Marvell’s $115 Billion Bet on Optical AI Networks

Marvell agreed to buy optical-components maker Lumentum for $115 billion including debt on Nov. 2—11.5 times what it paid for Inphi in 2020. Buyers are still weighing where optics belong in the AI cluster.

For AI infrastructure builders, bandwidth is harder to buy one component at a time. Electrical interfaces, optical links and packaging all have to fit the same power and deployment budget. AI networking is becoming a vertically integrated systems business: advantage shifts from optimizing a switch or a signal-processing chip alone to coordinating the silicon, photonics, packaging and physical supply that move data among accelerators. Chip designers were moving that boundary before Marvell’s deal was announced.

Chipmakers moved optics toward compute before Marvell’s deal

Intel demonstrated a bidirectional optical compute-interconnect chiplet in June 2024, designed for higher-bandwidth transfers in AI workloads. In April 2025, Lightmatter introduced a silicon-photonic interposer and chiplet to move data between chips using light. Neither demonstration establishes that optical links have displaced electrical ones throughout a cluster. Both show builders treating the connection between chips as part of the compute design rather than a cable choice made afterward.

A cluster must move data across several boundaries: out of a chip, through a package or board, and onward to other machines. Improving one link does little if another interface consumes the available bandwidth or power budget. Intel’s chiplet and Lightmatter’s interposer approach that problem from different positions in the path.

Packaging turns an optical purchase into a system decision

Moving photonics closer to silicon makes component selection harder to separate from assembly and operation. A co-packaged-optics analysis compares DSP-based transceivers, linear pluggable optics, near-packaged optics and co-packaged optics across power budgets and total cost of ownership. It identifies scale-up networking as an important application, not a universal winning architecture.

Those alternatives leave buyers with different choices about where to put optical functions and how tightly to couple them to switching or compute silicon. Tighter coupling can change power use and bandwidth at the interface, but it also makes manufacturability and deployment part of the design decision. Builders also need advanced-packaging capacity to fit chips and optical links together. Co-packaged optics has been promised for years; suppliers still have to prove that a particular design beats less integrated options in the system a customer will operate.

Marvell’s purchases trace the widening boundary

Marvell announced its $10 billion Inphi acquisition in 2020 to expand its cloud-data-center and 5G opportunity. It subsequently agreed to buy network-chip startup Innovium for $1.1 billion. In 2025, Marvell said it would acquire Celestial AI for at least $3.25 billion, with the price able to reach $5.5 billion if revenue targets were met. Marvell moved from networking silicon into photonic links with the Celestial AI deal.

On Nov. 2, Marvell agreed to acquire Lumentum in an all-stock transaction valued at $115 billion including debt. That stated value is 11.5 times the price of the Inphi deal. Marvell has agreed to the purchase, not completed the integration; the comparison measures how much larger the proposed step is, not what it will earn.

Nvidia had invested $2 billion in each company during 2026. After its Marvell investment, Nvidia also said the companies planned to work on silicon photonics for high-speed data transmission. Those ties put an accelerator supplier on both sides of the proposed combination. They do not establish which optical architecture Nvidia or Marvell will ultimately deploy.

Other suppliers are buying technology while customers buy capacity

Credo, which makes active electrical cables for AI servers, agreed to acquire DustPhotonics for up to $1.3 billion in April 2026. DustPhotonics targets faster, lower-cost transfer for next-generation AI clusters. Credo’s move approaches photonics from an electrical-connectivity business, while Marvell approaches it through networking silicon. Separate suppliers have found reason to reach across the same component boundary.

Ciena CEO Gary Smith said AI demand was outpacing Ciena’s supply of optical-networking chips. Reporting on the optical supply chain found shortages and price increases extending through lasers, substrates, fiber and connectors. Corning’s fiber business benefited from data-center demand and a $6 billion Meta deal, while Nvidia explored Corning’s co-packaged optics for servers. Acquiring Lumentum could give Marvell more control over one part of this chain; it cannot manufacture away constraints across every other part.

Customers now shape the supplier’s capital structure

Google expanded its chip-development agreement with Marvell and received a warrant to purchase up to $12.2 billion of Marvell shares. A warrant is an option to buy shares, not evidence that Google has spent that amount. It nevertheless makes the customer relationship relevant to Marvell’s ownership as well as its engineering work.

Broadcom has taken a different position in the buildout. It arranged a $63 billion financing vehicle for Anthropic chip purchases, drawing banks and other financial firms into the infrastructure order. Broadcom is organizing capital behind demand, making financing terms and dependence on major customers part of the competition alongside chip performance. Broadcom’s financing partners face a return hurdle on AI infrastructure even when chip orders are strong.

Lumentum shares were up 128% year to date when the agreement was reported. Buyers still have competing optical designs to choose from, and Marvell has yet to integrate the company. Marvell has agreed to put $115 billion behind Lumentum while customers still weigh where to put optics in the cluster.