Marvell Technology stock popped about 10% on Wednesday after Alphabet’s Google secured the right to buy up to $12.2 billion of its shares, while Broadcom shed roughly $87 billion in market value on the very same headline. This was never really about one chip contract. It was Google buying itself a second supplier, and telling the first one in public that its seat is no longer safe.
The Warrant Is the Whole Story
Google is not writing a check. Marvell granted a warrant covering 58,970,907 shares at $206.58 apiece, worth about $12.18 billion if fully exercised, which is what sent Marvell’s stock up 10% when CNBC broke down the terms. Most of it vests against purchase milestones: roughly 240,000 shares unlock for every $500 million in qualifying chip orders Google places, running through fiscal 2033.
Read the incentive structure and the deal explains itself. Google pays nothing up front. Marvell only collects if Google actually buys silicon, and the more Google buys, the more of Marvell’s own upside Google takes home. Hit every target and the relationship could carry something close to $120 billion in cumulative procurement, an arrangement the Futurum Group summarized as a warrant vesting toward $120 billion. Google ends up the fifth-largest shareholder in a company whose order book it also controls.
Google pays nothing today, Marvell gets paid only if Google keeps buying, and every dollar of that spending hands Google a bigger piece of its own supplier.
That is cheaper leverage than an acquisition and considerably more durable than a contract. It also does not require a single antitrust filing.
Why Broadcom Took the Hit
Broadcom closed down 4.61% at $362.48, and that single session erased about $87 billion of market value. The read was brutal in its simplicity: investors treated the news as a transfer of share rather than a rising tide, and priced Broadcom as the donor.
That reaction is probably too clean. Broadcom remains Google’s principal partner on the tensor processing unit itself, under a long-term agreement CEO Hock Tan signed in April covering multiple TPU generations and the AI networking around them, a point CNBC’s own analysts stressed while assessing where Broadcom stands after the dip. Broadcom still owns the core ASIC design, the advanced packaging, the high-bandwidth memory integration and the switching silicon. Marvell did not take the crown jewel.
What Marvell took was everything bolted onto it.
What Marvell Actually Won
The expanded agreement covers products that attach to the TPU ecosystem rather than replace it:
- AI inference accelerators
- Storage controllers
- Network interface controllers
- Memory interface controllers and near-memory compute
Individually, none of those is a TPU. Collectively, they are the connective tissue of a Google data center, and they are exactly the components a hyperscaler orders in volumes that dwarf the headline processor. TrendForce reported that Marvell and AMD are now reshaping the Google TPU supply race and putting pressure on Broadcom and MediaTek in the process. A supplier that owns the attach layer today is credentialed to bid on the core tomorrow.
The Real Trade Is Optionality
Strip away the silicon and this is a story about hyperscalers refusing to be captive. Google spent years with essentially one custom-chip partner for its most strategic hardware program, which meant one company’s pricing, one company’s roadmap and one company’s capacity constraints. The warrant fixes that without Google spending cash it would rather put into compute.
The structure is spreading. Nvidia agreed to backstop $105 billion in financing so OpenAI could build out an Ohio data center full of Nvidia chips, a deal that got scaled back sharply from its original size after investors started asking uncomfortable questions about circular demand. Anthropic has been building its own custom-silicon team. Equity, warrants and vendor credit are becoming the currency of AI infrastructure because ordinary purchase orders no longer buy the certainty these companies need.
There is a difference worth marking, though. Nvidia’s arrangement points its money downstream at a customer, which is how you manufacture demand for your own product. Google’s points upstream at a supplier, which is how you manufacture competition for your own spending. The first structure flatters a seller’s revenue. The second cuts a buyer’s costs. Only one of those gets more attractive when the AI capex cycle finally slows.
What to Watch Next
The tell will be Broadcom’s next disclosure of Google-related revenue and any language about renegotiated terms. Alphabet shares barely moved on the announcement, which is its own quiet verdict: the market thinks Google gave up very little to gain a great deal of room. Whether Marvell can convert an attach-layer beachhead into core TPU work is the question worth $120 billion, and nobody will know the answer until the first milestone tranche vests.
For now, one hyperscaler has arranged its supply chain so that two of the best custom-silicon shops in the industry compete for its business, and it did so without spending a dollar it had not already planned to spend.