Marvell to buy Lumentum for $115 billion as AI data centres move to light

Marvell puts $115 billion on light

Lumentum brings high-performance lasers, optical transceivers, optical circuit switching and compound semiconductor manufacturing. Marvell chief executive Matt Murphy cast the purchase as an architectural shift. "AI infrastructure is undergoing a fundamental architectural shift toward optical connectivity, creating an opportunity that comes along once in a generation," he said. Marvell had already bought Celestial AI to bolster its optical interconnect offerings. The two businesses sit on opposite sides of the same link. Marvell designs the DSPs, switches and custom accelerators that sit inside optical transceivers. Lumentum supplies the lasers those transceivers use to turn electrical signals into light.

Why it matters: every large AI cluster reaches a point where signals leave the network chip as electricity and become light inside a transceiver. Whoever owns the lasers and the switching owns a constrained part of the supply chain. Buyers of AI networking gear get a more vertically integrated Marvell competing with Broadcom. They also get fewer independent laser suppliers to play off. The closing timetable and any regulatory review remain unknown.

The bottleneck the deal is aimed at

In cloud networks, the bottleneck sits at the hand-off. Silicon-based network interface cards pass signals to laser-driven transceivers. Co-designing the electrical and optical signal chain more tightly could cut latency and power draw.

Lumentum chief executive Michael Hurlston puts the opportunity at $550 billion. Not everyone is convinced the price is right. Axel Rudolph is chief technical analyst at IG Group. He told Reuters that investors are right to ask whether Marvell is paying too much and taking on too much at once. The deal is all stock, so Lumentum holders end up owning a slice of the combined company and its execution risk.

What Marvell already ships

The deal builds on hardware Marvell already sells. In March 2026 the company expanded its 1.6T optical DSP portfolio. It said its Ara platform was shipping in mass volume to global customers. In September 2026 it demonstrated 2nm optical technologies including 400G-per-lane PAM4. These target 3.2T data centre connectivity.

Rivals are spending too. In March 2026 Nvidia announced a US$2 billion investment in Coherent to expand optics research and manufacturing in the United States. Coherent introduced its PhotonLink platform for co-packaged and near-packaged optics in September 2026. Co-packaged optics puts the light engine next to the switch chip, not at the front panel. Much of the promised power saving comes from that placement.

Forecasts, and what they leave out

Forecasts for the sector vary widely. One report puts optical interconnects at US$38.5 billion by 2033, growing at a 12.3% compound annual rate. A separate report forecasts the photonic computing market at US$23.60 billion by 2036 at a 43.3% compound annual rate.

The two cover different scopes. One covers optical interconnects, the other photonic computing. The figures are not directly comparable. Neither is a revision of the other.

Optical compute gets a rack-scale test

The Advanced Research and Invention Agency said on 3 November 2026 that it will deploy Lumai Iris inference servers in its Scaling Inference Lab. They will be the first optical inference platform in the lab. CommonAI CIC runs the lab. The lab tests next-generation compute on real workloads in a commercial data centre. It uses standard power, cooling and rack integration. ARIA is a UK public research funder. An Act of Parliament created it.

Lumai says its technology can lower data centre power consumption by 90%. It can also deliver 50x higher processing speeds. It also says its servers could cut total cost of ownership to around one-tenth of conventional GPU architectures.

Those are vendor claims until the lab publishes results. David Lewis is ARIA's chief science and technology officer. He said Lumai's technology was the most mature optical compute platform the agency evaluated. Lumai chief executive Xianxin Guo said power, not demand, now limits how fast AI can be deployed.

ARIA cites an estimate. Meeting one leading AI lab's compute ambition with conventional accelerators alone would reach a $100 trillion infrastructure bill within five years. The same estimate puts the additional energy requirement above 1,000 GW. That is a projection, not a measurement. It is also the kind of number that makes a 90% power saving worth checking at rack scale rather than in a slide deck.