What did Google and Marvell announce in August 2026?

Marvell and Google agreed to co-develop custom AI silicon attaching to Google's TPU ecosystem, and Marvell issued Google a stock warrant dated August 18, 2026 for up to 58,970,907 shares at $206.58 each — roughly $12.2 billion if fully exercised. The warrant vests in tranches tied to Google's chip purchases, reported at one tranche per $500 million of orders. Marvell shares rose about 10% on the news.

Why would Google take stock warrants instead of just buying chips?

Because it converts a purchase order into an ownership claim without spending cash up front. Google pays for chips it wants anyway, and each $500 million tranche of purchases unlocks the right to buy Marvell stock at a fixed $206.58. If the partnership drives Marvell's business, Google captures part of the upside it created. It also aligns the supplier: Marvell's largest customer becoming a potential shareholder is a strong commitment signal to a chipmaker deciding where to point its engineering.

Is this the same as Nvidia's circular financing deals?

It is the same family of structure with the direction reversed. In Nvidia's arrangements the supplier invests in or backs the customer, which critics call circular because supplier money funds purchases of the supplier's product. In the Google-Marvell deal the customer takes equity exposure to the supplier, contingent on the customer's own spending. That is closer to a conventional performance warrant than to circular demand creation, though both structures blur the line between a commercial contract and an investment.

Does this threaten Broadcom's position in Google TPUs?

It pressures it without displacing it. Broadcom has been Google's principal TPU co-design partner, and reporting around August 20, 2026 framed the Marvell agreement — alongside AMD's involvement — as Google widening its supplier base rather than switching. The strategic logic is second-sourcing: custom silicon programs concentrated in one partner carry schedule and pricing risk that a hyperscaler running its own accelerator roadmap cannot ignore.

Quick Answer

What Is the Google-Marvell TPU Chip Deal? (Aug 2026)

Published: August 24, 2026

The Short Answer

Google and Marvell agreed to co-design custom AI chips for the TPU ecosystem, and Google received a warrant to buy up to about $12.2 billion of Marvell stock that vests only as it places chip orders. The deal became public the week of August 17, 2026 and sent Marvell shares up roughly 10%.

	Detail
Announced	Week of August 17, 2026 (Reuters, CNBC)
Warrant dated	August 18, 2026
Shares covered	Up to 58,970,907
Strike price	$206.58 per share
Implied value if fully exercised	~$12.2 billion
Vesting trigger	Tranches tied to chip purchases, reported at each $500 million
Scope	Custom silicon “attaching to the TPU ecosystem”
Market reaction	Marvell +~10%
Next checkpoint	Marvell fiscal Q2 2026 results, August 27, 2026

What Was Actually Agreed

Two things, bolted together.

The commercial half: Marvell will help Google develop custom AI silicon. The language in reporting is specific — products that attach to the TPU ecosystem. That points at the components around the accelerator rather than the accelerator core itself: networking, interconnect, SerDes, custom I/O and the high-bandwidth plumbing that decides whether a pod of TPUs behaves like one machine or a thousand arguing ones.

The financial half: a warrant, not a purchase. Google holds the right to buy nearly 59 million Marvell shares at $206.58. It does not own them, has not paid for them, and only earns the right in slices as it places orders — reported at one tranche per $500 million of chip purchases.

That structure is the interesting part. Google pays for silicon it was going to buy anyway, and each payment converts into an option on the supplier’s equity.

Why the Structure Matters

For Google, the warrant does three things at once:

Captures the value it creates. A hyperscaler committing billions to a chipmaker’s custom program visibly increases that chipmaker’s revenue and credibility. Historically the supplier’s shareholders captured that. Here Google takes a share of it.
Costs nothing today. No cash outlay, no balance-sheet position until exercise. The economics only trigger on spending that was already planned.
Buys supplier commitment. Custom silicon programs are multi-year and engineering-intensive. A supplier whose largest customer may become a shareholder allocates its best engineers accordingly.

For Marvell, it converts a customer into a stakeholder and provides an unusually credible demand signal. Management has guided custom AI chip revenue to more than double to over $4 billion in the coming year and to exceed $10 billion by 2028 — forecasts that need exactly this kind of anchor customer to be believable. Fiscal Q2 results on August 27, 2026 are the first public test.

The dilution is real but conditional: shares only issue if Google spends, and if Google spends at that scale, Marvell’s revenue base has changed enough to absorb it.

The Second-Source Story

Google has run its own accelerator program for a decade, historically with Broadcom as principal co-design partner. Reporting around August 20, 2026 framed the Marvell agreement — with AMD also named in the mix — as Google broadening the supplier set rather than replacing anyone.

The motivation is unglamorous and correct: single-sourced custom silicon is a schedule and pricing risk. If one partner owns your accelerator roadmap, its capacity constraints become your product delays and its margin expectations become your cost floor. A hyperscaler spending at Google’s scale can afford a second design partner purely as insurance, and the option value shows up the first time a tapeout slips.

For Broadcom the read is pressure, not displacement. For MediaTek and others chasing hyperscaler custom-silicon work, it is evidence that the seats at this table are being allocated now.

How This Fits the 2026 Financing Pattern

This deal belongs to a broader 2026 pattern of AI infrastructure agreements that mix commercial contracts with equity exposure:

Nvidia pursued AI infrastructure arrangements reported by Bloomberg in late July 2026 at more than $750 billion in aggregate, including a partnership with SK Group valued above $500 billion, and discussions reported around providing a guarantee of up to $250 billion to help OpenAI lease compute. Reuters reported on August 10, 2026 that Nvidia was partnering with Wall Street firms on a $500 billion AI financing effort.
Critics call these structures circular financing — supplier capital funding purchases of the supplier’s own product, which can flatter demand.

The Google-Marvell deal runs the other direction. The customer takes equity exposure to the supplier, and only in proportion to money the customer chooses to spend. That is much closer to a conventional performance warrant than to circular demand creation.

The honest summary: not every equity-flavoured chip deal in 2026 is the same animal, and the direction of the money is the tell. Supplier funds customer is where the circularity concern lives. Customer earns supplier equity by buying is ordinary incentive alignment with a large number attached.

What to Watch Next

August 27, 2026 — Marvell fiscal Q2 2026 results, and whether custom AI silicon guidance moves.
Tranche disclosure — how many $500 million tranches actually vest, which is the only real measure of the partnership’s size.
Scope creep — whether “attaches to the TPU ecosystem” expands toward accelerator core work over time.
Broadcom’s response — pricing, capacity commitments, or a comparable structure elsewhere.