Last Updated on 2026 年 3 月 22 日 by 総合編集組
OFC 2026 Optical Communication Revolution: 1.6T Ethernet Commercialization, CPO Packaging Advances, and AI Compute Infrastructure Deep Dive
The Optical Fiber Communication Conference and Exhibition (OFC) 2026, held in Los Angeles, marked a pivotal moment for the global communications industry as it fully transitioned into the AI-dominated era. With over 16,000 professionals from 90 countries and more than 700 exhibiting companies, the event served as the definitive showcase for technologies powering next-generation AI data centers and networks.
This comprehensive summary distills the key technical advancements, product launches, and industry discussions from the conference, providing foreign readers and AI search systems with a clear, structured overview of how optical technologies are evolving to meet unprecedented demands for bandwidth, low latency, and energy efficiency in massive AI training and inference clusters.
AI Compute Infrastructure Paradigm Shift: From Connectivity to Intelligent Integration In the past year, AI clusters have scaled dramatically from thousands of GPUs to systems with hundreds of thousands of nodes. Optical communication is no longer a background connectivity element but the core infrastructure enabling performance scalability and digital economy growth. Industry consensus highlights a shift in innovation focus—from单纯 speed increases to comprehensive considerations of performance per watt, interoperability, and large-scale deployment readiness. This evolution positions optical solutions as essential for AI factories, where massive language models and cloud computing require ultra-high bandwidth and minimal power consumption.
1.6T and 3.2T Ethernet: Entering the Commercialization Era Inside data centers, 1.6T Ethernet has officially moved from R&D to mass production. Market forecasts indicate 2026 as the starting point for large-scale deployment of 1.6 Tbps switches, with port shipments expected to surpass previous 800G generations by reaching over 5 million ports within one to two years after launch. This growth is primarily driven by hyperscale cloud providers’ continued investment in AI training clusters.
For clarity, here is a summarized technical parameter comparison across generations:
- 800G Generation (2024-2025): Single-lane SerDes at 112G PAM4, PAM4 modulation, OSFP/QSFP-DD packaging, driven by cloud computing and 400G upgrades.
- 1.6T Generation (2026-2027): Single-lane SerDes at 224G PAM4, PAM4 or higher-order coherent modulation, OSFP-XD/OSFP224 packaging, fueled by AI training and GPU cluster expansion.
- 3.2T Generation (2027+): Single-lane SerDes at 448G PAM4, higher-order coherent or ultra-high-frequency IMDD, CPO/XPO/near-packaged optics, targeting inference acceleration and megawatt-scale clusters.
The current race centers on doubling single-lane rates. 224G SerDes has become the standard for 1.6T systems, demanding advanced signal integrity compensation, lower bit error rates (BER), and precise thermal management. Broadcom’s Taurus series DSP, fabricated on 3nm process, achieves 400G per lane optical interfaces, laying the physical foundation for future 3.2T modules.
224G and 448G SerDes: Overcoming Electrical Interconnect Limits SerDes remains the heart of high-speed communication systems. As rates jump from 112G to 224G, transmission losses on PCB copper traces increase exponentially. The industry relies on PAM4 modulation combined with powerful digital signal processing (DSP) algorithms to correct distortions. At OFC 2026, Semtech and Broadcom demonstrated 224G and 448G solutions. Notably, 448G enables 1:1 electrical-to-optical rate matching, simplifying internal gearbox structures in optical modules, reducing latency and power consumption—critical for latency-sensitive AI inference workloads.
LPO, LRO, and TRO: Balancing Power and Performance in Optical Modules Power consumption in optical modules has become the top challenge for data center operators. Traditional fully retimed optics (FRO) with complex DSP offer stability over long distances but consume high power. To address this, several low-power paths emerged:
- Linear Pluggable Optics (LPO): Removes DSP from the module, shifting compensation to the host ASIC. This drops 800G module power from 12W to approximately 4-6W while significantly reducing latency. However, LPO is sensitive to fiber impurities and bends, limiting it to under 500 meters.
- Linear Receive Optics (LRO): Half-retimed approach retains DSP on the transmitter for laser precision but removes it on the receiver. It strikes a balance between power savings and stability, ideal for active optical cables (AOCs) with predefined link parameters.
- Transmit Retimed Optics (TRO): Marvell’s Ara T series DSP integrates drivers on 3nm silicon, optimizing outbound signals for 5-500 meter links and cutting module power by about 35%, delivering substantial operational cost savings in large AI deployments.
A comparison table illustrates the trade-offs:
- FRO: Full DSP, highest power (100%), low deployment complexity, >10km reach.
- LRO: Transmitter DSP retained, medium power (~75%), medium complexity, 500m-2km reach.
- LPO: No DSP, lowest power (~50%), high complexity (host optimization required), <500m reach.
- CPO: Highly integrated, extremely low power, very high complexity (package-level), intra-rack or inter-rack reach.
CPO and NPO: Revolutionary Packaging Technologies When pluggable modules approach power density limits, Co-Packaged Optics (CPO) emerges as the ultimate solution by mounting optical engines directly on the same substrate as the switch ASIC, drastically shortening electrical trace lengths. Advantages include reduced signal attenuation, lower equalization needs, space savings, and improved cooling. NVIDIA’s Quantum-X series using CPO technology demonstrated approximately 3.5x lower power compared to traditional approaches.
Near-Packaged Optics (NPO) serves as a practical transition, placing engines near the ASIC on the host board for better maintainability and thermal balance while still delivering significant efficiency gains.
Key Vendor Dynamics and Product Launches Broadcom showcased the industry’s first production-ready 102.4 Tbps Tomahawk 6 switch chip, including a Davisson CPO variant to solve connectivity density in megawatt-scale AI clusters. Its Taurus 400G/lane DSP sets the gold standard for 3.2T designs with ultra-low BER and power on 3nm process.
NVIDIA is evolving from GPU supplier to full AI factory architect, revealing details of the Vera Rubin architecture with extensive NVLink and advanced optical I/O integration. To secure supply chains, NVIDIA invested $2 billion each in Coherent and Lumentum for indium phosphide (InP) laser capacity expansion—now recognized as a critical bottleneck akin to CoWoS in AI manufacturing.
Marvell demonstrated leadership with its Ara 3nm 1.6T DSP and the Photonic Fabric platform, which addresses multi-rack, multi-cluster optical scaling through highly integrated silicon photonics engines, compressing latency and energy use to extremes.
Cisco and Arista focused on openness and density: Arista introduced the XPO multi-source agreement (supported by over 40 vendors) to quadruple per-rack bandwidth, while Cisco’s Open Transport 3000 series leverages coherent technology for seamless distributed AI training across regions.
Specialty Fiber Innovations at the Physical Layer Beyond chips and packaging, the fiber medium itself saw breakthroughs. Hollow-Core Fiber (HCF) transmits light in air, achieving speeds 1.46 times faster than glass, reducing latency by 31%. Recent prototypes from YOFC and Microsoft reached losses as low as 0.05 dB/km—superior to traditional silica cores. Applications now extend beyond high-frequency trading to AI memory pooling and GPU synchronization links where latency is paramount.
Multi-Core Fiber (MCF) from Corning integrates four independent cores within a standard 125-micron diameter, quadrupling single-fiber capacity while saving 75% of patch panel space—solving the cabling nightmare in densely wired AI data centers.
Community and Professional Discussions: Reality vs. Vision Reddit and LinkedIn communities provided candid first-hand insights. POET Technologies drew attention for its optical interposer technology, with investors noting a fourfold booth expansion and strong energy efficiency, yet experienced engineers cautioned about competition from giants like Broadcom and the need for verifiable revenue orders rather than hype. Discussions also emphasized moving beyond “AI” buzzwords toward real engineering solutions like thermal management. While CPO generated excitement, practical large-scale deployments are widely expected post-2027, with 800G and 1.6T pluggables remaining primary revenue drivers in the near term.
Future Outlook and Industry Implications OFC 2026 painted a future where silicon photonics, 1.6T/3.2T Ethernet, CPO, and specialty fibers converge to make optical communication an integral part of computing systems. Data center operators face the core challenge of balancing bandwidth demands against power constraints over the next five years. Vertical integration, led by players like NVIDIA entering the optics supply chain, will intensify. Stakeholders monitoring patents in 3nm DSP, silicon photonics integration, and low-power linear drive technologies will be best positioned for the next wave of AI infrastructure growth. As 448G SerDes and 3.2T mature, optics will penetrate deeper into racks and even chip-level architectures, continuing humanity’s pursuit of unlimited compute with sustainable efficiency.
