Photonics by Celestial AI
Marvell Technology's acquisition of Celestial AI brings its Photonic Fabric technology to the forefront. This technology promises to revolutionize the scaling of AI compute and memory through optical interconnects, pushing the boundaries of electrical data transmission.
Source: Own Representation
Acquisition of Celestial AI
Marvell Technology, Inc. announced on December 2, 2025, the acquisition of Celestial AI for approximately $3.25 billion. The transaction includes $1 billion in cash and approximately 27.2 million Marvell shares valued at approximately $2.25 billion. Additionally, an earn-out component of up to another 27.2 million shares is contingent upon Celestial AI achieving defined revenue milestones by the end of Marvell's fiscal year 2029.
The closing of the acquisition is expected in the first quarter of 2026, subject to regulatory approvals and customary conditions. Marvell anticipates Celestial AI to deliver "meaningful revenue contributions" starting in the second half of fiscal year 2028. Konkret nennt Marvell einen annualisierten Run-Rate-Orientierungswert von $500 million by Q4 FY2028 and $1 billion by Q4 FY2029.
Concurrently, the deal is closely intertwined with hyperscaler relationships. Reuters reports, that in connection with the acquisition, Marvell granted Amazon a stock warrant tied to purchases of Photonic Fabric products through the end of 2030. In a In einer SEC filing regarding Marvell's AWS collaboration warrant mechanics and exercise price in this context are detailed, including an exercise price of $87.7706 and vesting-based tranches until 2030.
Photonic Fabric Technology
Celestial AI positions its "Photonic Fabric" as an optical interconnect platform designed to support the scaling of AI compute and memory from the package level to large clusters. The core approach is to transport data using light rather than purely electrical signals between compute and memory, pushing the limits of traditional copper pathways. Celestial AI highlights that its technology represents an optical interconnect platform .
Marvell highlights several specific differentiators in its announcement, including very low nanosecond-level latency, high thermal stability , and the ability to co-package photonics vertically in 3D packages with powerful XPUs and switches. This design is also intended to free up valuable "die edge" space, allowing for more room for HBM within the XPU package.
Marvell's accompanying information deck categorizes the Celestial portfolio into Connectivity Products (PFLink), System Products, PF Switches, and PF Appliances. It also mentions UALink compatibility at the package level , which is intended to create a clear bridge to open scale-up standards.
A notable technical anchor from Marvell's release is the claim that a Photonic Fabric scale-up chiplet will deliver 16 terabits per second of bandwidth within a single chiplet. Marvell presents this as a significant leap compared to current 1.6T ports in scale-out applications.
Source: convergedigest.com
A detailed overview of Celestial AI's Photonic Fabric scaling networks, illustrating the technological foundations of optical connectivity.
Relevance for AI Systems
In modern AI systems, the bottleneck is increasingly shifting from mere computational power to data movement and memory bandwidth. This "memory wall" is also directly addressed by Celestial AI itself. Celestial claims that Photonic Fabric can support HBM3E and HBM4 requirements for bandwidth and latency while aiming for very low pJ/bit values.
HBM4 has been officially released by JEDEC as a standard since April 2025, and reports mention up to 2 TB/s of bandwidth per stack with a 2048-bit interface and a doubling of independent channels from 16 to 32 compared to HBM3. This scale makes it clearwhy any technology that brings HBM closer to compute or meaningfully expands bandwidth outside the traditional on-package paradigm becomes interesting.
The Marvell presentation additionally shows two directions beyond simple XPU-to-XPU links: "pooled, disaggregated memory" and photonic interconnects within multi-chip packages via OMIB. This is relevant for data centers because it describes a real operational problem: when workloads fluctuate, expensive HBM in fixed GPU configurations often remains underutilized, while memory is scarce elsewhere. This statement is an architectural inference from the disaggregated memory narrative, not an officially quantified measurement. Marvell Presentation.
Source: highyieldmarkets.com
Photonic Fabric technology enables unprecedented acceleration of data transfer within AI systems by eliminating bottlenecks.
Ecosystem and Competition
Celestial AI and Marvell visibly position their roadmap within an environment where open scale-up standards are gaining importance. UALink 200G 1.0 is described in a white paper as a new industry standard for scale-up interconnects, initially developed by a promoter group including companies such as Alibaba, AMD, Apple, AWS, Cisco, Google, HPE, Intel, Meta, Microsoft, and Synopsys.
Simultaneously, several players are driving co-packaged optics as a fundamental architectural motif. Nvidia describes CPO in its technical blog as a way to overcome traditional electrical and pluggable architectures in AI fabrics, and names specific product families like Quantum-X and Spectrum-X Photonics. Imec, in turn, demonstratesthat the scaling of CPO is not just a matter of optics designs but heavily depends on packaging and interfacing yield.
These three lines align technically: open scale-up protocols, optical engines close to compute, and a packaging stack that enables physical integration in high volumes. This statement is a synthetic classification based on the cited sources, not a single official claim by a company.
Future Developments
Marvell communicates clear commercialization expectations starting in 2028, linking them with hyperscaler traction. Reuters also frames this as a strategic move intended to strengthen Marvell's position in the competition with Nvidia and Broadcom.
If the timeline is taken seriously, the most likely first broad deployment will not be a complete optical rebuild of a data center, but a gradual infiltration of optics into previously electrified zones: first in scale-up links within high-performance racks, then into tighter package topologies with more HBM per XPU. This assessment follows directly from Marvell's depiction of the "next inflection point" logic within rack, system, and package. Marvell Announcement.
A plausible deployment scenario is a training pod where multiple custom XPUs or GPUs are coupled via optical chiplets, as traditional copper connections become increasingly expensive in terms of reach and performance per bit. Marvell explicitly discusses a transition of high-speed XPU links from copper to optics for rack-scale architectures in its announcement.
A second scenario concerns memory pooling. The Marvell deck shows "pooled, disaggregated memory" as an explicit application track. . Here, the potential lies less in spectacular benchmark numbers and more in operational efficiency: more dynamic memory allocation, less over-provisioning, better utilization profiles for large clusters. This consequence is a business operational derivation from the technical concept, not a metric published by the manufacturer.Marvell Press Release and Information Deck.
Whether the ambitious bandwidth and efficiency promises are achieved at scale ultimately depends on packaging realities, thermal robustness, and standard-compliant system integration. This is precisely where the perspectives of Marvell, Celestial AI, UALink, and broader CPO research converge. Imec, UALink and NVIDIA Developer Blog.
Those who follow this topic further should look less at buzzwords and more at three indicators: initial design wins in hyperscaler roadmaps, concrete products with clear thermal and yield data, and the question of how much HBM can economically be integrated closer to compute thanks to optical integration. This is the point at which an exciting technology can become a new standard architecture. Celestial AI Blog and Marvell Investor Relations.