Google's sweeping custom-chip agreement with Marvell Technology (NASDAQ: MRVL) drove the stock up more than 5% Wednesday and forced a reset in how investors are reading Alphabet's (NASDAQ: GOOGL, GOOG) AI silicon strategy. The tension is real but specific: Marvell winning space inside Google's Tensor Processing Unit ecosystem is not the same as Marvell displacing Broadcom (NASDAQ: AVGO), which dropped more than 5% as if it were.

The deal's structure tells the story

At the center of the agreement is a warrant giving Google the right to buy nearly 59 million Marvell shares at $206.58 each, representing an aggregate exercise price of roughly $12.2 billion if fully exercised. Google does not hold those shares today. About 1.4 million vest automatically in the first year; the remaining 57.6 million are tied to qualifying revenue, with roughly 240,000 shares unlocking for every $500 million Google sends Marvell's way. Full vesting would imply about $120 billion in qualifying revenue from Google-related purchases through early 2033, though Marvell has made clear that figure is a vesting threshold, not a spending commitment.

The risk is that the headline number gets priced as a done deal. It is a ceiling on a performance-based relationship, and that is a materially different thing.

What Marvell will actually build covers a range of specialized silicon attached to Google's TPU ecosystem: AI inference accelerators, networking chips, storage controllers, memory interface controllers, and near-memory computing technology. Google's TPUs are its in-house alternative to Nvidia (NASDAQ: NVDA) GPUs for running AI workloads, and adding a second major supplier reflects how seriously it is investing in that architecture.

The counterargument

The counterargument deserves its due: Broadcom is not losing Google. Broadcom signed a long-term agreement in April to develop future TPU generations and supply components for next-generation AI racks through 2031. What Wednesday's move repriced is the possibility that Marvell captures a growing share of Google's future AI silicon spending. That is a share-of-wallet question, not an incumbent-replacement story, and the tape appears to be running ahead of what the facts resolve.

The broader context matters here. Polymarket gives Google only a 7% chance of having the world's best AI model at the end of 2026, compared with 66% for Anthropic, 14.5% for xAI, and 8% for OpenAI. A Google that is behind on model leadership has every reason to spend harder on infrastructure. Both Marvell and Broadcom are positioned inside that spending.

On balance, the line to watch is August 27, when Marvell reports fiscal second-quarter earnings and investors will press management on the scale and timing of revenue from the Google relationship.