NextFin News - Marvell Technology and Alphabet's Google expanded their chip-development partnership Wednesday, and the price of that commitment is written in Marvell stock: the semiconductor designer granted Google a warrant to buy as many as 58,970,907 shares at $206.58 apiece, a potential $12.2 billion equity stake that vests only if the products the two companies build together actually sell.
The structure is the story. Google does not hand over $12.2 billion for an option on Marvell's future. Instead, Marvell hands Google the right to buy shares at a fixed price, and most of those shares unlock only as revenue from jointly developed chips accumulates - one tranche for every $500 million in sales generated by the partnership's products from Marvell's third quarter of fiscal 2027 through the end of fiscal 2033. Nearly 1.4 million shares vest in equal quarterly installments during the first year after the warrant is executed, with the remainder tied to commercial results. If the chips flop, the equity largely never vests. If they become the backbone of Google's AI data centers, Google ends up owning a meaningful slice of the designer it hired.
Marvell shares jumped more than 10% in premarket trading Wednesday, with the intraday advance reaching as much as 12%, while Broadcom - Google's incumbent custom-chip partner - fell on the news. The move confirms what the semiconductor market has been pricing all year: the race to design AI accelerators in-house at hyperscalers is consolidating around a small group of specialists, and Marvell is forcing its way into the inner circle alongside Broadcom.
The Deal: A Warrant, Not a Check
Marvell disclosed the arrangement in a regulatory filing Wednesday. Under its terms, Google may purchase up to 58,970,907 Marvell shares at $206.58 each, for a maximum aggregate value of roughly $12.2 billion. The strike price matters: with Marvell's stock trading near the low $200s before the announcement, the warrant is issued close to the market rather than at a deep discount, which signals that this is less a financing event and more a long-term alignment mechanism.
The vesting schedule is where the incentives get interesting. The first year is mechanical - nearly 1.4 million shares release in equal quarterly installments regardless of outcomes. After that, vesting is explicitly contingent on "discretionary purchases" from Marvell by Google, tracked against revenue generated by the products the two companies co-develop. One tranche unlocks for each $500 million in revenue stemming from those products, running from Marvell's fiscal third quarter of 2027 through fiscal 2033. In effect, Marvell has written Google a call option on its own stock whose exercise value is backstopped by real commercial adoption.
The partnership itself covers "a comprehensive range of custom silicon programs" tied to Google's tensor processing units, Marvell said, including accelerators and storage controllers. That scope is broader than the two-chip program first reported in April - a memory processing unit designed to work alongside Google's TPUs and a new TPU optimized for AI inference - and it formalizes a relationship that had been in negotiation since spring.
This is not Marvell's first warrant tied to a hyperscaler relationship. In December 2025, as part of its acquisition of Celestial AI, Marvell issued Amazon a warrant to purchase up to 1,045,171 shares at $87.0029 each, subject to certain conditions. The Google warrant is roughly 56 times larger in share count and, more importantly, structured around a multi-year revenue ramp rather than a single transaction. Taken together, the two deals sketch a pattern: Marvell is using equity-linked instruments to lock in alignment with the cloud giants that dominate AI infrastructure spending.
The Market's Verdict: Long Marvell, Short the Incumbent
The tape read the announcement as a direct share grab from Broadcom. Marvell's double-digit premarket surge came paired with a decline in Broadcom shares, a mirror-image reaction that has played out before: when reports surfaced in April that Google was in talks with Marvell, Marvell gained roughly 5% in premarket trading while Broadcom and other Google suppliers slipped.
The symmetry is not accidental. Broadcom has been the dominant designer of custom AI accelerators for hyperscalers, and Google is its marquee client. Broadcom's management has said its AI semiconductor revenue reached $10.8 billion in its fiscal second quarter, and it guided the current quarter to about $16 billion - more than 200% year-over-year growth. The company also extended its long-term agreement to design and supply Google's TPUs and networking components through 2031. A warrant giving Google the right to buy $12.2 billion of Marvell stock does not tear up that contract, but it tells investors that Google is deliberately cultivating a second source.
For Marvell, the stock reaction is the latest chapter in a remarkable re-rating. The shares have more than tripled over the past year as the market recognized that custom silicon - application-specific integrated circuits designed for a single customer's AI workloads - is growing faster than the merchant GPU market in percentage terms. Marvell reported fiscal first-quarter revenue of $2.42 billion, up 28% year over year, with data center sales representing 76% of the total. Management has guided full-year fiscal 2027 revenue to nearly $11.5 billion, up about 40%, with data center revenue growing roughly 50%. The company now expects custom chip revenue to exceed $10 billion in fiscal 2029, and it raised its fiscal 2028 outlook to approximately $16.5 billion.
Why the Warrant Structure Changes the Incentives
Most chip-design contracts pay the designer for engineering effort - non-recurring engineering fees up front, then unit margins on every chip shipped. That aligns the designer to complete the tape-out, not necessarily to make the product a commercial success. Marvell's warrant flips part of the compensation into equity that pays off only if Google's products generate substantial revenue.
That is precisely why the market rewarded the structure. Investors have long worried that hyperscaler custom-chip programs are lumpy and politically fragile: a cloud giant can switch designers between generations, squeeze margins once a design is locked, or simply bring more work in-house. By taking equity that vests on revenue, Marvell signals confidence that it will not be swapped out after the first generation, and Google signals that it wants its designer to think like a partner rather than a vendor.
There is also a capital-allocation logic on Google's side. Alphabet is spending tens of billions annually on AI infrastructure - it guided 2026 capital expenditure to between $180 billion and $190 billion after earlier lifting its 2025 guidance to $85 billion and then to $92 billion at the midpoint. Buying Marvell shares through a warrant, rather than writing a giant check, lets Google keep cash inside the business while still binding Marvell's fortunes to the success of Google's silicon. If the chips succeed, Google participates in Marvell's equity upside; if they do not, Google's exposure is limited.
The dilution math, however, is not trivial. The full 58.97 million shares would represent a meaningful percentage of Marvell's current share count if fully exercised - an overhang that existing shareholders absorb in exchange for a deeper, longer-lived partnership with one of the world's largest compute buyers. The vesting schedule softens the blow, since most of those shares appear only if the partnership generates real revenue, but the potential dilution will sit in the background of every quarterly report.
The Second-Order Question: Who Designs the Hyperscalers' Chips?
The first-order reading of this deal is simple: Marvell wins business, Broadcom loses some. The second-order question is more consequential for the semiconductor industry's structure. As AI training and inference migrate from Nvidia's merchant GPUs to custom accelerators, the value in the AI stack is shifting from the chip vendor to the design partner and the customer that owns the workload.
Nvidia itself appears to have read the room. On March 31, Nvidia announced a $2 billion strategic investment in Marvell, wiring Marvell's custom silicon and networking gear more tightly into Nvidia's own AI stack. That deal was widely interpreted as an effort to make it easier for Nvidia's customers to combine Nvidia GPUs with custom accelerators - a defensive move against the very trend this Google warrant accelerates. Nvidia is effectively hedging: if hyperscalers keep buying GPUs for training, Nvidia wins; if they shift inference to custom silicon designed by Marvell, Nvidia still participates through its equity stake and interconnect technology.
The emerging architecture of AI computing is splitting into two lanes. In the first, Nvidia remains the default for training frontier models, with its CUDA software moat intact. In the second, hyperscalers run inference - the cheaper, higher-volume task of serving already-trained models - on their own TPUs and custom chips designed by Broadcom, Marvell, or increasingly their own teams. Marvell's new warrant with Google is a bet that the second lane will grow faster, and that Marvell will be one of the toll collectors.
This is where the cyclical-versus-structural call matters. The AI infrastructure buildout is cyclical in the sense that hyperscaler capital expenditure will eventually normalize after the current deployment wave - Alphabet's own guidance, raised three times in a year, cannot climb at that pace indefinitely. But the shift of AI inference onto custom silicon is structural: once a cloud provider has invested in a TPU fleet and rewritten its serving stack, it does not migrate back to merchant GPUs. The economics of inference reward specialization - lower cost per query, better power efficiency, tighter integration with the provider's software. That is a regime change, not a cycle, and it is why Broadcom and Marvell are being re-rated even as some investors worry about an AI capex peak.
The expanded partnership spans a comprehensive range of custom silicon programs related to Google's tensor processing units, including accelerators and storage controllers.
The Counter-Thesis: A Warrant Is Not a Revenue Contract
The strongest argument against reading this as a Marvell breakthrough is also the simplest: a warrant tied to revenue milestones is not a revenue contract. Google has committed no minimum purchase. If the co-developed chips underperform Nvidia's or Broadcom's offerings on price, power, or software, Google can simply not exercise - and the bulk of the 58.97 million shares never vest.
Broadcom's position is also far more entrenched than a single-day stock move suggests. Its TPU agreement with Google runs through 2031, and Broadcom has been designing Google's custom silicon for multiple generations. Switching a production TPU program to a new designer carries enormous execution risk - a single flawed tape-out can cost a hyperscaler an entire product cycle. Google's diversification into Marvell is real, but it is more likely to play out as a second-source strategy for specific accelerators than as a full migration away from Broadcom.
There is also a valuation argument. Marvell trades at a trailing price-to-earnings multiple well above Broadcom's, despite Broadcom generating more AI semiconductor revenue in a single quarter than Marvell generates in total company revenue. Part of that premium reflects Marvell's smaller base and faster percentage growth. But it also means the stock has less margin for error: any stumble in the Google program, or any sign that custom-chip growth is slowing, would hit Marvell harder than Broadcom. Broadcom, for its part, is not standing still - it has also expanded its AI deals beyond Google, including a commitment to supply Anthropic with approximately 3.5 gigawatts of TPU-based compute capacity starting in 2027.
The falsifying signal for the bullish read is specific and observable: if Marvell's data center revenue growth decelerates materially below its guided ~50% pace for fiscal 2027, or if the company fails to disclose meaningful revenue from the Google partnership by fiscal 2028, the thesis that this warrant represents the start of a multi-year revenue ramp - rather than a speculative option on a program that never scaled - would be wrong.
What to Watch Next
In the short term, the market will watch for two things: the formal execution of the warrant and the first disclosure of revenue attributable to the Google programs. Marvell reports its fiscal second-quarter 2027 results on August 27, having guided revenue of $2.7 billion, plus or minus 5% - about 35% year-over-year growth. That call is the earliest point at which management could begin updating investors on the partnership's progress, and any commentary on custom-silicon bookings will be parsed closely.
Over the medium term, the key question is whether other hyperscalers follow Google's lead. Amazon has long worked with Marvell on custom silicon, and the Amazon-Anthropic ecosystem is another potential customer. If Google's warrant structure becomes a template - hyperscalers taking equity in their chip designers, vested on revenue - it would deepen the alignment between cloud providers and the small group of designers capable of executing at the leading edge. It would also raise the barrier to entry for smaller design firms that cannot offer the same alignment terms.
In the long run, the structural question is whether custom silicon becomes the default for AI inference across the industry. If it does, Marvell and Broadcom are the primary beneficiaries among pure-play designers, and Nvidia's merchant GPU dominance in inference erodes even if its training business holds. If custom silicon proves harder to deploy than hoped - if software compatibility, tooling, and time-to-market keep workloads on GPUs - then today's re-rating of the custom-chip designers will look like a cyclical overshoot.
The Google warrant is Marvell's answer to that question, written in equity: the company is betting that the chips it builds with Google will generate tens of billions in revenue, and it is willing to let Google share in the upside if that bet pays off. For investors, the lesson is narrower but sharper. A warrant that vests on revenue is only as valuable as the revenue it tracks - and the next several quarters will show whether Google's $12.2 billion option is a down payment on a partnership or a bet that never comes due.
Explore more exclusive insights at nextfin.ai.
