An AI accelerator earns its cost back when it is doing useful work. Data has to reach it quickly, memory has to be available, and thousands of processors have to exchange information without consuming an unacceptable amount of power. Marvell sells silicon that helps determine how efficiently that happens.
That is why I find the investment case for $MRVL increasingly compelling. As cloud operators build larger AI systems, more of their spending must address the practical constraints that limit the return on those systems. Marvell has exposure to custom processors, optical links, switching and memory access. Each serves a different part of the same customer's infrastructure.
The September 17 agreement with GlobalFoundries adds a timely manufacturing commitment to that strategy. Its significance becomes clearer when viewed alongside Marvell's product portfolio and the customer relationships developing around it. The potential is substantial because Marvell can grow the amount of technology it supplies to an AI deployment as the deployment itself becomes larger.
The GlobalFoundries agreement makes capacity part of the investment case
On September 17, GlobalFoundries and Marvell expanded their multiyear agreement to add silicon-germanium, or SiGe, capacity at GF's Burlington, Vermont facility. The expansion builds on existing production and addresses components used in pluggable optical transceivers, near-packaged optics and co-packaged optics. GF says its current technology supports 200 gigabits per second per lane, with a roadmap for higher speeds.
My interpretation is that Marvell is preparing the manufacturing base for a broader optical business. Access to qualified capacity can influence whether an attractive design becomes a product that ships on the customer's schedule. Expanding an established production relationship should help with that requirement.
The announcement does not disclose wafer volumes, contract value or a revenue timetable. I therefore treat it as evidence of supply planning and strategic commitment. The commercial benefit will become visible as additional capacity supports shipments.
Optical networks still depend on difficult analog engineering
Light carries the information through an optical link, but electronic circuits still have to prepare and recover the signal. Drivers operate the transmitting device. At the receiver, a transimpedance amplifier turns the photodetector's small electrical current into a usable voltage signal. Noise, signal integrity and power consumption become demanding constraints at higher speeds.
SiGe is valuable in this part of the system because it supports high-frequency analog performance. GlobalFoundries specifically identifies optical amplifiers and drivers among the applications for its technology. Marvell's own TIA portfolio serves optical modules spanning short data-center connections and much longer links.
This gives the GF announcement a precise industrial context. A faster digital processor alone cannot solve every problem in an optical connection. The analog components must keep pace. For investors, that creates an opportunity to study the suppliers whose engineering becomes more demanding as bandwidth rises.
It also explains why a specialty manufacturing process can remain strategically important alongside advanced digital nodes. Different functions require different semiconductor technologies. Marvell's ability to assemble those capabilities into a qualified customer solution is a meaningful part of its value.
The optical opportunity extends across multiple architectures
Marvell already has products for the transition to 1.6-terabit optical modules. Its Ara platform uses a 3nm process and eight 200-gigabit electrical lanes paired with eight optical lanes. At its introduction, Marvell said the platform could reduce module power by more than 20% compared with its preceding approach. That is a company performance claim, with the final result dependent on the module implementation.
Lower power has an economic consequence. When a network needs thousands of links, improving the consumption of each link can reduce the burden on the overall installation. Customers evaluate bandwidth together with the electricity and cooling required to deliver it.
Marvell also offers a 200-gigabit-per-lane TIA and driver chipset for 800G and 1.6T linear-drive pluggable optics, or LPO, aimed at suitable short-reach connections. This matters because some optical designs reduce the need for a dedicated DSP inside the module. Marvell has products that participate in that alternative as well.
My bullish view depends on that breadth. Pluggables, LPO, near-packaged optics and co-packaged optics involve different design choices and can coexist. A broad portfolio gives Marvell more ways to compete as customers choose their architectures. The revenue and margin earned per link will still vary; an optical transition should never be modeled as the same chip content repeated everywhere.
Custom silicon gives Marvell a role earlier in the design process
A hyperscaler designing a custom accelerator must integrate much more than a compute engine. The design needs interfaces, memory access, packaging and a manufacturing plan that can survive qualification and operate at volume. Marvell's custom ASIC offering combines reusable intellectual property with engineering capabilities across those requirements.
Its disclosed capabilities include high-speed SerDes, die-to-die interconnects, advanced packaging and custom HBM integration work with Micron, Samsung and SK hynix. These functions matter because a customer can have a distinctive chip architecture while still needing an experienced partner to implement the surrounding system.
The commercial attraction is the possibility of repeated work across product generations. A successful program gives the supplier knowledge of the customer's technical requirements and operating priorities. That can support subsequent design opportunities, although each generation still has to be won.
There is also a plausible opportunity to supply additional connectivity and memory products around a custom processor. I view that as a source of potential expansion within customer accounts. A custom-chip win does not automatically award Marvell every adjacent component, so those opportunities need to be evaluated individually.
Google provides unusually specific evidence of the opportunity
Marvell's August 19 filing describes a Google agreement covering custom products attached to the TPU ecosystem. The named programs include AI inference accelerators and controllers for storage, networking and memory, together with near-memory compute. The breadth of that list supports the argument that custom silicon extends beyond a single accelerator.
The associated warrant allows Google to purchase up to 58.97 million Marvell shares at $206.58 each. Most of the warrant vests through 240 purchase-linked tranches, each tied to $500 million of custom-product revenue through fiscal 2033. Multiplying those thresholds produces $120 billion, but purchases are discretionary. That figure is a vesting framework, not contracted backlog.
I find the range of disclosed programs more useful than treating the maximum threshold as a forecast. It establishes a concrete relationship across several parts of Google's infrastructure. The investment question is how many of those programs reach production, how quickly they scale, and how much profit remains for Marvell shareholders after the associated costs and dilution.
NVIDIA broadens the route to commercial adoption
NVIDIA's March 31 announcement paired a $2 billion investment in Marvell with a partnership built around NVLink Fusion. Marvell is to provide custom XPUs and compatible scale-up networking, while NVIDIA supplies supporting components and its infrastructure ecosystem. The companies also announced collaboration on silicon photonics.
That relationship is strategically useful because compatibility can influence a customer's willingness to adopt a custom design. Integrating with an established infrastructure platform can reduce the amount of surrounding technology the customer must assemble and validate independently.
For Marvell, it creates another route for custom silicon to reach large deployments. It also allows the company to participate in infrastructure that combines products from multiple suppliers. Revenue will depend on the actual systems customers build; the investment itself should not be confused with a purchase order for Marvell chips.
Celestial AI expands where optical connectivity can earn revenue
Marvell completed the Celestial AI acquisition in February 2026. Its Photonic Fabric technology addresses optical connectivity within closely coupled AI systems, including links between processors and across racks. This gives Marvell exposure to a category of connections that can expand as customers build larger scale-up systems.
At the acquisition announcement, Marvell described a first-generation Photonic Fabric chiplet with 16 terabits per second of bandwidth. It also highlighted the ability to integrate optical connectivity with high-power compute packages. The technical objective is to support more bandwidth and reach without making the power required for communication prohibitive.
At closing, management outlined initial revenue in the second half of fiscal 2028, a $500 million annualized rate in that year's fourth quarter, and a $1 billion annualized rate in the fourth quarter of fiscal 2029. These were forward-looking targets; an annualized quarterly rate is different from revenue earned over a full year.
The investment appeal is the additional market that successful deployment would create for Marvell. The next evidence should be customer qualification, production schedules and commercial shipments. Those milestones will determine whether the engineering promise becomes a durable business.
Memory access could become a larger part of the growth story
Longer AI conversations and more demanding inference workloads require more stored context. A key-value cache retains information from earlier tokens so a model can use it when generating subsequent output. As those caches grow, the cost and accessibility of memory become consequential design issues.
Marvell's August memory announcement brought together Bravera SSD controllers, Structera CXL memory expansion and Photonic Fabric shared memory. The optical architecture is designed to share memory across multiple racks, with connections extending up to 50 meters and capacity for up to 32TB of warm KV-cache offload. Bravera SC6 sampling was scheduled to begin in the fourth quarter of calendar 2026.
The economic logic is attractive. If suitable data can be held in a shared memory tier and accessed efficiently, a customer may use expensive accelerator resources more effectively. HBM, expanded memory and SSD storage each have different latency, bandwidth and cost characteristics. A workable system assigns data to the tier appropriate for the workload.
I see this as a potentially valuable extension of Marvell's business. Products that improve utilization can justify spending even when a customer is scrutinizing the cost of adding more compute. The size of that benefit must be demonstrated in real workloads, with software support and customer deployment decisions determining adoption.
Switching connects the product portfolio to system performance
Marvell introduced its 102.4-terabit-per-second Teralynx T100 switch in June, with sampling planned that quarter. The company describes a design supporting up to 512 ports and multiple package options, including co-packaged optics. Higher connectivity can allow customers to simplify some network layouts and reduce the number of switching stages.
This creates an interesting relationship within the portfolio. Better switching can improve the economics of a larger cluster, while a more efficient network may need fewer links for a given configuration. An investor should evaluate the overall semiconductor content and profitability of the deployment rather than assume every product category rises in direct proportion to accelerator counts.
Marvell's September AI Infra Summit presentation also included PCIe switching, connectivity telemetry and CXL memory products. That breadth gives the company a practical reason to participate in discussions about the entire system. In my view, the opportunity is to win more of the components whose performance customers must optimize together.
The operating results already show meaningful demand
For the quarter ended August 1, Marvell reported $2.739 billion of revenue, up 37% year over year, with data-center revenue growing 46%. Operating cash flow was $605.5 million. Management guided the following quarter to $3.150 billion of revenue, plus or minus 5%, and expected the custom business to accelerate in the second half of fiscal 2027.
Those results give the longer-term product discussion a substantial operating base. Marvell is funding its development effort while an existing business is already producing revenue and cash. The next question is how effectively that business converts additional volume into earnings and free cash flow.
Product mix deserves close attention. Marvell's filings explain that some turnkey ASIC arrangements carry lower gross margins. A larger custom business can therefore increase gross-profit dollars without raising the consolidated gross-margin percentage. I would assess the outcome through operating profit, cash generation and the diluted share count.
The October 6 Investor Day is a useful upcoming checkpoint. I will be looking for clearer production schedules, updated optical expectations and evidence that the broader portfolio is generating additional customer programs.
The conditions that matter for shareholders
The principal execution risk is timing. Large custom-chip programs can shift with customer development schedules, and optical products must meet demanding qualification requirements. A delay can affect revenue well before it changes the long-term technical opportunity. Manufacturing access, packaging availability and competitive pricing also influence the returns Marvell can earn.
Supply commitments deserve particular attention after the GF announcement. Marvell reported approximately $8.52 billion of unconditional commitments to foundry, test and assembly partners as of August 1. That company-wide figure predates the September agreement and cannot be assigned to GF. It shows why securing supply must be matched with disciplined demand forecasting.
Valuation remains part of the decision. Strong company growth can coincide with disappointing share-price performance when investors have already priced in faster growth or better margins. I would require a credible path to higher cash flow per diluted share, supported by customer ramps that can be tracked.
The most convincing progress would be broader production across custom programs, reliable delivery of higher-speed optical products and commercial adoption of shared-memory technologies. Persistent delays, weak cash conversion or heavy concessions to win business would make the thesis less attractive.
Why I am bullish on the long term opportunity
Marvell is building a business around constraints that become harder to manage as AI infrastructure expands. Custom compute needs capable interfaces. Larger systems need efficient connections. More demanding inference workloads need better access to memory. The company has products and development programs addressing each requirement.
The GF capacity expansion strengthens the manufacturing preparation behind that effort. Google provides a concrete example of the range of custom programs Marvell can pursue. NVIDIA adds another commercial route, while Celestial AI and the memory portfolio extend the applications in which Marvell can participate.
My positive view rests on the combination of an established revenue base and credible opportunities to sell more technology into future systems. If Marvell executes, its role in AI infrastructure could become substantially larger over the next several years. For shareholders, the decisive outcome will be how much of that expansion becomes durable cash flow per share.
