Marvell Expands CXL Role into AI Infrastructure with Structera, Optics Accelerate in 2027 Q2

Architectural Update: CXL, Structera and Scale-Up Optics

CXL and the evolving memory architecture

Marvell’s CXL strategy appears to be expanding beyond its original server-centric memory architecture into a broader memory-expansion and inference-enablement layer for AI infrastructure. Management said the company’s multiyear CXL investment was initially aimed at traditional compute-centric server architectures but is now “ideal for memory expansion and inferencing.” ¹

The principal architectural drivers are:

• Inference economics: AI inference workloads require more flexible and efficiently deployed memory resources, increasing the relevance of CXL-based memory expansion. ¹
• Memory scarcity: Customers are modifying their infrastructure plans to incorporate more CXL-type technology in response to constrained memory availability. ¹
• Heterogeneous hyperscaler designs: The technology is being deployed at multiple hyperscalers, with different customers pursuing varying architectures and doing so in what management described as “extremely high volumes.” ¹
• Expanding design-win base: Marvell has secured additional CXL design wins over the preceding several quarters, although it did not quantify the programs in this transcript. ¹

Structera’s role

The excerpt does not provide specific technical or financial detail on Structera itself, so it would be premature to characterize its exact topology, protocol implementation, bandwidth, or revenue contribution from this disclosure alone. However, management’s comments place the Structera-related opportunity within a larger CXL architecture that is broadening from a conventional CPU-memory attachment model toward distributed memory expansion, inference optimization and XPU attach. ²¹

That broader ecosystem is also visible in Marvell’s custom-silicon portfolio. The company identified memory-interface controllers, near-memory compute, inference accelerators, storage controllers and network-interface controllers as examples of products attached to hyperscaler XPU ecosystems. ³ Management also stated that CXL and memory expansion are examples of customers rapidly adapting architectures to improve performance and cost relative to tokens generated in inference workloads. ⁴

The strategic implication is that CXL is becoming less of a standalone controller product and more of an architectural building block for fungible compute, memory and networking resources. Management explicitly described customers as optimizing across supply, power and architectural constraints and emphasized Marvell’s ability to customize or adapt products quickly as those requirements shift. ⁴

Broader scale-up and scale-out ecosystem

The networking architecture is separating into several related but simultaneous transitions rather than following a single replacement cycle:

1. Scale-out remains primarily pluggable and optical. Marvell expects pluggable modules to remain the primary form factor for scale-out networks. ⁵
2. Scale-up is moving optics closer to the XPU and switch. The much higher bandwidth density required in scale-up networks is driving optics closer to the compute and switching endpoints. ⁵
3. Copper and optics will coexist for several years. Customers are initially deploying copper, but increasing cluster size, reach requirements and bandwidth limitations are pushing the architecture toward optical interconnects. ⁵⁶
4. NPO and CPO will coexist rather than follow a simple handoff. Customers are evaluating both near-package optics and co-packaged optics, along with multiple modulator technologies and different cost, power and maturity trade-offs. ⁵
5. Scale-up switching is also becoming protocol-diverse. Marvell is addressing UALink and ESUN through internally developed switches and NVLink through its partnership with NVIDIA around NVLink Fusion. ⁵

Management’s central architectural message is therefore “and, not or”: customers are expected to deploy multiple optical and switching solutions simultaneously because their architectures differ by workload, system design and deployment timing. ⁷

Scale-up optics: progress versus the Celestial AI targets

The clearest quantified comparison is between the prior Celestial AI target and the company’s current scale-up optics outlook:

The important conclusion is that Celestial has progressed from being the principal quantified reference point to being one component of a much broader Marvell scale-up optics franchise. Marvell said the Celestial Photonic Fabric CPO solution was approximately $150 million for the following year, while the overall scale-up optics category—including NPO—had previously been framed at approximately $300 million. ¹

Since then, management has said that scale-up optics is accelerating beyond the prior-quarter estimate and is a significant driver of the $1.5 billion increase in the next-year outlook. ⁷ It also characterized scale-up optics, switching and related connectivity as among the most meaningful contributors to that increase, while noting that the overall increase is broad-based rather than attributable to one program. ⁹¹⁰

The company did not provide a new precise dollar figure for either Celestial or the total scale-up optics category in these excerpts. Accordingly, the defensible interpretation is directional: the opportunity has been materially upgraded, but the disclosure does not support replacing the earlier $150 million Celestial or $300 million category figures with a specific new number.

Why NPO is changing the opportunity size

The major change versus the original Celestial framing is the increasing importance of NPO. Marvell is participating in NPO through:

• Broadband analog silicon-germanium TIAs and drivers, where management cited significant market share and content. ⁷
• Marvell’s internally developed NPO “light engine,” which has been in development for several years and has been demonstrated at OFC. ⁷
• A broad set of customer and partner design wins across different optical architectures. ⁷

Management explicitly said that NPO activity is advancing in parallel with CPO rather than waiting for a wholesale transition to co-packaged optics. ⁷ The company also reported strong NPO adoption at multiple customers and increased its fiscal 2028 scale-up optics revenue outlook meaningfully versus prior expectations, positioning Marvell as one of the largest NPO enablers in AI infrastructure. ⁵

This creates a larger and more diversified revenue opportunity than a Celestial-only analysis would imply. The addressable content can include optical engines, silicon photonics, modulators, TIAs, drivers and associated switching, rather than only the Photonic Fabric solution. ⁷⁵

Marvell’s architectural positioning

Marvell is attempting to provide an end-to-end scale-up platform rather than a point component. The company is combining:

• Scale-up switches based on UAL and ESUN, together with NVLink Fusion participation. ⁵
• NPO and CPO optical solutions. ⁷⁵
• Broadband analog TIAs and drivers. ⁷
• Silicon photonics and multiple modulator technologies, including MZM, EAM and MRM. ⁵
• Integration with customer XPU and GPU architectures. ⁸
• Manufacturing, packaging and R&D capabilities spanning multiple optical technologies. ⁸

Management argues that customers need a partner capable of “bookending” the optical link and managing the full path between the XPU and switch, rather than relying on isolated point solutions. ¹¹ Marvell also believes that optical attach could nearly double the switching-related serviceable available market as scale-up networks transition from primarily electrical or copper links to optical links. ¹¹

The company’s existing switching execution is intended to reinforce that optical strategy. Marvell cited traction in its Teralynx Ethernet architecture, expansion of its customer base and volume delivery of complex switching products, while stating that the next requirement is a credible roadmap from electrical solutions through NPO and ultimately full CPO integration. ¹²

Financial and timing implications

The timing is early but commercially significant. Customers are planning initial scale-up optics deployments as early as the next year, while copper and optical technologies are expected to coexist for several years during the transition. ⁵ Marvell’s fiscal 2028 outlook was raised to approximately $18 billion, up $1.5 billion from the prior $16.5 billion outlook, with expected growth of approximately 50% versus the previous 45% expectation. ¹³

Connectivity is identified as probably the largest net contributor to the $1.5 billion increase, with scale-out, scale-up and switching all contributing. ¹⁰ Marvell also stated that scale-up opportunity remains largely ahead of the company and that its relevant businesses are on or ahead of the trajectory toward a $1 billion annualized revenue run rate, although the excerpt does not specify which individual business or businesses that statement covers. ¹⁴

Bottom line

• CXL: The architecture is expanding from traditional server memory expansion into a broader AI infrastructure role involving inference, memory scarcity, XPU attach and customized hyperscaler designs. Structera should be viewed within this expanding CXL platform, but the excerpt does not disclose enough detail to assess Structera’s standalone specifications or revenue. ²¹⁴
• Ecosystem: The market is not converging on one architecture. Scale-out pluggables, scale-up copper, NPO, CPO, multiple modulator technologies and several scale-up switching protocols are expected to coexist. ⁷⁵
• Celestial versus current optics outlook: The earlier Celestial Photonic Fabric reference was approximately $150 million for the following year, within an approximately $300 million total scale-up optics category including NPO. ¹ The current message is that the total opportunity is materially larger and accelerating, with NPO emerging as a parallel and potentially very substantial contributor rather than merely a bridge to CPO. ⁷⁸⁵
• Investment-quality interpretation: The strongest evidence of progress is not a newly disclosed standalone Celestial revenue number, which was absent, but the expansion of the opportunity from one CPO program into a broader NPO/CPO, analog, photonics and switching platform that management says is materially increasing the next-year and following-year outlook. ⁷⁸⁹