Marvell pays $115 billion for Lumentum in all-stock optical bet closing H1 2028

Marvell Technology is acquiring optical components maker Lumentum for $115 billion including debt, structured as an all-stock transaction with an H1 2028 close. The custom chipmaker is betting that optical interconnects will displace copper in hyperscale AI data centers, where latency and bandwidth now dictate rack economics. Lumentum manufactures coherent optics and high-speed transceivers used inside the clusters training frontier models. Marvell designs the adjacent silicon photonics and PAM4 DSPs. This is vertical integration at the photon layer.

The deal price represents roughly 8x Lumentum's trailing twelve-month revenue and is financed entirely through equity. Marvell will issue new shares to Lumentum holders at a ratio not yet disclosed, pending the usual shareholder votes and regulatory clearances across three continents. No cash changes hands. The structure echoes late-1999 telecommunications roll-ups, when inflated stock served as acquisition currency and integration risk lived on the balance sheet for years. The parallel is not lost on semiconductor analysts who lived through the fiber glut.

What matters is timing and substrate. AI training clusters are moving to 800G and 1.6T optical links to handle the east-west tensor traffic between GPU pods. Copper DAC cables cannot scale past seven meters at those speeds without signal integrity collapse. Marvell already sells the ASIC and SerDes chipsets that sit next to Nvidia's GPUs; owning the optical layer lets it sell a complete interconnect solution to Microsoft, Meta, and the cloud hyperscalers building out multi-hundred-thousand-GPU sites. Lumentum's coherent DSP technology also unlocks longer-reach connectivity for inference at the edge, where distributed model serving will require sub-five-millisecond round-trip guarantees. The operational synergy is tighter than it looks—Marvell can now control bill-of-materials cost, supply-chain lead times, and roadmap sequencing across the entire photonic signal path.

The risk is execution debt. Integrating two engineering cultures with different R&D cadences and customer bases typically takes eighteen months minimum, and that assumes no product roadmap collisions or talent attrition. Marvell is also assuming that optical disaggregation—splitting compute, memory, and storage across a photonic fabric—becomes the dominant AI architecture by decade-end. If hyperscalers instead double down on monolithic GPU nodes with tighter on-package interconnects, Lumentum's long-reach portfolio loses relevance. The all-stock structure protects Marvell's cash position but dilutes existing shareholders and ties equity value to integration milestones that won't hit until 2029.

Operators should watch three markers: regulatory filings in Beijing by Q1 2026, since both companies derive revenue from Chinese 5G and telecom infrastructure; Marvell's next quarterly guidance for custom silicon bookings, which will reflect whether hyperscalers are pre-committing to optical fabric builds; and Lumentum's Design Win announcements over the next six quarters, which indicate whether product roadmaps are staying on schedule during the merger hold period. The H1 2028 close assumes no extended antitrust review, but semiconductor M&A now averages 22 months from signing to completion.

Nvidia disclosed in its latest 10-Q that it is working with six optical partners on NVLink disaggregation prototypes. Lumentum was not named.