Marvell’s $115 Billion Lumentum Bet

In March 2026, Nvidia announced plans to invest $2 billion each in optical suppliers Coherent and Lumentum. Eight months later, Marvell agreed to acquire Lumentum for $115 billion including debt. A faster AI chip cannot recover the time it spends waiting for data.

As bandwidth, power and packaging constraints move nearer compute chips, lasers, transceivers and photonic packaging become harder to buy as interchangeable parts. Chip and network-platform owners have reason to finance capacity, co-design interfaces or acquire suppliers instead.

Builders moved optics toward the chip before buyers moved upstream

In June 2024, Intel demonstrated a bidirectional optical compute-interconnect chiplet for higher-bandwidth AI data transfers. In April 2025, Lightmatter introduced a silicon-photonics interposer and chiplet for moving data between AI chips. OpenLight, pursuing custom silicon-photonic chips for data-center interconnects, raised $34 million later that year. These were different products from different builders, but each addressed the boundary between compute silicon and its connections.

A demonstration is not a deployment, and a launch does not establish favorable manufacturing economics. Intel, Lightmatter and OpenLight treated the route between chips, packages and networks as part of the AI system rather than a specification for a removable module. Networking-chip designers are also bringing optical components closer to their silicon.

Cluster throughput depends on more than a transceiver

In May 2026, Nikkei Asia reported shortages and price increases across lasers, substrates, optical fiber and connectors as AI demand spread through the supply chain. A customer seeking more network capacity needs the materials and manufacturing steps behind the finished link, not merely a faster transceiver design.

By October, Ciena chief executive Gary Smith said AI demand was outpacing supply for Ciena’s optical-networking chips. His experience of the 2001 telecom crash matters here. A shortage can reveal a durable technical requirement while still producing a temporary pricing cycle. Buyers can respond by securing supply; investors cannot infer that today’s tight market gives every supplier lasting pricing power.

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. Cisco introduced routing systems designed for faster transfers across long optical cables. Links inside a compute system and links across a data-center network impose different engineering demands, but both consume capacity in the optical supply chain.

Closer packaging creates an architectural contest, not a settled winner

Bringing optics nearer a switch or compute package changes what engineers must optimize together: electrical reach, optical power, packaging and the network design around the chip. SemiAnalysis compared conventional DSP-based transceivers with linear-pluggable, near-packaged and co-packaged optics, examining their power budgets and implementation trade-offs. Co-packaged optics is a candidate for closer integration, not an announced replacement for every pluggable module.

A laser maker that can manufacture at scale still needs to qualify its products within several customers’ packaging and networking designs. Different architectures can preserve demand for different components. They can also shift which firm controls the specifications a supplier must meet. The commercial boundary follows the engineering interface.

Capital is shortening the distance between platforms and suppliers

In March 2026, Nvidia said its planned investments in Coherent and Lumentum would support U.S. research, development and manufacturing. Nvidia did not specify their instruments or conditions or confirm that either transaction had closed. It separately invested $2 billion in Marvell and said the companies planned to work on silicon photonics.

In November, Marvell agreed to acquire Lumentum in a Reuters-reported $115 billion all-stock transaction including debt, as Marvell sought to compete with Broadcom. Lumentum supplies products including indium-phosphide lasers, transceivers and optical circuit switches used to move data between AI chips. The agreement is not a completed acquisition. It is also a different mechanism from Nvidia’s planned supplier investments: Marvell would bring component capabilities inside a semiconductor and networking company, subject to the deal’s completion.

Nvidia’s activity sits within a broader shift toward financing AI hardware capacity rather than relying solely on arm’s-length orders. Such capital can help a platform owner secure development attention and future supply. It does not, by itself, show that the owner has exclusive access or that the supplier has improved its margins.

Coherent has strategic access, but its returns are unmeasured

Nvidia assigned Coherent the same planned $2 billion amount as Lumentum. That places Coherent alongside a direct optical peer in a capacity decision by a major AI-system owner. Marvell’s proposed purchase of Lumentum gives the comparison sharper stakes: one supplier could become part of a broader chip-and-network platform, while Coherent remains an independent supplier.

Independence offers a plausible advantage if Coherent can qualify components across competing customers and architectures. It also leaves Coherent responsible for funding and executing its own manufacturing and co-design work. The supplied record contains no comparable periods of Coherent AI revenue mix, backlog, pricing, margins or capacity utilization. Nvidia’s announcement establishes strategic interest; it does not establish that Coherent has already converted that interest into durable earnings.

Scarcity will reward the interface that survives competition

Public markets have priced optical exposure aggressively: Bloomberg reported that AI demand pushed Chinese optical-device maker Zhongji Innolight’s market capitalization close to $146 billion. Supplier selection has a policy dimension too. In August, Reuters reported that the FCC was drafting restrictions on imports of new Chinese data-center optical-transceiver models. A proposed restriction is not an enacted one, but it gives buyers another reason to examine where their links can be sourced.

Neither a high valuation nor a restricted supply pool settles which supplier will own the difficult work after today’s shortages ease. If independent makers can produce qualified lasers and packages for multiple chip and network designs, neutrality remains an asset customers can buy. If qualification and capacity become inseparable from a platform owner’s design, that owner captures more of the value. Coherent must show it can keep competing clusters connected at scale so their accelerators spend less time waiting for data.