Marvell Moves Into Datacenter Optics With Planned $80 Billion Lumentum Acquisition

Photonics moves closer to the center of datacenter design

For decades, silicon photonics has been expected to enter datacenters at scale and reduce the distance and latency between compute, memory, and storage. High-bandwidth, low-latency optical links could eventually allow operators to assemble systems from pools of components, configuring CPUs, GPUs, XPUs, DPUs, and memory for specific workloads rather than relying only on discrete, largely identical computers.

That vision did not become economical or sufficiently low-risk while electrical signaling could still be pushed toward the limits of physics. Photonics adoption was therefore delayed, and manufacturers faced the familiar chicken-and-egg problem: high-volume production required demand, while demand depended on lower costs and proven supply.

Generative AI changed the equation. Its bandwidth and latency requirements are substantially higher than those of earlier machine-learning systems. As hyperscalers, cloud providers, and AI model developers invest in increasingly capable hardware, photonics is moving beyond optical transceivers between systems and regions. It is also being considered for connections within systems and across the components of large computing complexes.

Composability is not necessarily the immediate objective. The more direct goal is to overcome what the industry calls the “copper wall.” However, optical connections could eventually make it practical to combine trays or racks of CPUs, GPUs, XPUs, DPUs, and memory, then configure their ratios as workloads change.

Marvell’s proposed Lumentum acquisition

Marvell plans to acquire photonics specialist Lumentum for $80 billion in stock and assume its debt, producing a transaction value of approximately $115 billion. The deal would be nearly twice the $67 billion Dell spent acquiring EMC and VMware in October 2015, and about four times the amount Avago Technology paid for Broadcom in May of that year.

Broadcom later spent $61 billion to acquire VMware in May 2022. Its enterprise portfolio also includes Brocade Fibre Channel switches and systems software from CA Technologies and Symantec.

Lumentum is one of the two major suppliers of indium phosphide lasers used in optical interconnects. These components support scale-out networks across racks and potentially rows, as well as scale-in networks connecting elements within a compute system. Coherent is the other major supplier. Both Lumentum and Coherent received $2 billion to expand manufacturing in March 2026.

Marvell itself received a $2 billion investment from Nvidia a few weeks later to help customers adopt NVLink ports on custom XPUs and potentially CPUs.

Nvidia could theoretically acquire one or both optical suppliers, but antitrust concerns would make that difficult. Acquiring only Lumentum or Coherent would provide roughly half of the relevant laser supply, while developing an internal operation would take too long. Multiple suppliers could help expand capacity and maintain competition, although supply remains constrained in markets such as high-bandwidth memory.

Optical interposers and lasers may become another limiting factor for the generative AI industry. That supply constraint is central to the strategic importance of the Lumentum transaction.

A supply advantage for Marvell’s custom silicon business

Marvell has a growing custom CPU and XPU business with Amazon Web Services, Microsoft, and Google. Owning Lumentum or Coherent could give Marvell greater influence over laser supply for that business while allowing it to continue selling components to Nvidia, AMD, Intel, hyperscalers, cloud providers, and AI companies developing their own chips with assistance from firms such as Broadcom, Marvell, Alchip, and MediaTek.

Marvell chief executive officer Matt Murphy described the transaction as a way to become “a category leader for the next generation of AI connectivity.” The timing also reflects the expectation that photonics laser orders are already booked through early 2029, making supply a potential constraint on Marvell’s custom-chip expansion.

Marvell’s market capitalization was approximately $150 billion when the deal was announced. The company would use $80 billion in stock for the acquisition, assume another $35 billion in debt and obligations, and add three Lumentum directors to its board.

The transaction follows Marvell’s acquisition of PHY manufacturer Inphi and its $3.25 billion acquisition of Celestial AI in December 2025. Celestial AI contributed photonics interposer technology intended to support optical on-package compute architectures.

The companies’ reporting quarters have been aligned for comparison, although they are one month out of phase. Lumentum reported a $7.16 billion loss two quarters earlier, largely associated with converting debt to stock. Excluding that accounting event, the company reported operating income of $279.3 million on $1.01 billion in sales. In its most recent quarter, Lumentum reported $360.3 million in profit on $1.26 billion in sales.

Marvell reported $308 million in net income on $2.74 billion in sales in its most recent quarter. Lumentum’s recent profitability coincides with increased demand expectations for lasers in future Nvidia and other AI systems.

Murphy said the Lumentum opportunity could be larger than the opportunity Marvell saw in Inphi. He described the industry as moving from scale-out optical connections toward scale-up and eventually scale-in connections as the “copper wall” moves closer to the computing elements.

According to Murphy, Lumentum’s business could reach a scale roughly an order of magnitude larger over the next several years than Marvell initially expected from Inphi. He also emphasized that Lumentum’s orders are already recorded rather than being only a forward projection.

Expanding Marvell’s connectivity market

Marvell also expects the acquisition to increase its position in the semiconductor industry, allowing it to compete more directly with MediaTek and Qualcomm while approaching the scale of Intel and AMD in selected areas.

Marvell describes the combined company as the largest pure-play provider of datacenter connectivity chips. That description includes products extending from compute engines to connections across datacenter regions. Nvidia and Broadcom, however, are expected to maintain datacenter connectivity businesses several times larger than the combined Marvell and Lumentum operation.

Marvell estimates a total addressable market of $400 billion in calendar 2030. The estimate is driven mainly by scale-out and scale-across chips for AI clusters, with additional opportunities in scale-up and scale-in connections inside nodescale and rackscale systems.

Lumentum previously estimated a $90 billion total addressable market. Marvell believes it can pursue an additional $60 billion, primarily through near-package optics (NPO) and co-packaged optics (CPO) chips and packaging. The estimate could expand if custom-chip designers seek suppliers capable of supporting the full optical fabric of an AI system.

Optical fabrics across multiple network levels

Larger compute and storage complexes require more components to be packed into greater physical areas. In this model, a row can take on the role previously associated with a rack, while a system board can become the functional equivalent of a compute engine. Copper signaling cannot support every part of this expansion.

Murphy described a roughly tenfold bandwidth increase for each level of the network hierarchy, moving inward from distributed datacenters to rackscale systems, then to the elements within those systems and finally to compute and storage components on a package or tray:

• Scale across: Connections between distributed datacenters.
• Scale out: Connections between rackscale systems within a datacenter.
• Scale up: Connections between elements in a rackscale system, including coherent memory connections.
• Scale in: Connections between compute and storage components on a package or within a tray.

The reverse tradeoff applies when moving outward through the hierarchy. Devices cannot provide enough electrical or optical lanes to maintain the same bandwidth from a package or tray to a global network. Instead, bandwidth generally decreases by an order of magnitude at each broader level, extending link reach and relying on successive layers of the fabric to exchange data.

Marvell and Lumentum already provide products for scale-across and scale-out networking. Near-package optical circuits can also support scale-up coherent networking for switches and XPU accelerators.

The next stage involves tighter integration between fabrics, accelerators, and electronic switches. The objective is to increase bandwidth while reducing latency and power consumption per bit per second.

Manufacturing and materials expertise

Large-scale deployment of optics for scale-up and scale-in systems requires experience designing and delivering chips in volume. Marvell and Lumentum have each delivered billions of chips over their histories.

The technology also requires expertise in several semiconductor materials:

• Indium phosphide, used for lasers.
• Silicon germanium, used for heterojunction bipolar transistors in the analog front ends of optical links.
• Gallium arsenide, used in another type of laser.

Combining Marvell and Lumentum would place these capabilities under one organization. That could make vertical integration and volume planning easier for customers, although it would also concentrate more supply-chain risk in a single company.

Optical circuit switches

The transaction could also give Marvell a stronger position in optical circuit switches, which may serve as alternatives to Ethernet switches in some AI backend networks.

Google first used its Palomar optical circuit switch in 2020 with TPUv4 accelerators. The system used reconfigurable arrays of micromirrors to connect nearly 10,000 TPU accelerators in changing network topologies. The project is also associated with the Apollo codename, although Apollo refers to the broader architecture and Palomar to the particular MEMS switch.

If optical circuit switches can reduce power consumption and latency, other hyperscalers, cloud providers, and AI developers may adopt similar approaches. Marvell also offers Innovium TeraLynx Ethernet switch ASICs for backend networks.

A lengthy approval process

The acquisition is expected to close in the first half of 2028, assuming it receives the required approvals. That timeline is long in the rapidly changing AI hardware market.

Regulators in China and Europe could review the transaction, and the extended closing period leaves time for competing offers. The deal also places Coherent in a more prominent strategic position as another potential acquisition target, similar to how Intel’s acquisition of FPGA maker Altera helped put Xilinx into play a decade ago.