- Marvell Technology has agreed to acquire Celestial AI, a semiconductor startup that developed Photonic Fabric, an optical interconnect technology targeting AI training cluster infrastructure.
- The deal expands Marvell’s AI silicon portfolio into scale-up networking — the high-bandwidth layer connecting AI accelerators within a single training cluster — where electrical interconnects face growing bandwidth and power limitations.
- Celestial AI’s Photonic Fabric uses wavelength-division multiplexing (WDM) over optical fiber, allowing multiple simultaneous data streams per fiber and enabling bandwidth densities that copper-based alternatives struggle to match at short reach.
- Financial terms were not disclosed; the transaction is subject to customary regulatory approvals and closing conditions.
What Happened
Marvell Technology announced an agreement to acquire Celestial AI, a San Jose-based semiconductor startup specializing in photonic interconnect technology for AI data center infrastructure, according to a Marvell press release. The acquisition brings Celestial AI’s Photonic Fabric architecture — a WDM-based optical interconnect system — into Marvell’s existing AI product lines, which include custom ASICs, ethernet controllers, and optical components for hyperscale deployments.
Celestial AI was founded to address what its team characterized as a structural mismatch between AI cluster compute density and the bandwidth ceiling of copper-based electrical interconnects. The company’s Photonic Fabric product targets the scale-up networking layer: the hardware responsible for moving data between AI accelerators — GPUs or custom AI chips — within a single server rack or training pod.
“This acquisition is a critical step forward in Marvell’s strategy to deliver the complete silicon solution for AI infrastructure,” said Matt Murphy, Marvell’s President and CEO, in the company’s announcement. Chi Lam, co-founder and CEO of Celestial AI, stated in the release that joining Marvell would provide the manufacturing relationships and customer reach required to bring Photonic Fabric to hyperscale AI deployments at global scale.
Why It Matters
Scale-up networking has become a priority for AI infrastructure investment as model training runs increasingly span tens of thousands of accelerators and require sustained, low-latency data movement between compute nodes. Nvidia has dominated this segment with NVLink, a proprietary interconnect technology embedded in its DGX and HGX server platforms — the H100 generation’s NVLink 4 provides 900 GB/s total bidirectional bandwidth per GPU, a figure that competing approaches must meet or exceed to earn hyperscale design wins.
Marvell has built a substantial custom AI ASIC business serving hyperscalers including Google, but its scale-up connectivity offering has remained limited relative to its switching and optical DSP product lines. The Celestial AI acquisition positions Marvell to offer a more complete AI silicon platform — one that spans compute and cluster interconnect — a portfolio configuration that competing vendors have not yet assembled in a single product family.
Technical Details
Celestial AI’s Photonic Fabric employs wavelength-division multiplexing to encode multiple independent data channels onto distinct wavelengths of light, all transmitted simultaneously over a shared optical fiber. Because each wavelength carries a separate data stream, aggregate throughput scales with the number of wavelengths deployed rather than requiring additional physical cabling — a meaningful advantage in dense rack environments where cable volume affects airflow and thermal management.
The architecture targets short-reach applications spanning chip-to-chip distances measured in centimeters through rack-to-rack links of several meters, the range where signal-integrity constraints on high-speed copper become a limiting factor at current and planned data rates. Celestial AI has claimed its design achieves higher bandwidth per watt than copper alternatives in this distance range, a claim consistent with published photonics literature showing optical links deliver increasingly favorable energy efficiency as signaling rates rise above approximately 100 Gbps per lane. The company has positioned Photonic Fabric as complementary to, rather than a replacement for, longer-haul optical transceivers already deployed in data center interconnect applications.
Who’s Affected
Hyperscale cloud operators designing next-generation AI training infrastructure are the primary intended customers for the combined Marvell-Celestial AI product roadmap. Microsoft, Google, Amazon Web Services, and Meta have each disclosed investments in custom AI silicon and evaluated alternatives to Nvidia’s vertically integrated stack; an optical scale-up interconnect from Marvell could enter hardware qualification pipelines for clusters planned for 2027 deployment cycles and beyond.
Nvidia’s NVLink business and Broadcom’s networking portfolio face a potential new competitor if Marvell productizes Photonic Fabric within its existing sales channels. Startups in the photonic interconnect space — including Ayar Labs — are also pursuing similar design wins at hyperscale, and Marvell’s entry consolidates one competitor while adding capital and manufacturing reach that could accelerate commercialization timelines for optical scale-up solutions across the sector.
What’s Next
Neither Marvell nor Celestial AI disclosed the transaction price or a target closing date; the deal remains subject to regulatory review in relevant jurisdictions. Marvell said Celestial AI’s engineering team will integrate into its electro-optics and interconnect product groups following close. The company is expected to discuss integration timelines and any updates to its AI silicon roadmap during its next quarterly earnings call.
