On November 2, 2026, Marvell Technology and Lumentum announced that they will combine. CNBC reports that "Lumentum shareholders will receive 4.5 Marvell shares for each Lumentum share" and that "The deal is expected to close during the first half of 2028." The companies said the transaction "has an implied enterprise value of $115 billion for Lumentum". Because the deal is all stock, its value moves with Marvell's share price, and CNBC reports that Marvell shares fell 13 percent on the day while Lumentum's fell nearly 19 percent.
Why a chip company wants a laser company
Marvell sells custom silicon, switching and electrical interconnect. Lumentum makes lasers, optical transceivers and optical circuit switches. Marvell says the combination lets it join electrical and optical technology to meet AI connectivity demands. CNBC puts the context plainly: "Companies are increasingly relying on optical connections to piece together chips used to run and train large language models." Marvell CEO Matt Murphy told CNBC: "You need scale-up technology that needs semiconductors and optical working seamlessly together." On supply, he said "the reason that they're sold out is because the demand is exploding, and we don't see that changing anytime soon." Earlier this year Marvell closed its acquisition of Celestial AI, a startup working on optical interconnect, so this is its second large bet on the same idea.
NVIDIA is part of the story too. CNBC notes NVIDIA's recent multibillion-dollar investments in both companies and the criticism that such stakes amount to funding one's own customers and suppliers. Murphy called the deal "a double down on supporting Nvidia, who is the ecosystem leader in this area."
What it means for people who run AI workloads
Training and serving large models increasingly means splitting one model across many accelerators. Tensor and expert parallelism send activations between chips on every layer, so the bandwidth and latency of those links bound how far a model can be split before communication dominates. Copper links work over short distances inside a rack. Beyond that, signals have to become light, which is why optics suppliers now matter to cluster design. When Murphy says parts are sold out, the practical consequence for buyers is that the supply of optical components, not only of GPUs, decides how large a cluster can be built and when.
What remains uncertain. The deal needs regulatory and shareholder approval and is not expected to close until 2028, so nothing changes for buyers soon. The announcement gives no product roadmap, prices or specifications, and the "sold out" remark is a chief executive's statement on television, not reported order data. Whether combining the companies relieves supply or simply moves margin from one to the other is not something the release can show.
Analysis: For teams that rent capacity rather than build clusters, the lesson is to ask about the network, not just the accelerator. A cluster with fast chips and slow interconnect may suit independent inference requests but not training or very large mixture-of-experts serving. At Belgavi's Book Store, Tanishka could add three questions when comparing GPU offers for fine-tuning the help-desk model: how many accelerators share a fast scale-up domain, what bandwidth links them beyond it, and whether a job can be placed entirely inside one domain. The cheapest price per GPU hour is not the cheapest run if half the hours are spent waiting on the network.
Sources and evidence
Marvell agreed on November 2 to buy Lumentum in an all-stock deal (4.5 Marvell shares per Lumentum share, implied enterprise value of $115 billion), expected to close in the first half of 2028. The companies argue that AI clusters need electrical and optical interconnect designed together. Marvell's CEO said optical parts are sold out on demand. Both companies have received NVIDIA investment, and nothing changes for buyers before the deal closes.