Marvell Technology recently announced a definitive agreement to acquire Celestial AI for $3.25 billion in upfront consideration ($1 billion cash plus $2.25 billion in stock), with potential earnout payments of up to an additional $2.25 billion based on revenue milestones through fiscal 2029.

Marvell projects Celestial AI will reach a $500 million annualized revenue run rate in Q4 fiscal 2028, doubling to $1 billion by Q4 fiscal 2029.

The acquisition puts Marvell in a strong position to address what the company characterizes as a new semiconductor TAM for optical interconnects.

Who is Celestial AI?

Celestial AI, founded in 2020 by CEO David Lazovsky and COO Preet Virk, developed the Photonic Fabric technology platform for optical interconnects in AI computing systems.

The company focuses specifically on optical scale-up interconnects that operate within and across racks in AI clusters, differentiating its approach from traditional co-packaged optics implementations.

Celestial’s first-generation Photonic Fabric chiplet integrates electrical and optical components into a compact form factor that Celestial AI claims delivers 16 terabits per second of bandwidth per chiplet (that’s about 10 times the capacity of current 1.6T ports used in scale-out applications).

The technology is designed for 3D co-packaging with XPUs and scale-up switches, allowing optical connections to be made directly into the XPU rather than from the die edge.

Celestial AI has said it’s engaged with multiple hyperscalers and has secured at least one major design win with a large hyperscaler planning to use Photonic Fabric chiplets in next-generation scale-up architecture.

Strategic Rationale and Fit Within Marvell

The acquisition addresses Marvell’s need to expand beyond its current connectivity portfolio focused on scale-out and scale-across networking. While Marvell has established positions in AI-related electro-optics (which generated $819 million in Q3 fiscal 2026, up 2.6x year-over-year) and custom XPU design services ($418 million in Q3, up 83.4%), the company lacks technology for optical scale-up interconnects within multi-rack AI clusters. Bringing Celestial in-house solves that.

Driving the adoption of silicon photonics is the technical limitations of copper interconnects at increasing bandwidths and distances. As XPU-to-XPU connections scale from rack-level to multi-rack configurations, copper’s signal integrity degrades significantly (bandwidth increases require proportional reductions in cable length due to noise characteristics).

Marvell is betting that optical interconnects become economically necessary despite higher upfront costs when cluster architectures require both high bandwidth and extended reach between processing elements. It’s a bet being similarly placed across the industry.

Celestial AI’s technology specifically targets scale-up fabrics where each XPU needs direct memory access to every other XPU in the cluster, requiring any-to-any connectivity with minimal latency.

Traditional scale-out optical networking addresses rack-to-rack and datacenter-to-datacenter connections, but scale-up applications require optical technology that can integrate more tightly with XPU packages.

Celestial’s 3D co-packaging approach employs allows the optical connection to bypass traditional die-edge constraints, potentially freeing perimeter space for additional HBM integration.

However, the acquisition carries execution risks. The technology remains in early commercial deployment, with meaningful revenue not expected until the second half of fiscal 2028, about three years post-acquisition.

Major platform vendors, including Nvidia and AMD, are evaluating multiple photonic solutions, while customer design cycles remain long. Marvell’s ability to capture projected market share depends on Celestial AI’s technology demonstrating clear advantages in power efficiency, thermal performance, and total cost of ownership compared to both copper alternatives and competing optical approaches.

Analysis

Marvell’s acquisition of Celestial AI is a strategic bet on the inevitable transition from copper to optical interconnects within AI compute clusters, driven by fundamental physics constraints on electrical signaling at increasing bandwidths and distances.

The deal positions Marvell to address an emerging market for optical scale-up interconnects, complementing its existing scale-out and scale-across connectivity portfolio with technology specifically designed for multi-rack AI training clusters requiring any-to-any XPU connectivity with minimal latency.

The acquisition’s success depends on multiple factors beyond Marvell’s direct control, including the pace of AI infrastructure growth, customers’ willingness to adopt complex 3D co-packaged photonic solutions, competitive offerings from established platform vendors, and the broader economics of optical versus copper interconnects.

For technology decision-makers evaluating AI infrastructure roadmaps, the acquisition aligns with broader industry momentum toward optical scale-up interconnects, though deployment timelines remain uncertain.

Implementors planning large-scale AI clusters should factor optical interconnect availability into long-term architecture planning while recognizing that copper-based solutions will remain dominant for near-term deployments.

The deal ultimately strengthens Marvell’s position in AI infrastructure markets at a critical inflection point where cluster architectures are evolving from single-rack to multi-rack configurations. Whether Celestial AI’s specific photonic approach captures significant market share remains to be determined, but Marvell has secured a credible technology position and customer relationships that provide meaningful optionality as the optical interconnect market develops over the next three to five years.

It’s a strong acquisition that plays directly to Marvell’s strategy and momentum.

Competitive Outlook & Advice to Buyers

Multiple well-funded competitors are pursuing alternative optical interconnect approaches, each with different technical tradeoffs:

• Lightmatter has developed photonic compute and interconnect technology with substantial venture backing and customer engagement.
• Ayar Labs focuses on in-package optical I/O technology that also targets short-reach, high-bandwidth applications.
• Avicena has developed micro-LED-based optical interconnect technology claiming advantages in power efficiency and integration density.
• Astera Labs, already a public company, has established positions in connectivity solutions and may expand into adjacent optical technologies.

Platform vendors pose perhaps the most significant competitive threat. TSMC’s COUPE program will provide integrated photonics capabilities to major customers including Nvidia and AMD, who command dominant positions in AI accelerator markets.

These vendors may prefer in-house or TSMC-integrated optical solutions rather than adopting third-party chiplets, particularly for differentiated architectures.

Intel has also demonstrated silicon photonics capabilities since 2013 and continues development of integrated optical solutions.

Any of these larger players could leverage superior manufacturing scale and vertical integration to achieve cost advantages Marvell cannot match.

The competitive landscape also reflects uncertainty about standardization and interoperability. UALink, CXL, and other emerging interconnect standards may evolve in ways that favor certain optical approaches over others.

All that said, let’s look at how a Marvell with Celestial AI’s IP might stack up..

These sections are only available to NAND Research clients and IT Advisory Members. Please reach out to [email protected] to learn more.