Google Hands Marvell a $12.2 Billion Warrant — and Reshapes the Custom Silicon Chessboard

On August 19, 2026, Marvell Technology filed paperwork that instantly became one of the most consequential AI-supplier stories of the year: Google now holds the right to buy up to $12.2 billion of Marvell stock — roughly 58.97 million shares at $206.58 apiece — as part of an expanded commercial agreement to develop custom chips that “attach to the TPU ecosystem.”

Marvell’s shares popped nearly 10% on the news. But the market reaction is the least interesting part of the story. What matters is the structure: a hyperscaler taking an equity warrant in a chip designer in exchange for a deep, multi-year custom-silicon partnership. That is a new pattern in the AI buildout — and it lands exactly where the industry’s center of gravity is shifting: away from merchant GPUs and toward custom accelerators.

What the deal actually covers

According to Marvell’s regulatory filing and reporting from Bloomberg, Reuters, CNBC, and the Financial Times, the agreement spans a broad range of technologies used alongside Google’s Tensor Processing Units (TPUs):

• AI inference accelerators — silicon purpose-built to run trained models at scale, which is where the majority of Google’s compute dollars increasingly go as inference overtakes training.
• A memory processing unit — a chip designed to work with TPU fabrication capacity, attacking one of the most stubborn bottlenecks in modern AI systems: moving data between memory and compute.
• Processors that run AI models, manage data storage, and handle networking — the supporting cast of components that surround TPUs inside Google’s data centers.

The equity component is the eye-catcher. Google’s warrant covers up to 58,970,907 shares at a strike price of $206.58. If exercised in full, Google would become one of Marvell’s largest shareholders — an extraordinary alignment mechanism between a customer and a supplier that also competes for the same design wins against Broadcom, the reigning king of custom AI silicon.

From rumor to contract in four months

The deal didn’t come out of nowhere. In April 2026, The Information reported that Google was in talks with Marvell to design two new AI inference chips: a memory processing unit and a new TPU variant targeting efficient inference. At the time, observers noted the strategic logic — for Marvell, a Google inference contract would validate its position as the second-most important custom AI chip designer in the world, behind Broadcom, which reportedly counts Google’s TPU program among its largest revenue streams.

Morningstar called Google Marvell’s “white whale” — the customer that custom-silicon investors had long hoped the company would land. Now it has, with an equity kicker that makes the commitment mutual: Google wins a cheaper, more capable supply chain; Marvell wins a hyperscaler with skin in the game.

Why hyperscalers are designing their own chips

The backstory is the slow-motion divergence between what Nvidia sells and what hyperscalers actually need. Google’s TPU program is the oldest and most mature custom-AI-silicon effort in the industry, and it has been expanding aggressively as AI workloads shift from training runs to always-on inference serving. Custom accelerators are cheaper at scale, more power-efficient for fixed workloads, and — critically — they reduce dependence on a single supplier whose data-center GPUs remain supply-constrained and expensive.

Marvell’s role in that shift is to be the design house: the company that turns a cloud provider’s architecture ideas into manufacturable silicon. Bloomberg reporting around the deal suggests custom AI chips could reach $30 billion to $50 billion in sales industry-wide in the coming years, and Marvell has now secured a direct claim on the largest and most experienced custom-chip buyer on the planet.

The timing is notable for another reason. Anthropic has been assembling its own in-house silicon team; Microsoft is preparing to unveil its Maia 300 chip; Amazon keeps iterating on Trainium. Every major AI player is hedging against Nvidia-dependence, and Google — which never fully depended on it — is deepening the moat around the ecosystem it built first.

The warrant as a new financing instrument

Strip away the chip specifics and the deal’s most novel feature is financial. Supplier financing is spreading through the AI economy: Nvidia mobilized over $500 billion of third-party capital with Wall Street partners; Broadcom is reportedly assembling a debt machine approaching $100 billion to fund chip purchases by Anthropic; and now Google is taking equity exposure in its own supplier.

There is a coherent logic here. Custom-chip deals demand years of engineering investment before revenue flows. A warrant structure lets Google lock in priority and alignment without an outright acquisition — and gives Marvell’s shareholders a powerful signal about demand durability. It also raises governance questions: when your biggest customer becomes one of your largest potential shareholders, negotiating leverage gets complicated in ways that will play out over years.

What to watch

Three things will determine whether this deal lives up to its headline number. First, whether the new inference chips and memory processing unit actually reach volume production and displace incumbent suppliers inside Google’s fleet. Second, whether Marvell can hold its second-place position in custom silicon as every hyperscaler shops for design partners. And third, whether Google exercises — or lets lapse — the warrant itself, which will tell the market exactly how the partnership’s economics are tracking.

For an industry that spent the last two years arguing about whether the custom-silicon threat to Nvidia is real, Google just put $12.2 billion of optionality behind the answer.

Sources are listed in the references section below.