The Photonic Revolution in the Package: An Analysis of Co-Packaged Optics for Next-Generation AI Data Centers

Section 1: Executive Summary

The relentless expansion of data generation and processing, catalyzed by the exponential growth of artificial intelligence (AI), machine learning (ML), and hyperscale cloud computing, has pushed conventional data center interconnect technologies to their fundamental physical limits. The traditional architecture, reliant on pluggable optical transceivers connected to switch Application-Specific Integrated Circuits (ASICs) via long copper traces on a printed circuit board (PCB), is encountering insurmountable barriers in power consumption, bandwidth density, and signal integrity. This report presents a comprehensive analysis of Co-Packaged Optics (CPO), an emerging architectural paradigm poised to resolve this impending interconnect bottleneck. CPO represents a fundamental shift, moving beyond incremental improvements to a revolutionary re-engineering of the relationship between silicon and light.

The central thesis of this analysis is that CPO—the heterogeneous integration of Photonic Integrated Circuits (PICs) and electronic ICs within a single package—is an inevitable, albeit challenging, technological transition. It is not merely a “better transceiver” but a necessary architectural response to the breakdown in the historical scaling relationship between compute performance and I/O capability. By drastically shortening the electrical path between the ASIC and the optical conversion engine, CPO directly addresses the physics of diminishing returns associated with high-speed electrical signaling.

This report quantifies the primary benefits of CPO, which include a dramatic reduction in interconnect power consumption by 30-50% or more, enabling a virtuous cycle of lower operational costs and reduced capital expenditure on data center cooling infrastructure.1 Furthermore, CPO unlocks unprecedented bandwidth densities exceeding 1 Tbps per millimeter of silicon edge, a critical enabler for the massive I/O requirements of next-generation switch ASICs and AI accelerators.2 These advantages in power and density, combined with lower latency, position CPO as a foundational technology for the AI “factories” of the future.

However, the path to widespread adoption is fraught with significant hurdles. This analysis provides a detailed assessment of these challenges, chief among them being thermal management. The co-location of heat-generating ASICs with temperature-sensitive photonic components creates a complex thermal crosstalk problem that demands advanced cooling solutions.4 Equally critical are the operational challenges related to serviceability; the integrated nature of CPO conflicts with the established “hot-swappable” maintenance model of pluggable optics, posing a significant barrier for data center operators.2 Manufacturing complexity, fiber attachment precision, and the nascent state of ecosystem standardization further compound the difficulties.

The competitive landscape is evolving rapidly, with CPO positioned against the incumbent pluggable optics and emerging alternatives like Linear Pluggable Optics (LPO), which offers an intermediate step by preserving the pluggable form factor while reducing power. The market is currently being shaped by two distinct strategic approaches: the vertically integrated, system-level model pursued by compute-centric companies like NVIDIA, and the merchant silicon model championed by component suppliers like Broadcom, which aims to enable a broader ecosystem.3

This report concludes with a strategic outlook, projecting a phased adoption timeline beginning with niche, high-performance AI applications and gradually expanding as the technology matures and the ecosystem solidifies. The transformative potential of CPO extends beyond the component level, enabling flatter network topologies and resource disaggregation that will fundamentally reshape next-generation data center architecture. For stakeholders across the value chain—from data center operators and hardware vendors to investors—understanding the nuanced interplay of CPO’s profound benefits and formidable challenges is paramount for navigating the next decade of data center evolution.

Section 2: The Interconnect Bottleneck: The Catalyst for a Paradigm Shift

The emergence of Co-Packaged Optics is not an isolated innovation but a direct and necessary response to a confluence of physical limitations and application-driven demands that are straining the traditional data center interconnect paradigm to its breaking point. The decades-long reliance on copper-based electrical signaling over PCBs is colliding with the laws of physics, while the explosive growth of artificial intelligence has fundamentally altered the nature and volume of data traffic, creating an urgent need for a new architectural approach.

The Physics of Diminishing Returns

For decades, the data center industry has scaled network bandwidth by increasing the data rate of electrical signals traveling from a central switch ASIC to pluggable optical modules at the faceplate of a switch or server. However, this approach is now encountering a wall of diminishing returns governed by fundamental electromagnetic principles.7 As data rates per electrical lane have increased from 25 Gbps to 50 Gbps, 100 Gbps, and now toward 200 Gbps, the physical properties of the copper traces on standard PCB materials like FR-4 introduce severe signal degradation, a phenomenon known as insertion loss.8

This loss increases dramatically with both frequency (data rate) and distance. To compensate for this degradation and ensure a clean signal reaches the optical module, complex and power-hungry Digital Signal Processors (DSPs) are required at both ends of the link.3 At speeds of 100 Gbps per lane and beyond, the power consumed by these DSPs and the SerDes (Serializer/Deserializer) drivers can become a substantial portion of the entire system’s power budget, in some cases threatening to exceed the power of the core switch logic itself.9 The power needed to transmit electrical signals becomes prohibitive at rates approaching 200 Gbps, marking a clear tipping point where the copper-based approach is no longer sustainable.8

This creates a fundamental imbalance. While Moore’s Law has continued to drive exponential increases in the processing capability of silicon ASICs, the ability to move data on and off these chips via electrical I/O over PCBs has not scaled at a commensurate rate.5 The result is a growing “I/O bottleneck,” where immensely powerful processors are often left idle, starved for data because the interconnect cannot keep pace.5 CPO is therefore not simply an incremental improvement in transceiver efficiency; it is a fundamental architectural re-engineering designed to circumvent the physical limitations of the PCB by moving the electrical-to-optical conversion point from the faceplate to within millimeters of the ASIC die. This strategic relocation shortens the high-speed electrical path, drastically reducing insertion loss and the corresponding power required for signal compensation, thereby re-aligning I/O scaling with compute scaling.

The AI Data Deluge

Compounding these physical limitations is a profound shift in data center workloads, driven primarily by the rise of artificial intelligence and machine learning. Traditional enterprise and cloud workloads were characterized by “north-south” traffic, where data flows primarily between users or other data centers and the servers within. In contrast, modern AI workloads, particularly the training of Large Language Models (LLMs), are dominated by “east-west” traffic.7 This involves massive, continuous, and low-latency communication among thousands of interconnected GPUs or custom AI accelerators working in parallel on a single task.2

In this context, the network is no longer just a means of connecting servers; it becomes the fabric of a single, massive, disaggregated supercomputer.10 The performance of the entire AI cluster is not solely dependent on the processing power of individual GPUs but is critically limited by the bandwidth and latency of the interconnect fabric that binds them together.6 This elevates the interconnect from a peripheral networking component to a core element of the computing architecture itself, as crucial as the processor’s clock speed or memory bandwidth.

This causal link explains the deep investment in CPO by compute-focused companies like NVIDIA. They recognize that the value of their next-generation, more powerful GPUs is contingent on solving the interconnect problem. Without a scalable, power-efficient way to connect thousands of these accelerators into a cohesive system, the full potential of the silicon cannot be realized. CPO is thus positioned as a critical enabling technology for the massive “AI factories” required to train and deploy the increasingly complex models of the future.6

The Inadequacy of Incumbent Solutions

The dominant interconnect technology for the past two decades has been the pluggable optical transceiver. These modules, available in standardized form factors like QSFP-DD and OSFP, have been instrumental in the growth of data centers, offering modularity, interoperability, and field serviceability.14 However, this very architecture is now becoming a primary constraint.

The physical size of pluggable modules limits the number of ports that can be placed on the front panel of a 1RU switch, creating a “faceplate density” bottleneck.5 As switch ASICs scale to capacities of 51.2 Tbps and beyond, it becomes physically impossible to present all of that bandwidth through standard pluggable cages. More fundamentally, the placement of these modules at the faceplate necessitates the long, power-hungry electrical traces from the central ASIC that are at the heart of the power consumption problem.15 While the pluggable architecture has served the industry admirably, its inherent separation of the ASIC and the optics is the very source of the inefficiency that CPO aims to eliminate. For the highest-performance applications, the pluggable model is approaching its practical and economic limits, creating the imperative for a more integrated solution.5

Section 3: Foundational Technologies: Photonic Integrated Circuits (PICs)

At the core of the Co-Packaged Optics revolution is the Photonic Integrated Circuit (PIC), a microchip that represents a paradigm shift from manipulating electrons to manipulating photons—particles of light. A PIC, also known as an integrated optical circuit or planar lightwave circuit, is a device that integrates two or more photonic components onto a single substrate to generate, guide, process, and detect light, thereby forming a functional optical circuit.17 By miniaturizing complex optical systems onto a semiconductor chip, PICs provide the high performance, small footprint, and power efficiency required to make CPO a viable technology.17

Principles of Operation: From Electrons to Photons

Unlike electronic integrated circuits (ICs) that use the flow of electrons to process information, PICs use photons. Information is encoded onto optical signals, typically in the visible or near-infrared spectrum (wavelengths from 850 nm to 1650 nm), and manipulated as it travels through the chip’s circuitry.17 This use of light as the information carrier enables significantly higher speeds and bandwidth compared to electronics, bypassing the traditional limitations of electrical signaling.19 The design and operation of these circuits involve complex physics, governing the behavior of photons and their interaction with various materials to modify a light signal’s frequency, magnitude, and phase.17

Anatomy of a PIC

A PIC is composed of several fundamental building blocks, analogous to the resistors, capacitors, and transistors of an electronic IC. These components are arranged to perform specific functions on the light signal traveling through the circuit.17

Light Sources: The circuit requires a source of coherent light, which is typically provided by a laser. In some PIC platforms, laser diodes can be integrated directly onto the chip as active components.17 In others, light from an external laser is coupled into the PIC.
Waveguides: These are the optical “wires” or interconnects of the PIC. They are structures, often a ridge or channel of high-refractive-index material, that confine and guide light from one component to another with minimal signal loss.17
Modulators: These are the most critical components for data transmission. Modulators encode electrical data onto the optical signal. This is typically achieved by using an electrical field to change the refractive index of a material in the waveguide’s path, which in turn modulates the phase or amplitude of the light passing through it.17 Common modulator designs include the Mach-Zehnder interferometer (MZI) and the highly compact micro-ring resonator.17
Couplers & Splitters: These components manage the flow of light within the circuit. Couplers combine optical signals from two or more input waveguides into a single output, while splitters do the reverse.17 More complex devices, such as an Arrayed Waveguide Grating (AWG), act as wavelength-division multiplexers/demultiplexers, splitting a multi-wavelength signal into its constituent colors or combining multiple single-wavelength signals into one fiber.17
Filters: To manage signals in multi-wavelength systems, filters are used to selectively pass or block specific wavelengths. Interferometric structures like ring resonators are commonly used for this purpose due to their sharp resonant peaks and compact size.17
Photodetectors: At the receiving end of an optical link, photodetectors convert the incoming optical signal back into an electrical signal that can be processed by standard electronic circuits.19

Material Platforms and Their Trade-offs

While electronic ICs are overwhelmingly dominated by silicon, PICs are fabricated from a variety of material systems, with the choice of material representing a fundamental trade-off between integration capability and manufacturing scalability. This choice directly influences the architectural decisions for the final CPO system.17

Indium Phosphide (InP): InP is a compound semiconductor material that possesses a direct bandgap, allowing it to efficiently generate, amplify, and detect light. This makes it the only mature platform capable of monolithically integrating all necessary active components—lasers, amplifiers, modulators, and photodetectors—onto a single chip.17 This high level of integration simplifies system design, but InP fabrication is more specialized and costly than silicon-based processes.
Silicon Photonics (SiPh): This platform leverages the mature and vast infrastructure of Complementary Metal-Oxide-Semiconductor (CMOS) manufacturing used for electronic ICs, enabling high-volume, low-cost production.18 Silicon is an excellent material for passive photonic components like low-loss waveguides, modulators, and filters. However, due to its indirect bandgap, silicon is an extremely inefficient light emitter. This fundamental limitation necessitates a hybrid approach for SiPh-based PICs, where light must be generated by a separate laser chip (typically InP) and coupled into the silicon chip. This can be done via an external laser source (ELS) or by heterogeneously integrating the InP laser die directly onto the silicon photonics wafer or chip.5 The ELS approach simplifies thermal management within the CPO package by moving the hot laser to a remote, cooler location, but it introduces significant optical coupling losses and adds the complexity of a delicate, polarization-maintaining fiber connection.5 The integrated laser approach is more optically efficient but exacerbates the thermal challenges within the package. Thus, the inherent properties of silicon photonics create the central design dilemma facing CPO engineers today.
Silicon Nitride (SiN): SiN is another CMOS-compatible material that offers distinct advantages, most notably ultra-low waveguide propagation loss and a very wide transparent spectral range.17 These properties make it highly suitable for applications requiring long on-chip optical paths, such as certain types of sensors, spectrometers, and quantum computers. However, like silicon, it cannot be used to create active components like lasers or amplifiers.18

Fabrication and Design Flow

The creation of a PIC is a complex, multi-step process that mirrors the sophistication of electronic IC manufacturing. The process begins at the design level, where engineers use specialized electronic/photonic design automation (EPDA) software, such as Synopsys OptoCompiler, to lay out the circuit.20 Foundries provide Process Design Kits (PDKs) that contain pre-characterized models of basic photonic components, allowing for circuit-level and system-level simulation to verify performance before fabrication.20 Once the design is finalized and verified, it is translated into a set of masks. The PICs are then fabricated in a foundry using standard semiconductor manufacturing techniques like photolithography to pattern wafers, followed by processes of etching and material deposition to create the various photonic structures.22 This reliance on established foundry processes, particularly for SiPh, is a key enabler for producing PICs at the scale and cost required for the data center market.

Section 4: Co-Packaged Optics (CPO): Architecture and Integration

Co-Packaged Optics represents a fundamental re-architecting of the interface between high-performance electronics and optical communication links. It is defined as an advanced, heterogeneous integration methodology that brings the optical engine, which contains the Photonic Integrated Circuit (PIC), and the primary electronic IC, such as a switch ASIC, into immediate proximity within a single, unified package mounted on a common substrate.1 The foundational principle of the CPO paradigm is the aggressive minimization of the physical distance over which high-speed electrical signals must travel—from centimeters on a PCB in a traditional pluggable system to mere millimeters within the package substrate.1 This reduction in electrical path length is the key to unlocking significant gains in power efficiency, bandwidth density, and latency.

This technology marks the convergence of two historically separate engineering domains: high-speed semiconductor packaging and optical communications. In the past, switch vendors focused on ASIC and board design, while a distinct industry of optical specialists developed pluggable modules that interfaced at a standardized electrical socket.14 CPO collapses this boundary, creating a new, deeply interdisciplinary challenge where the design of the ASIC, the optical engine, the package substrate, and the system’s thermal solution are inextricably linked and must be holistically co-designed from the outset.5 This convergence requires R&D teams to possess a broad range of expertise spanning digital and analog IC design, silicon photonics, optical physics, advanced materials science, and system-level thermal mechanics, representing a significant barrier to entry and a primary reason for the technology’s measured adoption rate.3

System Architecture Breakdown

A typical CPO-based network switch system is composed of several tightly integrated components, representing a departure from the modular, distributed architecture of its predecessors.1

Central ASIC: At the heart of the package is a large, high-performance ASIC, such as an Ethernet switch chip or an AI accelerator (XPU). This is the primary heat source and the originator/terminator of the high-speed data streams.
Optical Engines (OE) and Electrical Engines (EE): Surrounding the central ASIC on the package substrate are multiple chiplets known as optical engines and electrical engines.23 The OE consists of the PIC, which performs the electro-optical conversion (modulators) and opto-electrical conversion (photodetectors). The EE consists of companion electronic ICs (EICs), such as laser drivers and transimpedance amplifiers (TIAs), which provide the high-speed analog interface between the ASIC’s digital SerDes and the PIC’s optical components. These OE/EE chiplets are the functional equivalent of a traditional optical transceiver, but miniaturized and brought into the package.
Fiber Optic Interface: Instead of connecting to pluggable modules on the faceplate, optical fibers connect directly to the package. This is typically achieved using high-density fiber arrays that are precisely aligned to the optical engines, coupling light into and out of the PICs via structures like edge couplers or grating couplers.17
External Laser Source (ELS): Most first-generation CPO systems employ an external laser source. In this architecture, one or more high-power lasers are housed in a separate, often pluggable, module located elsewhere in the system where cooling is more manageable.5 Light from the ELS is then “piped” into the CPO package and distributed to the various optical engines via specialized, polarization-maintaining optical fiber. This approach decouples the primary heat source (the laser) from the already thermally-congested ASIC package but introduces optical power loss and additional system complexity.14

Heterogeneous Integration Methodologies

The physical assembly of these disparate silicon and photonic components within a single package relies on advanced semiconductor packaging technologies. The choice of integration method represents a critical trade-off between performance, manufacturing complexity, and thermal management.9

2D Integration: This is the most straightforward approach, where the ASIC and the optical engine chiplets are placed side-by-side on a conventional organic package substrate.9 Interconnections are made through the routing layers of the substrate. While relatively simple and cost-effective to manufacture, this method offers the least performance improvement, as the electrical trace lengths on the organic substrate are still significant, leading to higher parasitic inductance and capacitance.9
2.5D Integration: This more advanced technique utilizes an interposer, which is a thin slice of silicon or glass with extremely fine-pitch wiring, placed between the active chips and the main package substrate.11 The ASIC and optical engines are flip-chip bonded onto this interposer, allowing for much shorter, denser, and higher-performance interconnects than are possible on an organic substrate. Connections from the interposer down to the package are made using Through-Silicon Vias (TSVs). This 2.5D approach represents the current “sweet spot” for high-performance CPO, offering a substantial performance gain over 2D integration with manageable manufacturing processes, though large, high-yielding interposers can be costly and have size constraints.9
3D Integration: Representing the ultimate frontier in packaging, 3D integration involves stacking the PICs and EICs vertically.5 This is achieved using technologies like micro-bumping and direct copper-to-copper hybrid bonding, which enable direct vertical connections between the stacked dies.26 This architecture offers the highest possible interconnect density and the lowest electrical parasitics, promising the best performance and smallest footprint. However, it also presents immense challenges. The primary obstacle is thermal management, as stacking a heat-producing EIC directly on top of a temperature-sensitive PIC creates a severe thermal dissipation problem, often referred to as a thermal “nightmare”.4 Furthermore, the manufacturing complexity and yield challenges of 3D stacking are substantial. The industry’s progression along this 2D-to-3D roadmap will be paced by the co-evolution of packaging technologies and innovative thermal management solutions.

Section 5: A Quantitative Leap in Performance: The Benefits of CPO

The architectural shift to Co-Packaged Optics is motivated by a set of compelling and quantifiable performance benefits that directly address the critical bottlenecks facing next-generation data centers. By fundamentally re-engineering the physical relationship between the processing silicon and the optical interface, CPO delivers transformative gains in power efficiency, bandwidth density, latency, and system-level cost.

Power Efficiency Gains

The most significant and widely cited advantage of CPO is its dramatic improvement in power efficiency. In a conventional system with pluggable optics, high-speed electrical signals must traverse many centimeters of lossy copper trace on the PCB, requiring powerful, energy-intensive DSPs and SerDes drivers to maintain signal integrity.3 CPO collapses this electrical path to a few millimeters within the package, drastically reducing the power needed for the electrical I/O. This can lead to a 30-50% reduction in interconnect power consumption compared to traditional DSP-based pluggable optics.1

Leading industry players have reported even more substantial figures. Broadcom and NVIDIA, for example, have claimed power consumption savings of up to 3.5 times over previous architectures.3 A significant portion of this saving comes from the potential to eliminate one or more layers of power-hungry retimers or DSPs that are necessary in pluggable systems to compensate for board losses.5 IBM researchers have found that moving optical connections all the way to the chip can result in an energy consumption reduction from 5 picojoules per bit ($pJ/bit$) down to less than 1 $pJ/bit$.31

These power savings at the component level create a powerful virtuous cycle at the data center facility level. Every watt of power consumed by IT equipment must be removed by the data center’s cooling infrastructure. Due to inefficiencies in power delivery and cooling, characterized by a metric called Power Usage Effectiveness (PUE), a one-watt saving at the chip can translate to a saving of 1.5 to 2 watts at the facility level. The reduced heat generation from CPO-based systems allows for higher compute density within a server rack, meaning more processing power can be installed in the same physical footprint.11 This increased density and lower overall power draw can significantly reduce the capital expenditure (CapEx) required for cooling and power distribution infrastructure in new data center builds, as well as lowering the ongoing operational expenditure (OpEx) related to electricity costs. CPO is therefore a key enabling technology for building more powerful and sustainable data center infrastructure.11

Unprecedented Bandwidth Density

CPO fundamentally overcomes the physical limitations of the standard pluggable module form factor, which restricts the I/O density on a switch’s front panel. By integrating the optical interfaces directly along the perimeter of the ASIC die, CPO enables a massive increase in bandwidth density. This metric, often expressed in terabits per second per millimeter (Tbps/mm) of silicon edge, is a critical measure of a system’s ability to move data on and off a chip.2 CPO architectures have been demonstrated to achieve bandwidth densities greater than 1 Tbps/mm.3 This capability is essential for scaling the total I/O bandwidth of massive switch ASICs and AI accelerators to 51.2 Tbps, 102.4 Tbps, and beyond, a feat that is physically challenging, if not impossible, with faceplate-mounted pluggable optics.

Latency Reduction

For tightly-coupled, high-performance computing (HPC) and AI training clusters, interconnect latency is a critical performance parameter. CPO reduces latency through two primary mechanisms. First, the signal propagation delay is inherently minimized by replacing the long electrical path on the PCB with a much shorter path inside the package.1 Second, the potential elimination of DSPs, which are required in high-speed pluggable modules for signal reconstruction, removes the processing delay associated with these complex chips.5 While the savings may be on the order of nanoseconds, in massive parallel computing systems where thousands of processors are constantly communicating, these small reductions accumulate and can lead to significant improvements in overall application performance and faster model training times.

Cost and System Simplification

While the initial development and manufacturing costs of CPO are high, the technology holds the promise of a significantly lower cost-per-bit at scale. Projections suggest a potential cost reduction of up to 40% compared to pluggable optics.2 This cost advantage stems from several factors. The optical engines can be manufactured using high-volume, wafer-level semiconductor processes, benefiting from economies of scale. Furthermore, CPO eliminates the need for numerous discrete components required in a pluggable-based system, such as the mechanical housing and connectors for the modules, retimer chips on the main board, and potentially the DSPs themselves.2 This component integration also simplifies the design of the main system PCB, which may require fewer layers and less complex routing, further contributing to a reduction in overall system cost.1

Section 6: The Path to Adoption: Overcoming Critical Challenges

Despite the compelling performance advantages of Co-Packaged Optics, its path to widespread adoption is impeded by a formidable set of technical, operational, and ecosystem-level challenges. These hurdles must be systematically addressed before CPO can transition from a niche, high-performance solution to a mainstream data center technology. The primary obstacles revolve around thermal management, reliability and serviceability, manufacturing complexity, and the maturation of a standardized, multi-vendor ecosystem.

Technical Hurdles

Thermal Management

The co-location of a large, high-power ASIC, which can operate at temperatures exceeding 100°C, with highly temperature-sensitive photonic components within a single package creates the most significant technical challenge for CPO: thermal management.2 Photonic devices, particularly lasers and silicon-based modulators, exhibit performance characteristics that are highly dependent on temperature. For instance, the wavelength of a laser and the resonant peak of a ring resonator can drift with temperature changes, potentially causing a link to fail. The heat generated by the ASIC can conduct through the package substrate and create thermal crosstalk, degrading the performance and long-term reliability of the adjacent optical engines.5 This necessitates a holistic, system-level approach to thermal design, often requiring advanced cooling solutions such as direct liquid cooling, which brings its own complexity and cost to the data center environment.6 Sophisticated multiphysics simulation tools are essential during the design phase to accurately model heat flows and develop effective mitigation strategies.4

Fiber Attachment and Packaging

The manufacturing process for CPO involves the precise alignment and robust attachment of hundreds, or even thousands, of individual optical fibers to the optical engines within the package.5 This is an exceptionally demanding task, as the cores of single-mode fibers must be aligned to the on-chip waveguides with sub-micron accuracy to minimize optical coupling loss.8 Achieving this level of precision in a high-volume, high-yield manufacturing environment is a significant challenge. The development of reliable Fiber Array Units (FAUs) and automated, high-throughput alignment and bonding processes is critical for making CPO commercially viable at scale.8

Signal and Power Integrity

Integrating high-speed digital, sensitive analog, and optical components in such close proximity creates a complex electromagnetic environment. Ensuring clean power delivery (power integrity) to all components and preventing electromagnetic interference (EMI) and crosstalk between the massive digital switching fabric of the ASIC and the sensitive analog driver circuits of the optical engines is a non-trivial multi-physics problem.5 This requires sophisticated co-design and co-simulation of the electrical and photonic domains to account for all parasitic effects introduced at the packaging stage and ensure the integrity of the entire system.

Operational and Ecosystem Hurdles

Reliability and Serviceability

Perhaps the most significant business and operational barrier to CPO adoption is the radical departure from the established maintenance model of data centers. Pluggable optical modules are designed to be field-replaceable and hot-swappable. If a module fails, a technician can replace it in minutes with minimal disruption.5 In a CPO system, the optics are permanently integrated into the switch package. A failure in a single optical port could necessitate the replacement of the entire line card or switch chassis—a far more complex, time-consuming, and expensive procedure often referred to as a “rip-and-replace” scenario.2 This lack of serviceability is a major concern for data center operators, who place a premium on uptime and operational simplicity.

This fundamental conflict between the technological optimization offered by CPO and the operational pragmatism of pluggables is a primary driver for a bifurcated market. Vertically integrated companies building massive, homogenous AI clusters, such as NVIDIA, can engineer their systems and operational procedures to absorb the risks of CPO in order to achieve maximum performance for their critical “scale-up” fabrics.25 In contrast, general-purpose cloud and enterprise data centers, which prioritize flexibility, multi-vendor sourcing, and simplified maintenance, are likely to be more cautious, potentially favoring intermediate solutions like LPO that preserve the familiar pluggable model.33

To mitigate this issue, many first-generation CPO designs use an external laser source (ELS). As the laser is often the most failure-prone and heat-sensitive component, moving it to a separate, pluggable module addresses both the serviceability and thermal challenges for that specific component.5 However, this is a pragmatic but imperfect workaround. The ELS architecture introduces its own complexities, including the need for expensive and delicate polarization-maintaining fiber to pipe light into the package, which results in significant optical power loss (up to 50%) and adds multiple new potential points of failure.5 While ELS is a crucial enabler for today’s CPO systems, the long-term industry goal remains the development of highly efficient and thermally robust integrated lasers that can be reliably co-packaged, creating a more truly integrated and efficient system. Advances in materials like quantum dot lasers are a key area of research to achieve this goal.24

Vendor Lock-in and Standardization

The current CPO market is dominated by proprietary, vertically integrated solutions from a small number of large vendors.5 The lack of industry standards for the electrical and optical interfaces between the ASIC and the optical engine chiplets prevents the development of a multi-vendor “mix-and-match” ecosystem.2 Hyperscale data center operators, in particular, strongly resist vendor lock-in, as they rely on a competitive, multi-source supply chain to drive down costs and ensure supply continuity. The maturation of the CPO market will depend heavily on the establishment of standards that can foster a broader, more open ecosystem.

Manufacturing Complexity and Cost

The advanced packaging and testing procedures required for CPO are inherently more complex and costly than those for traditional systems.2 Achieving high manufacturing yields for these complex, multi-chiplet modules is essential to realizing the promised long-term cost-per-bit savings.9 The industry is still in the early stages of scaling these manufacturing processes, and the high initial development costs remain a barrier for many potential players.36

Section 7: The Competitive Landscape: CPO in Context

Co-Packaged Optics does not exist in a vacuum. It is an evolutionary technology entering a complex and dynamic data center interconnect market. Its adoption and ultimate success will be determined by its relative strengths and weaknesses when compared against the entrenched incumbent technology—pluggable optics—as well as other emerging, less disruptive alternatives like Linear Pluggable Optics (LPO). Furthermore, looking beyond networking, the principles of CPO are being extended to even tighter forms of integration, such as In-Package Optical I/O (OIO), for direct chip-to-chip communication. Understanding this competitive context is crucial for assessing the specific applications and timelines for CPO’s deployment.

CPO vs. Pluggable Optics

The comparison between CPO and traditional pluggable optics is a classic case of revolutionary versus evolutionary design, highlighting a fundamental trade-off between peak performance and operational flexibility.

Pluggable Optics: The primary advantages of pluggable transceivers are their unparalleled modularity, field serviceability, and the existence of a mature, multi-vendor ecosystem built around well-defined standards (MSAs).15 This allows data center operators to easily replace failed units, upgrade to new technologies, and source components from multiple competing vendors, ensuring price competition and supply chain resilience.5 However, as previously detailed, their architecture is the source of the power consumption and bandwidth density bottlenecks that CPO is designed to solve.15
Co-Packaged Optics: CPO offers objectively superior performance in the critical metrics of power consumption, bandwidth density, and latency.37 By integrating the optics and electronics, it provides a technically optimized solution for the most demanding interconnect challenges. However, this performance comes at the cost of operational rigidity. The lack of serviceability and the current proprietary nature of CPO systems are significant drawbacks for many deployment scenarios.15

CPO vs. Linear Pluggable Optics (LPO)

Linear Pluggable Optics has emerged as a clever and pragmatic intermediate step between traditional pluggables and fully integrated CPO. LPO seeks to capture a significant portion of the power and latency benefits of CPO while retaining the crucial pluggable form factor.

LPO Architecture: LPO modules are physically identical to traditional pluggable transceivers but are internally simplified by removing the power-hungry DSP chip.33 The task of compensating for signal impairments is shifted from the module to the host switch ASIC, which must be equipped with more powerful SerDes and advanced signal processing capabilities.33
Pros and Cons: The key advantage of LPO is that it reduces module power consumption by approximately 50% and eliminates DSP-induced latency, all while preserving the hot-swappable, field-serviceable nature of pluggables.30 This makes it a much less disruptive upgrade path for data center operators. However, LPO is not a universal solution. It is primarily suited for shorter-reach applications (typically up to a few hundred meters) where signal degradation is less severe.33 Furthermore, it breaks the “plug-and-play” interoperability of traditional modules, as it requires tight co-design and validation between the LPO module vendor and the switch ASIC vendor to ensure the link will function correctly.30

In-Package Optical I/O (OIO)

While CPO is primarily focused on networking applications (connecting switches to servers or other switches), the next logical step in integration is In-Package Optical I/O (OIO), also referred to as on-package optical I/O. This technology applies the same principles of co-packaging to enable direct, ultra-high-bandwidth, die-to-die communication within a single server or compute node.15

OIO Architecture: Championed by companies like Ayar Labs, OIO involves integrating optical I/O chiplets directly with processors (CPUs, GPUs) or memory.25 The goal is to replace the electrical traces on a PCB or package substrate that connect these components with optical links. This can enable massive increases in chip-to-chip bandwidth, which is a critical bottleneck in modern AI accelerators that need to access large pools of high-bandwidth memory (HBM).5
CPO vs. OIO: CPO is an evolution of the network transceiver, designed for “scale-out” networking between systems. OIO is a revolution in computer architecture, designed for “scale-up” connectivity within a system. OIO promises even greater bandwidth density and power efficiency than CPO but is an even more deeply integrated and proprietary technology.15

Comparative Technology Matrix

The following table provides a consolidated comparison of these interconnect technologies across key architectural and performance metrics, offering a framework for understanding their respective roles in the evolving data center landscape.

Feature	Traditional Pluggable Optics	Linear Pluggable Optics (LPO)	Co-Packaged Optics (CPO)	In-Package Optical I/O (OIO)
Architecture	Modular, external transceiver	DSP-less pluggable module	Optics integrated with switch/XPU on-package	Optical chiplets for die-to-die interconnect
Primary Application	General-purpose data center networking (Scale-out)	Short-reach, latency-sensitive networking	Ultra-high-performance AI/HPC fabrics (Scale-out & Scale-up)	Chip-to-chip, memory disaggregation (Scale-up)
Power Consumption	High (DSP-intensive)	Medium (~50% less than DSP-based)	Low (system-level optimization)	Lowest (shortest reach, no networking overhead)
Latency	High (DSP processing delay)	Low (no module DSP)	Lower (shorter electrical path)	Lowest (direct die-to-die link)
Bandwidth Density	Limited by faceplate	Limited by faceplate	High (Tbps/mm at package edge)	Highest (Tbps/mm² area density)
Serviceability	Excellent (hot-swappable)	Excellent (hot-swappable)	Poor (entire unit replacement)	N/A (integrated into compute SoC)
Ecosystem Maturity	Fully mature, standardized	Emerging, requires co-design	Nascent, largely proprietary	Highly proprietary, emerging standards (UCIe-O)

This matrix clearly illustrates the fundamental trade-offs at play. As the level of integration increases from left to right, performance metrics like power efficiency and bandwidth density improve dramatically. However, this comes at the direct expense of operational flexibility, serviceability, and ecosystem maturity. This clarifies why there is no single “best” solution; the optimal choice is highly dependent on the specific application’s priorities, whether they be the raw performance required for an AI supercomputer (favoring CPO/OIO) or the operational simplicity and flexibility needed for a general-purpose cloud data center (favoring Pluggables/LPO).

Section 8: The CPO Ecosystem: Key Innovators and Their Contributions

The development and deployment of Co-Packaged Optics are being driven by a diverse and dynamic ecosystem of companies, from semiconductor behemoths and networking leaders to specialized startups and foundational technology providers. The market is currently characterized by a competitive tension between two primary strategic models: the vertically integrated, system-level approach and the merchant silicon, component-level approach. The interplay between these players will shape the technology’s roadmap and the pace of its adoption.

Semiconductor and Switch Giants

These are the large, vertically integrated companies with the deep R&D budgets and broad expertise required to tackle the interdisciplinary challenges of CPO. They are the primary drivers of the technology today.

Broadcom: As a leading merchant silicon vendor for network switches, Broadcom has been a key pioneer in CPO. Their Tomahawk switch series and “Bailly” CPO platform are designed to provide hyperscale data center operators with high-density, power-efficient Ethernet switching solutions.2 Broadcom’s strategy is to sell CPO-enabled switch ASICs and reference platforms to a wide range of system builders (such as Arista, Cisco, and others), aiming to enable and foster a broader ecosystem around their technology. They have announced platforms with 51.2 Tbps switching capacity, claiming significant reductions in power and cost-per-bit.3
NVIDIA: A dominant force in AI computing, NVIDIA is aggressively pushing CPO as an essential component of its end-to-end AI data center solutions. Their CPO technology is integrated into their Quantum-X (InfiniBand) and Spectrum-X (Ethernet) platforms, which are designed to create the massive, scalable fabrics needed for their “AI factories”.6 Unlike Broadcom, NVIDIA’s approach is highly vertically integrated; their CPO solutions are tightly coupled with their GPUs, software stack (e.g., NVLink), and system designs. Their goal is not to sell CPO components, but to sell complete, optimized AI supercomputing clusters where CPO is a critical enabling technology. This “walled garden” approach allows for maximum system-level optimization at the expense of open interoperability.28
Intel: Leveraging its decades of expertise in silicon manufacturing, advanced packaging, and its pioneering work in silicon photonics, Intel is another key player.24 The company is pursuing a strategy that involves both developing its own CPO-enabled products and collaborating with key ecosystem partners. A notable example is their strategic partnership with foundry giant TSMC to develop advanced CPO solutions that combine Intel’s silicon photonics technology with TSMC’s advanced packaging capabilities, targeting the needs of hyperscale data centers.38

This competitive dynamic between the “Android model” of Broadcom (enabling an ecosystem) and the “Apple model” of NVIDIA (a closed, optimized system) will be a defining feature of the CPO market’s evolution. The long-term success of CPO in the broader data center market will depend on whether the open ecosystem can match the performance and integration advantages of the proprietary, vertically integrated solutions.

Networking Leaders

Traditional networking equipment manufacturers are navigating the transition to CPO, balancing the need to innovate and meet the demands of their most advanced customers with the need to support their large enterprise base that values stability, interoperability, and backward compatibility.

Cisco Systems, Arista Networks, Juniper Networks: These companies are the primary customers for merchant silicon like Broadcom’s Tomahawk series. They are actively developing their own CPO-enabled switch platforms.38 For instance, Arista has introduced switches with native CPO support delivering 28.8 Tbps of capacity.38 Cisco has been strengthening its CPO capabilities through strategic acquisitions, such as Luxtera, to bring more optical expertise in-house.38 Their challenge is to integrate this new, complex technology while maintaining the reliability and operational simplicity their customers expect.

Specialists and Enablers

The CPO ecosystem is supported by a range of other critical players who provide essential components, technologies, and manufacturing services.

Marvell: A key competitor to Broadcom in the switch ASIC space, Marvell is also developing ASICs, such as its Prestera DX series, that are optimized for CPO integration. They provide the advanced signal integrity and thermal management features necessary to support 800G and 1.6T optical connectivity in CPO-based systems.38
Ayar Labs: This company is a leading proponent of In-Package Optical I/O (OIO), representing the next frontier of integration. While their primary focus is on chip-to-chip optical interconnects rather than networking, their TeraPHY™ optical engine chiplets are a key example of the technology that could eventually merge with or succeed CPO for certain applications.25
Foundries (e.g., TSMC): Semiconductor foundries are the bedrock of the CPO ecosystem. Advanced packaging technologies, such as TSMC’s Chip-on-Wafer-on-Substrate (CoWoS), are the foundational manufacturing processes that enable the 2.5D integration of ASICs and optical engines.28 These foundries are developing detailed roadmaps for optical engines and advanced 3D bonding techniques that will be critical for future generations of CPO and OIO products.28

Section 9: Strategic Outlook and Recommendations

The trajectory of Co-Packaged Optics is set to be one of the most consequential technological shifts in data center hardware over the next decade. While the immediate path is laden with challenges, the long-term drivers are compelling, suggesting a future where the tight integration of photonics and electronics becomes standard for high-performance systems. This concluding section synthesizes market forecasts and technology trends to project a likely adoption roadmap, explores the transformative architectural impacts, and provides actionable recommendations for key stakeholders.

Technology Roadmap and Adoption Timeline

The adoption of CPO will not be a sudden, market-wide replacement of pluggable optics but rather a phased progression, driven by the cadence of network speed upgrades and the escalating demands of specific applications.

Near-Term (2024-2026): CPO adoption will remain nascent and focused on niche, high-performance applications where the benefits of power and density outweigh the risks of a new technology. The primary drivers will be massive “scale-up” AI training clusters, where CPO provides a superior alternative to copper for dense, short-reach interconnects between accelerators.35 These initial deployments will be largely proprietary and vertically integrated, led by players like NVIDIA and hyperscalers with the scale and technical expertise to manage the integration risks.35 The market during this phase will be small but strategically significant, proving the technology’s viability.
Mid-Term (2027-2030): As the industry transitions to switch generations of 102.4 Tbps and beyond, enabled by 200 Gbps-per-lane SerDes technology, the power and density limitations of pluggable optics will become increasingly acute. At these speeds, the signal integrity challenges for electrical traces may become so severe that CPO transitions from a performance advantage to a technical necessity for “scale-out” networking.29 During this period, broader adoption is expected, contingent on the maturation of the manufacturing ecosystem and the emergence of initial industry standards to foster multi-vendor interoperability. The market is projected to grow significantly, with global sales of CPO ports potentially reaching 4.5 million by 2027.30
Long-Term (2030 and beyond): CPO is expected to become a mainstream technology for high-end data center switches and compute interconnects. The market is forecast to exceed $1.2 billion by 2035, growing at a compound annual growth rate (CAGR) of nearly 29% from 2025.28 The technology itself will continue to evolve, with a likely progression towards more advanced 3D integration and the eventual on-package integration of laser sources, moving away from the interim ELS architecture.

Transformative Impact on Data Center Architecture

The widespread adoption of CPO will have profound, second- and third-order effects on how data centers are designed and operated, enabling new architectural possibilities.

Flattened Network Topologies: The immense bandwidth density of CPO enables the design of switches with a much higher “radix” (i.e., a greater number of high-speed ports).37 High-radix switches can support flatter network topologies, such as single-tier “spine” networks, which reduce the number of hops a data packet must take to traverse the data center. This directly translates to lower overall network latency and simplified network management.
Resource Disaggregation: CPO is a key enabler for the concept of resource disaggregation. In this architecture, the traditional monolithic server is broken down into independent, network-addressable pools of resources—compute (CPUs/GPUs), memory, and storage. These pools can then be interconnected via a high-bandwidth, low-latency optical fabric enabled by CPO.10 This allows for the dynamic composition of “virtual servers” tailored to specific workloads, dramatically improving resource utilization, as components are no longer stranded within underutilized physical servers.10

Recommendations for Stakeholders

Navigating the transition to CPO requires a strategic and forward-looking approach from all participants in the data center ecosystem.

For Data Center Operators:

Begin TCO Analysis Now: Operators should move beyond comparing the upfront cost of CPO versus pluggables and develop comprehensive Total Cost of Ownership (TCO) models. These models must factor in the significant OpEx savings from reduced power and cooling against the potential increase in maintenance costs and complexity associated with the “rip-and-replace” service model.
Drive Standardization: Actively participate in and support industry forums and standards bodies. A strong, collective customer voice is essential to push vendors towards open, interoperable standards that will prevent vendor lock-in and foster a healthy, competitive ecosystem.
Pilot and Learn: For hyperscale operators, initiating pilot deployments in non-critical environments can provide invaluable hands-on experience with the operational challenges of CPO, informing future large-scale deployment strategies.

For Hardware Vendors (OEMs and ODMs):

Invest in Interdisciplinary Talent: The convergence of silicon and photonics requires breaking down traditional R&D silos. Building and retaining teams with deep, cross-functional expertise in electronics, photonics, packaging, and thermal engineering will be a key competitive differentiator.
Innovate on Serviceability: The serviceability problem is the single greatest barrier to CPO adoption in the mainstream market. Vendors who develop innovative solutions—be it more reliable components, modular CPO designs, or advanced diagnostics that can predict failures—will have a significant advantage.
Forge Deep Ecosystem Partnerships: Success in the CPO era will require closer collaboration than ever before across the supply chain, from ASIC designers and photonics specialists to packaging houses and cooling solution providers.

For Investors:

Adopt a Long-Term Perspective: Recognize that CPO is not an overnight revolution but a long-term architectural transition. The path will be incremental, with significant revenue growth likely back-loaded toward the end of the decade.
Identify Key Technology Inflection Points: The companies that provide breakthrough solutions to the key technical challenges will be the long-term winners. Key indicators to monitor include advances in thermally robust integrated lasers (e.g., quantum dot lasers), high-yield advanced packaging techniques, and novel micro-cooling technologies suitable for 3D-stacked architectures.
Look Beyond the Obvious Players: While semiconductor giants currently lead the charge, the CPO transition will create opportunities for a wide range of companies, including those in specialized materials, manufacturing and test equipment, and advanced simulation software. A diversified investment strategy that captures value across the enabling ecosystem is prudent.

Cutting-edge Technology Courses by Uplatz