Marvell’s $115 Billion Bet on the Space Between AI Chips

A giant deal for a quieter bottleneck

An AI processor can be extraordinarily fast and still spend too much of its time waiting. Once a model is spread across many chips, the connections between them help determine how much useful work the whole system can deliver. That makes an apparently unglamorous part of the AI supply chain worth watching. On November 2, Marvell announced an agreement to acquire optical-components company Lumentum in an all-stock transaction with an implied enterprise value of $115 billion. The proposed combination would bring more of the electronics that handle data and the optical hardware that carries it under one roof. The companies’ announcement describes a business spanning connections between data centers through to links inside advanced semiconductor packages.

The size of the deal invites a familiar argument about whether AI infrastructure has become too expensive. There is also a more useful technical question: how much performance is being left between the processors, and who can recover it? Marvell is betting that customers increasingly need electrical and optical components designed together, rather than a collection of individually impressive parts. That proposition has implications well beyond this transaction. It concerns the shape of future AI machines, the suppliers capable of building them, and the practical work needed to keep them running. A promising architecture, however, does not by itself establish that an acquisition will earn its price.

Why faster chips create a networking problem

Large AI workloads divide work among processors. Those processors then need to exchange information, sometimes frequently enough that communication becomes a constraint on the computation. NVIDIA’s NCCL documentation, for example, describes a library devoted to operations such as gathering, distributing and combining data across GPUs. These operations are part of making many accelerators cooperate, not an optional convenience added after the real work is finished.

It helps to distinguish two networking jobs. Scale-up links connect accelerators into a tightly coupled computing domain. Scale-out networks connect systems across a larger cluster. NVIDIA’s rack-system documentation illustrates the distinction with NVLink inside a rack and a separate compute network between racks. Other architectures draw the physical boundaries differently, but the distinction between tightly coupled computation and a broader cluster remains useful.

Consider a simplified example: a group of processors finishes its assigned calculations, but the next step depends on data arriving from elsewhere. Faster arithmetic does not eliminate that wait. The relevant questions include how much data the links can carry, how long each transfer takes, and how consistently the network behaves under load.

This is why counting installed chips can give an incomplete picture of an AI system. The operational goal is to finish useful work within a power and cost budget. A networking improvement can help achieve that without making any individual processor faster. Conversely, a poorly matched network can limit the value of a very expensive processor upgrade.

Optics is moving closer to the processor

Fiber already carries data through networks. The interesting change is where the system converts electrical signals into light and back again. In a conventional pluggable design, optical transceivers sit at the equipment’s edge, with electrical signals traveling between them and the main chip. Co-packaged optics brings optical engines much closer to that chip, inside the same package. Shortening the electrical path can reduce the burden of moving high-speed signals across a board.

NVIDIA’s technical explanation of silicon-photonics switches describes integration at the packaging level and the removal of separate signal-processing components used in traditional designs. The important point is architectural: the optical connection becomes something engineers design alongside the chip and package, rather than treat as a distant attachment. That changes the opportunities for optimization and the practical constraints on assembly.

It does not follow that every copper connection is suddenly obsolete. Distance, speed, power, cost and serviceability still shape the choice. Broadcom’s Tomahawk 6 announcement explicitly supports both passive copper connections and a co-packaged-optics option. A sensible system can use different approaches for different links.

Nor should a power improvement measured at an optical port be read as the same improvement for an entire data center. Accelerators, memory, cooling and other equipment remain in the energy budget. The useful comparison is a complete system doing the same work, with the same reliability requirements. Moving optics inward creates an opportunity to improve that result; it does not make the result automatic.

What Lumentum adds to Marvell

The companies’ investor presentation shows how the proposed combination joins different kinds of engineering expertise. Marvell brings interconnect electronics, switching and custom-silicon capabilities. Lumentum contributes lasers, optical transceivers, optical circuit switches and specialized semiconductor manufacturing. The companies’ case is that coordinating these elements can produce better integrated products and simplify the customer’s path to deployment. That is a plausible direction, though it remains a plan rather than a demonstrated benefit of the merger.

Marvell has already been moving toward this territory. It completed its acquisition of Celestial AI on February 2, adding the company’s Photonic Fabric optical-interconnect technology. At the time, Marvell said it expected initial revenue contributions in the second half of its fiscal 2028. That forecast is a reminder that acquiring technology and turning it into a substantial shipping business run on different clocks.

Lumentum also supplied evidence of commercial demand on November 2. In its first-quarter fiscal 2027 results, the company reported $1.26 billion in revenue for the quarter ended September 26. It said a second major near-packaged/co-packaged optics win carried purchase commitments exceeding $5 billion. Those commitments should not be confused with revenue already earned, or with an independent guarantee of future deliveries.

The strategic appeal is easier to understand when viewed from the customer’s side. If a link underperforms, finding the cause across several suppliers can be difficult. A more integrated supplier could take responsibility for a larger part of the signal path. The commercial test is whether that reduces engineering time and operating cost enough to outweigh any loss of flexibility.

The hard part is building and servicing it

Integrating optics more closely with expensive silicon creates a demanding manufacturing problem. Components have to work together before the finished package becomes valuable. Assembly quality, testing and repair arrangements therefore belong in the discussion alongside bandwidth. A fast demonstration is encouraging; a repeatable production process is what lets customers build a fleet.

TSMC’s April 2026 technology-symposium announcement described a co-packaged-optics implementation using its COUPE optical engine on a substrate, with production planned to begin in 2026. This provides a useful view of the underlying work: optical integration depends on packaging technology as well as the electronics and light sources. The announcement is evidence of a manufacturing roadmap, not proof that every customer design has reached volume production.

Repairability deserves equal attention. NVIDIA’s description of its co-packaged-optics ecosystem explains why its external laser sources are field-replaceable and thermally separated from the main switch. Keeping the light source modular allows a failed unit to be replaced without disturbing the surrounding switch infrastructure. Its design also addresses temperature variation that can shorten laser life.

That example shows why “more integrated” should not simply mean “everything permanently fused together.” Different components have different service needs. An operator wants the performance advantage of close integration without turning a routine repair into the replacement of a much larger, costly assembly. The winning product has to reconcile both demands. Ownership of more of the supply chain may help engineers do that, but it cannot remove the need to prove reliability in use.

A market already being contested

Marvell and Lumentum would enter this next phase alongside serious competitors and experienced customers. Broadcom announced shipments of its Tomahawk 6–Davisson co-packaged-optics switch in October 2025. NVIDIA has developed photonics-based networking as part of its broader AI-system portfolio. The merger should therefore be understood as an attempt to strengthen a position in an active architectural transition, rather than the arrival of optical networking itself.

Customers have also been doing sophisticated optical work for years. Google’s 2022 Jupiter Evolving research paper described production experience using optical circuit switches and software-defined control to change the connections within its data-center network. The paper reported improvements from a combination of topology, traffic engineering and operational changes. Optical circuit switching is a different function from placing optical engines beside a processor, but it demonstrates how much value can depend on the network’s overall design.

The implication is that customers are unlikely to evaluate this deal solely as a shopping list of components. They will want to know how the combined portfolio fits their architectures, their software and their maintenance practices. A supplier that can reduce integration work has something valuable to offer. A customer that needs interchangeable parts and multiple sourcing options may assign a different value to the same proposal.

Both positions are reasonable. Greater supplier integration can make a system easier to deliver while also concentrating more engineering decisions in one company. Whether that trade is attractive depends on the product, the buyer and the alternatives available when it ships.

The headline price has a long timetable

The transaction is not complete. The companies expect it to close in the first half of calendar 2028, subject to shareholder approvals, regulatory approvals and other customary conditions. Under the merger agreement described in Marvell’s SEC filing, Lumentum common shareholders would receive 4.5 Marvell shares for each Lumentum common share. The $115 billion figure is an implied enterprise value, not a $115 billion cash payment. A fixed exchange ratio also means the value of the shares offered moves with Marvell’s stock.

These distinctions matter because the architecture will continue changing while the deal is pending. Customers will make procurement decisions, competing products will evolve, and each business must continue executing. A merger announcement cannot pause those processes. An integrated product plan needs to remain useful by the time engineers can deliver it at scale.

The investor presentation also projects more than $1 billion in annual run-rate cost synergies. That is a management expectation about future benefits, not an achieved reduction in costs. It is separate from the question of whether customers will see lower prices or better performance. Shareholder economics and customer economics can improve together, but one does not establish the other.

It is therefore possible to find the technical rationale persuasive while remaining unconvinced about the price or execution risk. The optical transition can be real without every acquisition made around it being successful. Keeping those judgments separate makes the deal easier to assess as evidence arrives.

What would make this bet pay off

The most revealing evidence will come from working systems. Customers need to see how much useful throughput an architecture delivers under realistic communication loads, rather than only the maximum speed of an individual link. Comparisons should keep the workload and operating conditions clear. An improvement that appears during a short test may matter less if it disappears when a cluster is busy.

Manufacturing and operations provide the next tests. Can the supplier deliver enough qualified parts on time? Can technicians isolate failures and replace the right components without long interruptions? Does performance remain predictable as the deployment grows? These are less eye-catching measures than a record bandwidth figure, but they are closely connected to the value of the equipment already installed.

The final question is how much freedom the customer retains. An integrated offering could shorten development schedules and reduce the number of interfaces a buyer must manage. Buyers will still need credible upgrade paths, support commitments and a clear understanding of where they depend on a particular supplier. Those details determine whether buying a more complete system makes future changes easier or harder.

Marvell’s proposed acquisition puts an enormous valuation on the effort to move data more effectively. It also offers a useful way to think about the next stage of AI infrastructure. Faster processors create their fullest value when the rest of the machine can keep them supplied and coordinated. The space between the chips is becoming a central engineering problem. The evidence that matters next is how reliably, and at what total cost, suppliers can solve it.

11/03/2026

• AI Chip Design Is Moving Beyond Code Generation
• The AI Spending Scoreboard Is Counting Different Things