Illustration for: Marvell to buy Celestial AI for USD 3.5 B, boosting XPU memory access fabrics
Business & Startups

Marvell to buy Celestial AI for USD 3.5 B, boosting XPU memory access fabrics

2 min read

Marvell’s $3.5 billion purchase of Celestial AI signals a clear push into the next generation of data‑centre interconnect. The deal, announced in early March, pairs Marvell’s silicon‑based networking portfolio with Celestial’s XPU‑focused memory‑fabric technology. For operators juggling ever‑larger AI workloads, the ability to move data across multiple processors without a bottleneck has become a top priority.

Yet stitching together dozens of racks isn’t as simple as adding more cables; it demands hardware that can handle higher power draws, tighter latency windows and a surge in bandwidth requirements. That’s why the acquisition emphasizes specialised switches and protocols designed to keep the fabric humming efficiently. As the market looks to scale beyond single‑node solutions, the question becomes whether this integrated approach can deliver the performance edge that large‑scale AI clusters need.

The company added that this architecture enables each XPU to directly access the memory of others. These advanced fabrics require specialised switches and protocols such as UALink to achieve optimal performance and efficiency. As multi‑rack scale‑up fabrics demand power, bandwidth, latency, and reac…

The company added that this architecture enables each XPU to directly access the memory of others. These advanced fabrics require specialised switches and protocols such as UALink to achieve optimal performance and efficiency. As multi-rack scale-up fabrics demand power, bandwidth, latency, and reach, interconnects will increasingly shift to all-optical connections. Sandeep Bharathi, president of the Data Centre Group, said, "By combining our UALink scale-up switch roadmap and Celestial AI's breakthrough optical scale-up interconnect, we will enable customers to build AI systems that scale beyond the limits of copper and redefine what's possible in AI data centre architecture." Celestial AI's photonic fabric technology enables scaling of AI clusters with high-bandwidth, low-latency optical connections.

Related Topics: #Marvell #Celestial AI #XPU #memory fabric #UALink #data centre #AI workloads #optical interconnect

Will Marvell’s purchase translate into faster data‑centre links? The $3.25 billion deal brings Celestial AI’s photonic fabric technology under Marvell’s umbrella, promising direct XPU‑to‑XPU memory access. Yet the announcement offers few details on how the specialized switches and UALink protocol will integrate with existing infrastructure.

Because multi‑rack scale‑up fabrics consume significant power, bandwidth and latency, the combined effort must address those constraints. And while the acquisition is framed as a boost to next‑generation AI and cloud workloads, the actual performance gains remain unclear. Marvell’s statement that “next‑generation accelerated systems are no longer confined” hints at broader ambitions, but no roadmap is provided.

Consequently, analysts will watch whether the photonic interconnects can deliver the efficiency claimed, especially as data‑centre operators weigh cost against potential benefits. In short, the purchase expands Marvell’s portfolio, but whether it will materially reshape connectivity for AI workloads is still uncertain. Some customers have expressed interest, yet contracts have not been disclosed.

Further Reading

Common Questions Answered

What is the total value of Marvell's acquisition of Celestial AI, and what strategic capability does it aim to enhance?

Marvell agreed to acquire Celestial AI for approximately $3.5 billion, a move designed to strengthen its data‑centre interconnect portfolio. The deal focuses on integrating Celestial’s X‑processor‑focused memory‑fabric technology to enable direct XPU‑to‑XPU memory access across racks.

How does the combined architecture allow each XPU to access the memory of other XPUs, and which protocol supports this functionality?

The architecture creates a shared memory fabric where each XPU can directly read and write to the memory of neighboring XPUs, eliminating traditional bottlenecks. This capability is enabled by specialized switches that implement the UALink protocol, which optimizes performance and efficiency.

What role do photonic fabrics play in Marvell’s plan for multi‑rack scale‑up interconnects?

Photonic fabric technology, brought in from Celestial AI, uses light‑based signaling to reduce latency and power consumption in large‑scale rack deployments. By leveraging these all‑optical connections, Marvell aims to meet the high bandwidth and reach requirements of future AI workloads.

What challenges does the article highlight regarding the integration of Celestial AI’s switches and UALink protocol into existing data‑centre infrastructure?

The article notes that while the acquisition promises faster XPU‑to‑XPU links, it provides limited details on how the specialized switches and UALink protocol will mesh with current hardware. Addressing power, bandwidth, and latency constraints in multi‑rack fabrics will be essential for successful integration.