Marvell’s Photonic Fabric Displaces Copper in AI Scale-Up Era—Optical Infra Alpha Emerges

The exponential growth of AI is hitting a physical wall. As enterprises deploy larger models requiring thousands of interconnected processors, the traditional copper wiring that moves data between them has become a critical bottleneck. Electrical signals degrade over distance, consume excessive power, and simply cannot provide the bandwidth modern workloads demand. This creates a scaling paradox: the more compute you add, the more power you waste just moving data around.

The economic and thermal math is unsustainable. Copper interconnects use about twice the power of optical alternatives and offer much less reach. As AI accelerators approach multi-kilowatt power levels, adding copper connections becomes a drain on both budgets and cooling systems. This forces a costly trade-off between performance and efficiency.

The solution lies in a paradigm shift to optical interconnects and a new architectural approach. Next-generation systems are moving beyond single-rack limits to distribute hundreds of AI accelerators across multiple racks. These form what are known as "scale-up fabrics." The key advantage is direct memory access between processors, which dramatically improves resource utilization and removes the need to duplicate high-bandwidth memory across systems. In practice, this means a cluster of AI chips can act as a single, unified supercomputer, unlocking far greater efficiency.

Marvell's strategic acquisition of Celestial AI for $3.25 billion is a foundational step in building this new infrastructure layer. Celestial's Photonic Fabric technology integrates optical components directly into processor packages, effectively removing copper from the critical scale-up connections. This delivers a massive leap in performance, with 16 terabits per second of bandwidth per chiplet-ten times the capacity of current leading optical ports. By addressing the power and bandwidth constraints at the physical layer, Marvell is positioning itself to enable the next phase of AI scaling.

Exponential Growth Drivers: From 800G to 1.6T and the Scale-Up Paradigm

The transition from 800G to 1.6T connectivity is not just an incremental upgrade; it is the next major inflection point on the AI infrastructure S-curve. Marvell is not merely participating in this shift-it is driving it with a multi-pronged portfolio designed to capture the entire exponential growth curve. The company's 3nm 1.6T Ara platform is now in mass production, enabling hyperscalers to deploy the first generation of 1.6T pluggable connectivity. This isn't a niche product; it's the foundational hardware for the next wave of AI data center builds, directly addressing the bandwidth explosion fueled by larger models and distributed workloads.

Marvell is expanding this leadership into the critical inter-rack connectivity layer with new coherent optical solutions. The COLORZ 1600, the industry's first 1.6T ZR/ZR+ pluggable powered by a 2nm DSP, is engineered for secure, high-bandwidth links between data center pods. This is essential for the scale-up paradigm, where AI accelerators are no longer confined to a single rack. These new ZR/ZR+ platforms integrate MACsec for encrypted optical transport, a non-negotiable feature as AI workloads become increasingly distributed and sensitive. This moves the security layer down into the physical interconnect, simplifying the overall architecture.

The true power of Marvell's strategy lies in the integration of these pieces. The company is building the complete stack-from the industry-leading DSPs and SerDes to the optical engines and the CXL and UALink interconnects acquired through strategic buys. This integrated portfolio is what enables the scale-up fabric architecture. By combining optical interconnects for long-haul, high-bandwidth links with CXL for memory disaggregation and UALink for accelerator communication, Marvell provides a single-vendor solution that simplifies data center design. This integration directly tackles the adoption friction that often slows new tech rollouts.

The result is a system that improves resource utilization at scale. Instead of each server needing its own expensive, high-bandwidth memory, the scale-up fabric pools resources across racks. This architectural shift, powered by Marvell's integrated silicon, is the key to unlocking exponential growth in AI compute efficiency. The company is positioned not just to sell components, but to provide the fundamental infrastructure layer that makes the next phase of AI scaling economically viable.

Financial Impact and Market Positioning

The technological positioning is now translating into explosive financial momentum. Marvell's data center business is the clear engine of growth, with revenue of $1.518 billion last quarter, up 38% year-over-year and representing 73% of total sales. This drove a record $2.075 billion in total revenue for the period. The growth is concentrated in custom AI accelerators, a segment where inference demand is beginning to scale rapidly and where Marvell's silicon is seen as a strong fit.

This isn't just a quarterly beat; it's a fundamental re-rating of the company's potential. Management is guiding for sales to exceed $3 billion in the near term, with next year's sales approaching $11 billion in 2028. The forward outlook calls for sales of $15 billion versus the Street's $13 billion estimate, a massive gap that signals accelerating demand. CEO Matt Murphy noted that the data center revenue growth forecast for next year is now higher than prior expectations, a clear signal of confidence in the adoption curve.

Marvell's pure-play exposure to this growth is defined by its addressable market. The company is positioned at the heart of an optical networking market that is set to quadruple. This isn't a niche opportunity; it's the foundational infrastructure layer for the AI scale-up paradigm. By integrating optical interconnects, CXL, and UALink, Marvell is building the complete stack for this exponential shift. The financial impact is twofold: it captures the revenue from selling the components and secures its role as the essential vendor for the next generation of AI data centers.

The competitive context is clear. Marvell is in a race with other custom AI chip leaders like Broadcom, which reported its own blowout quarter with AI revenue growing 106%. Yet Marvell's strategic acquisitions, like Celestial AI, are explicitly aimed at owning the next bottleneck-optical interconnects. This moves the company beyond selling discrete chips into providing the critical infrastructure that enables the entire scale-up fabric. In the long run, that infrastructure layer is where the most durable value is created.

Catalysts, Risks, and What to Watch

The thesis for Marvell's optical infrastructure is now set on a clear path, but its validation hinges on a few key forward-looking events. The next phase is about translating technological leadership into commercial adoption and integrated execution.

The primary catalyst is the commercial rollout of Celestial AI's Photonic Fabric. This technology is the linchpin for the scale-up fabric architecture, moving optical interconnects from the rack edge to the core of the processor package. Its success will be measured by the speed at which hyperscalers adopt it to build multi-rack AI clusters. A rapid ramp would confirm the market's acceptance of the new paradigm and accelerate Marvell's capture of the projected semiconductor TAM for optical interconnects.

Another near-term catalyst is the adoption of Marvell's 1.6T ZR/ZR+ solutions, like the COLORZ 1600, by major cloud providers. These platforms are designed for the secure, high-bandwidth links that connect data center pods in a scale-up fabric. Widespread deployment would validate the company's leadership in the next-generation coherent optical market and drive revenue from its 2nm DSPs and SerDes portfolio.

Yet a major risk looms over this growth story: execution on integrating Celestial AI's technology. The $3.25 billion acquisition is a transformative bet, but integrating a complex photonic platform into Marvell's existing silicon and software stack is a multi-year challenge. The projected revenue from this new market is still years away, and any delay in productization or performance issues could slow the adoption curve. The company must seamlessly blend Celestial's innovation with its own execution discipline to avoid becoming a victim of its own ambition.

What investors should watch is Marvell's guidance on two fronts as adoption accelerates. First, the company's outlook for custom AI accelerator revenue will signal whether inference demand is scaling as expected. Second, its share of the optical interconnect market will reveal how effectively it is capturing the shift from copper to all-optical fabrics. These metrics will provide the clearest signal of whether Marvell is successfully building the fundamental rails for the next AI paradigm or merely participating in a crowded field.