Marvell Signs $115B Lumentum Deal for Optical AI Switches

Updated on November 4, 2026

Marvell Technology has entered into a definitive agreement to acquire optical and photonic hardware leader Lumentum in an all-stock transaction valued at approximately $115 billion. Announced on November 2, 2026, the blockbuster semiconductor consolidation unites Marvell’s high-radix data center switch silicon, digital signal processors (DSPs), and Co-Packaged Optics (CPO) platforms with Lumentum’s Indium Phosphide (InP) laser fabrication plants and MEMS-based Optical Circuit Switches (OCS).

As hyperscale AI clusters scale past 100,000 accelerators across multi-building campuses, electrical copper cables hit a severe physical barrier known across infrastructure engineering as the “copper wall.” At 224G-per-lane signaling and beyond, passive direct-attach copper (DAC) cables lose signal integrity over distances exceeding a few feet, forcing data centers to rely on power-hungry pluggable optical transceivers and electrical packet switches.

By vertically integrating Lumentum’s compound semiconductor fabs and optical switching subsystems, Marvell positions itself as a vertically integrated alternative to Broadcom and NVIDIA’s proprietary networking stacks. Rather than sourcing external continuous-wave (CW) lasers for its Co-Packaged Optics engines, Marvell gains direct control over the scarce Indium Phosphide wafer supply chain that governs next-generation 1.6T and 3.2T optical interconnect production.

Equally strategic is Lumentum’s R300 Optical Circuit Switch portfolio. Unlike traditional electronic spine switches that convert incoming light into electrical signals, buffer packets in silicon, and re-modulate lasers on the egress port, MEMS optical circuit switches bounce photonic beams directly off microscopic mirrors, eliminating optical-electrical-optical (OEO) conversion power and latency.

Breaking the Data Center Copper Wall with vertically Integrated Photonics

Inside a gigawatt-scale AI training or disaggregated inference cluster, networking optics and switch silicon now account for 20% to 30% of total system power and capital expenditure. Every time a 800G or 1.6T optical link passes through a conventional spine switch, two transceivers and an electrical ASIC burn dozens of watts performing OEO conversion.

Google pioneered MEMS optical circuit switching inside its TPU v4 and TPU v5p pods to dynamically route around failed nodes and reconfigure 3D torus topologies in milliseconds. By bringing Lumentum’s R300 OCS hardware under the same roof as Marvell’s Teralynx switch ASICs and custom XPU interconnect IP, Marvell can now offer AWS, Microsoft, Meta, and sovereign cloud builders a complete optical pod fabric out of the box.

What a $115B Valuation Signals for AI Infrastructure Roadmaps

Valuing an optical components and photonic systems manufacturer at $115 billion underscores a fundamental shift in AI hardware economics: compute FLOPs are no longer the sole determinant of cluster scaling. When a single GPU or custom XPU die is surrounded by seven or eight HBM4 stacks, the limiting factor on cluster efficiency is how quickly and cheaply bits move between racks and across data center campuses.

Competitors including Broadcom (with its Bailly CPO platform) and NVIDIA (with Quantum-X and Spectrum-X photonics) now face a rival that owns both the high-speed CMOS SerDes/DSP design and the underlying Indium Phosphide laser physics.