Marvell Technology has taken a significant step in the race for AI infrastructure. The company has announced a definitive agreement to acquire startup Celestial AI for approximately $3.25 billion, in a deal that reinforces its commitment to high-speed optical interconnections in next-generation data centers.
The acquisition, which still requires applicable regulatory approvals, is expected to close in the first quarter of 2026. This will allow Marvell to consolidate a technology considered key to the next wave of AI architectures: fully optical scale-up networks connecting hundreds of accelerators (XPUs) across multiple racks.
A billion-dollar deal for a new kind of “data pipelines”
The agreed package includes $1 billion in cash and approximately 27.2 million Marvell shares, valued at around $2.25 billion based on recent average share prices. Additionally, Celestial AI shareholders could receive up to another $2.25 billion in additional shares if certain revenue milestones are met by the end of Marvell’s fiscal year 2029.
In total, the deal could exceed $5.5 billion if the most ambitious targets are achieved. Currently, the confirmed amount stands at $3.25 billion, placing this acquisition among the most significant of the year in the semiconductor segment for AI data centers.
The market’s reaction has been positive: following the announcement, Marvell’s shares rose notably, supported by strong quarterly earnings and growth forecasts driven precisely by demand for AI infrastructure.
What Celestial AI does and what their Photonic Fabric is
California-based Celestial AI has become one of the most cited names in optical interconnects for AI, thanks to its Photonic Fabric platform. This technology aims to replace short-distance electrical links—within the rack, between systems, and even at the package level—with high-capacity, low-latency, low-power optical links.
Their first chiplet for Photonic Fabric scale-up interconnection integrates all electrical and optical components needed for the link into a single package, offering up to 16 Tb/s bandwidth, ten times more capacity than the 1.6 Tb/s ports commonly used in many scale-out applications. These chiplets can be co-packaged with XPUs and scale-up switches, drastically increasing the total bandwidth of a system.
One of Celestial AI’s technological differentiators is its thermal stability. The company states it can operate in extreme environments generated by kilowatt-scale XPUs and switches, allowing the integration of photonics directly into high-power 3D packages. By entering from the “center” of the chip rather than from the edge, this co-packaged optical component frees up silicon perimeter space, which can be used for additional HBM memory within the same package.
From copper cables to light: the next leap in AI
The strategic logic is clear: AI data center architectures are evolving rapidly. Systems are no longer limited to a single rack; they are deploying as multi-rack clusters where hundreds of accelerators must communicate seamlessly as a single logical system with direct access to each other’s memory.
In this context, copper interconnects—limited by reach, bandwidth, and energy efficiency—are becoming insufficient. To go further with less power and more capacity, the industry is shifting toward solutions that are completely optical both within racks and at the system level.
Celestial AI aims precisely at this inflection point: its optical links promise more than double the energy efficiency of equivalent copper interconnects, with latencies below nanoseconds and spans of up to tens of meters within the data center.
How Celestial AI fits into Marvell’s strategy
Marvell already held a significant position in data center connectivity: Ethernet solutions, optical components for data center interconnect (DCI), PAM4 DSPs, CXL, and a prominent role in the UALink ecosystem for scale-up networks connecting accelerators within the same domain.
With the acquisition of Celestial AI, the company aims to offer a comprehensive connectivity portfolio:
- Scale-out (between racks and data centers) using existing Ethernet and optical solutions.
- Scale-across and long-distance DCI.
- And now scale-up optical connectivity within and between racks, enabled by Photonic Fabric, where the majority of XPU-to-XPU traffic in large AI clusters is concentrated.
Marvell executives have emphasized that this move expands their addressable market for AI connectivity and accelerates their roadmap to become a leading provider for major cloud providers and other high-performance data center operators.
Celestial AI, on its part, already collaborates with several hyperscalers and ecosystem partners on next-generation architectures, and it is anticipated that its Photonic Fabric chiplets will be co-packaged with custom XPUs and scale-up switches in the first large-scale commercial deployments of optical interconnects for AI.
Projected timeline and economic impact
Marvell expects Celestial AI’s contribution to revenue to become meaningful in the second half of its fiscal year 2028, reaching an annualized $500 million by that quarter, doubling to $1 billion a year later in the fourth quarter of fiscal year 2029.
In the shorter term, the company projects total revenues of around $10 billion in the upcoming fiscal year, with double-digit growth driven largely by data center business and solutions related to the AI race.
If forecasts hold, the acquisition of Celestial AI will not only strengthen Marvell’s position in the AI connectivity ecosystem but also cement the role of co-packaged optical interconnects as a cornerstone of future data center design.
Frequently Asked Questions about Marvell’s acquisition of Celestial AI
What exactly has Marvell bought in this deal?
Marvell acquires Celestial AI, a startup specializing in the Photonic Fabric platform, which enables ultra-high-capacity optical interconnections between accelerators (XPUs) and switches within and between racks. Besides technology and equipment, Marvell gains a customer and partner base—including hyperscalers—that were already testing this solution in next-gen AI architectures.
Why is optical interconnection so critical for AI data centers?
Current AI models require clusters with hundreds of accelerators working in parallel. Traffic between XPUs dominates bandwidth consumption, and latency becomes a critical factor. Copper electrical connections are hitting limits in reach, bandwidth, and power. Optical solutions offer higher capacity, greater range within data centers, and lower energy per bit—making them essential for scaling next-generation AI systems.
When will Celestial AI start delivering significant revenue to Marvell?
According to company forecasts, Photonic Fabric solutions are expected to generate meaningful revenue starting in the second half of fiscal year 2028, reaching about $500 million annually by that time and up to $1 billion a year later, assuming adoption by hyperscalers and other large operators meet expectations.
How does this deal affect competition in the AI market?
The acquisition strengthens Marvell’s position against competitors like Broadcom and NVIDIA, also developing optical and high-speed connectivity solutions. While these players push their own co-packaged optical and high-speed networking solutions, Marvell aims to be an integrated provider of both electrical and optical connectivity for scale-out and scale-up data centers, with proprietary photonics technology ready for integration into the chips and systems of major cloud providers.
via: investor.marvell

