Marvell disclosed a commercial agreement with Google covering custom silicon programs that attach to the TPU ecosystem, paired with a warrant on 58.97 million shares that vests one tranche per $500 million of custom product revenue. Futurum examines what the deal does to Marvell’s XPU outlook, how it compares to the 50-opportunity pipeline the company has reported, and why Marvell’s custom silicon capabilities won the socket.
What Is Covered in This Article:
- Marvell’s warrant to Google: 58.97 million shares at $206.58, roughly $12.2 billion and 7% of the company
- Vesting mechanics of 240 tranches, one per $500 million of custom product revenue, a $120 billion ceiling through fiscal 2033
- The commercial agreement scope across inference accelerators, storage, networking, and memory controllers, and near-memory compute attached to the TPU ecosystem
- The impact on Marvell’s XPU outlook for over $10 billion in custom revenue in fiscal 2029
- The competitive field: Broadcom through 2031, MediaTek on cost-optimized TPUs, and Google’s reported Frozen v2 inference chip
The News: Marvell Technology (NASDAQ: MRVL) disclosed in an August 19 SEC filing that it entered a commercial agreement with Google on July 29, 2026 for the development of custom semiconductor products, and issued Google a warrant on 58,970,907 shares of common stock on August 18. The filing describes the partnership as covering “a comprehensive range of custom silicon programs that attach to the TPU ecosystem, including AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute.”
The warrant has an exercise price of $206.58 per share and expires August 18, 2033. A time-based tranche of 1,360,867 shares vests in equal quarterly installments over the first year. The remaining shares vest in 240 equal tranches of roughly 240,000 shares, one tranche per $500 million of Custom Products revenue recognized from the third quarter of fiscal 2027 through fiscal 2033. Full exercise would cost Google roughly $12.2 billion and deliver about 7% of Marvell’s shares outstanding. The warrant is non-transferable without Marvell’s consent, and Google receives customary registration rights. Marvell shares rose roughly 10% on the disclosure. Neither company issued a press release, and the deal reached the market through the 8-K filing on the Marvell investor relations site.
Marvell Attaches Across Google’s TPU Stack With a Warrant Vesting Toward $120B
Analyst Take: Marvell’s custom silicon story has run for a year on KPIs: an 18-project engagement list, a 50-opportunity pipeline, a 20% share target. The Google agreement replaces KPIs with a name. Google paid nothing for a claim on 7% of its supplier, and Marvell in exchange secured a contracted position across the Google TPU stack, the highest-volume custom accelerator platform in production. The vesting schedule turns the relationship into a public meter, one tranche of roughly 240,000 shares per $500 million of custom product revenue, so the market gains a quarterly odometer on a hyperscaler procurement relationship that would otherwise stay hidden.
Futurum’s view is that the agreement is the strongest external validation yet of Marvell’s custom silicon model. It converts an XPU outlook built on modeled TAM into named, metered revenue, and it rescales a disclosed pipeline that now looks conservative. The socket landed at Marvell because the capabilities the TPU roadmap now demands, memory expansion and near-memory compute for decode-heavy inference, PCIe Gen 6 storage, and the controllers that move tokens between accelerators, already existed at Marvell, qualified and shipping, while Broadcom’s design capacity stays committed to the TPU generations themselves through 2031.
The Agreement Converts Marvell’s XPU Outlook Into Named, Metered Revenue
Marvell’s XPU outlook has rested on modeled markets. At its June 2025 custom AI event, the company sized its 2028 data center TAM at $94 billion, with custom compute at $55.4 billion of that total: $40.8 billion of custom XPU growing at a 47% CAGR, and the remaining $14.6 billion in XPU-attach categories such as co-processors, retimers, CXL controllers, and co-packaged optics components. Management targets 20% of the custom market by 2028, up from under 5% a few years ago, and said on its most recent earnings call that it remains confident in its target model of delivering over $10 billion in custom revenue in fiscal 2029. The strain on that outlook has been visibility. Custom silicon is guided to grow above 20% in fiscal 2027 while electro-optics grows above 70%, so the segment carrying the company’s valuation narrative has been its slower disclosed grower.
The Google agreement changes the character of that outlook without changing a single estimate. The $120 billion vesting ceiling requires Google alone to average nearly $18 billion of custom product purchases annually through fiscal 2033, nearly double the $10 billion the entire custom business targets across all customers in fiscal 2029, so the ceiling is aspirational by construction, sized so vesting never runs out before the relationship does. Read the other way, a Google relationship running at even half of the ceiling’s pace would reach the fiscal 2029 target on its own.
The per-tranche economics keep the incentive honest. At the roughly $230 the stock reached after the announcement, one tranche is worth about $5.6 million intrinsic, an effective rebate near 1.1% on each $500 million of purchases. At Marvell’s June high of $329.88 the same tranche is worth roughly $29.6 million, a 5.9% rebate. That figure is Futurum’s arithmetic and it means Google’s discount deepens as its own purchasing drives Marvell’s stock. AMD set the template in October 2025 with warrants to OpenAI on up to 160 million shares vesting against 6 GW of deployments.
The Google TPU Stack Rescales a Pipeline Marvell Sized at $75 Billion
The comparison to Marvell’s reported XPU pipeline shows the quantum leap this deal presents. The company has disclosed 18 custom projects in flight: 12 devices at the four major US hyperscalers, split 3 XPUs and 9 XPU-attach devices, plus 6 devices at emerging AI customers. Beyond the committed projects sits a pipeline of more than 50 opportunities across over 10 customers that management sizes at roughly $75 billion in lifetime revenue. Set the Google warrant against those numbers and the rescaling is visible: the $120 billion vesting ceiling from this single customer exceeds the lifetime value of the entire 50-opportunity pipeline by 60%. Both figures are ceilings, and neither is bookable revenue. The relative size still matters, because it shows Marvell and Google contemplating a relationship larger than everything else in the custom funnel combined.
The agreement’s five product categories also imply several sockets rather than one. Earlier reporting had Google in talks with Marvell on an inference-focused TPU and a memory processing unit designed to work alongside existing TPUs. The signed scope’s inference accelerators and near-memory compute track those descriptions closely. The inference accelerator category is the one that reaches beyond attach economics into the $40.8 billion XPU slice of Marvell’s TAM, and Google’s roadmap suggests why it wants a second design bench there. Reporting from The Information, summarized by Tom’s Hardware, describes a project called Frozen v2 that etches Gemini’s architecture into silicon while keeping weights updatable, with engineers projecting 6 to 10x more tokens per unit of power than the newest TPUs and deployment targeted for 2028 at volumes below TPU production. Those projections are sourced to two people familiar with the project and await silicon. A trial-run, model-specific chip is a poor fit for Broadcom’s high-volume TPU cadence and a natural fit for a design partner paid per program. Broadcom keeps its agreement to design and supply TPU generations runs through 2031 while MediaTek can supply cost-optimized variants. Marvell’s win likely fits in the disaggregated complements Google is deliberately creating around that core.
The Socket Went to Marvell Because the Capabilities Already Existed
What hyperscalers buy from Marvell is a specific bundle of attributes. The customer keeps the architecture and the differentiation while Marvell contributes the silicon plumbing that takes years to qualify, including, SerDes and a 64 Gbps/wire bidirectional die-to-die interface delivering over 3x UCIe bandwidth density, an in-house 2.5D advanced packaging platform, a published 2nm IP platform at TSMC, PCIe Gen 6 and CXL controllers, electro-optics, and custom HBM integration work with all three memory suppliers. The model tolerates lower gross margin than merchant chips in exchange for sockets that renew every generation, and Marvell accepts the attach sockets that larger rivals treat as beneath the accelerator itself.
The portfolio was announced before the contract was. Futurum’s coverage of the FMS 2026 launches described an AI memory infrastructure line spanning the Bravera SC6 PCIe 6.0 SSD controller, Structera CXL memory expansion and near-memory acceleration, and Photonic Fabric optical shared memory, aimed at the KV cache bottleneck in agentic inference; the Google TPU stack scope maps onto that portfolio almost item for item. Near-memory compute is the tell, because decode-heavy agentic workloads are memory bound, and Futurum has argued that memory silicon is becoming a distinct procurement category. The deal completes Marvell’s hyperscaler sweep with Trainium programs at Amazon , the Maia accelerator ecosystem at Microsoft, and now Google. Futurum believes that Marvell is the only merchant designer with signed programs at all three.
The $120 Billion Requires Purchase Orders That Do Not Yet Exist
The warrant is an option against commitments Google has not made. The 8-K discloses no minimum purchase, no volume floor, no exclusivity, and no statement on expected financial impact. Custom silicon programs run on multi-year design cycles, so meaningful revenue sits in fiscal 2028 at the earliest, which leaves the custom segment trailing the optics segment in growth for now. Google gains a second source and pricing power over Broadcom, and Marvell adds a customer that can redirect any program to MediaTek, Broadcom, or an internal team at each design generation, exactly the flexibility a five-category scope preserves. Custom programs also come with structurally lower gross margins than merchant silicon, a trade Marvell has defended as accretive operating income and critics have called margin dilution with extra steps. Additional clarity on Google’s roadmap and infrastructure plans will shape how commitments become purchase orders.
What to Watch:
- Whether Marvell raises its 20%+ custom silicon growth outlook at the August 27 earnings call
- Whether performance tranches begin vesting in fiscal Q3 2027 filings, the first public record of Google’s actual purchasing
- Whether Broadcom expands its Google agreement into attach categories or concedes the periphery
- Whether Marvell is named the design partner for Frozen v2 or other model-specific inference silicon
- Whether Amazon and Microsoft negotiate warrant structures of their own, making performance-vesting equity standard in hyperscaler procurement
See more details on the Marvell website.
Declaration of generative AI and AI-assisted technologies in the writing process: This content has been generated with the support of artificial intelligence technologies. Due to the fast pace of content creation and the continuous evolution of data and information, The Futurum Group and its analysts strive to ensure the accuracy and factual integrity of the information presented. However, the opinions and interpretations expressed in this content reflect those of the individual author/analyst. The Futurum Group makes no guarantees regarding the completeness, accuracy, or reliability of any information contained herein. Readers are encouraged to verify facts independently and consult relevant sources for further clarification.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Read the full Futurum Group Disclosure.
Other Insights From Futurum:
COMPUTEX 2026: Are Agentic CPUs Rivals or Complements?
Marvell’s Teralynx T100 Puts Power Efficiency at the Center of AI Networking
Marvell Q1 FY 2027 Raises Full-Year Outlook on AI Data Center Demand
Author Information
Brendan is Research Director, Semiconductors, Supply Chain, and Emerging Tech. He advises clients on strategic initiatives and leads the Futurum Semiconductors Practice. He is an experienced tech industry analyst who has guided tech leaders in identifying market opportunities spanning edge processors, generative AI applications, and hyperscale data centers.
Before joining Futurum, Brendan consulted with global AI leaders and served as a Senior Analyst in Emerging Technology Research at PitchBook. At PitchBook, he developed market intelligence tools for AI, highlighted by one of the industry’s most comprehensive AI semiconductor market landscapes encompassing both public and private companies. He has advised Fortune 100 tech giants, growth-stage innovators, global investors, and leading market research firms. Before PitchBook, he led research teams in tech investment banking and market research.
Brendan is based in Seattle, Washington. He has a Bachelor of Arts Degree from Amherst College.
