OFC 2026: AI Data Centers Reshape Optical Networks, CPO and Pluggable Optical Modules in a Showdown

On March 17, 2026 local time, OFC 2026, the world’s premier optical communication conference, officially opened at the Los Angeles Convention Center. As a key industry bellwether amid the AI computing boom, this edition has shifted entirely away from traditional communication topics, with the reshaping of optical interconnects in AI data centers as the absolute main theme. The two technical routes of CPO (Co-Packaged Optics) and pluggable optical modules are in direct competition, with 1.28T ultra-high-speed solutions, liquid cooling integration, silicon photonics and other technologies launched intensively, announcing that the next-generation data center optical-electrical architecture has entered a decisive commercial phase.

The exponential growth of AI large model training and inference is pushing data center optical communications to physical limits. Clusters of hundreds of thousands of GPUs and EB-scale interconnection bandwidth have left traditional electrical interconnects and conventional optical modules facing a triple bottleneck of power consumption, density and latency. Data shows that today, intra-data center chip-to-chip transmission energy consumption accounts for more than 40%, and the single-port power consumption of 1.6T pluggable modules is approaching 30W, making thermal management and signal integrity for high-density racks unsustainable. At OFC 2026, the entire industry chain has reached a consensus: the shift from copper to optics is irreversible, optical-electrical networks must be redesigned around AI computing power, and the choice between CPO and pluggable solutions will directly determine data center architecture trends over the next five years.

The CPO route centers on ultimate performance and is regarded as the long-term optimal solution for large-scale AI clusters. Its core logic is to co-package optical engines with switching ASICs and GPU chips, reducing the electrical signal transmission distance from centimeter-scale to hundreds of microns, drastically cutting loss and power consumption. At this exhibition, NVIDIA demonstrated a prototype of the Rubin platform and Spectrum-X Photonics CPO, with measured power consumption reduced by 40%, latency entering the picosecond range, and bandwidth density increased threefold; Coherent launched a 6.4T slot-type CPO solution to advance high-level integrated mass production; Broadcom released data showing that CPO power consumption at 800G ports is approximately 5.5W, only one-third of traditional pluggable modules. CPO’s advantages directly address AI pain points: higher density, lower power consumption, and shorter latency, suitable for million-card supercomputing clusters. However, the industry chain also faces challenges: packaging yield, cross-vendor collaboration, thermal management and operational complexity remain to be overcome; replacement requires full-board swapping, leading to significant maintenance costs and supply chain lock-in risks.

The pluggable camp holds onto the mainstream market with maturity and flexibility, breaking bottlenecks through technical iterations. Relying on standardization, hot-plug capability and easy maintenance, traditional pluggable modules remain the current mainstream deployment choice, accounting for over 90% of the market in 2026. Facing competition from CPO, the pluggable route is evolving rapidly: Arista, together with 45 vendors, launched the 1.28T XPO liquid-cooled pluggable module, with a built-in cold plate supporting 400W heat dissipation, four times higher density than existing solutions, balancing high speed and maintainability; LPO (Linear Pluggable Optics) eliminates DSP, cutting power consumption by about 50%, with Google and Meta announcing large-scale 1.6T LPO deployments in 2026; the 6.4T NPO (Near-Packaged Optics) debuted as a compromise solution, striking a balance between integration and maintainability. The pluggable camp emphasizes that a standardized ecosystem, legacy compatibility, and minute-level fault replacement are still essential for cloud vendors and enterprise-level markets, and will not be completely replaced by CPO.

OFC 2026 sends a clear signal: the two routes are not a zero-sum game, but scenario-based differentiation and long-term coexistence. Ultra-large AI clusters and scenarios prioritizing ultimate energy efficiency will gradually adopt CPO; general cloud computing, enterprise data centers and legacy upgrade markets will remain dominated by pluggable modules; mid-range scenarios will be filled by LPO and NPO. The industry chain no longer struggles with "either-or" choices, but builds layered solutions around speed, power consumption, cost and operation and maintenance: 800G/1.6T in commercial use, 3.2T in large-scale verification, 1.28T prototypes launched, indium phosphide capacity expansion, silicon photonic integration, and liquid cooling supporting facilities advancing simultaneously to sustain the growth of AI computing power.

In terms of industrial impact, this conference has reshaped the competitive landscape of optical communications. Leading vendors focus on high-end solutions and standard setting, while small and medium-sized manufacturers concentrate on mass production and cost control; optical chips, packaging, materials, liquid cooling and other segments are embracing incremental growth. Capital markets responded quickly, with optical communication stocks such as Lumentum, Applied Optoelectronics and Coherent rising collectively during OFC, confirming that AI optical interconnects have entered a high-boom cycle.

OFC 2026 is more than a technology exhibition; it is a declaration of reconstruction for AI infrastructure. The competition between CPO and pluggable modules is essentially a balance between performance and engineering, between the future and the present. With the commercialization of 1.28T, improved CPO yield and enhanced ecological collaboration, optical-electrical networks will evolve from supporting systems to core computing components. It is foreseeable that 2026–2027 will be a critical period for the large-scale deployment of both routes, and more efficient, denser and more flexible optical interconnects will lay a solid digital foundation for the era of artificial general intelligence.