marvell granted Google a warrant to buy 58.97 million shares at $206.58 each — $12.18 billion if fully exercised — in exchange for building custom silicon around Google's TPUs. Most of the warrant only vests if Google hits purchase targets running through fiscal 2033. Broadcom, which had been Google's primary custom-chip partner, dropped more than 5% on the news. This is what dual-sourcing looks like when the buyer is large enough to pay in equity.

What actually happened

Per Reuters, the deal covers a broad slice of the stack that sits around Google's tensor processing units: processors that run the models, silicon that manages storage, and the networking parts that move data between them. Analysts put the potential revenue at roughly $120 billion through fiscal 2033 if Google hits the targets the warrant is tied to.

The structure is the interesting part. Google does not pay $12.2 billion. Google gets the right to buy at a fixed strike, and that right unlocks in proportion to how much it actually buys from Marvell. Marvell gets a customer with a contractual reason to keep ordering. Google gets a supplier with a contractual reason to prioritize it — and a hedge against being single-sourced on Broadcom.

Marvell rose roughly 8%. Broadcom fell more than 5%. Morningstar's William Kerwin framed it as "a growing pie at Google for new sources, rather than a competitive displacement of Broadcom" — Google adding capacity, not swapping vendors.

Why this matters for your business

You are not negotiating warrants. But you are buying the output of this supply chain every time you send a token, and the shape of these deals tells you where inference prices go.

The signal here is that the largest buyer of AI compute on earth just decided single-sourcing its accelerator supply was too risky to keep doing. Google has more leverage over Broadcom than you will ever have over any vendor, and it still paid — in equity, in an eight-year commitment — to have a second option. Read that as a floor on how seriously to take your own concentration risk, not a ceiling.

Two things follow for a small team:

Cheap inference is a build-out, not a trend. Deals structured on 2033 purchase targets mean capacity keeps arriving for years. Price cuts on frontier models are likely to continue, but they are the byproduct of a capex race — not a promise in your contract. Do not build a margin model that requires next year's price.

Second-sourcing is a design decision you make once. The reason Google can add Marvell is that it controls the interface — the TPU architecture is theirs, the suppliers plug into it. Your equivalent is keeping your model calls behind an interface you own, so a provider change is a config edit. Teams that hard-coded one SDK spend the savings from the price war on a migration.

Sources: Reuters via Yahoo Finance, CNBC.