Marvell Expands CXL Strategy, Structera Link, and Scale-Up Optics Growth in Q2 2027

Architectural Update: CXL, Structera, and Scale-Up Optics

CXL evolution and Structera

Marvell’s CXL strategy appears to be evolving from a conventional server-attached memory-expansion architecture into a broader infrastructure layer for memory expansion, inference, and XPU attach. Management said the original server-centric investment is now well suited to both memory expansion and inferencing, with deployments at multiple hyperscalers using different architectures and at “extremely high volumes.” ¹

The most important architectural driver is that CXL is solving two problems simultaneously:

• Inference economics and performance: Increasing inference demand is creating a need for additional, flexible memory capacity and bandwidth. ¹
• Memory scarcity: Customers are modifying their plans to use more CXL-type technology because of constrained memory availability. ¹

This suggests CXL is becoming less of a single standardized server feature and more of a fungible architectural tool that hyperscalers can deploy differently depending on memory availability, workload characteristics, power constraints, and accelerator design. Management separately emphasized that customers are adapting architectures rapidly because of supply, power, and architectural constraints, and that inference optimization is becoming increasingly important as AI infrastructure shifts from training toward monetized inference workloads. ²

Structera

The excerpt does not provide a separate technical description, product specification, revenue target, or design-win count specifically for Structera. The relevant update is therefore best interpreted at the CXL-platform level: Marvell’s CXL investment is gaining traction in multiple hyperscaler architectures, with additional design wins secured during the last two quarters. ³¹

Management characterized the overall memory-expansion opportunity as “massive” and said that a comprehensive memory-expansion discussion, presumably including the company’s CXL offerings, would be provided at Investor Day. ¹ The company also identifies CXL as part of its broader XPU-attach strategy, alongside custom NICs and other specialized components that sit around the accelerator rather than being the accelerator itself. ⁴

Broader CXL and AI-memory ecosystem

The ecosystem is broadening beyond conventional CPU memory attachment. Marvell described custom attach products spanning AI inference accelerators, storage controllers, network-interface controllers, memory-interface controllers, and near-memory compute, with these products increasingly connected to the TPU ecosystem of a major hyperscaler. ⁵

That broadening matters because CXL and related memory-interface products can be deployed as part of a larger optimization stack rather than as isolated components. Marvell’s stated value proposition is the ability to customize or adapt products quickly as hyperscalers rebalance compute, networking, and memory to maximize performance and cost per token. ²⁶

The strategic implication is that CXL may benefit from several overlapping trends:

1. More inference-oriented infrastructure, which tends to increase the importance of memory capacity and data movement. ¹²
2. Hyperscaler-specific architectures, rather than one uniform industry design. ¹⁷
3. Memory scarcity and supply flexibility, which can encourage customers to redesign systems around memory expansion. ¹
4. Higher attach content around XPUs, including memory controllers, near-memory compute, and other specialized silicon. ⁵

Marvell’s custom business is also expected to benefit from this attach expansion: management said the company is seeing strong design activity across current and next-generation hyperscaler programs, with XPU attach benefiting from increasing demand for CXL and custom NICs. ⁴

Scale-up optics: architectural transition

Scale-out networks continue to rely primarily on pluggable optical modules, but Marvell expects scale-up networks to require optics much closer to the XPU and switch because of their substantially higher bandwidth density. ⁸

The transition is not expected to be an immediate copper-to-optics replacement. Management said copper and optics are likely to coexist for several years, although customers are already aggressively planning scale-up optical deployments beginning as early as the next fiscal year. ⁸

The architecture is therefore becoming a portfolio rather than a single technology decision:

• NPO, or near-package optics, for bringing optics close to the switching or accelerator package. ⁸
• CPO, or co-packaged optics, for tighter integration of optics and switching. ¹⁸
• Multiple modulator technologies, including MZM, EAM, and MRM, each with different trade-offs in cost, power, and maturity. ⁸
• Advanced silicon photonics supported by Marvell’s broadband analog portfolio of TIAs and drivers. ⁷⁸
• Marvell’s internally developed NPO “light engine,” which has been under development for several years and has been demonstrated at OFC. ⁷

The company’s core architectural message is explicitly “and, not or.” Customers are expected to deploy several optical and electrical approaches in parallel rather than move through a simple, sequential replacement cycle from one technology to the next. ⁷⁹

Progress versus the Celestial AI targets

The prior Celestial AI framing was relatively specific: Marvell had described the Celestial AI Photonic Fabric CPO solution as approximately $150 million of revenue for the following year, while total scale-up optics—including NPO—was estimated at approximately $300 million. ¹

The current update is substantially more positive, but less precise at the individual Celestial-program level:

The key conclusion is that the original $150 million Celestial target has not been presented as the ceiling for the business. Celestial’s Photonic Fabric remains strategically important and should contribute materially to the next-year CPO ramp, but management now views the total opportunity as much broader because NPO, CPO, broadband analog components, internally developed optical engines, and scale-up switching are developing concurrently. ⁷⁹

Management specifically said that CPO would be important in the next year and the year after, but that NPO activity is accelerating in parallel through multiple design wins and customer partners. ⁷ Consequently, the upside relative to the original Celestial framing is primarily a portfolio expansion rather than simply a larger forecast for one CPO product.

Why the opportunity is expanding

Several factors explain the increased confidence:

• Marvell is participating in NPO through both its silicon-germanium TIAs and drivers, where management cited significant market share and content, and through its own internally developed light-engine solution. ⁷
• The company is coupling optical products with its scale-up switches, including UAL and ESUN, while also integrating with customer XPU and GPU architectures. ⁹
• Marvell has experience in packaging, modulator technologies, silicon photonics, and optical development extending over roughly a decade, according to management. ⁹
• Customers are evaluating multiple solutions simultaneously because the trade-offs among cost, power, and technology maturity remain meaningful. ⁸
• Marvell is also positioned across three purpose-built scale-up switching protocols: internally developed UAL and ESUN, plus its partnership with NVIDIA around NVLink Fusion. ⁸

The company’s strategy is therefore to sell more than an optical component. Management emphasized the importance of supplying the full end-to-end link, from one end of the interconnect to the other, rather than offering a narrow point solution. ¹⁰ Marvell believes this broader portfolio could materially increase the economic content associated with switching; management said optical attachment could almost double the switching-related serviceable available market over time. ¹⁰

Financial significance

Scale-up optics is now one of the principal contributors to Marvell’s raised outlook. The company increased its fiscal 2028 revenue outlook by approximately $1.5 billion, to approximately $18 billion, and management identified scale-up optics, switching, and other connectivity businesses as major contributors to that increase. ¹¹¹²

Management also said connectivity was probably the largest net contributor to the $1.5 billion increase, with the bucket including scale-out, scale-up, and switching. ¹¹ This is important because the upside is not dependent solely on Celestial CPO revenue; it is distributed across optical DSPs, broadband analog, NPO, CPO, switching, and scale-across connectivity. ¹¹¹³

The company’s broader connectivity portfolio is already showing significant momentum: 800G optical demand remains strong, 1.6T is ramping rapidly, scale-out switching is expected to more than double in the current year, and broadband analog demand for TIAs and drivers continues to exceed expectations. ¹³ These existing franchises provide a commercial and technical base from which scale-up optics can ramp.

Overall assessment

CXL: The architecture is broadening from traditional server memory expansion into a flexible memory-and-inference infrastructure layer. Structera should be viewed within that larger CXL and XPU-attach strategy, but the excerpt does not provide enough product-specific information to assess Structera independently. The strongest evidence is the combination of hyperscaler deployments, multiple architectural implementations, additional recent design wins, and demand supported by both inference growth and memory scarcity. ¹

Scale-up optics: Progress is ahead of the original Celestial-centered framing. The earlier benchmark was approximately $150 million for Celestial Photonic Fabric and approximately $300 million for total scale-up optics in the following year; management now describes the total category as materially larger, with a particularly significant contribution to the $1.5 billion increase in the next-year outlook and an even larger opportunity in the subsequent year. ¹⁷⁹

The central architectural takeaway is that Celestial is becoming an important component of a much broader Marvell platform rather than the entirety of the scale-up optics thesis. NPO and CPO are advancing together, copper will coexist with optics during the transition, and Marvell is attempting to capture value across the complete chain: analog optical components, photonics, packaging, optical engines, switching silicon, and XPU/GPU integration. ⁷⁹⁸¹⁴