Google, Marvell Sign $12.2B Custom AI Chip Warrant

Marvell Technology has handed Alphabet’s Google a stock warrant worth as much as $12.18 billion in exchange for a multi-year commitment to build custom AI silicon, deepening a partnership that both sides now describe as a cornerstone of Google’s next generation of AI infrastructure. Marvell disclosed the deal in an 8-K filing on Tuesday, August 19, 2026, and the company’s shares jumped more than 11% in premarket trading before opening up about 10% on the news; rival Broadcom — long the incumbent partner on Google’s in-house TPU program — fell more than 3%.

The structure is unusual for a chip-supply agreement. Rather than a straight commercial contract, Marvell issued Google a warrant to buy up to 58.97 million common shares at an exercise price of $206.58 each, roughly a 6.7% stake in the company if fully exercised. Vesting is tied almost entirely to Google actually spending on Marvell silicon: one tranche unlocks for every $500 million in qualifying custom-chip revenue Marvell books from Google, with the full package earning out only if the relationship generates on the order of $120 billion in cumulative sales through the end of Marvell’s fiscal 2033. The warrant expires August 18, 2033.

What Google is buying

Marvell said its scope of work under the expanded partnership goes well beyond a single accelerator part. The company will develop AI inference accelerators, along with storage, networking, and memory interface controllers and near-memory computing technologies for Google, positioning Marvell as a broad component supplier across the rack rather than a point solution.

That framing matters. For Google, the calculation is that AI serving cost is now driven not just by the accelerator die but by everything around it — the interconnect fabric that binds chips into a coherent training or inference pod, the memory hierarchy that feeds the compute, and the storage tier that moves parameters and activations in and out. Owning custom versions of those blocks, rather than buying merchant parts, lets a hyperscaler squeeze more of the total system cost onto its own roadmap. Google has been leaning into that logic for years with its TPU program; the Marvell deal is a bet that a specialized silicon partner can compress that roadmap further and faster than an internal team on its own.

For Marvell, the incentive structure is the point. Only about 1.36 million of the shares vest in equal quarterly installments during the first year, guaranteeing at least some benefit up front. The remaining 57.61 million shares vest across 240 tranches through fiscal 2033, each unlocked only by hitting a $500 million revenue milestone. If Google hits the top end of the plan, that stake makes it the fifth-largest investor in Marvell. If Google diverts custom-chip spend elsewhere, most of the warrant simply never vests and Marvell dilutes far less. The deal is written to align interests over an eight-year horizon in a market where accelerator sourcing decisions are anything but stable.

Why this rattles Broadcom

Google’s custom-chip supply chain has been dominated for most of the TPU era by Broadcom, which co-designs and manufactures Google’s tensor processors and books billions in annual revenue from the account. Broadcom’s shares fell more than 3% on the Marvell news, and the reason is straightforward: an eight-year, hundred-billion-dollar potential commitment to a competing custom-silicon partner is exactly the kind of announcement that recasts what “the Google TPU business” is worth to Broadcom.

Nothing in the Marvell filing suggests Broadcom is being displaced from the existing TPU program in the near term. What it does suggest is that Google is deliberately building a two-supplier posture for the workloads that surround the accelerator — inference-optimized ASICs, custom networking, custom memory controllers — so that no single partner controls the price and roadmap of the whole system. For Broadcom, that changes the negotiating asymmetry going forward. For Marvell, it changes the size of the addressable pipeline in a way its own earnings so far have not reflected.

The timing is pointed. Marvell reports its Q2 fiscal 2027 earnings on August 27, 2026 — eight days after the announcement — one day after Nvidia. For a company whose stock has trailed the AI-chip leaders for stretches of this cycle, framing an eight-year customer relationship as a warrant deal ahead of the print puts the strategic argument in front of investors before the quarterly numbers do.

The hyperscaler custom-silicon race

The Marvell announcement is the latest data point in a broader shift the hyperscalers have been executing since AI serving costs stopped being a rounding error. Amazon has Trainium and Inferentia. Microsoft is scaling its own family of accelerators — see our coverage of the Microsoft Maia 300 order at TSMC — as it works to cut Nvidia dependency. Meta has moved production of its in-house Iris AI chip forward as it builds out its own training and inference stack. Anthropic is reportedly in talks with Samsung for a custom AI chip of its own.

The pattern is the same across all four: as accelerator spend approaches the largest single line item in a hyperscaler’s AI capex budget, owning more of that silicon roadmap becomes a first-order strategic question rather than a marginal cost optimization. Merchant chips from Nvidia stay in the mix — they are still the fastest way to stand up frontier training runs — but the steady-state inference workloads that generate most of the tokens are increasingly a home-cooked business.

What Google is buying with the Marvell warrant, in effect, is a way to accelerate that internal roadmap without absorbing all of it in-house. Marvell brings decades of custom-ASIC design experience, established relationships with TSMC’s advanced nodes, and a portfolio of interconnect and memory-controller IP that would take years to build from scratch. Google supplies the workload knowledge and the volumes that make a custom design pay for itself. The warrant aligns the two companies to keep grinding on that co-design over an unusually long horizon.

The financial mechanics

At a strike of $206.58 per share, the warrant is priced roughly at Marvell’s recent trading range, meaning the value Google captures is entirely a function of how much the stock rises between now and each tranche vesting. If Marvell trades flat, the warrant is worth little; if the eight-year plan works and the AI-silicon franchise re-rates, Google’s warrant becomes a multi-billion-dollar unrealized gain layered on top of the savings it books from a cheaper, more integrated silicon supply.

That structure is closer to a venture-style equity stake than to a traditional supplier discount. It transfers some of the payoff from Marvell’s shareholders to Google, but only in the scenarios where Marvell itself wins — that is, where Google’s custom-chip purchases actually materialize at the promised scale. It also creates a public marker of the relationship’s size, which is why the reaction ran through Broadcom’s stock before Marvell’s earnings call.

The $120 billion cumulative sales figure through fiscal 2033 is the number that will get quoted most often, and it deserves care. It is a ceiling — the aggregate spend required for the full warrant to vest — not a floor Marvell has committed to hit. Investors will price the deal at some fraction of that ceiling; how large a fraction depends on how quickly the first vesting tranches roll in and how visibly Marvell breaks out custom-silicon revenue in its quarterly disclosures.

What it means

Who wins if it works: Marvell most directly — a warrant-linked, eight-year commitment from a hyperscaler is the kind of demand signal chip designers rarely land, and it validates a business model that has been mostly overshadowed by Broadcom’s TPU relationship. Google wins if the partnership actually collapses the accelerator supply-chain cost the way TPU has collapsed training cost, and if a second custom-silicon partner reduces Broadcom’s leverage on price and roadmap in the process.

Who’s exposed: Broadcom, most obviously — not because it loses today’s TPU work, but because the strategic value of being Google’s sole custom-chip partner has just been publicly reduced. The read-across to Nvidia is subtler: another hyperscaler locking in a large custom-silicon supply chain is another slice of AI serving demand routing around merchant GPUs, in line with the trend we traced in coverage of Microsoft’s Maia 300 order at TSMC. Nvidia’s frontier training business is unlikely to be dented; its inference share at Google over the long run is the question.

What to watch next: three things. First, Marvell’s August 27 earnings call, where management will be pushed to describe how the Google agreement is expected to phase in and how much of the ceiling is realistic on any near-term horizon. Second, whether Google formally moves any part of its TPU program to Marvell versus keeping the deal to inference accelerators and system components around Broadcom-designed TPUs — the language in Marvell’s filing leaves both readings open. Third, whether Broadcom counters, either with a similarly structured incentive deal for another hyperscaler or with a public restatement of its own Google roadmap. On an eight-year contract, the first year’s disclosures will set expectations for the whole thing.