Marvell Reports a Record Quarter, but AI Growth Brings a Margin Test

Marvell entered Google News after reporting record quarterly revenue of $2.739 billion, driven by a 46% rise in data center sales. The headline number was strong, but the company’s changing revenue mix creates a more complicated test.

Data center revenue reached $2.17 billion during Marvell’s second fiscal quarter of 2027. That business generated 79% of company revenue, compared with 74% one year earlier. Total revenue grew 37% year over year and 13% from the previous quarter.

The result confirms that Marvell has become far more dependent on AI infrastructure spending. It also places the company against Broadcom in custom accelerators, networking silicon, and the connections surrounding large AI clusters.

Marvell is not trying to replace Nvidia’s flagship processors. Its opportunity sits around those processors, where data must move between chips, racks, and data centers. The company also designs custom computing silicon for cloud operators seeking specialized alternatives.

That positioning explains the record quarter. It also explains the central risk. Custom silicon can deliver substantial revenue, but its margins can trail those of standard products sold across many customers.

Google News Focuses on Marvell’s 46% Data Center Surge

The quarter matters because Marvell’s AI business accelerated while becoming an even larger share of the entire company.

Marvell reported the results on August 27, 2026, for the quarter that ended August 1. Its quarterly results showed revenue exceeding management’s previous midpoint by $39 million.

The company recorded GAAP net income of $308 million, equal to $0.33 per diluted share. Non-GAAP net income reached $865.9 million, while non-GAAP earnings were $0.94 per diluted share.

Operating cash flow totaled $605.5 million. GAAP gross margin was 53.1%, while the company’s adjusted calculation produced a 58.9% gross margin.

The data center business supplied the clearest growth signal. Revenue there increased from approximately $1.49 billion one year earlier to $2.17 billion in the latest quarter.

That represents growth of roughly $680 million within twelve months. It also means the data center segment generated nearly four dollars from every five dollars of Marvell’s quarterly revenue.

Chairman and CEO Matt Murphy said AI-related bookings remained exceptionally strong. He also said revenue growth should accelerate through the remainder of fiscal 2027.

Those statements remain management forecasts, not independently verified outcomes. However, Marvell supported them with a higher near-term target.

For the third fiscal quarter, management forecast revenue of $3.15 billion, plus or minus 5%. The midpoint implies sequential growth of approximately 15% from the latest record.

Marvell also projected GAAP earnings of $0.53 per diluted share, plus or minus $0.05. Its non-GAAP forecast was $1.10, using the same range.

The outlook indicates that the second quarter was not presented as a temporary peak. Management expects another substantial increase as connectivity products and custom programs expand together.

This acceleration separates the latest result from the company’s earlier AI recovery. Marvell’s fiscal first-quarter revenue was $2.418 billion, up 28% year over year.

Three months later, total growth had reached 37%. Data center growth reached 46%, matching the annual growth rate that segment delivered across fiscal 2026.

The unusual point is not simply that Marvell set another record. The company had already reported several record quarters during the AI infrastructure expansion.

The important change is the composition of that growth. Connectivity remains strong, while management expects custom silicon to accelerate significantly during the second half of fiscal 2027.

That combination gives Marvell two ways to participate in an AI deployment. It can help design specialized processors, then sell components that move data around those processors.

This is why the Google News headline deserves more than a quick earnings summary. Marvell is becoming a concentrated bet on how hyperscalers build their next generation of AI systems.

AI Clusters Need More Than Expensive Accelerators

Marvell’s opportunity grows when networking and data movement become constraints on useful AI computing capacity.

An AI cluster combines thousands of processors, memory devices, switches, and optical links. Performance depends on how efficiently those components exchange data, not only on each processor’s individual speed.

Scale-out networking connects servers or racks across a cluster. Scale-up networking links processors more tightly, allowing them to operate on a large workload with lower communication delays.

Scale-across technology connects separate data centers or computing campuses. This layer becomes more important when a single facility cannot hold the processors, power capacity, or cooling equipment required.

Marvell sells products across these connection layers. Its portfolio includes optical digital signal processors, Ethernet switches, interconnect chips, storage controllers, and custom application-specific integrated circuits.

An application-specific integrated circuit, or ASIC, is designed for a defined workload. Cloud operators use these custom chips to improve efficiency or reduce dependence on general-purpose processors.

Marvell does not manufacture those chips itself. It develops designs and intellectual property while relying on outside foundries and manufacturing partners for production.

The company’s custom platform combines high-speed interfaces, Arm-based computing, security, storage, advanced packaging, and optical technologies. Customers can select components while retaining control over the final system architecture.

Connectivity provides a broader product base. An optical digital signal processor converts and conditions data as it travels through fiber links inside or between data centers.

Marvell cited continued demand for 800-gigabit optical products and a growing contribution from 1.6-terabit equipment. Higher link speeds allow more information to move without multiplying physical connections at the same rate.

The company also sells 51.2-terabit Ethernet switching silicon. These switches direct traffic between servers and racks, helping large clusters avoid communication bottlenecks.

Management expects its 51.2-terabit products to more than double scale-out switching revenue during fiscal 2027. That expectation depends on customer deployments, product availability, and continued infrastructure spending.

Marvell has also expanded through acquisitions. It completed its acquisition of Celestial AI on February 2, 2026, adding an optical interconnect platform designed for scale-up connections.

Celestial AI’s Photonic Fabric uses optical communication to move data between computing and memory resources. Marvell says the approach can support high bandwidth with lower latency and improved power efficiency.

The Celestial AI transaction added technology that targets future systems, rather than most of the revenue reported in the latest quarter.

Marvell completed its XConn acquisition eight days later. XConn develops switches for PCI Express and Compute Express Link connections, which move data between processors, accelerators, and memory.

These acquisitions show where Marvell expects pressure to move. AI systems face increasing constraints from memory bandwidth, electrical reach, power consumption, and communication delays.

Adding more accelerators does not solve those limitations automatically. A processor waiting for data is expensive capacity that cannot complete useful work.

This creates a market around Nvidia, AMD, and cloud-designed processors. Marvell can benefit from accelerator growth even when another company supplies the central computing engine.

Marvell’s first-quarter filing also disclosed a strategic partnership with Nvidia. The arrangement connects Marvell’s custom processors and compatible networking technologies with Nvidia’s AI infrastructure environment.

That relationship makes Nvidia supporting context, rather than Marvell’s primary opponent. The sharper competitive comparison is Broadcom, which also combines custom accelerator programs with AI networking products.

Broadcom reported $10.8 billion in fiscal second-quarter AI semiconductor revenue, up 143% year over year. Its AI semiconductor results demonstrate the scale of the opportunity and the competition facing Marvell.

Marvell therefore needs more than expanding AI capital expenditure. It must convert its technical portfolio into durable customer programs while competing against a much larger custom silicon supplier.

Marvell and Broadcom Are Competing for the Architecture Around AI

The central contest is over who becomes the preferred design and connectivity partner for hyperscalers building specialized AI infrastructure.

Broadcom and Marvell both occupy positions beyond merchant processors. Each company supplies networking products and helps large customers develop chips designed for specific workloads.

These engagements differ from selling the same processor to thousands of buyers. A custom program can require years of engineering before generating meaningful production revenue.

Once deployed, the resulting chip can ship in significant volumes. It can also become deeply connected to the customer’s software, networking, and data center design.

That creates potentially durable relationships. However, each major program can materially affect quarterly growth, making timing and customer concentration important.

Marvell said its custom silicon business had grown from almost nothing to approximately 25% of data center revenue over several years. That statement appeared in its 2026 proxy materials.

The company expects custom programs to accelerate during the second half of fiscal 2027. It also sees processor-attached chips becoming another multibillion-dollar opportunity over time.

An XPU is a general label for specialized computing processors, including AI accelerators. An XPU-attached chip handles supporting functions placed close to that main processor.

These supporting devices can manage connectivity, memory movement, security, or other system tasks. Their importance increases as cloud operators build more specialized architectures.

Marvell’s advantage is the breadth of intellectual property it can combine inside a customer design. High-speed serial interfaces, optical links, packaging, and storage can all influence system performance.

Broadcom offers a similar combination at greater scale. Its reported AI revenue includes custom accelerators and networking, giving it established relationships with major cloud operators.

The competition is not decided by a single benchmark. It depends on design wins, production execution, manufacturing access, software compatibility, and each customer’s internal roadmap.

Customer identity also remains partly obscured. Semiconductor suppliers often avoid naming hyperscale clients because their contracts and product plans are confidential.

That limits outside verification of individual custom programs. Investors usually see aggregate bookings, revenue guidance, and management descriptions before seeing customer-level evidence.

Marvell’s latest quarter provides aggregate proof that demand is growing. It does not reveal how much future growth depends on one customer, one processor generation, or one deployment schedule.

Its annual filing identifies customer concentration, manufacturing dependence, and rapid product transitions among material business risks.

Those issues matter more as the data center share approaches four-fifths of total revenue. Weakness in another segment can no longer offset a large interruption in AI infrastructure demand.

At the same time, Marvell’s concentration is not accidental. The company has sold businesses outside its preferred infrastructure strategy and acquired technologies aimed directly at AI connectivity.

Fiscal 2026 revenue reached $8.2 billion, up 42% from fiscal 2025. Data center sales increased 46%, while communications and other revenue rose 31%.

The latest quarter extends that shift. Data center revenue alone now exceeds the company’s total annualized revenue when Murphy became CEO in 2016.

This transformation gives Marvell clearer exposure to AI investment. It also reduces the diversification that previously softened changes within any single end market.

Broadcom’s scale raises another challenge. It can fund several custom programs, absorb development costs, and spread reusable intellectual property across a larger revenue base.

Marvell must counter with specialized expertise and execution. Its optical portfolio, Ethernet switching products, and acquired scale-up technologies broaden the value it can bring to one customer design.

The company can also participate when customers do not select it for the main accelerator. Networking and optical components can appear in systems built around competing processors.

That distinction keeps the contest from becoming a simple winner-take-all race. A hyperscaler can use one supplier for an accelerator and another for switching or optical connections.

However, the most valuable relationships extend across multiple layers. A supplier involved in the processor, attached devices, switching, and optics gains more revenue from each deployed cluster.

Marvell’s record therefore shows progress, not competitive resolution. Broadcom’s larger AI business remains the reference point that Marvell must confront as custom revenue accelerates.

Custom Silicon Growth Comes With a Margin Tradeoff

The strongest part of Marvell’s growth forecast is also the part most likely to pressure gross margin.

Marvell produced a 58.9% non-GAAP gross margin in the latest quarter. Its third-quarter guidance places that measure between 57.5% and 58.5%.

The midpoint represents a decline of 0.9 percentage points. Management linked part of that pressure to the faster growth of custom products.

A standard semiconductor product spreads design costs across multiple customers. The supplier controls the product roadmap and can capture more value when demand rises.

A custom chip serves one customer’s requirements. Large customers can negotiate aggressively because they contribute substantial volume and participate closely in the design process.

This does not make custom silicon unattractive. It means revenue growth and gross margin expansion do not always arrive together.

Custom programs can still produce operating leverage. Engineering costs are incurred before production, while higher shipment volume can lift revenue faster than ongoing operating expenses.

Marvell expects to reach a non-GAAP operating margin between 38% and 40% during its fourth fiscal quarter. That goal relies on strong revenue growth offsetting pressure at the gross-margin level.

The difference matters because a record top line can hide changing economics. Investors should watch how much additional operating profit Marvell retains from each new dollar of AI revenue.

The latest quarter offers encouraging evidence, but not a complete answer. Non-GAAP earnings per share grew 40% year over year, slightly faster than total revenue.

GAAP net income also increased substantially from the previous year. However, comparisons between GAAP and non-GAAP results require care because adjustments exclude several real expenses.

Marvell’s operating costs are rising as it pursues the larger opportunity. Management increased its fiscal 2027 non-GAAP operating expense forecast to support additional programs and revenue.

Supply is another constraint. Marvell plans substantial capacity prepayments to manufacturing partners during fiscal 2027.

These payments help secure access to wafers, packaging, and other constrained resources. They also commit cash before customer demand becomes reported revenue.

Operating cash flow was $605.5 million during the second quarter, down slightly from the first quarter’s record. Management attributed part of that movement to higher supplier prepayments.

Inventory ended the quarter near $1.36 billion. A nearly flat sequential balance suggests Marvell was not simply building a large stockpile to create current revenue.

Still, supply commitments expose the company if customer schedules change. A delayed program can leave components, capacity reservations, or engineering resources without their expected near-term return.

Marvell reported total debt of approximately $5 billion at the quarter’s end. It also issued new senior notes during 2026 while financing acquisitions and future capacity.

Its quarterly filing describes integration, supply, concentration, and forecasting risks in greater detail than the earnings headline.

Acquisition integration adds another uncertainty. Celestial AI and XConn expand the product roadmap, but their technologies still need to reach customer qualification and production.

Optical scale-up connections remain an emerging market. Marvell’s technology could gain adoption, yet customers are evaluating several electrical and optical approaches.

Standards can also change purchasing decisions. Hyperscalers may favor open interfaces, proprietary connections, or a mixture depending on their processor and network architecture.

The company’s forecast assumes broad demand across connectivity and custom products. A slowdown in either category would weaken the argument that Marvell has two independent AI growth engines.

Customer concentration deserves equal attention. A small number of cloud operators account for much of global AI infrastructure spending.

Winning one major program can accelerate revenue rapidly. Losing a socket, facing a delayed deployment, or encountering an internal customer redesign can reverse that effect.

The reported 46% data center growth is verified at the segment level. Claims about future bookings, product ramps, and market size remain forward-looking statements from management.

That does not invalidate the forecast. It defines what readers should treat as evidence and what still requires confirmation.

The margin guide is the first visible pressure test. If custom revenue accelerates while operating margin improves, Marvell will show that scale can outweigh the product-mix drag.

If margins keep declining, record revenue will look less impressive. The company would then be growing quickly while retaining a smaller portion of each sale.

Three Signals Will Test Marvell’s AI Growth Story

The next three checkpoints will show whether Marvell’s record reflects a durable platform shift or an unusually strong stage of the spending cycle.

The first signal is Marvell’s third-quarter revenue and gross margin. Management’s $3.15 billion midpoint requires approximately 15% sequential growth.

Reaching that target would confirm that connectivity demand and custom silicon are accelerating together. Missing it would raise questions about program timing or supply availability.

Gross margin will provide the second half of the answer. A result within the guided range would support management’s explanation that mix pressure remains controlled.

A result below that range would suggest that custom growth, manufacturing costs, or other factors are reducing the value captured from rising sales.

Readers should judge both measures together. Revenue acceleration without acceptable margin performance would weaken the economic case behind the Google News headline.

The second signal is Marvell’s Investor Day, scheduled for October 6, 2026. Management has promised more detail about its long-term AI infrastructure strategy.

The useful disclosures will concern customer diversity, custom program timing, optical scale-up adoption, and the expected contribution from acquired technologies.

Broad market estimates will matter less. Investors already understand that AI infrastructure spending is large.

The harder question is how much of that spending Marvell can convert into recognized revenue. Another key question is whether those sales produce durable operating profit.

Specific production schedules would strengthen confidence in the fiscal 2028 outlook. Vague market-size claims without customer or deployment evidence would leave the current uncertainty intact.

The third signal is competitive execution from Broadcom and major cloud operators. Broadcom’s custom accelerator and networking revenue provides the closest public comparison.

Continued rapid Broadcom growth would confirm that hyperscaler demand remains strong. It would also raise the standard Marvell must meet to gain share.

Cloud companies’ capital expenditure plans offer another part of this signal. Persistent spending supports demand for accelerators, Ethernet switches, optical links, and custom processors.

A shift toward slower infrastructure growth would test every supplier. Marvell would face added exposure because data centers now contribute 79% of its revenue.

Product announcements also deserve scrutiny. New custom accelerators reveal which cloud operators are internalizing more chip design and which semiconductor partners support them.

Deployment details matter more than prototype announcements. Production volume, network architecture, and connection speeds determine how much revenue reaches suppliers.

Developers and enterprise AI buyers do not purchase most of Marvell’s components directly. However, these infrastructure choices influence the cost and availability of the computing services they use.

Better networking can raise accelerator utilization, reducing the time expensive processors spend waiting for data. Optical links can also help clusters expand beyond electrical connection limits.

Those improvements can affect model training times, inference capacity, and cloud service economics. They will not guarantee lower customer bills, because providers control pricing and capacity allocation.

Knowledge workers following these developments through Google News should separate demand evidence from company forecasts. The latest revenue is reported performance, while future program ramps remain expectations.

The record quarter establishes that Marvell has secured a meaningful position in AI infrastructure. Its data center business is no longer a secondary source of growth.

The harder phase begins now. Marvell must scale custom silicon, preserve profitability, integrate new optical technologies, and compete with Broadcom for major designs.

Watch the third-quarter revenue target first, then the October Investor Day, followed by competing custom chip deployments. Together, those signals will test whether Marvell’s acceleration can last.

For readers tracking AI infrastructure, the practical question is no longer whether Marvell benefits from the buildout. The latest results answer that clearly.

The question is whether Marvell can turn concentrated demand into a balanced, profitable platform before spending patterns or customer roadmaps change. Keep that distinction in view as the next Google News headline arrives.