Marvell Near Deal to Buy Celestial AI for Multiple Billions

Marvell Technology is advancing toward a potential acquisition of Celestial AI in a transaction that could be announced imminently, according to a report by The Information that was cited by Reuters. The deal is expected to involve a mix of cash and stock, with total consideration, including product milestone earnouts, potentially exceeding $5 billion.

Marvell, a networking chipmaker with a market capitalization of about $78 billion according to LSEG data, competes with larger rivals for custom chip work and data center networking business. Company officials and Celestial AI did not immediately respond to requests for comment. Reuters was not able to independently verify the report.

The talks reflect how strategic buyers are valuing technologies that address the most pressing bottleneck in modern artificial intelligence systems, memory bandwidth. Celestial AI, backed in part by a unit of Advanced Micro Devices, has focused on photonics to create fast optical links between compute chips and memory modules. The startup raised $250 million in March, bringing its total venture capital raised to $515 million, and counts Intel board veteran Lip Bu Tan among its directors.

Photonics promises lower latency and reduced power consumption compared with electrical interconnects, an attribute that matters as AI models grow in size and data movement costs increasingly dominate total system energy. For an acquirer like Marvell, which supplies switching and networking silicon to cloud providers and enterprise customers, integrating optical interconnects could help it offer more complete solutions for large scale AI deployments.

Market implications are immediate and broader. In the short run a deal would reinforce Marvell's competitive stance against Broadcom for custom silicon and networking contracts with hyperscalers. Over the longer term it signals a new wave of consolidation as established chipmakers move to internalize specialized startups that can plug performance gaps in AI systems. Valuations above $5 billion would also underline how investors are assigning high multiples to firms that can materially reduce the memory wall for AI workloads.

Regulatory questions may follow. Large semiconductor transactions often attract close review from competition authorities and security regulators in the United States and other jurisdictions, particularly when they touch critical infrastructure for cloud computing. The precise composition of the deal, including any equity component and investor identities, will influence the review pathway.

From an industry trend perspective, the reported talks fit into a pattern of incumbents buying capabilities rather than building them from scratch to keep pace with rapid advances in AI hardware. Whether through optics, specialized accelerators, or closer chip to memory integration, the next phase of hardware innovation is likely to be defined by technologies that reduce data movement costs. For customers and investors, the transaction would be a clear sign that the market for AI infrastructure is maturing into a battleground for vertically integrated platforms.