The Optical Reality Check: Why Near-Packaged Optics and Circuit Switching Are Defeating CPO to Capture the AI Networking Boom
The Mathematics of the Co-Packaged Optics Yield Crisis
The initial euphoria surrounding Co-Packaged Optics was grounded in sound physics but fatally flawed in its manufacturing assumptions. The core premise of integrating the optical engine directly onto the switch application-specific integrated circuit via advanced 3D packaging was meant to aggressively collapse power consumption. However, the realities of sub-micron optical alignment and component calibration have triggered a system-level yield collapse, effectively pushing mass commercialization of scale-up Co-Packaged Optics out to 2028 or later. The math governing this delay is brutal: integrating 32 optical engines onto a single switch package, assuming a highly optimistic 95 percent yield per individual engine, results in compounding defect rates that drag the overall system yield down to approximately 19 percent. When factoring in the high insertion loss of external lasers—often exceeding 3.5 decibels in system tests—and hyperscaler anxieties over multi-million-dollar switches needing total replacement due to single-engine failures, the Co-Packaged Optics timeline fundamentally breaks down.
In response, hyperscalers are aggressively pivoting toward Near-Packaged Optics as the pragmatic standard. By moving the optical engine off the main chip package and placing it on a high-performance substrate mere centimeters away, Near-Packaged Optics completely avoids foundry-level compounding yield penalties. This architecture relies on ultra-short copper traces for the final millimeter hop but still dramatically drops power consumption to between 4.4 and 6.9 picojoules per bit—roughly half the energy required by traditional digital signal processor-heavy pluggable modules. Crucially, Near-Packaged Optics utilizes socketable components, salvaging system serviceability and providing hyperscalers with the operational reliability required for clusters running tens of millions of compute hours.
Broadcom's Structural Dominance in the Switch Tier
This architectural pivot structurally favors Broadcom over Nvidia in the core AI networking layer. Broadcom has embraced a pragmatic, open-ecosystem approach that natively supports Near-Packaged Optics, allowing hyperscalers to avoid the severe manufacturing risks associated with advanced 3D optical packaging. Broadcom's Tomahawk 5, operating at 51.2 terabits per second, and the recently deployed Tomahawk 6, operating at 102.4 terabits per second, are currently dominating the hyperscaler Ethernet fabric market with essentially zero near-term competition at the 100-terabit threshold.
Broadcom's financial metrics validate this architectural victory. Based on the company's recent fiscal reporting, Broadcom has amassed an artificial intelligence backlog exceeding 73 billion dollars, with expectations for its AI networking revenues to double year-over-year. By focusing on supplying the core switch silicon and 400G-per-lane optical digital signal processors, Broadcom allows cloud giants like Google, Meta, and Amazon to source their optical engines independently. This fosters a deep bilateral lock-in where hyperscalers rely entirely on Broadcom's silicon supremacy but maintain supply chain flexibility on the optical periphery, isolating Broadcom's earnings from the yield drag currently destroying the Co-Packaged Optics roadmap.
Optical Circuit Switching Becomes the Scale-Out Standard
In parallel with the transition to Near-Packaged Optics, Optical Circuit Switching is rapidly becoming the gold standard for hyperscale data center network architectures, driven initially by massive internal validation at Google. Traditional electrical packet switching requires power-intensive optical-to-electrical-to-optical conversions at every network hop. Optical Circuit Switching bypasses this inefficiency entirely by utilizing Micro-Electro-Mechanical Systems mirrors to physically redirect light beams between fiber pathways, delivering near-zero latency interconnects.
Google's deployment of this technology within its TPU v4, v5p, and v6e architectures has proven that Optical Circuit Switching can enable dynamic, sub-10-nanosecond cluster reconfiguration while reducing overall data center network power consumption by 30 to 40 percent compared to legacy topologies. The sustained success of this architecture has forced industry research firms, including Cignal AI, to triple their 2026 forecast for the total addressable market of Optical Circuit Switches. While initially captive to Google's internal Palmetto designs, the multi-billion-dollar Optical Circuit Switching market is rapidly expanding to external suppliers as other cloud providers seek to replicate these massive power and latency efficiencies.
Component Chokepoints: The Continuous Wave Laser Duopoly
Regardless of whether a hyperscaler utilizes Near-Packaged Optics today or attempts Co-Packaged Optics tomorrow, both architectures mandate the separation of heat-sensitive light sources from the switch processor. This requires External Laser Sources, driven by ultra-high-power Indium Phosphide Continuous Wave lasers. As the AI industry aggressively scales from copper to optical interconnects, the market for Continuous Wave lasers has emerged as a critical supply chain bottleneck.
This segment operates as an effective duopoly controlled by Lumentum and Coherent, who collectively hold the vast majority of global high-power laser production capacity. Industry checks indicate both companies are currently sold out, with capacity booked years in advance. The strategic importance of this chokepoint was recently underscored by AMD, which has been aggressively negotiating large-scale purchase agreements for Continuous Wave lasers to secure its own optical interconnect supply chain and bypass Nvidia's component ecosystem entirely. For investors, Lumentum and Coherent represent pure toll-bridge assets in the AI optical buildout, capturing revenue regardless of which specific packaging architecture ultimately wins the data center floor.
Module Integrators Evade Disintermediation
The delay in Co-Packaged Optics offers a massive, multi-year lifeline to traditional optical module integrators. Under a pure Co-Packaged Optics paradigm, much of the module value chain was expected to be absorbed by advanced foundries and chip designers, severely disintermediating traditional assemblers. Near-Packaged Optics permanently reverses this threat. Because the optical engine remains an independent, socketable module on the baseboard, optical module manufacturers retain their lucrative design, testing, and assembly roles.
Industry checks reveal that under the Near-Packaged Optics architecture, optical module makers capture roughly double the value content compared to what they would earn merely supplying raw components for Co-Packaged Optics. Companies like Zhongji Innolight and Eoptolink are the primary beneficiaries of this extended lifecycle. Innolight, for instance, is projected by market analysts to capture approximately 80 percent of Google's orders for optical modules above 800G through 2026. With Google scaling its TPU deployments into the millions, Innolight is positioned to ship over six million high-speed modules to a single customer, sustaining gross margins that rival top-tier semiconductor designers.
The Threat Vector: Hyperscalers Reject Nvidia's Proprietary Lock-In
At its core, the rapid adoption of Near-Packaged Optics and Optical Circuit Switching is a strategic defensive maneuver by Alphabet, Amazon, and Meta against vendor lock-in. Nvidia has aggressively pushed a vertically integrated roadmap, attempting to tether hyperscalers to its proprietary InfiniBand fabrics, custom Co-Packaged Optics designs like Quantum-X, and specialized 800-volt direct current power delivery systems. The manufacturing delays hitting this highly integrated supply chain have given hyperscalers the perfect opening to diverge.
By adopting open-standard Ethernet switches from Broadcom, combining them with Near-Packaged Optics engines from independent module makers like Innolight, and routing traffic through Optical Circuit Switches using Lumentum lasers, cloud providers are successfully modularizing the AI data center. This architectural fracturing restricts Nvidia's pricing power strictly to the core compute node while allowing hyperscalers to commoditize the network transport layer. Supply chain players overly exposed to Nvidia's proprietary networking ecosystem face significant downside risk as the industry standardizes on the pragmatic, multi-vendor reality of Near-Packaged Optics.