Marvell Agrees to Buy Lumentum for $115 Billion in Record Optical Tech Deal

Marvell Technology has agreed to acquire optical components manufacturer Lumentum Holdings in an all-stock transaction valued at an implied enterprise value of $115 billion, uniting two critical pillars of data infrastructure to build an integrated silicon-and-photonics giant. Under the terms announced Monday, Lumentum shareholders will receive 4.500 shares of Marvell common stock for each share of Lumentum they own, representing an equity value of approximately $102.5 billion and an implied price of $1,132 per share. The price represents a 23.9% premium over Lumentum’s unaffected closing price on October 28, before word of negotiations leaked.

The transaction, unanimously approved by both boards of directors, is structured to leave Marvell shareholders owning roughly 67% of the combined entity, while Lumentum shareholders will hold approximately 33% on a fully diluted basis. Matt Murphy will remain chairman and chief executive officer of the merged semiconductor company, with Lumentum CEO Michael Hurlston and two additional Lumentum directors joining an expanded Marvell board. The companies expect the deal to close in the first half of 2028, subject to shareholder approvals, international regulatory scrutiny, and customary closing conditions.

The Marvell Lumentum acquisition marks the largest single transaction in the history of the optical and networking semiconductor industry, surpassing Marvell’s previous record purchase of Inphi for $10 billion in 2021. It comes during a pivotal moment for artificial intelligence infrastructure, where conventional copper wiring is hitting physical limits of speed, distance, and heat, forcing hyperscale operators to replace electrical interconnects with light.

+---------------------------------------------------------------------------------+
|                       MARVELL-LUMENTUM TRANSACTION AT A GLANCE                  |
+---------------------------------------------------------------------------------+
| Enterprise Value            | $115 Billion                                      |
| Equity Value                | $102.54 Billion                                   |
| Consideration               | All-stock (Fixed exchange ratio of 4.500:1)       |
| Implied Per-Share Value     | $1,132 per Lumentum share                         |
| Unaffected Premium          | 23.9% (vs. October 28, 2026 close)                |
| Pro Forma Ownership         | 67% Marvell | 33% Lumentum                        |
| Projected Cost Synergies    | >$1 Billion annual run-rate within 24-36 months   |
| Combined Addressable Market | $550 Billion (projected through early 2030s)      |
| Targeted Closing Window     | First Half of 2028                                |
+---------------------------------------------------------------------------------+

“AI infrastructure is undergoing a fundamental architectural shift toward optical connectivity, creating an opportunity that comes along once in a generation,” Murphy told analysts during a joint conference call. “The companies that move first and help define this transition will set the standard for decades to come. By bringing together Marvell’s advanced semiconductor design, custom compute, and optical digital signal processing with Lumentum’s specialized compound semiconductor manufacturing, high-power lasers, and optical switching, we are building an end-to-end platform capable of powering compute fabrics of unprecedented scale.”

Wall Street’s initial reaction reflected unease over the massive equity dilution and macroeconomic jitters across tech. Shares of Lumentum dropped 18.6% following the announcement, while Marvell tumbled 12.7%, as institutional investors absorbed the scale of share issuance and an extended 18-month regulatory timeline. Despite the market pullback, both leadership teams defended the valuation as a defensive and offensive necessity to counter rival Broadcom and monopolize the physical layer of next-generation AI clusters.

The Physical Wall: Why Copper Is Dying Inside the Data Center

To understand why Marvell is spending $115 billion on an optical components firm, one must look at the physical limitations of sending electricity through copper wire. For decades, modern computing scaled according to Moore’s Law, which allowed chipmakers to double the number of transistors on a die every two years. As processors grew faster, the communication pipes between them—printed circuit board (PCB) copper traces and Direct Attach Copper (DAC) cables—scaled in tandem.

That trajectory has stalled. The culprit is not the compute silicon itself, but high-frequency electromagnetism.

When data travels across an electrical conductor at gigabit speeds, it does not flow smoothly through the center of the wire. Instead, an electromagnetic phenomenon known as the skin effect forces the alternating current to crowd into the outer perimeter of the conductor. As signal speeds rise from 25 gigabits per second (Gbps) to 56 Gbps, 112 Gbps, and now 224 Gbps per lane, the effective cross-sectional area of the wire shrinks dramatically. Resistance spikes, turning electrical energy into waste heat.

ELECTRICAL CHANNEL (COPPER):
[GPU / ASIC] ---> [Retimer] ---> [Thick Copper Cable] ---> [Retimer] ---> [Switch]
   - High attenuation at 224G (>40 dB loss per meter)
   - Heavy thermal dissipation (skin effect, dielectric absorption)
   - Usable passive reach: < 1 meter

OPTICAL CHANNEL (PHOTONICS):
[GPU / ASIC] ---> [Optical DSP] ---> [Laser / Modulator] === Glass Fiber ===> [Photodetector]
   - Attenuation in silica fiber: < 0.2 dB per kilometer
   - Zero electromagnetic crosstalk; near-zero thermal loss along fiber
   - Reach: 100 meters to several kilometers without intermediate retimers

Compounding the skin effect is dielectric loss. The insulating materials that surround copper traces on a circuit board absorb electrical energy, dissipating it before the signal reaches its destination. At 224 Gbps, the attenuation (signal loss over distance, measured in decibels per meter) is so severe that passive copper cables cannot reliably transmit data farther than one meter without unmanageable bit error rates.

Ten years ago, passive copper cables could easily span five to seven meters across an entire data center row, linking servers directly to top-of-rack switches. Today, at 1.6 terabits per second (1.6T) aggregated bandwidth across eight 200G lanes, passive copper can barely stretch across a single equipment rack. Active Electrical Cables (AECs), which embed tiny retimer chips in each connector head to clean and boost degraded signals, extend that reach slightly. Yet AECs introduce heavy thermal penalties, drawing substantial wattage merely to push electrons across two or three meters.

Data center energy budgets cannot sustain this paradigm. An AI supercluster containing 100,000 accelerators operating concurrently on a massive language model spends up to 20% to 30% of its total operational energy simply moving data between memory, processing nodes, and network switches. The energy required to drive an electrical bit through a complex PCB trace and a thick copper cable sits between 15 and 30 picojoules per bit (pJ/bit).

Photons behave entirely differently. When data is encoded as pulses of light traveling through ultra-pure silica glass fiber:

Near-Zero Dielectric Loss: Optical signals experience loss rates of roughly 0.2 dB per kilometer at standard infrared wavelengths (such as 1310 nm or 1550 nm), compared to 40 dB per meter for electrical signals running at 224 Gbps.
No Skin Effect: Photons do not interact electrically with the medium, eliminating high-frequency surface crowding and localized resistive heating.
Immunity to Crosstalk: Tightly packed copper traces suffer from capacitive and inductive coupling, bleeding electrical noise into adjacent lanes. Optical fibers carry light beams in isolation, regardless of physical proximity.
Wavelength Division Multiplexing (WDM): Multiple distinct wavelengths (colors) of laser light can travel simultaneously down a single strand of fiber without interference, multiplying the data capacity of a single physical link by factors of 4, 8, or 16.

Optical transmission can reduce the energy per bit to less than 3 to 5 pJ/bit, and early co-packaged optics architectures target sub-1 pJ/bit metrics. Marvell recognized that without directly controlling optical laser technology, its custom processors and high-speed networking chips would inevitably become trapped behind an insurmountable copper wall.

Anatomy of an Optical Engine: Inside Lumentum’s Laser and Photonics Arsenal

While silicon is an extraordinary material for building digital logic gates, it is physically incapable of emitting light efficiently. Silicon has an indirect bandgap, meaning that when an excited electron falls back to its lower energy state, it cannot shed its energy as a photon without transferring momentum to the crystal lattice, dissipating almost all of its energy as heat (phonons).

To create light, engineers must turn to compound semiconductors—specifically Group III-V materials from the periodic table, such as Indium Phosphide (InP) and Gallium Arsenide (GaAs). These elements feature a direct bandgap, allowing electrons to recombine with holes and release energy directly as coherent photons.

Lumentum is one of the world's few industrial-scale fabricators of these compound semiconductor devices. The acquisition gives Marvell direct ownership of multiple critical optical technologies:

+---------------------+---------------------------------------------------------------+
| COMPONENT           | FUNCTION & STRATEGIC APPLICATION                              |
+---------------------+---------------------------------------------------------------+
| Indium Phosphide    | Direct bandgap semiconductor wafers used to fabricate high-   |
| (InP) Lasers        | speed 1310nm/1550nm lasers for single-mode fiber links.       |
+---------------------+---------------------------------------------------------------+
| Electro-absorption  | Monolithically integrates a Continuous Wave laser with an     |
| Modulated Lasers    | electro-absorption modulator on a single die, reaching        |
| (EMLs)              | 200 Gbps per lane modulation for 800G and 1.6T transceivers.  |
+---------------------+---------------------------------------------------------------+
| Continuous Wave     | Ultra-high-power, low-noise laser sources mounted externally  |
| (CW) Light Sources  | to feed silicon photonic integrated circuits and CPO engines. |
+---------------------+---------------------------------------------------------------+
| Vertical-Cavity     | Surface-emitting lasers built on GaAs for low-cost, short-    |
| Surface-Emitting    | reach (multimode fiber) links inside individual server racks. |
| Lasers (VCSELs)     |                                                               |
+---------------------+---------------------------------------------------------------+
| Optical Circuit     | Directs physical beams of light using 2D/3D MEMS micro-       |
| Switches (OCS)      | mirrors, eliminating electrical conversion at the spine.      |
+---------------------+---------------------------------------------------------------+

Electro-absorption Modulated Lasers (EMLs)

In high-speed single-mode data center networks, directly modulating a laser—turning its electrical bias current on and off at hundreds of gigahertz—causes the output wavelength to fluctuate uncontrollably, an imperfection called "chirp." Chirp smears optical pulses as they traverse fiber, distorting data.

To solve this, Lumentum manufactures Electro-absorption Modulated Lasers (EMLs). An EML integrates two distinct structures onto a single InP die: a continuous-wave distributed feedback (DFB) laser that runs at a constant current and steady temperature, and an integrated electro-absorption modulator (EAM). The EAM acts like an ultrafast optical shutter. When a small reverse voltage is applied to the modulator, the material's bandgap shifts via the quantum-confined Stark effect, absorbing light; when the voltage drops, light passes unhindered. Lumentum’s EMLs operate at modulation speeds of 200 Gbps per lane, enabling compact 1.6-terabit optical transceivers running over eight optical channels.

Continuous Wave (CW) Laser Sources

For advanced Silicon Photonics (SiPh), where optical waveguides and modulators are etched directly into silicon wafers, the light must still be generated externally. This requires high-power Continuous Wave (CW) laser engines. These laser diodes emit an unbroken beam of high-intensity, narrow-linewidth light. The light is coupled via fiber into the silicon chip, where the silicon photonics circuitry splits, modulates, and routes it.

Silicon photonics chips run hot because they are situated right next to power-hungry compute processors. Because laser efficiency plummets as temperatures rise, standard lasers degrade quickly if placed directly onto a 90°C computing die. Lumentum’s high-power CW lasers are engineered to operate reliably as detached, field-replaceable external laser sources (ELS), pumping light into the system while remaining isolated in cooler airflow zones.

Optical Circuit Switching (OCS)

Beyond lasers, Lumentum has pioneered commercial Optical Circuit Switches (OCS), including its flagship R300 platform. Traditional network switches are electrical: they intercept an optical signal from a fiber, convert it into an electrical signal using a photodetector, amplify and retime it, push it through an enormous silicon crossbar switch to inspect packet headers, and finally convert it back into an optical signal using a laser. This process, known as Optical-Electrical-Optical (O-E-O) conversion, burns significant electricity and introduces microsecond latency delays at every network hop.

TRADITIONAL ELECTRICAL SWITCH (O-E-O):
Optical In -> [Photodiode] -> [DSP/SerDes] -> [Switch ASIC] -> [DSP/Driver] -> [Laser] -> Optical Out
* Incurs high power dissipation, latency penalties, and data-rate lock-in.

LUMENTUM OPTICAL CIRCUIT SWITCH (OCS):
Optical In -> [Collimator Array] -> [2D/3D MEMS Mirrors] -> [Collimator Array] -> Optical Out
* Zero O-E-O conversion, sub-microsecond physical switching, bit-rate agnostic.

An OCS abandons O-E-O conversion entirely. Instead, incoming fibers pass light into an array of microscopic, computer-controlled 2D or 3D Micro-Electro-Mechanical Systems (MEMS) mirrors suspended on silicon hinges. By applying precise electrostatic voltages, these MEMS mirrors pivot across multiple axes, bouncing light beams directly from an input fiber to any output fiber through free space.

Because the data never converts into electricity, an OCS consumes a fraction of the power required by electrical packet switches, introduces virtually zero transmission latency, and is entirely agnostic to protocol and data rate. A switch installed for 800G links can route 1.6T or 3.2T data streams without replacing internal components.

The Silicon Brain: Marvell’s Optical DSPs and Custom ASIC Pipeline

While Lumentum manufactures the physical components that produce and steer photons, Marvell designs the complex silicon engines that organize, clean, and drive the electronic data feeding into those optical systems.

At data rates of 100 Gbps or 200 Gbps per electrical lane, digital signals degrade rapidly. By the time high-speed bits travel across a PCB from an AI accelerator to an optical module, they emerge distorted by jitter, inter-symbol interference (ISI), and attenuation. The waveform resembles noisy static rather than sharp binary ones and zeros.

RAW TRANSMISSION:
Analog Jitter / Attenuation ---> [ MARVELL OPTICAL DSP ] ---> Clean Optical Signal
                                  - ADC (Analog-to-Digital)
                                  - FFE / DFE Equalization
                                  - PAM4 Symbol Decoding
                                  - Laser Driver Signaling

This is where Marvell’s Optical Digital Signal Processors (DSPs)—anchored by product lines like the Ara 1.6T PAM4 DSP and Aquila coherent-lite DSP families—become essential. The DSP is a piece of mixed-signal silicon that performs high-speed mathematical reconstruction on distorted analog waves:

Analog-to-Digital Conversion (ADC): Massive arrays of interleaved ADCs sample incoming analog electrical voltages tens of billions of times per second.
Adaptive Equalization: Sophisticated digital filters (Feed-Forward Equalization and Decision-Feedback Equalization) run continuous mathematical algorithms to model and cancel out physical line distortion and echo reflections.
PAM4 Encoding/Decoding: Instead of simple Non-Return-to-Zero (NRZ) signaling—where a high voltage is a "1" and a low voltage is a "0"—modern high-speed links use Pulse Amplitude Modulation 4-Level (PAM4). PAM4 uses four distinct voltage thresholds (00, 01, 10, 11) within each clock cycle, doubling throughput at the exact same baud rate. However, PAM4 slashes the amplitude difference between voltage states by two-thirds, making the signal far more vulnerable to background noise. Marvell’s DSPs provide the forward error correction (FEC) engines needed to recover lost bits.
Modulator Driving: The DSP outputs an equalized, clean signal directly into driver circuits that modulate Lumentum’s EML lasers or silicon photonics modulators.

Marvell has also expanded aggressively into Coherent Optical DSPs through its COLORZ line. Unlike PAM4, which only modulates the intensity of light, coherent optics modulate both the amplitude and the phase of the light wave across two orthogonal polarization states (such as Dual-Polarization Quadrature Phase Shift Keying, or DP-QPSK, and 16-QAM). Coherent transmission extracts orders of magnitude more bandwidth out of a single fiber over long distances, such as campus and data center interconnect (DCI) networks spanning up to 120 kilometers.

+---------------------------------------------------------------------------------+
|                       THE VERTICALLY INTEGRATED VALUE CHAIN                     |
+---------------------------------------------------------------------------------+
|                                 MARVELL (Logic)                                 |
|  - Custom AI Accelerators (XPUs / ASICs)                                        |
|  - Ethernet Switching Silicon (Teralynx 51.2T / 102.4T)                         |
|  - High-Speed SerDes (224G / 448G Electrical Interfaces)                        |
|  - Optical DSPs (Ara PAM4, Aquila Coherent-Lite, COLORZ DCI)                    |
|  - Celestial AI Photonic Fabric Architecture                                    |
+---------------------------------------------------------------------------------+
                                       |
                     Heterogeneous Co-Design & Integration
                                       v
+---------------------------------------------------------------------------------+
|                                LUMENTUM (Optics)                                |
|  - Indium Phosphide (InP) Wafer Fabs (Caswell, UK & North America)              |
|  - High-Power Continuous Wave (CW) Laser Sources                                |
|  - Electro-absorption Modulated Lasers (EMLs - 200G/lane)                       |
|  - Micro-Electro-Mechanical (MEMS) Optical Circuit Switches (OCS R300)          |
|  - Advanced Transceiver Packaging & Subsystems                                  |
+---------------------------------------------------------------------------------+

Beyond merchant DSPs, Marvell is a primary supplier of custom AI ASICs (Application-Specific Integrated Circuits). Hyperscalers like Google and Amazon increasingly design their own proprietary AI accelerators (such as Google’s TPU) to avoid paying high margins for general-purpose GPUs. In August 2026, Marvell landed a landmark supply agreement with Google valued at up to $120 billion through fiscal 2033 to supply custom companion processors and interconnect silicon.

Before this transaction, Marvell lacked internal laser fabrication capabilities. Every optical system it designed required third-party lasers and discrete packaging from outside suppliers. By bringing Lumentum in-house, Marvell can co-design electrical DSP logic, silicon photonics waveguides, and compound semiconductor laser sources simultaneously.

From Pluggables to Co-Packaged Optics: The Architectural Evolution

The computer architecture inside AI clusters is moving through an evolutionary sequence, stepping away from traditional pluggable transceiver modules toward tight physical integration. The Marvell Lumentum acquisition is engineered to capture market share across every step of this migration.

EVOLUTION OF OPTICAL INTEGRATION:

Step 1: Standard Pluggable (Current)
[ Switch ASIC ] ===== Copper PCB Trace (10-12") =====> [ Optical Transceiver (DSP + Laser) ]
* Max flexibility, but high trace loss and energy draw.

Step 2: Linear Pluggable Optics / LPO (Transitional)
[ Switch ASIC with Enhanced SerDes ] ===== Copper PCB (6-8") =====> [ Optical Module (No DSP) ]
* Removes module DSP to save power, but creates strict channel margin limits.

Step 3: Near-Package Optics / NPO (Emerging)
[ Substrate: Switch ASIC  --- short trace (<2") --->  Optical Engine ]
* Shortens electrical path, placing optical modules directly onto the host board.

Step 4: Co-Packaged Optics / CPO (Target Architecture)
+----------------------------------------------------------+
| Multi-Die Substrate (2.5D/3D Advanced Packaging)         |
|   [ Switch ASIC / GPU ] <== Dense Die-to-Die ==> [ SiPh Engine ] <== Fiber |
+----------------------------------------------------------+
                      ^ External CW Laser Source (Lumentum)
* Eliminates long electrical traces entirely; sub-1 pJ/bit efficiency.

1. Pluggable Optical Transceivers (The Current Standard)

In standard data center racks, switches feature metal cages on their front faceplates. Pluggable optical transceivers (such as OSFP or QSFP-DD form factors) slide into these slots. Inside each transceiver module sits a dedicated optical DSP, driver chips, photodetectors, and discrete lasers.

While pluggables offer operational flexibility—if a laser fails, a technician simply pulls out the bad module and inserts a replacement—they force high-speed electrical signals to travel 10 to 12 inches across a switch motherboard, pass through high-loss mechanical connectors, and enter the module. This long copper pathway consumes substantial power before light is ever produced.

2. Linear Pluggable Optics (LPO)

To cut power consumption, data center operators began experimenting with Linear Pluggable Optics. LPO removes the DSP chip from the pluggable module entirely. Instead, the main network switch ASIC uses an ultra-powerful, high-precision Serializer/Deserializer (SerDes) to drive the optical modulators directly over an analog electrical channel.

LPO cuts transceiver power consumption by roughly 20% to 30%. However, without a dedicated DSP inside the module to reconstruct the signal, LPO systems suffer from narrow operating margins, fragile interoperability, and extreme sensitivity to cable lengths and temperature fluctuations.

3. Co-Packaged Optics (CPO)

Co-Packaged Optics bypasses the faceplate transceiver entirely. In a CPO architecture, the silicon photonics engine is moved directly inside the processor package, mounted onto the exact same organic substrate or silicon interposer as the main switch ASIC or AI compute die.

By shrinking the electrical distance between the compute logic and the optical modulators from 300 millimeters down to less than 10 millimeters, CPO eliminates the need for power-hungry electrical SerDes drivers. Communication between the compute engine and the network takes place almost entirely through photons.

CPO introduces a critical engineering problem: laser thermal sensitivity. Silicon compute dies regularly run at temperatures exceeding 85°C to 100°C. If an Indium Phosphide laser diode is co-packaged directly next to an AI processor, the intense heat rapidly degrades its optical output, destabilizes its wavelength, and cuts its operational lifespan.

The industry’s answer is the Detached External Laser Source (ELS). The high-power continuous-wave lasers reside in an external, field-replaceable module mounted in the cool airflow zone of the server rack. Thin glass fibers route unmodulated light into the co-packaged optical engine on the main chip substrate, where silicon modulators slice the beam into data pulses.

Lumentum’s command over high-power, low-noise CW lasers gives Marvell a vital component for scaling CPO architectures. When coupled with Marvell’s $3.5 billion acquisition of optical interposer pioneer Celestial AI in February 2026, Marvell now controls both the photonic fabric that routes light into the package and the laser engines that supply that light.

4. Photonic Fabrics and Memory Disaggregation

The final stage of this architectural evolution extends beyond networking switches into compute-to-memory interconnects. Modern artificial intelligence training clusters are fundamentally choked by memory bandwidth. Large language models require terabytes of weights to be distributed across thousands of accelerators. When a GPU runs out of high-bandwidth memory (HBM) capacity, it must idle while fetching parameters over the network.

By integrating optical engines directly alongside compute dies, optical interconnects enable photonic memory pooling. Instead of routing memory requests through layers of PCIe switches and network protocol stacks, an optical fabric lets hundreds of compute nodes access a shared pool of memory over direct light paths. Latency drops from microseconds to nanoseconds, effectively turning an entire data center row into a single unified computing node.

Economic Architecture: Synergies, Margins, and the $550 Billion Addressable Market

The financial scale of the Marvell Lumentum acquisition reflects the shifting economics of the artificial intelligence supply chain. Historically, optical component manufacturing was viewed by Wall Street as a cyclical, capital-intensive, low-margin hardware business subject to severe price erosion. Semiconductor logic designers like Marvell routinely commanded non-GAAP gross margins above 60%, while optical module builders struggled in the 25% to 35% range.

THE SHIFTING PROFIT POOL:

Historical Model (2015-2022)
[ ASIC Design (Marvell) ]        ---- Gross Margin: 62-65%
         |
[ Merchant Optical Module ]      ---- Gross Margin: 25-32% (Commoditized assembly)

Vertically Integrated Model (Post-Acquisition / AI Era)
+-------------------------------------------------------------------------------+
| Integrated Optical-Electronic Subsystem                                       |
| (ASIC + DSP + InP Laser + Silicon Photonics + Packaging)                      |
|                                                                               |
| Combined Target Gross Margin: 58-62%                                          |
| Non-GAAP Operating Margin Target: >45%                                         |
+-------------------------------------------------------------------------------+
* Controlling internal laser fabrication recaptures upstream module margins.

That economic relationship has flipped. Because optical components have become the definitive physical bottleneck for AI scaling, high-power lasers, coherent optical subsystems, and optical circuit switches now command high pricing power.

Lumentum demonstrated this economic surge in its fiscal first-quarter 2027 earnings, reported alongside the merger announcement:

Quarterly Revenue: $1.26 billion, up 136% year-over-year from $533.8 million.
Adjusted EPS: $4.57 per share, outperforming consensus estimates of $4.21 and up from $1.10 a year earlier.
Non-GAAP Operating Margin: Tracking toward 45% to 46% for the subsequent quarter, approaching traditional semiconductor software and fabless logic margins.
Quarterly Guidance: Fiscal Q2 revenue projected between $1.575 billion and $1.625 billion, far exceeding analyst expectations of $1.48 billion.

+---------------------------------------------------------------------------------+
|                       LUMENTUM STANDALONE FINANCIAL TRAJECTORY                  |
+---------------------------------------------------------------------------------+
| Metric                        | Q1 Fiscal 2026         | Q1 Fiscal 2027 (Reported)
+-------------------------------+------------------------+------------------------+
| Net Revenue                   | $533.8 Million         | $1.26 Billion (+136%)  |
| Non-GAAP Operating Margin     | ~22.4%                 | ~43.5% (Guiding 45-46%)|
| Diluted Non-GAAP EPS          | $1.10                  | $4.57 (+315%)          |
| Primary Growth Engines        | Telecom / Industrial   | EML Lasers, OCS R300,  |
|                               | Transceivers           | CW Light Sources for AI|
+---------------------------------------------------------------------------------+

By acquiring Lumentum, Marvell projects over $1 billion in annual run-rate cost synergies within 24 to 36 months following the deal’s close. These synergies will stem from:

Supply Chain Verticalization: Eliminating third-party laser markups by capturing InP fabrication in-house.
Elimination of Merchant Margins: Marvell previously purchased external components for its optical modules; Lumentum similarly purchased discrete electronic chips. Merging eliminates internal transfer margins.
Consolidated Packaging Infrastructure: Heterogeneous integration of DSPs and optical engines within shared cleanroom and automated testing facilities across North America and Southeast Asia.
Expanded Total Addressable Market (TAM): Marvell projects that the combined corporate market opportunity will expand to $550 billion by the early 2030s. The data center interconnect market alone is forecast to grow at an annualized rate of 65%, reaching $65 billion by 2030.

For Marvell shareholders, the initial 12.7% selloff reflects the mathematics of an all-stock transaction. Issuing hundreds of millions of new shares to finance an implied equity valuation of $102.5 billion introduces immediate equity dilution.

However, Marvell management projects that due to Lumentum’s rapid margin expansion and strong cash flow generation, the transaction will be accretive to Marvell’s non-GAAP earnings per share in the first full year following closing. The long-term upside hinges on whether Marvell can run a vertically integrated business model that captures the full structural value of optical compute links.

The Titan Clash: Broadcom, Nvidia, and the Customer Channel Dilemma

The corporate consolidation behind the transaction redraws competitive boundaries across the semiconductor landscape, intensifying Marvell's rivalry with Broadcom while introducing strategic tension with Nvidia.

+---------------------------------------------------------------------------------+
|                     AI NETWORKING COMPETITIVE LANDSCAPE                         |
+---------------------------------------------------------------------------------+
| Capability               | Broadcom            | Combined Marvell-Lumentum      |
+--------------------------+---------------------+--------------------------------+
| Custom AI Silicon (ASIC) | Dominant (Google,   | Strong (Google, Tier-2 CSPs,   |
|                          | Meta, ByteDance)    | Amazon Custom Silicon)         |
| Merchant Switch Silicon  | Dominant (Tomahawk, | Challenger (Teralynx           |
|                          | Jericho 3AI)        | 51.2T/102.4T platform)         |
| Optical DSPs             | Category Leader     | Category Leader                |
|                          | (Sian, Centauri)    | (Ara, Aquila, COLORZ)          |
| In-House Compound Fabs   | Captive Internal    | Comprehensive InP / GaAs       |
| (InP / GaAs Lasers)      | Laser Capacity      | Merchant & Captive Fabs        |
| Optical Circuit Switching| Limited / Research  | Market Leader (OCS R300,       |
|                          |                     | Google Apollo Partner)         |
| Photonic Interposers     | Strategic Alliances | Proprietary (Celestial AI      |
|                          | (TSMC COUPE)        | Photonic Fabric)               |
+---------------------------------------------------------------------------------+

The Broadcom Playbook

For years, Broadcom CEO Hock Tan executed a disciplined strategy: dominate the most defensible, mission-critical nodes in the data center stack. Broadcom controls the high-end merchant switching silicon market with its Tomahawk and Jericho switch lines and holds commanding market share alongside Marvell in optical DSPs. Broadcom also maintains internal optical component manufacturing capabilities, which it has quietly leveraged to build co-packaged optical switch prototypes.

However, Broadcom historically operated its optical component unit at an arm's-length distance, choosing not to produce full merchant pluggable transceiver modules to avoid directly competing with its primary DSP customers—transceiver makers like Coherent, Innolight, and Eoptolink.

With the Lumentum deal, Marvell is breaking that unwritten industry rule. It is moving into complete, vertically integrated optical subsystem manufacturing. Marvell will now design the digital switch ASIC, engineer the optical DSP, fabricate the InP laser source, manufacture the MEMS optical circuit switch, and assemble the final optical engine.

The Customer Channel Conflict

This complete integration introduces severe channel conflict across the optical networking industry.

Historically, Marvell acted as a neutral merchant supplier of optical DSPs. Companies like Coherent, Cisco, and major Asian module manufacturers bought Marvell’s DSP silicon to build finished transceivers. By purchasing Lumentum, Marvell becomes a direct competitor to its own largest customers.

A transceiver company buying DSPs from Marvell must now weigh whether it is subsidizing an adversary that sells identical end-to-end optical modules to hyperscalers. If those third-party module makers defect and source their DSP silicon exclusively from Broadcom or emerging mixed-signal startups, Marvell risks losing merchant DSP market share.

Marvell management has sought to counter this concern by insisting it will maintain an open-ecosystem merchant strategy, continuing to supply standalone DSPs and raw Lumentum laser components to external module manufacturers. Yet historical precedent in the semiconductor sector suggests that preserving structural neutrality after a vertical acquisition is difficult when internal product groups compete for identical customer sockets.

THE CHANNEL FRICTION:
             +-----------------------+
             | Marvell Semiconductor |
             +-----------------------+
              /                     \
  Supplies DSPs                Acquires 100% of
            /                         \
           v                           v
+----------------------+     +----------------------+
| Independent Module   |     |  Lumentum Subsystems |
| Makers (e.g. Cisco,  |     |   (In-House Optics)  |
| Coherent, Innolight) |     +----------------------+
+----------------------+                 |
           \                             /
            \                           /
      COMPETE FOR THE EXACT SAME HYPERSCALE CONTRACTS
                            v
            +-------------------------------+
            |  Cloud Providers (CSP Titans) |
            +-------------------------------+

The Nvidia Factor

Nvidia occupies a complex position within this transaction. In March 2026, Nvidia signed an agreement to invest $2 billion into Lumentum to secure dedicated domestic United States laser manufacturing capacity, paired with a multibillion-dollar future purchase commitment. Nvidia required guaranteed access to high-power CW laser sources to feed its optical NVLink roadmaps for future GPU architectures.

Now, Nvidia's primary laser supplier will be owned by Marvell—a company aggressively trying to dethrone Nvidia's compute dominance by designing custom AI accelerators for hyperscalers.

Nvidia cannot easily switch suppliers. Laser foundries require years of cleanroom construction, specialized metal-organic chemical vapor deposition (MOCVD) reactors, and proprietary process recipes that cannot be replicated overnight. Nvidia must navigate a reality where its critical optical supply chain runs through a key custom-silicon rival.

Regulatory Scrutiny, Geopolitics, and Fab Dependencies

The unprecedented scale of the transaction ensures it will face strict regulatory reviews across three continents before reaching its targeted close in early 2028. The deal arrives amid intense antitrust oversight of artificial intelligence infrastructure and growing geopolitical competition over semiconductor manufacturing supply chains.

REGULATORY APPROVAL PATHWAYS:
+---------------------------------------------------------------------------------+
| JURISDICTION | REGULATORY BODY | PRIMARY SCRUTINY AREAS                         |
+---------------------------------------------------------------------------------+
| United       | FTC / DOJ       | - Vertical foreclosure of optical DSPs         |
| States       |                 | - Monopolization of high-power laser sources   |
|              |                 | - Protection of Nvidia strategic investment    |
+---------------------------------------------------------------------------------+
| European     | European        | - Channel access for European telecom OEMs     |
| Union        | Commission      | - Impact on competitive transceiver supply     |
+---------------------------------------------------------------------------------+
| China        | SAMR            | - Access to InP laser components for domestic  |
|              |                 |   optical module manufacturers (e.g. InnoLight)|
|              |                 | - Long-term behavioral supply commitments      |
+---------------------------------------------------------------------------------+

The regulatory review process for the Marvell Lumentum acquisition faces a multifaceted international gauntlet:

United States (FTC / DOJ): Regulators will focus on vertical foreclosure. If Marvell controls both the optical DSP and the underlying laser supply, can it starve competing transceiver manufacturers or prioritize its internal custom ASICs over open merchant ecosystems?
European Commission: European regulators will scrutinize access for European telecom infrastructure vendors (such as Nokia and Ericsson) that rely on Lumentum's optical subsystems for telecommunications networks and coherent fiber backbones.
China (SAMR): The State Administration for Market Regulation represents the highest hurdle. China is the global hub for optical transceiver assembly; companies like Innolight, Eoptolink, and H3C rely heavily on Lumentum's laser dies and Marvell's DSPs. SAMR has previously delayed, conditioned, or blocked major cross-border semiconductor deals (including Qualcomm's bid for NXP and Applied Materials' planned purchase of Kokusai Electric). To win approval in Beijing, Marvell will likely have to agree to strict, multi-year behavioral remedies, guaranteeing continued component supply to Chinese optical manufacturers on fair, reasonable, and non-discriminatory (FRAND) terms.

The Fab Realities: Fabless Silicon Meets Foundries of Fire

The merger fundamentally changes Marvell's corporate operating structure.

Throughout its history, Marvell operated as a purely fabless semiconductor firm. It designed complex microchips, wrote software stacks, and outsourced 100% of physical wafer fabrication to external foundries, primarily TSMC in Taiwan. Marvell owned no furnaces, no chemical vapor deposition chambers, and zero cleanrooms.

Lumentum operates on a completely different physical model. Compound semiconductor lasers cannot be manufactured on standard silicon CMOS foundry lines. Indium Phosphide is a delicate, brittle crystalline substrate. Growing the atomic-scale quantum wells required for high-speed lasers demands specialized internal manufacturing facilities.

FAB ARCHITECTURE DICHOTOMY:

MARVELL (Pure Fabless CMOS Model):
[ Chip Design ] ---> Sent to [ TSMC Foundry (Taiwan/US) ] ---> Outsource Assembly & Test

LUMENTUM (Internal Specialty Fab Model):
[ InP Wafer Growth ] ---> [ Internal Fabs (UK / North America) ] ---> [ Assembly (Thailand/Taiwan) ]
* Brittle Group III-V compound wafers; customized epitaxial crystal growth.

Lumentum owns and operates compound semiconductor wafer fabrication facilities in Caswell in the United Kingdom, alongside facilities in North America, backed by packaging operations in Thailand and Taiwan. Lumentum recently completed a fourfold capacity expansion of its primary Indium Phosphide fab lines to satisfy surging demand for AI laser components.

Managing this internal physical manufacturing footprint requires distinct operational disciplines. Marvell must master wafer yields, cleanroom capacity utilization, capital depreciation schedules, and environmental hazardous chemical processing. The deal transforms Marvell from an asset-light chip design firm into a hybrid enterprise balancing software design with compound semiconductor fabrication.

The Next Decade of AI Compute: What to Watch on the Road to 2028

The multi-year runway separating the announcement of the acquisition from its projected completion in the first half of 2028 creates an extended window during which optical technology will shift rapidly. As both engineering teams begin joint planning under interim operating protocols, the industry will watch several key milestones to measure the transaction's real-world impact.

THE ROAD TO 2028: CRITICAL MILESTONES
+-------------+-------------------------------------------------------------------+
| TIMEFRAME   | TECHNICAL & MARKET MILESTONE                                      |
+-------------+-------------------------------------------------------------------+
| 2026 - 2027 | Commercial ramp of 1.6T EML optical transceivers based on         |
|             | 200G-per-lane DSPs in hyperscale clusters.                        |
+-------------+-------------------------------------------------------------------+
| Mid-2027    | First volume shipments of Co-Packaged Optics (CPO) network        |
|             | switches integrating external CW laser sources.                   |
+-------------+-------------------------------------------------------------------+
| Late 2027   | Regulatory decisions from the US FTC, European Commission, and   |
|             | China's SAMR on vertical supply remedies.                         |
+-------------+-------------------------------------------------------------------+
| H1 2028     | Formal closing of the acquisition; integration of Lumentum        |
|             | photonics with Marvell custom silicon and Celestial AI fabric.    |
+-------------+-------------------------------------------------------------------+
| 2028 - 2029 | Commercial debut of 3.2T optical links and fully optical spine    |
|             | data center networks utilizing Dynamic Optical Circuit Switching. |
+-------------+-------------------------------------------------------------------+

The 1.6T to 3.2T Transition

The speed of modern AI networks doubles every two years. Hyperscale operators are currently transitioning from 800-gigabit to 1.6-terabit optical interconnects, powered by 200 Gbps per lane optical DSPs and high-speed EML lasers. By the time this transaction closes in 2028, the industry will already be deploying early 3.2-terabit systems running across 400 Gbps lanes.

At 400 Gbps per lane, traditional copper wiring becomes functionally obsolete inside compute clusters. Every centimeter of interconnect between switches, servers, and storage will rely on optical interfaces. How effectively Marvell and Lumentum align their silicon DSP roadmaps with compound laser capabilities will determine whether their combined architecture becomes the standard for 3.2T links.

The Rise of the All-Optical Spine

Network architectures within mega-data centers are undergoing a structural shift. Pioneered by Google through its proprietary Jupiter and Apollo fabrics, major hyperscalers are systematically replacing electrical spine switches with Optical Circuit Switches.

In a traditional spine-leaf topology, if an AI cluster finishes a training epoch and must synchronize parameters across 50,000 GPUs, network traffic congests at the electrical packet switches. By swapping electrical switches for Lumentum’s MEMS-based OCS, the data center can dynamically alter its physical network topology in real time. If an AI training run requires dedicated, high-bandwidth pipelines between specific server rows, the mirrors tilt to establish a direct light path between those clusters, bypassing intermediate network tiers entirely.

If Marvell pairs its switching silicon directly with Lumentum’s OCS platforms, it can offer an end-to-end dynamically reconfigurable optical data center fabric that Broadcom cannot easily match.

DYNAMIC OCS-RECONFIGURED CLUSTER:
Traditional: [GPUs] ---> [Leaf Switch] ---> [Electrical Spine Switch] ---> [Leaf] ---> [GPUs]
             * High latency, rigid topology, fixed bandwidth.

Dynamic OCS: [GPUs] ---> [Leaf Switch] === (Direct MEMS Mirror Path) ===> [Leaf] ---> [GPUs]
             * Topology dynamically changes based on AI workload communication phases.

Unresolved Strategic Questions

As the acquisition proceeds toward completion, several unresolved issues will determine whether Marvell’s $115 billion gamble succeeds:

Customer Defection and Dual-Sourcing: Hyperscalers maintain strict rules against single-source dependencies. Will Google, Meta, and Microsoft actively fund alternative optical component makers (such as Coherent or Fabrinet) to prevent a combined Marvell-Lumentum from exerting pricing power over high-speed optical links?
Preserving Innovation Under Deal Uncertainty: An 18-month regulatory closing window is a long time in semiconductor engineering. Can Lumentum retain its top photonic material scientists and laser packaging engineers while operating in corporate limbo?
Execution on Heterogeneous Packaging: Merging CMOS silicon, silicon photonics, and Indium Phosphide lasers into unified co-packaged modules requires extraordinary manufacturing precision. Yield losses in complex packaging can erase hardware margins quickly.

By agreeing to purchase Lumentum, Marvell made the defining structural wager of the artificial intelligence buildout. The acquisition recognizes that the computing revolution is no longer constrained merely by how fast transistors can switch inside a processor die.

The battle for computing supremacy has shifted to the spaces between the chips. In that high-stakes arena, electrons are giving way to light, and control over the lasers that generate those photons is now worth $115 billion.