In-Depth Research Analysis:
Executive Summary:
As AI infrastructure evolves from individual chip performance toward system-level efficiency, demand for custom computing, high-speed networking, and memory interconnects is becoming increasingly interconnected. Marvell Technology is positioned across both custom compute and data connectivity, allowing it to participate in a broader range of semiconductor applications within AI infrastructure.
This report examines three core growth drivers: the expansion of custom silicon and XPU-attached products, the growing importance of optical and scale-up connectivity, and the potential for these opportunities to translate into sustainable earnings and cash flow.
Our central view is that Marvell’s long-term opportunity extends beyond individual AI chips to the broader semiconductor content of AI computing systems. However, the ultimate economic value of this opportunity will depend on project commercialization, product profitability, R&D efficiency, and the impact of potential share dilution.
1 From Traditional Data Infrastructure to a Core AI Semiconductor Supplier
1.1 Data Centers Have Become Marvell’s Core Business
Marvell Technology is a fabless semiconductor company specializing in data processing, networking, storage, high-speed signal transmission, and custom computing.
Unlike semiconductor companies that primarily sell standardized computing products, Marvell provides both standard products and chips designed for specific customer requirements. Its underlying technology capabilities include high-speed SerDes, analog and mixed-signal design, digital signal processing, system-on-chip integration, and networking protocols.
These technologies have long been used in enterprise networking, communications, and storage infrastructure. The expansion of AI data centers is now extending their applications into computing systems that require higher bandwidth, lower latency, and greater energy efficiency.
This transition is increasingly visible in Marvell’s revenue composition.
In FY2026, Marvell generated revenue of USD 8.20 billion, representing year-over-year growth of approximately 42%. In FY2027 Q2, data center revenue reached approximately USD 2.17 billion, up 46% year over year and accounting for roughly 79% of total revenue. Communications and other businesses contributed approximately USD 568 million, representing year-over-year growth of 10%.
Sources: Marvell FY2026 financial results and FY2027 Q2 financial results.
The rising contribution from data centers indicates that Marvell’s operating performance is becoming increasingly linked to hyperscalers’ AI capital expenditure, custom silicon production ramps, and data center networking upgrades.
However, this transition should not be interpreted as a complete replacement of traditional businesses by AI-related revenue. Communications and other end markets continue to contribute revenue and cash flow, while demand recovery in legacy products, customer inventory adjustments, and business divestitures can also affect overall growth.
Marvell completed the sale of its automotive Ethernet business in August 2025. Consequently, changes in the business portfolio should be considered when comparing revenue across recent fiscal years.
Source: Marvell FY2027 Q2 Form 10-Q.
From a business positioning perspective, Marvell is increasingly centered on AI data centers while maintaining a portfolio of communications, storage, and other infrastructure products.
1.2 Reusable Core Technologies Underpin Marvell’s Competitive Position
Marvell’s competitive capabilities should not be assessed solely by the number of products it offers.
More importantly, several of its underlying technologies can be applied across different products and computing architectures.
For example, high-speed SerDes enables data transmission between chips and external devices. Digital signal processing helps improve the quality of high-speed electrical and optical transmission, while complex SoC design capabilities allow different computing, networking, and storage functions to be integrated into customized products.
These technologies can support custom computing chips, optical DSPs, Ethernet switches, and next-generation scale-up interconnects.
As a result, a new customer design program may create opportunities beyond the revenue generated by a single chip. Marvell may also be able to supply additional supporting products within the same computing platform.
However, technology reuse does not automatically translate into commercial success. Different products still require independent customer qualification, manufacturing arrangements, and production ramps, while competing in different market environments.
We believe Marvell’s business positioning is evolving from that of a supplier of individual semiconductor products toward a broader provider of chips spanning computing, connectivity, and storage within AI infrastructure.
This positioning expands the company’s addressable business opportunities while making its operating performance increasingly dependent on the long-term investment plans of major data center customers.
2 AI Architecture Is Evolving — Why Compute and Connectivity Must Be Analyzed Together
2.1 AI Computing Is Shifting from Individual Chip Performance to System-Level Efficiency
Historically, AI computing capacity expanded primarily through improvements in individual accelerator performance and increases in the number of deployed chips.
However, as computing systems scale to thousands of accelerators or more, effective system performance no longer increases simply in proportion to the number of chips deployed.
During AI training, accelerators must frequently exchange parameters and intermediate results. Inference systems also require continuous data movement between computing chips, memory, networking equipment, and storage systems.
If data transmission is insufficient, computing resources may remain idle while waiting for data, reducing actual utilization. Adding more accelerators may therefore fail to deliver a proportional improvement in effective system performance.
Consequently, large-scale AI infrastructure increasingly requires a balance between computing throughput, memory bandwidth, interconnect latency, network capacity, and overall power consumption.
For Marvell, this means potential demand is not limited to customers purchasing more computing chips. It also comes from the additional communication and data-processing requirements created by increasingly complex computing systems.
2.2 Custom Compute and General-Purpose GPUs Serve Different Commercial Requirements
General-purpose GPUs offer substantial computing flexibility and mature software ecosystems, allowing them to support a broad range of AI models and workloads.
Hyperscale cloud providers, however, operate large and relatively consistent workloads. This allows them to develop custom computing architectures optimized for their own models, software frameworks, and deployment requirements.
For example, training workloads place significant emphasis on computing throughput and multi-chip communication efficiency, while certain inference workloads may prioritize response latency, memory access, and cost per request.
The commercial value of custom silicon lies in optimizing performance, power consumption, and total cost of ownership for specific workloads.
Nevertheless, custom chips are not inherently more economical than general-purpose GPUs. Their economic attractiveness also depends on development expenses, production scale, software compatibility, and product iteration cycles.
A more balanced industry outlook is therefore one in which general-purpose GPUs and custom computing platforms coexist over the long term.
For Marvell, this creates opportunities to participate in selected custom computing programs while supplying networking, storage, and memory-related products around different computing platforms.
2.3 Data Connectivity Is Developing Across Three Distinct Application Layers
AI connectivity requirements can be divided into three principal layers.
Scale-up primarily addresses high-speed communication within a computing system, including connections among accelerators, CPUs, and memory resources. As computing systems expand across multiple racks, interconnect bandwidth, latency, and power consumption become increasingly important.
Scale-out focuses on network expansion between servers and computing nodes. It involves high-speed Ethernet switching, optical modules, and related signal-processing chips.
Scale-across primarily involves data transmission between geographically separate data centers, where transmission distance, capacity, power consumption, and network reliability must be considered together.
Marvell’s product portfolio spans several of these application areas, and the company continues to develop interconnect solutions for different transmission distances and network architectures.
Sources: Marvell Optical DSP and AI Connectivity product disclosures.
From a financial perspective, however, these three markets should not be combined into a single growth assumption.
Different applications require different technologies, product pricing, competitive structures, and commercialization timelines. Moreover, higher network speeds do not necessarily result in a proportional increase in semiconductor content, as more highly integrated products may reduce network layers or eliminate certain components.
Connectivity revenue growth must therefore be evaluated through network scale, transmission speeds, semiconductor content per system, and architectural changes.
3 Core Investment Thesis — From Custom Silicon to System-Level AI Semiconductor Content
Marvell’s investment thesis extends beyond the expansion of individual AI chip markets.
As AI infrastructure becomes more complex, hyperscalers require not only greater computing capacity but also a broader range of supporting semiconductor functions to connect, manage, and utilize that capacity.
Marvell’s combination of custom silicon and data connectivity capabilities creates opportunities across multiple parts of the same computing system.
However, the expansion of addressable opportunities is only the first stage of the investment thesis. The more important question is whether this broader product coverage can translate into sustainable revenue, operating profit, and per-share economic value.
We therefore evaluate Marvell’s long-term development through three interconnected drivers: increasing semiconductor content around custom computing platforms, the growing importance of data connectivity in system-level performance, and the conversion of business growth into sustainable financial returns.
3.1 The First Driver: AI Semiconductor Demand Is Expanding Beyond Core Accelerators
Traditional AI semiconductor analysis often focuses on the computing performance, shipment volumes, and unit value of GPUs and custom accelerators.
In practice, however, an AI computing system requires substantially more than its core computing chips.
Networking interfaces, storage controllers, memory interfaces, data offload engines, and other supporting components are necessary to enable the computing platform to operate efficiently.
As hyperscalers expand their proprietary computing architectures, demand for custom semiconductor products can therefore extend well beyond the core XPU itself.
Marvell’s distinguishing characteristic is that it can participate in selected core custom computing programs while also supplying supporting chips around different computing platforms.
The expanded Google partnership illustrates this broader business model. By participating in additional functions within the TPU ecosystem, Marvell’s potential revenue opportunities are not necessarily limited to securing a design win for the core computing chip.
The first component of our investment thesis is therefore:
Marvell’s long-term revenue opportunity depends not only on the number of AI accelerators deployed by its customers, but also on the semiconductor content required by each computing platform and the share of those functions that Marvell can supply.
This framework shifts the analysis from individual chip shipments toward semiconductor content across the broader computing platform.
Nevertheless, a larger potential market does not automatically translate into higher company revenue. Customer design decisions, competitive positioning, qualification requirements, and production ramps remain critical determinants of actual commercial returns.
3.2 The Second Driver: Connectivity Is Becoming a More Important Determinant of AI System Performance
While custom silicon demand is driven by the development of proprietary computing platforms, connectivity demand is also influenced by the expansion and increasing complexity of AI systems.
The two businesses are related, but their commercial drivers are not identical.
As AI clusters grow, accelerators must exchange larger volumes of data, servers require greater network bandwidth, and computing systems need more efficient interconnections among processing and memory resources.
Effective computing capacity therefore depends not only on the number of installed accelerators, but also on how efficiently those accelerators operate together.
Marvell’s capabilities in high-speed optical DSPs, networking switches, and next-generation scale-up interconnects allow it to participate in this evolution from individual servers toward larger, more complex computing systems.
The second component of our investment thesis is:
As AI infrastructure shifts from individual chip performance toward system-level efficiency, data connectivity may become a source of semiconductor demand that is not solely dependent on growth in core accelerator shipments.
This distinction matters because Marvell does not necessarily need to secure a greater number of core computing chip designs to expand its business.
Different customers may adopt different computing architectures, but those architectures still require networking, storage, and interconnect functions. Marvell can therefore potentially participate across multiple AI infrastructure configurations.
However, different connectivity technologies may compete with or replace one another, and higher network bandwidth does not necessarily translate into proportional semiconductor revenue growth.
The commercial implications of this thesis must ultimately be validated through actual customer deployment of optical DSPs, switching products, and next-generation interconnect solutions.
3.3 The Third Driver: Broader Product Coverage Must Translate into Sustainable Per-Share Value
The first two drivers explain why Marvell may gain access to a broader range of business opportunities.
They do not, by themselves, establish the economic value of that growth.
The central financial question is:
Can the company convert broader product coverage into higher-quality earnings and cash flow?
Custom silicon and connectivity products may benefit from shared technological capabilities. Marvell’s expertise in high-speed SerDes, signal processing, and system integration can be applied across multiple product categories, creating opportunities to reuse certain underlying technologies and engineering resources.
If the company can leverage existing technology platforms to secure additional customer designs and achieve scaled production, revenue growth may support more efficient utilization of its R&D investment.
However, this outcome is not guaranteed.
Custom silicon programs require substantial upfront engineering expenditure, while major hyperscale customers possess significant purchasing power. New connectivity products may also require extended commercialization periods.
At the same time, acquisition-related expenses, stock-based compensation, and customer warrant arrangements may cause net income growth and earnings per share to diverge.
The third component of our investment thesis is therefore:
Marvell’s long-term economic value depends not only on whether custom compute and data connectivity can grow simultaneously, but also on whether the company can leverage its shared technology capabilities to expand product coverage and convert incremental revenue into sustainable per-share earnings and cash flow after accounting for R&D expenditure, customer pricing pressure, and potential dilution.
This provides the link between Marvell’s industry opportunities and its financial and valuation framework.
The following sections examine custom silicon and data connectivity separately before assessing whether the two businesses can generate sustainable improvements in profitability and shareholder economic value.
4 Custom Compute — Expanding from Core XPUs to the Broader Computing Platform
4.1 Custom Silicon Opportunities Extend Beyond Core Computing Chips
In AI semiconductors, an application-specific integrated circuit, or ASIC, refers to a chip designed for a particular application.
From a commercial perspective, however, the broader custom silicon business should be divided into distinct product categories.
The first category is the Custom XPU, which performs the principal computing tasks. Revenue from these products is closely linked to the customer’s computing architecture, model requirements, and deployment scale.
The second category is XPU Attach, which includes customized products supporting the computing platform, such as networking interfaces, storage controllers, memory interfaces, data offload engines, and other supporting functional chips.
Although both categories fall within custom silicon, their growth mechanisms differ.
Core XPU programs typically involve substantial development complexity and close customer integration. Revenue primarily depends on the deployment of the corresponding computing platform.
XPU Attach demand depends not only on the number of accelerators deployed, but also on platform complexity and changes in system architecture.
For example, as a computing platform processes more data, customers may require additional dedicated networking, storage management, or memory access functions, creating new opportunities for custom silicon.
Marvell’s custom computing business should therefore not be evaluated simply by the number of AI accelerator orders it secures.
The more important questions are which functions the company can supply within a customer’s computing platform and whether successive platform generations can generate additional design opportunities.
4.2 The Google Partnership Expands Product Coverage but Does Not Establish a Core TPU Design Win
On August 19, 2026, Marvell disclosed an expanded long-term custom silicon partnership with Google.
According to the company’s Form 8-K, the programs cover a range of products supporting Google’s TPU ecosystem, including AI inference accelerators, storage controllers, networking interface controllers, memory interface controllers, and near-memory computing.
Source: Marvell Form 8-K, August 19, 2026.
The significance of this partnership lies in Marvell’s ability to participate in more functional components surrounding Google’s core computing platform.
However, TPU-adjacent products must be distinguished from the core TPU computing chip.
The disclosed partnership does not, by itself, establish that Marvell has secured the design of Google’s core TPU accelerator. AI inference accelerators and other supporting chips may perform important computing or data-processing functions, but their product positioning, technical requirements, and revenue structures are not necessarily identical to those of the core TPU.
Accordingly, the partnership should not be analyzed by directly attributing Google’s entire TPU expenditure to Marvell’s potential revenue.
Instead, the relevant questions concern the actual product scope, deployment scale, and purchasing schedule of the supporting chips.
Over the longer term, Marvell may expand its revenue opportunity within a single computing platform by supplying a broader range of products. Nevertheless, individual programs still require independent design, qualification, and production, creating differences in revenue realization timing.
4.3 The Warrant Agreement Establishes Long-Term Commercial Incentives, Not Guaranteed Orders
Another important element of the Google partnership is the warrant agreement.
According to Marvell’s August 19, 2026 Form 8-K, Google received warrants to purchase up to 58,970,907 Marvell common shares at an exercise price of USD 206.58 per share, subject to the relevant terms and conditions.
Approximately 1.36 million shares are subject to time-based vesting, while the remaining approximately 57.61 million shares primarily vest based on qualifying custom-product revenue.
The agreement establishes 240 revenue-based vesting tranches, with each tranche corresponding to USD 500 million of qualifying revenue.
Source: Marvell Form 8-K and related warrant agreement, August 19, 2026.
Based on these contractual revenue milestones, full performance-based vesting would correspond to approximately USD 120 billion of cumulative qualifying revenue.
However, this figure requires careful interpretation.
First, it represents the revenue threshold associated with full performance-based warrant vesting. It does not mean Google has committed to purchasing USD 120 billion of chips, nor should it be treated as an existing order backlog.
Second, actual revenue realization depends on Google’s future purchases, production ramps, and the continuation of the commercial relationship.
Third, even if the associated business ultimately generates substantial revenue, the warrant structure may affect Marvell’s future per-share economic value.
Revenue growth can increase total earnings, but if the number of common shares also rises, earnings-per-share growth may be lower than net income growth.
The actual financial impact of the warrants will depend on vesting, exercise, applicable accounting treatment, and other conditions, including future share prices. It would therefore be inappropriate to treat all potentially issuable shares as immediately outstanding.
We believe the partnership provides greater visibility into Marvell’s long-term custom silicon opportunities. However, its ultimate economic value must be assessed through actual customer purchases, project profitability, and the resulting impact on shareholders.
4.4 Custom Silicon Competitiveness Depends on Design Capabilities and Sustained Production
A custom silicon program typically involves several stages, including requirements definition, chip design, qualification, manufacturing, packaging, and system deployment.
Securing a design program is only the starting point of commercialization.
For complex AI chips, development timelines, advanced manufacturing processes, packaging capabilities, and customer system validation can all affect the actual production schedule.
Custom silicon analysis should therefore distinguish between securing a design program, completing product qualification, beginning volume production, and generating stable revenue.
Marvell’s high-speed interconnect IP, complex SoC design capabilities, and system integration expertise help it address customer-specific requirements.
At the same time, the company must manage substantial development expenses, competition for advanced manufacturing capacity, and the purchasing power of major customers.
Large hyperscalers may select different design partners for different product functions or platform generations. Securing one generation of a custom product does not guarantee participation in subsequent iterations.
Sustained growth in custom silicon therefore requires not only new customer programs, but also successful production ramps of existing projects and the ability to maintain technological and commercial competitiveness across product generations.
4.5 Revenue Growth and Profitability May Not Move in Parallel
Custom silicon programs can generate substantial revenue from individual customers, but their profitability may differ from that of standardized semiconductor products.
The development cost of a standardized product can be spread across a broad customer base. Custom silicon, by contrast, may require significant engineering resources dedicated to a specific customer, with cost recovery depending on contractual arrangements, pricing, and subsequent production volumes.
If a program achieves sustained volume production over an extended period, higher shipment volumes may allow the initial development investment to be absorbed more efficiently.
However, project delays or weaker-than-expected customer deployment may create pressure from development costs and production preparation expenses.
Consequently, expansion of the custom silicon business does not necessarily result in a proportional improvement in company-wide gross margin.
In evaluating Marvell’s custom computing business, three indicators are particularly important: the number of programs entering volume production, the resulting revenue growth, and changes in overall gross margin and operating expenses.
Only when revenue expansion translates into sustainable operating profit and cash flow can the long-term economic value of custom silicon growth be more fully demonstrated.
5 AI Data Connectivity — A Second Core Growth Driver
5.1 Optical DSPs: Moving from 800G to 1.6T
Data connectivity is an important business area that differentiates Marvell from suppliers focused primarily on custom computing chips.
As AI clusters expand, networks must transmit more data within constrained space and power budgets. This is driving optical communication products toward higher transmission speeds, greater bandwidth density, and lower power consumption per unit of data.
Optical DSPs are important signal-processing components within high-speed optical modules. They help improve signal quality, compensate for transmission impairments, and support high-speed communication.
From a commercial perspective, optical DSP demand depends not only on the number of optical modules deployed, but also on per-port transmission speed, module architecture, and signal-processing complexity.
In March 2026, Marvell announced an expanded 1.6T optical DSP portfolio covering AI data center connectivity applications with different reach and power requirements.
Source: Marvell optical DSP product announcement, March 12, 2026.
The economic significance of this upgrade is that customers moving from 800G to 1.6T must increase transmission capacity while managing power consumption, signal integrity, and deployment costs.
If Marvell continues to deliver products that meet the requirements of next-generation optical modules, network upgrades may create additional revenue opportunities.
However, higher transmission speeds do not necessarily translate into proportional semiconductor revenue growth.
Greater integration, improvements in module architecture, and the adoption of alternative signal-processing solutions for certain short-reach applications may change the number and value of discrete chips required per optical port.
Assessing the long-term growth of optical DSP revenue therefore requires consideration of actual port deployment, adoption rates across transmission speeds, product mix, and customer purchasing patterns.
Network bandwidth growth alone is not sufficient to estimate company revenue.
5.2 Data Center Switching: Greater System Efficiency Can Also Change Semiconductor Content
In addition to optical DSPs, networking switches represent another important part of Marvell’s AI data center portfolio.
Switching chips are responsible for forwarding data across large numbers of servers and computing nodes. Their performance affects network capacity, latency, and effective cluster utilization.
As AI clusters grow, switching chips must support higher aggregate throughput while controlling equipment requirements, network layers, and power consumption.
In June 2026, Marvell introduced its 102.4Tbps Teralynx T100 switching chip for high-speed AI and cloud data center networking applications.
Source: Marvell Teralynx T100 product announcement, June 1, 2026.
Higher-bandwidth switching chips can support larger networks and may reduce system complexity by eliminating certain network layers.
From a revenue perspective, however, improvements in individual chip performance must be distinguished from expansion of the company’s overall market opportunity.
Higher-capacity products may increase the value of an individual switching chip while reducing the number of switches required for a particular network configuration.
Final revenue therefore depends on the customer’s network architecture, product pricing, and overall deployment scale.
Moreover, the switching market includes established competitors with mature technologies. Product launches and performance improvements do not automatically translate into market-share gains. Customer qualification, software ecosystems, reliability, and system compatibility also influence commercialization.
The key factor for Marvell’s switching business is therefore whether higher-bandwidth products can achieve sustained customer adoption and generate actual revenue, rather than improvements in technical specifications alone.
5.3 Scale-Up: Optical Connectivity Is Moving Closer to the Computing System
Compared with conventional data center networking, scale-up connectivity focuses more directly on low-latency, high-bandwidth communication within computing systems.
As AI platforms expand from individual racks to multi-rack configurations, some connections traditionally implemented through electrical signaling face increasingly demanding constraints related to transmission distance, signal loss, and power consumption.
Conventional electrical interconnects retain established cost and integration advantages in short-reach, high-density applications. However, signal compensation and power requirements become more challenging as transmission speeds and distances increase.
Optical interconnects provide an alternative technical approach for selected applications, potentially improving high-speed transmission over longer distances and enabling more flexible deployment of computing resources.
In February 2026, Marvell completed its acquisition of Celestial AI, gaining access to the Photonic Fabric optical interconnect technology platform and expanding its capabilities in next-generation scale-up networking.
Source: Marvell Celestial AI acquisition completion announcement, February 2, 2026.
The acquisition has two principal implications.
First, it expands Marvell’s ability to participate in the evolution of interconnect technologies within computing systems, rather than focusing solely on conventional networking between servers.
Second, the acquired technology may complement Marvell’s existing optical DSP, switching, and custom computing businesses, allowing the company to offer a broader range of system-level solutions.
However, scale-up optical connectivity still faces commercialization uncertainty.
Actual adoption will depend on electro-optical conversion costs, system power consumption, packaging reliability, switching architecture, and customers’ existing technology roadmaps.
Importantly, optical connectivity does not need to replace every electrical interconnect application.
It should therefore be viewed as a potential source of future incremental growth rather than a business opportunity that has already been fully realized.
5.4 Celestial AI: Revenue Guidance Must Be Distinguished from Recognized Revenue
The commercialization timeline of Celestial AI has direct implications for Marvell’s future financial performance.
At the completion of the acquisition in February 2026, Marvell expected Celestial AI to begin contributing initial revenue in the second half of FY2028.
The company also outlined expectations for an annualized revenue run rate of approximately USD 500 million by FY2028 Q4, increasing to approximately USD 1 billion by FY2029 Q4.
Marvell additionally expected the acquisition to increase annual Non-GAAP operating expenses by approximately USD 50 million, alongside financial effects associated with the cash consideration and share issuance.
Source: Marvell Celestial AI acquisition completion announcement, February 2, 2026.
These figures represent forward-looking expectations disclosed by the company, rather than revenue that has already been recognized.
An annualized revenue run rate must also be distinguished from actual revenue generated over a full fiscal year.
For example, if a business reaches a USD 500 million annualized run rate during the fourth quarter, this does not mean that it has already generated USD 500 million in revenue during the fiscal year, nor does it imply that preceding quarters achieved an equivalent shipment level.
Financial projections should therefore reflect actual commercialization timing, customer deployment progress, and the expected quarterly revenue ramp.
We believe Celestial AI expands Marvell’s potential market opportunity, but its long-term value must still be validated through customer adoption, volume production, and the profitability of the resulting business.
5.5 XConn: Extending from Network Switching to Computing Resource Interconnects
In February 2026, Marvell completed its acquisition of XConn Technologies, expanding its portfolio of PCIe and CXL switching products.
Source: Marvell XConn acquisition completion announcement, February 10, 2026.
Unlike conventional Ethernet switching, which primarily connects servers and networking nodes, PCIe and CXL technologies are more closely associated with high-speed connections and resource sharing among computing devices, accelerators, and memory.
As AI computing systems expand, customers may require more flexible management of memory capacity and computing resources rather than simply adding more hardware within individual servers.
The acquisition of XConn allows Marvell to participate more extensively in the development of these system architectures and complements its scale-up switching capabilities.
However, PCIe, CXL, and other scale-up connectivity solutions do not represent a single interchangeable market.
Different protocols serve different applications, performance requirements, and customer ecosystems. Their commercialization timelines may also differ.
These businesses should therefore be evaluated through actual product deployment and customer adoption rather than by simply combining the potential market sizes of different interconnect protocols.
5.6 The NVIDIA Partnership: Greater Architectural Compatibility Does Not Equal Guaranteed Revenue
On March 31, 2026, Marvell and NVIDIA announced a strategic partnership involving custom XPUs compatible with NVLink Fusion, scale-up networking, and silicon photonics.
NVIDIA also announced a USD 2 billion investment in Marvell.
Source: Marvell and NVIDIA joint announcement, March 31, 2026.
The industry significance of this partnership lies in providing more system design options for customers seeking to deploy different computing chips while utilizing relevant NVIDIA infrastructure technologies.
For Marvell, participation in the ecosystem may create additional applications for its custom computing and connectivity products across different AI system architectures.
Nevertheless, ecosystem participation and future revenue are separated by several commercialization stages.
Products must satisfy relevant technical and compatibility requirements, customers must select specific architectures and procurement plans, and associated programs must enter production.
The partnership announcement alone should therefore not be treated as evidence of guaranteed incremental orders or revenue.
The USD 2 billion strategic investment must also be evaluated according to the nature and terms of the relevant financing instruments, including its effects on cash flow, capital structure, and potential dilution.
It should not be treated solely as a conventional cash inflow without considering the associated economic terms.
5.7 Supply Chain Capabilities and Product Iteration Remain Important
In AI connectivity, chip performance, power consumption, and cost are important competitive factors, but sustained production also depends on advanced manufacturing, analog and radio-frequency processes, packaging, testing, and customer qualification.
Different products may have different manufacturing requirements.
In addition to advanced digital logic, high-speed optical communication chips may require appropriate analog and mixed-signal manufacturing capabilities.
In September 2026, Marvell and GlobalFoundries announced an expanded collaboration to increase silicon-germanium technology capacity for next-generation optical connectivity, supporting applications such as high-speed pluggable optical modules, NPO, and CPO.
Source: GlobalFoundries and Marvell joint announcement, September 17, 2026.
This collaboration illustrates that successful commercialization of high-speed connectivity products depends not only on chip design, but also on whether the supply chain can meet performance, capacity, and yield requirements.
However, capacity expansion does not directly establish that corresponding customer orders have already been secured.
Technical and supply chain investment must ultimately translate into customer qualification, product deliveries, and recognized revenue.
Overall, we believe data connectivity provides Marvell with a source of potential growth distinct from custom computing, although optical DSPs, networking switches, and scale-up optical interconnects are at different stages of commercial maturity.
High-speed optical and networking products with established deployment bases remain important indicators of current business growth, while next-generation scale-up technologies are more relevant to the longer-term evolution of Marvell’s product portfolio and revenue composition.
6 Financial Analysis — Can Revenue Growth Translate into Sustainable Per-Share Economic Value?
6.1 Data Centers Drive Revenue Growth, but Product Mix Requires Continued Monitoring
In FY2027 Q2, Marvell generated revenue of USD 2.739 billion, representing year-over-year growth of 37% and sequential growth of approximately 13%.
Data center revenue reached approximately USD 2.172 billion, up 46% year over year, while communications and other businesses contributed approximately USD 568 million, up 10%.
Data centers accounted for the majority of total revenue and were the principal contributor to quarterly revenue growth.
Source: Marvell FY2027 Q2 Financial Results.
This revenue composition indicates that Marvell’s operating performance is becoming increasingly dependent on semiconductor deployment by major data center customers.
However, recent data center growth rates should not be mechanically extrapolated into future periods.
First, AI capital expenditure is project-driven, and actual semiconductor procurement can fluctuate with data center construction, computing platform transitions, and inventory adjustments.
Second, custom silicon programs can generate substantial revenue from individual customers. Changes in production schedules may therefore have a material effect on quarterly results.
Third, optical connectivity, networking switches, and custom silicon have different growth drivers, meaning that overall data center revenue may conceal important changes in the underlying product mix.
Future financial analysis should therefore use total data center revenue as a starting point while separately evaluating the growth contribution from Custom and Connectivity.
The data center business should not be treated as a single segment with a uniform growth rate and margin profile.
6.2 Revenue Guidance Indicates Continued Growth Expectations, but Remains Forward-Looking
In its FY2027 Q2 financial results, Marvell provided FY2027 Q3 revenue guidance of USD 3.150 billion at the midpoint, plus or minus 5%.
The company also guided to a GAAP gross margin of 52.9%–53.9%, a Non-GAAP gross margin of 57.5%–58.5%, and Non-GAAP diluted EPS of approximately USD 1.10, plus or minus USD 0.05.
Source: Marvell FY2027 Q2 Financial Results and Financial Outlook.
This guidance reflects management’s expectations for near-term revenue growth, but it remains forward-looking financial information.
Importantly, higher revenue does not necessarily translate into proportional earnings-per-share growth.
If incremental revenue comes primarily from products with relatively lower gross margins, or if new businesses require substantial R&D investment, revenue expansion may be accompanied by changes in profitability.
Additional shares and potential dilution may also affect the future EPS trajectory.
Financial projections must therefore evaluate revenue growth, gross margin, operating expenses, and diluted share count within the same framework.
Revenue guidance alone is insufficient to establish future earnings-per-share growth.
6.3 Gross Margin Is a Key Indicator of Earnings Quality
In FY2027 Q2, Marvell reported a GAAP gross margin of 53.1% and a Non-GAAP gross margin of 58.9%.
GAAP diluted EPS was USD 0.33, compared with Non-GAAP diluted EPS of USD 0.94.
Source: Marvell FY2027 Q2 Financial Results.
The difference between GAAP and Non-GAAP results reflects the impact of stock-based compensation, amortization of acquired intangible assets, and other adjustments.
Non-GAAP measures can help investors assess certain aspects of underlying operations, but they should not be treated as equivalent to the full economic earnings attributable to common shareholders.
For Marvell, future gross margin performance will depend on three principal factors.
The first is the business mix between custom silicon and standardized connectivity products. Different customers and products may have different pricing arrangements, design-fee structures, and production volumes, which can influence company-wide gross margin.
The second is technological evolution in high-speed optical DSPs and switching chips. New products may create incremental revenue opportunities, but advanced manufacturing processes, packaging, and development requirements can also increase costs.
The third is the integration and production efficiency of acquired businesses. Next-generation connectivity products may require substantial development and commercialization expenditure before reaching meaningful revenue scale.
Their profitability will need to be evaluated as commercial deployment progresses.
The key issue is therefore not simply whether Marvell can sustain high revenue growth, but whether changes in the business mix support relatively stable gross margins and continued operating profit expansion.
6.4 R&D Expenditure and Acquisition Costs Will Influence Operating Leverage
Marvell’s technological competitiveness requires sustained investment in research and development.
Custom compute, high-speed optical communications, and next-generation scale-up interconnects all involve complex chip design, advanced manufacturing processes, and extended customer qualification.
Unlike manufacturing businesses that primarily expand through additional production capacity, Marvell must continue investing in engineering resources, semiconductor design, and underlying intellectual property.
Consequently, high R&D expenditure does not necessarily indicate declining operating efficiency.
Its economic significance depends on the sustainable revenue ultimately generated by that investment.
In FY2027 Q2, Marvell reported GAAP R&D expenses of approximately USD 741 million, illustrating the importance of technology development within its operating cost structure.
Source: Marvell FY2027 Q2 Financial Results.
For custom silicon programs, sustained volume production may allow development expenses to be absorbed more efficiently over time.
For next-generation connectivity products that are still in the early stages of commercialization, however, R&D expenditure may precede revenue generation and weigh on near-term operating profit.
It is therefore important to distinguish operating leverage in mature businesses from upfront investment in new businesses.
We believe that more substantial operating leverage will depend on data center revenue growing faster than the associated operating expenses while gross margin remains relatively stable.
6.5 Cash Flow Must Be Evaluated Alongside Capital Requirements and Stock-Based Compensation
Cash flow is an important indicator of earnings quality, but operating cash flow should not be assessed solely in absolute terms.
During the first half of FY2027, Marvell generated approximately USD 1.244 billion in operating cash flow and reported approximately USD 282 million in purchases of property and equipment.
Using a simplified calculation of operating cash flow less this capital expenditure, free cash flow was approximately USD 962 million for the period.
The company also recognized approximately USD 534 million in stock-based compensation expense during the same period.
Source: Marvell FY2027 Q2 Form 10-Q. Free cash flow is calculated using the simplified definition of operating cash flow less purchases of property and equipment.
From an operating perspective, these figures indicate that Marvell has an established capacity to generate cash.
However, the fabless business model does not eliminate all manufacturing-related capital requirements.
Advanced chip production may involve supplier capacity arrangements, prepayments, technology licensing, and other contractual obligations.
At the same time, stock-based compensation generally does not create an immediate operating cash outflow, but it may affect existing shareholders’ per-share economic interests through an increase in outstanding shares.
Operating cash flow less capital expenditure therefore does not fully capture shareholder economic returns after dilution and other capital commitments.
Future financial analysis should consider free cash flow alongside stock-based compensation, share repurchases, acquisition expenditure, and changes in capital structure.
An improvement in cash flow should not automatically be equated with an equivalent increase in per-share value.
6.6 The Google Warrants and Acquisition Financing Make Per-Share Analysis More Important
Marvell has expanded its AI infrastructure portfolio through acquisitions and strategic partnerships, but these transactions also make the analysis of share count and capital structure more complex.
The Celestial AI acquisition involved both cash and share consideration. At the completion of the acquisition, Marvell expected the associated share issuance to increase its diluted weighted-average share count by approximately 27 million shares.
In addition, the Google partnership includes warrants subject to revenue-based vesting conditions, while the NVIDIA strategic investment must be assessed according to the terms of the relevant financing instruments.
Sources: Marvell acquisition announcements, Google warrant agreement, and related regulatory filings.
The financial effects of these arrangements should not be calculated by simply adding all potentially issuable shares to the current share count.
Different instruments have different exercise, conversion, and vesting conditions, as well as different accounting treatments.
The number of shares included in diluted EPS calculations may therefore change over time and as the relevant conditions are satisfied.
Nevertheless, these arrangements demonstrate why forecasting net income alone is insufficient to evaluate Marvell’s future financial performance.
The more important question is how much earnings per share can be generated as revenue and net income grow, after accounting for the economic costs of additional capital and equity arrangements.
We therefore believe Marvell’s financial analysis should extend beyond revenue growth to operating profit, cash flow, and per-share economic value on a fully diluted basis.
7 Valuation Framework — Separating Existing Earnings from Long-Term Growth Expectations
7.1 Not All Potential Business Opportunities Have the Same Degree of Visibility
Marvell currently operates businesses that generate established revenue, custom silicon programs entering volume production, and next-generation connectivity technologies that remain at earlier stages of commercialization.
These businesses differ in their development stages and in the visibility of their future revenue and profitability.
We therefore believe an appropriate valuation framework should distinguish the existing earnings base from longer-term opportunities that have yet to be fully realized.
The expansion of a potential market does not automatically mean that a company will capture a corresponding share of that market.
Even after securing a customer design program, revenue realization depends on production timing, actual deployment scale, and customer purchasing activity.
Valuation analysis should therefore account for differences in business maturity and the time required for potential revenue opportunities to translate into earnings.
7.2 A Three-Layer Earnings Framework
We propose dividing Marvell’s future operating performance into three layers.
The first layer comprises existing businesses and products with established revenue contributions. These include conventional data center connectivity, communications, and other infrastructure businesses. Their financial contribution should primarily be assessed through actual revenue, gross margin, and cash flow.
The second layer comprises custom computing and high-speed connectivity products that have secured customer programs and are progressing toward volume production. These businesses require continuous evaluation of production schedules, customer deployment, and actual financial performance.
The third layer comprises incremental opportunities that remain in technical qualification or early commercialization. These include certain scale-up optical interconnect products, future computing-platform support chips, and emerging network architectures.
These opportunities can be incorporated into longer-term operating scenarios, but they should not be assigned the same level of certainty as established businesses when assessing near-term earnings.
The three layers therefore require different assumptions and risk adjustments.
Existing businesses provide the current earnings foundation, production-stage programs influence revenue realization over the next one to two years, and longer-term technology opportunities affect the potential evolution of Marvell’s product coverage and growth profile.
7.3 Valuation Changes Depend on Both Earnings Realization and Market Expectations
For high-growth semiconductor companies, revenue and earnings expansion do not necessarily produce a proportional increase in market valuation.
If market prices already reflect substantial future growth, continued earnings expansion must be evaluated against the expectations embedded in the valuation.
Conversely, if commercialization of a particular business is delayed, expectations for earnings in the relevant periods and the associated valuation assumptions may need to be reassessed.
Marvell’s valuation should therefore be evaluated through three questions.
First, how much operating profit can future revenue growth generate?
Second, how might R&D investment, acquisition costs, and equity dilution affect the growth of per-share earnings?
Third, to what extent does the prevailing valuation already reflect the potential contributions from existing customer programs and longer-term connectivity opportunities?
Without a complete and independently developed financial model covering the relevant businesses, it would be inappropriate to derive a precise valuation conclusion solely from a single forward P/E multiple or an estimate of the potential addressable market.
This report instead focuses on the principal variables that could influence future earnings realization and valuation, using different operating scenarios to assess their potential financial implications.
7.4 Three Operating Scenarios
Base Case: Existing programs enter production gradually, while connectivity continues to grow.
Under this scenario, data center demand continues to support revenue, custom silicon programs progress broadly according to existing plans, and high-speed optical DSPs and networking switches remain important growth contributors.
Next-generation scale-up technologies gradually enter commercialization but do not become major revenue sources in the near term.
Earnings growth would primarily reflect revenue expansion and operating leverage within established businesses.
More Constructive Scenario: Custom Compute and Next-Generation Connectivity Commercialize in Parallel.
If existing customers expand custom silicon deployment, additional supporting-chip programs enter volume production, and optical connectivity and switching products gain broader customer adoption, revenue growth could accelerate.
However, the extent of profitability improvement would still depend on product mix, development expenditure, and the gross margins of incremental business.
Adverse Scenario: Customer Deployment Is Delayed and Business Expansion Does Not Produce Corresponding Profit Growth.
If major customers adjust capital expenditure plans, certain semiconductor programs experience production delays, or next-generation connectivity technologies commercialize more slowly than expected, revenue growth could moderate.
At the same time, existing R&D expenditure and acquisition-related costs could constrain operating profit improvement, while share dilution may affect earnings-per-share performance.
Overall, Marvell’s long-term business opportunities are closely linked to the evolution of AI infrastructure systems. However, their economic value still requires validation through production ramps, profitability, cash flow, and sustained per-share earnings growth.
8 Key Risks and Future Validation Indicators
Marvell’s future operating performance remains subject to several important risks.
Lower-than-expected hyperscaler capital expenditure. Changes in AI infrastructure investment plans could affect demand for custom silicon and connectivity products, altering the company’s revenue growth trajectory.
Custom silicon program delays and customer concentration. Architectural changes, project delays, or procurement adjustments by major customers could have a material impact on revenue and increase uncertainty around future operating performance.
Connectivity competition and technology transitions. The markets for optical DSPs, networking switches, and scale-up interconnects are highly competitive. Changes in technology roadmaps or weaker-than-expected customer adoption could affect addressable market opportunities and profitability.
R&D expenditure and supply chain constraints. Advanced manufacturing processes, packaging, capacity availability, and sustained development expenditure may increase operating costs. Production delays or lower-than-expected yields could also affect margins.
Acquisition integration and equity dilution. Celestial AI, XConn, and related strategic arrangements may generate additional operating expenses, integration costs, and potential dilution, affecting future per-share economic value.
Key indicators to monitor include revenue growth in Custom and Connectivity, production ramps of custom silicon programs, customer adoption of 1.6T optical DSPs and next-generation switching products, and commercialization progress at acquired businesses such as Celestial AI.
Marvell has also announced an Investor Day scheduled for October 6, 2026. Any additional information regarding long-term growth objectives, product roadmaps, and financial outlook will provide further reference points for evaluating the operating scenarios discussed in this report.
Source: Marvell Investor Relations announcement.
Research Conclusion
Marvell’s long-term business opportunities are expanding beyond demand for individual chips toward a broader range of custom computing and data connectivity functions within AI infrastructure.
On one side, hyperscalers’ continued development of proprietary computing architectures creates opportunities for custom XPUs and supporting semiconductor products. On the other, the expansion of AI clusters and the evolution of network architectures are increasing the importance of high-speed optical DSPs, networking switches, and next-generation scale-up interconnects.
These two growth drivers are interconnected. Custom computing platforms require efficient data transmission and resource management, while larger AI systems create additional applications for different connectivity technologies.
However, broader product coverage does not automatically translate into higher profitability. Customer purchasing schedules, product-specific margin structures, commercialization timelines, sustained R&D investment, and equity arrangements may all affect how effectively revenue growth translates into earnings per share and cash flow.
Our central research conclusion is therefore:
Marvell’s long-term growth opportunity depends not only on the increasing deployment of AI computing chips, but also on its ability to expand semiconductor content across different computing platforms through custom silicon, supporting chips, and high-speed connectivity. The key determinant of its long-term economic value will be whether these opportunities translate into sustained production revenue and, after accounting for profitability and equity costs, durable growth in per-share earnings and cash flow.
Within this framework, custom silicon revenue realization, data connectivity commercialization, and earnings conversion efficiency represent the three principal variables for evaluating Marvell’s future business development and valuation.
Disclaimer:
Sahm is subject to the supervision and control of the CMA, pursuant to its license no. 22251-25 issued by CMA.
The Information presented above is for information purposes only, which shall not be intended as and does not constitute an offer to sell or solicitation for an offer to buy any securities or financial instrument or any advice or recommendation with respect to such securities or other financial instruments or investments. When making a decision about your investments, you should seek the advice of a professional financial adviser and carefully consider whether such investments are suitable for you in light of your own experience, financial position and investment objectives. The firm and its analysts do not have any material interests or conflicts of interest in any companies mentioned in this report.
Performance data used in this report are obtained or compiled from company filings and disclosures, official stock exchanges and regulatory authorities, recognized industry associations, reputable financial databases, and other third-party sources that Sahm Capital Financial Company believes to be reliable.
IN NO EVENT SHALL SAHM CAPITAL FINANCIAL COMPANY BE LIABLE FOR ANY DAMAGES, LOSSES OR LIABILITIES INCLUDING WITHOUT LIMITATION, DIRECT OR INDIRECT, SPECIAL, INCIDENTAL, CONSEQUENTIAL DAMAGES, LOSSES OR LIABILITIES, IN CONNECTION WITH YOUR RELIANCE ON OR USE OR INABILITY TO USE THE INFORMATION PRESENTED ABOVE, EVEN IF YOU ADVISE US OF THE POSSIBILITY OF SUCH DAMAGES, LOSSES OR EXPENSES.