On August 19, Marvell Technology filed a Form 8-K that read, at first glance, like ordinary corporate housekeeping. Buried in Item 1.01 was a number that made traders sit up: the company had issued Google a warrant to purchase up to 58,970,907 shares of its common stock at an exercise price of $206.58 apiece. Fully exercised, that is roughly $12.2 billion of Marvell equity — enough to make Google the chipmaker's fifth-largest shareholder.
Google is not writing a check for that stake. It is earning it. According to the filing, the warrant sits on top of a commercial agreement the two companies signed on July 29, 2026, covering what Marvell describes as custom silicon programs that "attach to the [tensor processing unit] ecosystem." The list of products is specific: AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute. These are not Google's marquee TPUs themselves — Google designs those — but the surrounding hardware that feeds them data, moves results between racks, and keeps memory close enough to the math to matter.
The vesting schedule is where the deal becomes unusual. Only 1,360,867 of the warrant shares vest on time alone, in equal quarterly installments over the agreement's first year. The remaining 57.6 million vest in 240 equal tranches, and each tranche unlocks only when Google spends another $500 million on Marvell's custom products. That schedule runs from Marvell's third fiscal quarter of 2027 through the end of fiscal 2033, and the warrant itself expires on August 18, 2033. Multiply 240 tranches by $500 million and the arithmetic gives you the implied ceiling: roughly $120 billion in custom-chip purchases over about seven years.
Markets did the multiplication quickly. Marvell rose as much as 14 percent intraday before settling up roughly 8 to 10 percent. Broadcom — the company that has supplied Google's custom accelerators for most of the past decade — fell about 5 percent. Alphabet barely moved.
Why It Matters
For a decade, the phrase "Google's custom chip partner" had a single answer. Broadcom co-designed the TPU line, and the two companies extended that relationship in April 2026 in an agreement reported to run through 2031. Broadcom now claims something north of 70 percent of the custom AI accelerator design-services market and has guided to more than $100 billion in AI semiconductor revenue for fiscal 2027. When a customer of that size adds a second named supplier, the market reads it as a crack in an exclusive.
The more careful reading is that Google's silicon budget grew faster than any one vendor could absorb. William Kerwin, an analyst at Morningstar, told Reuters the agreement was "a big win for Marvell" but described it as "a growing pie at Google for new sources, rather than a competitive displacement" of the incumbent. That distinction matters. Broadcom's TPU work is the compute die itself — the hardest, highest-margin piece. Marvell's list is the connective tissue around it. Both can expand at once if Google is building enough racks.
And Google is. Alphabet's TPU-related infrastructure business was projected by Citizens analyst Andrew Boone earlier this year to generate around $3 billion in 2026 and roughly $25 billion in 2027 — an eightfold jump driven largely by selling TPU capacity to cloud customers rather than consuming it internally. Once a chip becomes a product you rent to other companies, the economics change. Every incremental customer needs storage controllers, network interface cards, and memory paths, and those are commodities only in the sense that someone has to build billions of dollars of them.
The strategic backdrop is familiar by now. Google, Amazon, Meta and Microsoft have all spent years designing their own accelerators to reduce what they pay for merchant GPUs, whose top-end parts sell for tens of thousands of dollars each. Demand has skewed hardest toward inference — the stage where a finished model is actually run, over and over, at scale. Marvell's product list reads like a bill of materials for exactly that workload.
What Marvell gets in return is not just orders but alignment. A supplier holding no equity has an incentive to maximize margin per unit. A supplier whose customer holds a warrant tied to volume has an incentive to keep prices low enough that the customer keeps buying. The warrant is, in effect, a rebate paid in stock — and it only costs Marvell dilution if the revenue arrives first.
The Reaction
The immediate market response was a clean three-way split: the new entrant up, the incumbent down, the buyer flat. That pattern told you how investors scored the news — as a transfer of expected future revenue from Broadcom's column to Marvell's, with Google paying nothing today for the option to profit from the transfer.
But the sharper conversation was about structure, not scoreboards. Bloomberg reported that arrangements like this one — where a chipmaker's sales are tied to its customer's stock, or a customer's spending is backstopped by its supplier — have started to worry investors who see circular financing inflating the whole sector's valuations. The Marvell warrant landed days after Nvidia agreed to backstop up to $105 billion of residual value on an Ohio data center campus that OpenAI is leasing. It also rhymes with an October 2025 arrangement in which AMD agreed to supply OpenAI and handed the ChatGPT maker an option on a stake of up to roughly 10 percent.
Investor Jeff Gundlach recently warned that turning AI chips into an asset class has the texture of a market top. It is a fair worry, and worth stating plainly: when supplier and customer each hold paper in the other, revenue and equity value can appear to validate one another without any outside party checking the price.
There is a reasonable counterargument, and Marvell's structure supports it better than most. This warrant is narrower than an open equity investment. It vests only against realized revenue, in fixed $500 million increments, and it cannot be transferred outside Google's controlled affiliates without Marvell's consent. Nothing vests on a press release. If Google's TPU buildout stalls, 97 percent of the warrant simply never comes into existence — which is precisely the opposite of the vendor-financing arrangements that made the late 1990s telecom bust so expensive.
Even so, the disclosure timing invites a squint. The commercial agreement was signed July 29. The warrant was issued August 18. The 8-K appeared August 19. Three weeks is well inside the ordinary reporting window, but it means the market learned about a $120 billion revenue framework only after the equity terms were locked.
What Comes Next
The first hard test arrives on August 27, when Marvell reports second-quarter fiscal 2027 results. Management has guided to roughly $2.7 billion in revenue, about 35 percent growth year over year, and analysts are looking for earnings near $0.65 per share. None of the Google tranches vest until the third fiscal quarter, so this report will not show warrant-driven revenue. What it can show is whether the custom silicon business is scaling on the trajectory management has promised — data center revenue passed $6 billion in fiscal 2026 out of about $8.2 billion in total sales, and the company has guided custom silicon to grow more than 20 percent this fiscal year before accelerating in fiscal 2028.
The second test is slower and harder to observe. Each $500 million tranche is a checkpoint, and the pace at which they vest will become the cleanest public signal anyone has about how fast Google's TPU fleet is actually growing. Marvell will have to disclose vesting activity; Google will not have to explain it. Analysts who currently triangulate hyperscaler capex from earnings calls and satellite photos of construction sites will suddenly have a counter that ticks in half-billion-dollar units.
Watch, too, for what Broadcom says. Its next earnings call will be the first opportunity for management to characterize the Google relationship after the Marvell announcement, and the framing — expanding pie or contested ground — will carry more information than the stock's one-day move did.
Closing Thoughts
The skeptical read on this deal is not hard to construct. A chipmaker gave away twelve billion dollars of potential dilution to secure orders it might have won anyway. A customer with more cash than almost any company in history arranged to be paid in equity for the privilege of spending money. And the whole thing was papered during a stretch when nearly every large AI transaction has come with some financial instrument attached to it, which is usually a sign that ordinary commercial terms have stopped being sufficient to close deals.
The more generous read is that this is what a supply chain looks like when it is maturing rather than overheating. Long-term supply agreements with volume-linked incentives are old technology in aerospace, in autos, in pharmaceuticals. What is new is the scale and the speed, and the fact that the incentive is denominated in shares rather than in unit discounts — a choice that arguably makes the arrangement more transparent, since a warrant has to be disclosed in a filing while a volume rebate does not.
Both readings can be true. The structure is disciplined; the environment it sits in is not. What the next seven years will actually measure is whether Google needed $120 billion of this hardware because inference demand was real, or because everyone in the industry had already agreed to act as though it was. The tranches will answer that question one half-billion at a time, and unlike most things in this cycle, they will do it in public.
한글 요약
마벨 테크놀로지는 8월 19일 SEC에 제출한 8-K를 통해, 구글에 자사 보통주 최대 58,970,907주를 주당 206.58달러에 매입할 수 있는 워런트를 8월 18일 발행했다고 공시했습니다. 전량 행사 시 약 122억 달러 규모로, 구글은 마벨의 5대 주주가 됩니다. 이 워런트는 7월 29일 체결된 커스텀 반도체 공급 계약에 부수된 것으로, 대상 제품은 구글 TPU 생태계에 붙는 AI 추론 가속기, 스토리지 컨트롤러, 네트워크 인터페이스 컨트롤러, 메모리 인터페이스 컨트롤러, 니어메모리 컴퓨트입니다. TPU 본체가 아니라 그 주변부 실리콘이라는 점이 핵심입니다.
주목할 부분은 베스팅 구조입니다. 1,360,867주만 1년간 분기 균등 시간 조건으로 확정되고, 나머지는 구글의 실제 구매액 5억 달러마다 1트랜치씩, 총 240트랜치로 나뉘어 2027 회계연도 3분기부터 2033 회계연도 말까지 확정됩니다. 240 × 5억 달러는 약 1,200억 달러로, 이 계약이 암시하는 매출 상한선입니다. 발표 당일 마벨 주가는 장중 최대 14% 올랐다가 8~10%대로 마감했고, 10년간 구글의 커스텀 칩 파트너였던 브로드컴은 약 5% 하락, 알파벳은 거의 움직이지 않았습니다. 모닝스타의 윌리엄 커윈은 이를 브로드컴의 교체가 아니라 구글 쪽 파이 자체가 커진 것으로 해석했습니다.
다만 공급사 매출이 고객사 주식과 얽히는 이른바 순환형 거래가 최근 AI 업계에 잇따르면서, 이런 구조가 밸류에이션을 부풀린다는 우려도 함께 제기됩니다. 마벨의 워런트는 실현 매출에만 연동되고 양도도 제한돼 상대적으로 보수적인 편이지만, 공시 시점이 지분 조건 확정 이후라는 점은 짚어둘 만합니다. 1차 검증대는 8월 27일 마벨의 2027 회계연도 2분기 실적 발표로, 회사는 약 27억 달러 매출을 가이던스로 제시한 상태입니다. 참고: SEC Form 8-K, CNBC, The Next Web.