When Power Consumption Becomes a More Urgent Challenge Than Bandwidth
The scale and complexity of AI models are growing exponentially. Industry analysis indicates that the computing power required to train leading AI models is doubling approximately every 3–4 months, far outpacing Moore‘s Law. Yet a critical but widely overlooked issue is that the energy consumption of optical interconnect networks supporting these computing clusters is rising at an alarming rate. In an AI data center housing tens of thousands of GPUs, optical interconnect energy consumption already accounts for a significant proportion of total network power. Consequently, industry consensus is undergoing a fundamental shift: energy efficiency, rather than raw bandwidth alone, has become the decisive metric guiding next-generation optical module technology roadmaps.
LPO: The Linear-Drive Energy Efficiency Revolution
In traditional optical module architectures, the Digital Signal Processor (DSP) chip performs critical signal compensation and recovery functions but is also the largest contributor to module power consumption. In modules operating at 800G and above, DSP power consumption can account for 40–50% of the total. The core breakthrough of LPO (Linear Pluggable Optics) technology is eliminating this power-hungry DSP chip entirely, using linear analog techniques to drive optoelectronic devices directly.
The benefits are immediate and substantial: significantly reduced system power consumption and transmission latency. Adtran’s LiteWave800 ultra-low-power 800G LPO module sets a new industry benchmark—operating at just 1 picojoule per bit (1pJ/bit) and consuming only 0.8 watts overall, establishing a new power class for 800G optics. In comparison, traditional DSP-based pluggable modules typically consume 12–15 watts. This means that simply by adopting LPO technology, a data center deploying thousands of optical modules can save hundreds of kilowatts in power consumption and significantly reduce cooling loads.
LPO technology demonstrates strong adaptability across application scenarios. Currently, electrical chip products for 100G and 200G single-lane applications have widely adopted linear-drive technology, with solutions compatible across next-generation optical interconnect architectures including LPO, NPO, and CPO. Semtech‘s CEI-224G-Linear IC family fully supports linear-drive requirements for 800G, 1.6T, and even 3.2T transceivers and optical engines, covering the complete range from compact NPO/CPO to various LRO/LPO/XPO interfaces. In 2025–2026, LPO is widely regarded as the most promising incremental technology for addressing data center challenges of high power consumption and high latency.
CPO and NPO: A Paradigm Shift from “Pluggable” to “Co-Packaged”
If LPO represents incremental innovation within the existing pluggable architecture, CPO (Co-Packaged Optics) and NPO (Near-Packaged Optics) represent a fundamental transformation of optical interconnect architecture. In traditional architectures, optical modules are independent pluggable devices connected to ASIC chips via SerDes channels. In CPO/NPO architectures, optical engines are integrated directly onto the same or adjacent package substrate as the switch chip, dramatically shortening electrical interconnect distances.
The core advantage of CPO technology lies in fundamentally addressing signal loss and reflection issues inherent in PCB traces. As signal rates increase to 224Gbps and beyond, electrical signal transmission distance on PCBs shrinks dramatically, pushing traditional pluggable architectures toward physical limits. CPO compresses electrical interconnect distances from tens of centimeters to millimeters or less by co-packaging optical engines with ASICs, significantly enhancing bandwidth density and energy efficiency. Industry projections indicate CPO technology will initiate deployment from 800G/1.6T port scenarios, with initial deployment expected between 2026 and 2027, followed by gradual volume ramp-up.
Yole Group analysts noted at Photonics West 2026 that CPO adoption is no longer a question of photonic technology feasibility—“Indeed, it already works!” The real challenges to overcome are industrial-scale hurdles: yield rates, thermal management, testing strategies, and manufacturing process scalability. These challenges are driving the supply chain toward system-level co-design approaches rather than discrete component models.
NPO, as an intermediate form of CPO, offers improved serviceability and upgrade flexibility while maintaining high integration levels. GIGALIGHT successfully launched its first-generation XT-1.6T DR16 NPO linear silicon photonics engine in 2026. The product adopts a socket-based form factor compliant with the OIF Co-Packaging-3.2T-Module standard, combines LPO linear direct-drive technology, and maintains power consumption below 16 watts. This marks the NPO/CPO ecosystem‘s transition from proof-of-concept to commercial deployment.
Silicon Photonics: The “Chip-ification” Revolution in Optical Communications
Silicon photonics technology represents a paradigm revolution in the optical communications industry, shifting from an “assembly model” to “chip-based manufacturing.” Traditional optical modules employ discrete device approaches, where lasers, modulators, detectors, and other optical components are manufactured separately and then precisely coupled and assembled. Silicon photonics integrates all these critical optical components onto a single silicon chip, leveraging mature CMOS semiconductor processes to enable large-scale, high-consistency, low-cost manufacturing.
Photonic Integrated Circuits (PICs) form the core of silicon photonics technology. PICs achieve highly integrated and modular optical signal processing by integrating multiple optical devices onto a single chip. This technology roadmap delivers not only significant size reduction and power savings but, more importantly, shifts the core industry value toward design and process capabilities—companies with autonomous PIC design capabilities can define optical path architectures and dominate process roadmaps, thereby capturing high-value-added segments of the supply chain.
From an application perspective, PICs are driving optical interconnects beyond data center scale-out scenarios into scale-up scenarios, opening larger markets spanning medium-to-short distances and device-level to chip-level applications. Driven by AI demand for optical communications, optical module/engine volumes have escalated from millions to tens of millions of units, with future potential reaching hundreds of millions. Given considerations of capacity, cost, and power consumption, the industry urgently needs silicon photonics as a new capacity paradigm to meet growing demand.
Silicon photonics development is also catalyzing the “chipletization” of optical I/O. At Photonics West 2026, demonstrations based on advanced architectures pointed toward targets of 100 Tb/s per accelerator and hundreds of optical ports per package. This evolution marks a fundamental shift from discrete optical modules toward integrated optical subsystems that are increasingly deeply embedded in advanced packaging, becoming part of the computing system itself.
Synergistic Breakthroughs in Emerging Materials and Advanced Architectures
In the next-generation optical module technology roadmap, breakthroughs in emerging materials play an equally critical role. Thin-Film Lithium Niobate (TFLN) has garnered significant attention for its low drive voltage and high bandwidth characteristics. HyperLight demonstrated a 1.6T-DR8 optical module based on its TFLN Chiplet Platform, capable of operating directly from the native low-swing electrical output of the DSP, further reducing system-level power consumption.
Optical Circuit Switching (OCS) technology is also gaining increasing importance in AI data center architectures. Cignal AI projects OCS system revenue to exceed $3.5 billion by 2029, more than doubling from 2025 levels. Google’s Apollo OCS architecture enables direct fiber-to-fiber connections via MEMS micro-mirrors, avoiding the power and latency penalties of repeated optical-electrical-optical conversions in traditional switching. A single OCS switch consumes only about 100 watts—approximately a 95% reduction compared to traditional switches drawing around 3,000 watts.
The 2026–2028 period has been defined by industry experts as a critical window for optical communication technology: silicon photonics is evolving from pluggable to CPO/NPO architectures; CPO deployment will enhance energy efficiency and bandwidth density; and new materials such as thin-film lithium niobate and BTO electro-optic modulation are achieving breakthroughs. Together, these advances will propel optical communications into a new era.
The development of next-generation optical module technology extends far beyond simply doubling data rates. It represents a systemic transformation encompassing architecture, materials, manufacturing processes, and supply chain collaboration. From the immediate energy savings delivered by LPO, to the architectural restructuring enabled by CPO/NPO, to the “chip-ification” manufacturing paradigm shift realized through silicon photonics—each technology pathway addresses the same fundamental question: how to minimize energy consumption and cost per transmitted bit while continuously doubling bandwidth. For optical transceiver suppliers and procurement enterprises alike, understanding and navigating these technology trends is essential not only for current product selection but also for securing long-term competitiveness in the future AI computing landscape.
