Marvell’s $18 Billion Revenue Target Puts Its AI Chip Forecast on Trial

Marvell raised its fiscal 2028 revenue forecast to roughly $18 billion, creating a striking google news headline about its data center chip business. However, the headline compresses an important distinction. The target covers Marvell’s entire company, while data center products provide most of the expected growth.

That distinction does not make the forecast less consequential. Marvell expects fiscal 2028 data center revenue to grow more than 60% from the preceding year. It also expects its custom silicon business to more than double during that period.

The forecast places Marvell against a demanding reality. Broadcom already holds deep custom-chip relationships with major cloud operators, while Nvidia controls the general-purpose AI accelerator market. Marvell must convert new designs, including an expanded Google partnership, into high-volume production without missing customer schedules.

The result is more than another optimistic AI projection. Marvell is betting that cloud companies will buy increasingly specialized processors, networking components, optical links, and memory controllers as their computing clusters expand.

What the $18 Billion Marvell Forecast Actually Means

Marvell forecast roughly $18 billion in total fiscal 2028 revenue, not $18 billion from data center chips alone.

The distinction matters because the original google news framing can leave readers with an inflated view of one business segment. Marvell’s management presented the figure during its fiscal second-quarter 2027 earnings call on August 27, 2026.

The company had previously projected roughly $16.5 billion for fiscal 2028. Its revised target added $1.5 billion and implied approximately 50% annual growth from the fiscal 2027 outlook.

Marvell also raised its fiscal 2027 revenue expectation to roughly $12 billion. That forecast was up from approximately $11.5 billion one quarter earlier.

The reported quarter provides evidence behind those revisions. Marvell generated $2.739 billion in total revenue during the quarter ending August 1, according to its quarterly filing. Revenue increased 37% from the comparable period one year earlier.

Data center revenue reached approximately $2.17 billion. It grew 46% annually and 18% sequentially, accounting for 79% of company revenue.

Those figures show why data center demand dominates the story. Still, they do not turn the fiscal 2028 company forecast into a segment-specific target.

Marvell’s remaining operations include enterprise networking, carrier infrastructure, consumer products, and other semiconductor categories. Their contribution is smaller, but it remains part of the $18 billion total.

Management’s segment forecast is separate. Marvell expects data center revenue to grow approximately 60% in fiscal 2027 and more than 60% in fiscal 2028.

That outlook is aggressive because the growth would build upon an already larger base. A company can post a high percentage increase when starting from limited sales. Sustaining that rate after data center revenue reaches billions each quarter is more difficult.

Marvell’s third-quarter guidance reinforces the expected near-term acceleration. The company projected approximately $3.15 billion in total revenue, subject to its stated guidance range.

Management also expected data center revenue to increase more than 20% sequentially and roughly 75% annually during that quarter. Those shorter-term numbers provide a bridge between current sales and the longer fiscal 2028 projection.

The correction therefore changes the interpretation, not the central conclusion. Marvell has not promised an $18 billion data center unit. It has tied almost its entire growth story to that unit’s performance.

Why Google News Put Marvell’s Custom Silicon Bet in Focus

The expanded Google relationship gives Marvell a route into more parts of the TPU system, but it does not guarantee the forecast.

Google and Marvell signed a commercial agreement on July 29, 2026, covering custom semiconductor development. Marvell disclosed the arrangement through a regulatory filing on August 19.

The partnership includes programs attached to Google’s Tensor Processing Unit, or TPU, ecosystem. A TPU is Google’s specialized processor for training and running machine-learning models.

According to Marvell’s commercial agreement filing, the work spans AI inference accelerators, storage controllers, network interface controllers, and memory interface controllers. It also covers near-memory computing.

Near-memory computing places processing closer to stored data. The approach can reduce the energy and time spent moving information between processors and memory.

The agreement therefore extends beyond building one headline processor. Marvell is positioning itself across the supporting silicon required to move, store, and process data inside Google’s AI infrastructure.

Marvell also issued Google a warrant covering up to 58,970,907 company shares. Most of those shares depend on commercial performance rather than the passage of time.

The structure aligns Google’s potential equity benefit with purchases of qualifying custom products. It also reveals the scale of business that Marvell hopes the relationship can support over several years.

However, the warrant is not equivalent to a committed purchase order. Vesting conditions can indicate ambitious commercial thresholds without establishing when every program will ship.

Custom silicon development follows a long sequence. The supplier and customer define the design, complete verification, manufacture initial chips, validate systems, and then increase production.

Revenue remains limited during much of that process. A delayed design decision or manufacturing issue can move meaningful sales between fiscal years.

Marvell says its Google programs are included in the fiscal 2028 forecast. Investors must therefore judge whether the company’s production schedules are conservative enough for a relationship covering several new product categories.

The partnership also does not remove Broadcom from Google’s supply chain. Broadcom has worked with Google on custom AI processors for years and announced another long-term agreement in 2026.

Google can use multiple suppliers across processor generations and supporting components. Marvell’s expansion signals diversification within Google’s infrastructure, not a clean replacement of an incumbent.

That is the real importance behind the google news attention. Marvell has gained a larger role inside a major AI platform. It still has to turn that role into shipped systems and recognized revenue.

Marvell Is Challenging Broadcom for More Custom AI Work

Marvell’s main contest is not directly against Nvidia’s GPU sales. It is against Broadcom for custom cloud silicon and the infrastructure surrounding it.

Nvidia sells programmable accelerators that can support many customers and workloads. Custom application-specific integrated circuits, or ASICs, are designed for narrower requirements defined by one customer.

Cloud operators pursue custom chips for several reasons. They can tailor compute, memory, networking, and power characteristics around workloads running at enormous scale.

A custom accelerator can reduce dependence on a single merchant-chip supplier. It can also give a cloud operator tighter control over its hardware and software roadmap.

Those benefits come with tradeoffs. A cloud company must commit engineering resources, predict future workload needs, and support a design that cannot easily serve another customer.

Broadcom has built its position through long-running design relationships with hyperscalers. Hyperscalers are cloud operators that deploy computing infrastructure across very large global networks.

Marvell is pursuing the same customers through a broader data center portfolio. Its products include custom processors, optical components, Ethernet switching chips, storage controllers, and other interconnect technology.

Interconnects move data among processors, memory, storage, and networking equipment. Their importance rises as AI clusters combine more accelerators into one computing system.

Marvell argues that this breadth lets it participate even when a customer chooses another company’s main accelerator. A data center still needs switching, optical links, memory interfaces, and storage connectivity around that processor.

That model helps explain why data center sales can grow before every custom accelerator reaches full production. Marvell’s interconnect products already benefit from larger and denser AI clusters.

The company’s 2025 custom AI presentation estimated that its addressable data center semiconductor market would reach $94 billion by calendar 2028. It identified custom accelerated computing as the largest opportunity within that total.

An addressable market estimate does not predict a supplier’s revenue. It describes the pool of potential spending that products might serve.

Marvell’s fiscal 2028 projection requires more than a growing market. The company must capture enough designs, deliver them on schedule, and retain meaningful content within each system.

Broadcom faces the same execution demands, but it starts with established customer positions. Marvell’s Google expansion increases competitive pressure without eliminating that advantage.

Nvidia remains part of the surrounding tension. Its GPUs offer a mature software environment and flexibility that custom chips often cannot match.

Cloud providers do not need to choose only one route. They can use GPUs for rapidly changing models and custom processors for stable workloads deployed at enormous volume.

Marvell’s opportunity therefore depends on a mixed market. Custom chips must take a larger share of cloud computing, while general-purpose accelerators continue serving workloads that value flexibility.

That makes Broadcom the clearer primary opponent. Both companies sell design expertise and supporting infrastructure to customers building their own AI processors.

The contest will not be settled through one benchmark. It will be decided through design wins, production volumes, customer spending, and the number of product generations each supplier retains.

The Mechanism Behind Marvell’s Accelerating Forecast

Marvell expects several product ramps to overlap, turning years of engineering work into a concentrated period of revenue growth.

Management said more than ten XPU-related attachment programs were moving toward higher-volume production. XPU is a broad industry label for specialized processors used in accelerated computing systems.

An attachment program supplies another chip used alongside the primary processor. Examples include networking, memory, storage, and connectivity components.

These programs matter because each accelerator creates demand beyond the processor itself. Larger clusters require faster links, more switching capacity, and better coordination between compute and memory.

Marvell expects its custom business to accelerate during the second half of fiscal 2027. It then expects custom revenue to more than double during fiscal 2028.

A new Tier 1 customer program is also scheduled to enter production. Marvell uses Tier 1 to describe a leading cloud or technology customer, although it does not publicly identify every program.

The forecast combines these custom ramps with continued interconnect growth. That combination spreads the opportunity across several product categories, but it does not remove customer concentration.

The company’s latest quarter shows how the model is changing. Data center revenue represented 79% of sales, compared with 74% in the prior-year period.

That concentration works in Marvell’s favor while cloud infrastructure spending rises. It also means a slowdown in AI deployment would have a larger effect on the company than before.

Supply preparation provides another clue about management’s confidence. Marvell expects approximately $1 billion in supplier capacity prepayments during fiscal 2027.

A capacity prepayment reserves manufacturing resources before products ship. Semiconductor companies use such arrangements when expected volume exceeds readily available supply.

Prepayments can support growth, but they also create risk. Marvell is committing capital based on forecasts for customer demand and production timing.

Advanced custom chips depend on outside manufacturing and packaging partners. Marvell designs the products but does not own the fabrication plants that produce them.

That arrangement lets the company access advanced processes without building factories. It also leaves Marvell exposed to shared supply constraints across the semiconductor industry.

The Google agreement adds another layer to the mechanism. It covers multiple products connected to one platform rather than a single design.

If several programs succeed, Marvell can earn revenue from different locations inside the same data center architecture. A weaker result in one category might then be offset by growth elsewhere.

The opposite is also true. Programs tied to one customer can share deployment schedules, capital budgets, and architectural decisions.

A change in Google’s infrastructure plan could affect several Marvell products at once. Diversification by product does not always equal diversification by customer.

Marvell’s $18 billion target therefore rests on synchronized execution. Custom processors must pass validation, attachment chips must arrive with them, and interconnect demand must keep expanding.

This is why the forecast is larger than a simple extrapolation from one strong quarter. Management is expecting multiple ramps to reinforce each other over the next several reporting periods.

What Marvell’s $18 Billion Target Does Not Prove

The forecast shows management’s expectations, but it does not establish customer commitments, future market share, or flawless production execution.

Marvell’s strongest evidence is current revenue. The company has reported record sales, expanding data center contribution, and rising near-term guidance.

The fiscal 2028 figure is less certain because it sits several product cycles ahead. The company must estimate volumes for chips that are still entering production or completing development.

A design win is not the same as sustained revenue. Customers can change deployment schedules, reduce order volumes, or move later generations to another supplier.

Custom silicon also creates concentration. A small number of large cloud operators account for much of the available market and can influence commercial terms.

Losing one major program can therefore matter more than losing many smaller customers. The effect becomes larger when each program spans several related components.

Marvell must also manage gross margin pressure. High-volume custom products can carry different economics from standard products sold across many customers.

The company reported a 36.6% non-GAAP operating margin in its latest quarter. Management expects scale to support profitability, but the future product mix will determine how much revenue growth reaches earnings.

Another uncertainty concerns the meaning of Google’s warrant. The arrangement signals a deep relationship and provides an incentive linked to commercial activity.

It does not independently verify future purchase volumes. Investors should separate the maximum number of warrant shares from expected semiconductor revenue.

The broader AI investment cycle creates a second risk. Cloud companies currently spend heavily on data centers, power, networking, and accelerators.

Those budgets depend on demand for AI services and the economics of running them. A slower return on infrastructure spending can delay new clusters or reduce their size.

Marvell’s data center concentration would amplify that change. Its smaller businesses would not automatically offset a broad pause in AI infrastructure orders.

Competition remains another pressure point. Broadcom can defend existing custom relationships, while Nvidia can make merchant accelerators more attractive through faster products and software improvements.

Cloud operators also maintain internal chip teams. Their engineering capabilities give them greater control over design choices and supplier negotiations.

Marvell’s own filings describe additional operating risks. These include dependence on third-party manufacturing, uncertain customer forecasts, rapid product transitions, and substantial development costs.

The latest financial results establish the current baseline. They cannot remove uncertainty from a forecast covering the next fiscal year.

Readers should also treat the original headline carefully. It describes the data center business as reaching $18 billion, while management described approximately $18 billion in overall company revenue.

The error is material because segment and company forecasts answer different questions. One measures the scale of Marvell’s central growth engine. The other includes every reporting category.

The most defensible conclusion is narrower. Marvell expects data center growth above 60% to drive company revenue toward $18 billion during fiscal 2028.

That claim is still ambitious. It simply avoids attributing the full target to one segment before Marvell has reported such a breakdown.

Three Signals That Will Test the Google News Narrative

The next tests are production growth, greater customer detail, and evidence that revenue can expand without weakening execution or profitability.

The first signal is Marvell’s fiscal third-quarter data center performance. Management expects the segment to grow more than 20% sequentially and approximately 75% annually.

Meeting that outlook would show that the acceleration is appearing in reported sales. A miss caused by program timing would weaken confidence in the fiscal 2028 path.

One quarter cannot validate a multiyear forecast. However, the expected increase is large enough to reveal whether current product ramps are moving as management described.

The second signal is Marvell’s Investor Day on October 6, 2026. The company has said it will discuss the long-term growth drivers behind custom silicon and data center infrastructure.

Investors need clearer bridges between current revenue and the $18 billion target. Useful details would include program timing, product mix, customer diversification, and anticipated margin effects.

Management does not need to identify every customer. It does need to explain how many programs contribute meaningful volume and when those contributions begin.

The event should also clarify how Marvell separates custom compute from attachment products and interconnect revenue. Those categories face different competitors, schedules, and economic profiles.

More detailed disclosure would strengthen the forecast by making its assumptions testable. Repeating the headline target without those connections would leave the largest uncertainties intact.

The third signal is the conversion of the Google partnership into qualifying product revenue. Marvell’s regulatory documents describe the scope of the relationship and its performance-linked warrant.

Future filings can reveal whether related commercial conditions are being satisfied. Marvell’s reported custom growth can also indicate whether Google-linked programs are reaching production.

The warrant agreement provides a formal framework, but the operating results will determine its importance.

These signals should be read together. Strong quarterly growth without customer diversification can increase concentration risk. More design wins without production revenue can extend the waiting period.

Revenue growth also needs to arrive with acceptable margins and supply discipline. Large sales gains lose some value if prepayments, development costs, or unfavorable product mix absorb the benefit.

For developers and enterprise buyers, the outcome affects more than Marvell’s financial results. A larger custom silicon market changes which processors support future AI services.

More specialized hardware can improve efficiency for stable, high-volume workloads. It can also produce greater differences among cloud platforms and their software environments.

Teams evaluating AI infrastructure should track processor availability, networking architecture, memory capacity, and workload portability. The accelerator name alone provides an incomplete view of system performance.

Knowledge workers following these programs can maintain a searchable knowledge base containing filings, product documents, benchmarks, and deployment announcements. That record makes changing corporate forecasts easier to evaluate.

The google news headline captured the size of Marvell’s ambition but blurred what management actually forecast. The company is targeting roughly $18 billion in total fiscal 2028 revenue, powered mainly by data center growth.

Now the burden shifts from projection to execution. Watch the next quarter, the October Investor Day, and evidence of Google-related production. Together, those signals will show whether Marvell’s forecast is becoming an operating result or remaining an ambitious target.