Marvell Grants Google a $12.2 Billion Warrant in Landmark Custom AI Chip Deal

What Happened

On August 19, 2026, Marvell Technology disclosed in a securities filing that it has issued Google a warrant to purchase up to 58,970,907 shares of its common stock at an exercise price of $206.58 per share. Fully exercised, that right is worth approximately $12.18 billion — and if Google eventually takes every tranche, it could own roughly 6.4% of the chip designer, measured against Marvell’s current diluted-share guidance of around 915 million shares.

The warrant is the financial anchor of a much bigger story: an expanded commercial agreement, dated July 29, 2026, under which Marvell will develop a wide range of custom semiconductor products for Google’s AI infrastructure. The stock market’s verdict was immediate. Marvell shares rocketed as much as 13% to around $243 in early trading, while Broadcom — Google’s incumbent custom-chip partner for the better part of a decade — slid 3–5% on the same headline.

This is not a routine supplier contract. It is one of the largest customer warrant structures in the semiconductor industry’s history, second in recent memory only to AMD’s October 2025 deal with OpenAI, and it signals a structural shift in who designs the silicon behind Google’s TPU empire.

The Anatomy of the Deal

According to Marvell’s disclosure, the agreement covers products that “attach to the TPU ecosystem” — that is, the custom silicon surrounding Google’s tensor processing units rather than (necessarily) the TPU die itself. The disclosed scope includes:

• AI inference accelerators — dedicated silicon for running trained models at scale
• Storage controllers — managing the fast storage layers that feed accelerators
• Network interface controllers (NICs) — the connective tissue between racks of compute
• Memory interface controllers — silicon that sits between processors and DRAM
• Near-memory compute — architectures that move processing closer to memory to dodge the memory-bandwidth wall

The warrant’s vesting mechanics are the most revealing part of the structure. Around 1.3 million shares vest on a time-based schedule over the first year — a small good-faith slice. The overwhelming majority of the remaining ~57.7 million shares vest only as Google actually purchases qualifying products from Marvell, beginning in Marvell’s fiscal Q3 2027 and running through fiscal 2033. Each additional $500 million of cumulative qualifying revenue unlocks another revenue milestone. Run the math to the very end and the ceiling works out to roughly $120 billion of cumulative qualifying revenue over about six and a half years.

The warrant was issued on August 18, 2026, and remains exercisable until August 18, 2033 — a seven-year horizon.

What It Is — and What It Isn’t

The $120 billion figure is getting the headlines, but it deserves careful reading. As several analysts quickly pointed out, Google has not received $12.2 billion of Marvell stock today, and Google has not committed to purchasing $120 billion of Marvell products. There is no $120 billion backlog. The warrant structure simply creates room for that much business — a framework whose maximum size tells you the ambition, while the vesting schedule tells you the reality: Google only gets shares as it actually buys chips.

Even under conservative assumptions, the numbers are striking. If Google hits just 50% of the purchase milestones, that implies roughly $60 billion of cumulative qualifying revenue — around $9 billion per year spread mechanically across the deal window, which is nearly the size of Marvell’s entire fiscal 2029 custom-silicon revenue target. Full vesting would imply roughly $18 billion per year, above Marvell’s current total-company fiscal 2028 revenue outlook.

For context on Marvell’s baseline: the company generated $8.2 billion of revenue in fiscal 2026, guided fiscal 2027 to approximately $11.5 billion and fiscal 2028 to $16.5 billion, and has set a target of exceeding $10 billion in custom silicon revenue by fiscal 2029 — based on a $55 billion custom-silicon TAM and a targeted 20% share.

Why Google Is Diversifying Its Custom Silicon

For over a decade, Google’s TPU program has been developed largely in partnership with Broadcom, whose custom-silicon (XPU) business is one of its most prized franchises. The companies expanded that arrangement in April 2026 with a long-term agreement covering future TPU generations and components for Google’s next-generation AI racks through 2031.

So why add Marvell now? Three pressures converge:

1. Nvidia substitution economics. Google, Amazon, Meta, and Microsoft are all racing to build custom accelerators as cheaper, more controllable alternatives to Nvidia’s GPUs. Every hyperscaler wants multiple capable design partners to compress costs and timelines.
2. Capacity and leverage. Dual-sourcing the TPU ecosystem gives Google negotiating power over its incumbents and hedges against any single partner’s roadmap slippage.
3. Scope beyond the accelerator. Marvell’s portfolio — optical DSPs, Ethernet switching, PCIe/CXL, data-center interconnect, and the Celestial AI photonics platform acquired in December 2025 — lets it sell not just a chip but a constellation of surrounding silicon. The Google deal is the first public proof that Marvell can win a custom accelerator program and pull the rest of its portfolio into the same customer relationship.

Notably, Marvell is no stranger to this warrant playbook: in December 2025 it issued Amazon a warrant tied to purchases of Celestial AI photonic-fabric products, vesting through 2030. The Google deal follows the AMD–OpenAI template (warrant for up to 10% of AMD tied to OpenAI’s chip purchases) at even larger absolute scale.

Market Reaction: A Zero-Sum Read

The trading tape told a clean story. Marvell surged as much as 13% to about $243.66, rebounding from a sharp Tuesday slide and extending a 155% year-to-date advance. Broadcom fell 3–5% — with the market reading Marvell’s gain as Broadcom’s loss in a fixed-size custom-silicon pie. Alphabet shares were essentially unchanged around $342, suggesting investors view this as a supplier reshuffle rather than a strategic pivot.

Adjacent connectivity names caught a bid: Astera Labs gained about 1% and Credo Technology rose 3%, as Marvell’s win was read as a rising-tide signal for custom-silicon-adjacent suppliers. The SPDR S&P Semiconductor ETF added roughly 1%.

The sharpest divergence — Marvell up double digits, Broadcom down — is the market’s clearest statement yet that hyperscalers are willing to rebalance decade-old supplier relationships in the space of a single filing.

What It Means for the AI Chip Landscape

The implications stretch beyond two tickers:

• Custom silicon is the battleground. Hyperscaler XPUs are no longer side projects; they are the primary threat to Nvidia’s margins and the primary growth engine for design partners like Broadcom and Marvell.
• Warrant-based customer alignment is now a pattern. AMD–OpenAI, Marvell–Amazon, and now Marvell–Google show hyperscalers and AI labs taking contractual equity stakes in their suppliers, aligning incentives over 5–10 year horizons without outright acquisition.
• Broadcom’s moat is deep but no longer exclusive. Broadcom keeps the core TPU franchise (its April 2026 agreement runs through 2031), but Marvell now owns meaningful TPU-ecosystem territory — inference accelerators, controllers, and near-memory compute — that used to default to the incumbent or to Google’s internal teams.
• Watch fiscal 2027–2029. The vesting schedule begins in Marvell’s fiscal Q3 2027. Each quarterly filing from that point becomes a progress report on Google’s actual purchase volume — and a real-time indicator of whether the $120 billion ceiling remains a marketing number or becomes an operating reality.

Sources