In-depth Observation of OFC 2026: AI-Driven Comprehensive Leap in Optical Interconnection

OFC 2026, the world's top optical communication event, concluded in Los Angeles from March 15 to 19, 2026. The strongest signal emanating from this conference is that optical communications has completely evolved from a "telecommunications exhibition" into an "AI exhibition". The bandwidth and power consumption challenges posed by AI computing power are driving optical interconnection technologies to move from laboratory research to large-scale commercial application.

AI Reshapes the Evolution Logic of Optical Communication Technologies

With the continuous expansion of AI cluster scale, high-speed interconnection is an essential requirement for accelerators (GPUs, DPUs, LPUs). As Jensen Huang put it at the GTC Conference, "We are transforming from a chip manufacturer into a full-stack AI infrastructure service provider. The AI factory—where energy is input and Tokens are output—will define the future of data centers." In this factory, optical interconnection acts as the conveyor belt. Traditional electrical interconnection has reached its limits in terms of bandwidth, density, latency and power consumption, with electrical-to-optical conversion emerging as the core bottleneck. Thus, optical interconnection has become the only solution to break through the "I/O wall".

Note: What is the I/O wall?

The I/O wall refers to a phenomenon where the growth rate of computing capability (computing power) far outstrips that of data input/output (I/O) bandwidth, leading to the underutilization of computing power.

Without optical interconnection to break the I/O wall, the computing power growth brought by Moore's Law cannot be fully harnessed. The core topic of this year's OFC Conference revolves around the next-generation evolution of optical interconnection technologies, which are accelerating along six key directions:

Layer 1: Underlying Specifications

1. The underlying rate is leaping to 400G per lane: 1.6T modules are scaling up in production, and 3.2T prototypes are making their debut.
2. Three material routes advance in parallel: EML, silicon photonics and thin-film lithium niobate underpin higher bandwidth transmission.

Layer 2: Packaging Architecture

3. Multiple packaging architecture routes coexist hierarchically: Pluggable, NPO, CPO, LPO and XPO form a layered coexistence pattern.

Layer 3: System Applications

4. OCS optical switching emerges as a new hotspot: Power consumption is reduced by nearly 95%, serving Scale-out scenarios.5. The Rubin Platform is on the verge of commercial deployment: NVLink 6 (3.6TB/s) leads the accelerated iteration of Scale-up solutions.

Industrial Ecology

6. Industrial ecological restructuring: AI giants engage in in-depth layout and participation, and Chinese manufacturers highlight their competitive advantages.

Trend 1: Leaping to 400G per Lane – Accelerated Evolution of PAM-4

To understand this trend, a fundamental concept must be clarified: PAM-4 is the most fundamental coding specification in the current optical interconnection industry. Its core value lies in doubling the data transmission rate without increasing the physical bandwidth (symbol rate).

The current evolution direction is the upgrade of single-lane rate from 200G/lane to 400G/lane, which means:

1. A 1.6T module can be implemented with 4 lanes (4×400G), cutting the number of lanes by half compared with the 8×200G solution.
2. A 3.2T module can be realized with 8 lanes, laying the technical foundation for the next-generation 6.4T modules.

This evolution direction has been fully verified at OFC 2026. Coherent showcased a hierarchical portfolio of pluggable optical technology products at the conference:

1. 1.6T level: It adopted a variety of optical technologies, including silicon photonic integrated circuits (PICs), high-power InP continuous-wave lasers, 200G InP electro-absorption modulated lasers (EMLs) and 200G gallium arsenide (GaAs) VCSELs, and adapted to DSP solutions from three different suppliers. This demonstrates Coherent's execution capabilities on three technical platforms: silicon photonics (SiPh), indium phosphide (InP) and VCSEL.
2. 3.2T level: It presented optical links based on 400G per lane PAM-4, including implementation schemes for silicon photonic PICs with 400G differential EMLs and 400G pure silicon PN-junction Mach-Zehnder modulators, verifying its leading position in the next-generation 3.2T pluggable architecture field.
3. 12.8T and above level: It exhibited a new multi-lane XPO pluggable MSA packaging form, designed to enhance the flexibility of system design and optimize power consumption and performance.

This marks the maturity of both physical implementation devices and diversified technical platforms for PAM-4, paving the way for the commercialization of 400G/lane technology.

Trend 2: Three Material Routes Advancing in Parallel

The evolution of optical chips follows the principle of "faster, higher and more powerful". The three material routes are not substitutive but form a parallel development pattern to meet different rate and power consumption requirements.

表格

1. Coherent showcased a full range of 1.6T/3.2T/12.8T solutions, covering silicon photonics, InP and VCSEL multi-technical platforms.
2. Accelink Technologies launched the world's first 3.2T silicon photonic single-mode NPO module, which has completed verification by leading cloud manufacturers.
3. Lumentum clearly pointed out at the conference that in the 8×400G era, EML will become the de facto industry standard, as silicon photonics solutions are unable to meet the bandwidth requirements of 400G/lane.

Trend 3: Multiple Packaging Architecture Routes Advancing in Parallel

One of the core topics of this year's OFC is the restructuring of the physical distance between optical engines and switch chips. Five technical routes have formed a clear hierarchical development pattern.

Key Breakthroughs at OFC 2026

1. Accelink Technologies: Its 3.2T silicon photonic NPO module completed full-system verification by leading domestic CSPs (covering optical engine OE, external light source ELSFP and fiber management module), making it the world's first manufacturer to achieve this breakthrough and marking the official shift of NPO technology from laboratory research to large-scale engineering deployment.
2. Lightmatter: Released the world's first detachable fiber array unit, solving the core pain points of CPO yield and maintainability.
3. Gigalight: Launched the industry's first 12.8Tbps liquid-cooled pluggable optical module, adopting a 64-lane ×200G design with cold plate heat dissipation technology.
4. Linktel: As a core founding member of the XPO MSA, it publicly displayed the 12.8T liquid-cooled XPO optical module for the first time.
5. Yangtze Optical Fibre and Cable (YOFC): Showcased the OIO solution, realizing high-density interconnection based on multi-core optical fibers and achieving precise coupling between multi-core optical fibers and array light sources.
6. Yilutong: Launched a pluggable FAU for CPO, achieving a structural upgrade from "fixed coupling" to "flexible plug-and-play" and supporting independent plugging and quick replacement.
7. Changxin Bochuang: Exhibited a full portfolio of FAU products for CPO, including sealed FA, detachable FA, mode field conversion FA and multi-core FA.
8. TFC Optical: Exhibited FAU components, POSA high-speed optical engines (including EO integration solutions) and packaging technologies for NPO/CPO/OIO scenarios, covering 800G/1.6T/3.2T high-speed optical transmission scenarios. With its low power consumption and low latency characteristics, the company's products support the upgrade of optical communication systems towards a faster, more efficient and green development direction.

Trend 4: OCS Optical Switching Emerges as a New Hotspot

Optical Circuit Switches (OCS) have become one of the most watched technical hotspots in the optical interconnection field in 2026. OCS performs physical optical path switching in the optical domain without electro-optical conversion, featuring ultra-low power consumption and nanosecond-level latency. With the rapid development of large language models and ultra-large-scale computing power clusters, traditional electrical switching architectures have encountered insurmountable bottlenecks in terms of power consumption, latency and scalability.

Equipped with the characteristics of all-optical transparent transmission, protocol independence and ultra-low latency, OCS has become the universally recognized inevitable evolution direction of the industry. It boasts remarkable power consumption advantages: industry research has confirmed that OCS technology can help reduce the overall power consumption of AI computing power clusters by more than 30%, which is of great strategic significance for alleviating the power supply pressure of data centers.

Industrial Background: Google's TPU architecture has long been deeply integrated with OCS and has been actually deployed in multiple generations of TPU clusters, fully verifying the feasibility and practical value of OCS in large-scale AI computing power application scenarios.

Note: Google's mainstream Palomar OCS at present is based on the MEMS (Micro-Electro-Mechanical Systems) solution, with 136 optical path channels (128 channels in actual use). Its core working principle is to reflect optical signals through a 2D MEMS micromirror array to achieve millisecond-level optical path switching, eliminating the need for electro-optical conversion via optical transceivers.

Key Updates at OFC 2026:

1. Accelink Technologies released a 320×320 MEMS OCS all-optical switch, with power consumption reduced by approximately 95% compared with traditional solutions.
2. ARK Photonics and Calient.AI jointly showcased an OCS all-optical interconnection solution, achieving a full-process manufacturing breakthrough from MEMS chip packaging and testing to system assembly. The solution features high product consistency and flexibly scalable production capacity.
3. Changxin Bochuang fully demonstrated the full-link technology from substrate to 2D FAU array and optical engine, and its independently developed 2D FAU design is deeply optically matched with mainstream MEMS solutions.

Application Scenarios: OCS serves the Scale-out layer, dynamically connecting cabinets to form a unified computing power pool for the entire cluster. With the continuous growth of AI computing power demand, OCS will become a key core device for internal interconnection in data centers.

Trend 5: The Rubin Platform is on the Verge of Deployment – 1.6T/3.2T Modules Scaling Up at an Accelerated Pace

Intensive information released at the GTC and OFC Conferences indicates that NVIDIA's Rubin Platform is on the verge of commercial deployment, driving optical interconnection technologies into a period of accelerated iteration. The core indicators of the Rubin Platform—NVLink 6 with a single GPU bandwidth of 3.6TB/s (14 times that of PCIe 6.0), the NVL72 cabinet with a total bandwidth of 260TB/s, and the mass production of Spectrum-X CPO switches—have put forward higher technical requirements for optical interconnection, directly driving the large-scale production of 1.6T optical modules and the debut of 3.2T prototypes.

Key Updates at OFC 2026:

1. 1.6T scaling up: Coherent, Lumentum, Linktel and other manufacturers exhibited their 1.6T optical module solutions; NVIDIA's expected global order volume for 1.6T optical modules in 2026 reaches 20 million units.
2. 3.2T prototype debut: Accelink's 3.2T silicon photonic NPO module has completed verification by leading CSPs; Coherent showcased a 3.2T prototype based on 400G per lane.
3. Coexistence of copper and optical solutions: The parallel application of multiple solutions has become the industry's first choice; copper cables are retained for Scale-up scenarios, while the shift to optical interconnection is accelerating for Scale-out scenarios.
4. Supporting industrial chain upgrades: Broadcom is mass-producing the 102.4T switch chip Tomahawk 6 series; Marvell launched the industry's first 1.6T ZR/ZR+ pluggable coherent module.

The large-scale deployment of the Rubin Platform is entering the countdown stage, and 1.6T/3.2T optical interconnection products are moving from laboratory research to engineering deployment, providing a solid connection foundation with higher bandwidth and lower power consumption for AI computing power clusters.

Trend 6: Industrial Ecological Restructuring – Joint Innovation by AI Giants and Cloud Providers, Chinese Manufacturers Highlight Their Advantages

• NVIDIA: The company sent its Senior Vice President to deliver a plenary session speech at OFC. With the Rubin Platform on the verge of deployment, it is directly driving the accelerated large-scale production of 1.6T/3.2T optical interconnection products.
• Alibaba Cloud & HGTECH: The two parties jointly realized the successful lighting-up of the world's first 3.2T NPO module and dynamically showcased the 3.2T NPO optical engine + ELSFP light source module solution on site. Based on the AI cluster Scale-up all-optical interconnection strategy, this solution features core advantages such as ultra-high integration (the rate capacity per unit area reaches 11.4 times that of 800G OSFP), ultra-low power consumption (50% reduction in power consumption per bit), detachable design, external light source, no need for DSP chips, industrial chain decoupling and compatibility with diverse application scenarios.

Chinese Manufacturers Highlight Their Competitive Advantages

Chinese optical interconnection enterprises delivered an impressive performance at this year's OFC, covering the entire industrial chain from optical chips, passive devices and optical modules to system equipment.

Enterprises such as YOFC, Accelink Technologies and FiberHome Telecommunication have ranked among the first echelon of the global optical communication industry, becoming the "main force" for China to participate in the global competition of optical interconnection technologies.

Summary of the Industrial Pattern: AI giants define the technological development directions of the industry, while Chinese optical interconnection enterprises transform cutting-edge technologies into large-scale commercial products by virtue of their full-industry-chain layout and rapid implementation capabilities. The two sides together form the current collaborative pattern of "direction leadership + implementation execution" in the global optical interconnection industry.

Conclusion: AI Defines Optical Interconnection, the Global Industry Jointly Builds the Computing Power Foundation

OFC 2026 fully demonstrates that optical communications has completely transformed from a "telecommunications exhibition" into an "AI exhibition". From the leap from 200G/lane to 400G/lane, to the parallel advancement of multiple packaging architectures, and then to the large-scale production of 1.6T modules and the debut of 3.2T prototypes, the main line of technological evolution remains unchanged: enabling faster, more energy-efficient and denser data transmission between accelerators.

Chinese optical interconnection enterprises delivered an outstanding performance at this year's OFC, covering the entire industrial chain from optical chips to system equipment, and have become an important main force in the construction of global AI computing power. Grasping the core main line of "AI defines optical interconnection" and closely following the commercialization progress of 400G/lane technology and the deployment rhythm of CPO/NPO solutions will be the core development propositions for the optical interconnection industry in the coming years.