As GPU performance growth shifts from "Moore's Law" to the "New Moore's Law," the efficiency of inter-chip communication has become a core variable determining the performance of intelligent computing clusters. On May 6th, NVIDIA and Corning entered into a multi-year commercial and technical collaboration. According to the agreement, Corning will build three new advanced manufacturing facilities in North Carolina and Texas, dedicated to producing optical connectivity solutions for NVIDIA. NVIDIA, in turn, invested $500 million to acquire two immediately exercisable warrants for Corning stock. This marks NVIDIA's latest investment move in the optical interconnect field this year. In March 2026, NVIDIA invested a total of $6 billion within 30 days, allocating $2 billion each to Lumentum, Coherent, and Marvell. These concentrated investments all target the optical interconnect and AI computing infrastructure sectors—a technology that uses light waves as the carrier and optical fibers as the channel is becoming key to breaking through AI computing bottlenecks with its "ultra-high bandwidth, ultra-low power consumption" advantages.
The bottlenecks are clear: the ceiling of electrical interconnect—bandwidth walls and power consumption walls—is hindering further upgrades to AI infrastructure. Over the past two years, global AI computing power has increased approximately 60,000-fold, but the "blood vessels" supporting it have severely lagged: traditional copper cable electrical interconnect bandwidth has only increased 30-fold, creating a widening gap. Meanwhile, as GPU power consumption continues to rise, the proportion of power consumed by electrical interconnects may increase further. For the industry, this will not only drive up data center operating costs but also increase cooling pressure. More seriously, issues like significant long-distance transmission loss, frequent electromagnetic interference, and reaching integration density limits are posing unprecedented transmission challenges for training trillion-parameter large models and parallel computing in clusters with hundreds of thousands of cards.
Technological Breakthrough: Optical Interconnect Enters an Accelerated Development Phase While electrical signals struggle through copper wires, photons travel through optical fibers at nearly the speed of light. Leveraging its physical nature, optical interconnect technology demonstrates outstanding advantages such as ultra-high bandwidth density, extremely low power consumption, ultra-low transmission latency, and strong anti-interference capabilities. Tech giants have keenly recognized these advantages and are closely engaged in commercial deployment.
● AMD announced the development of a CPO (Co-Packaged Optics) solution based on MRM for its next-generation Instinct MI500 AI accelerator, establishing an industry chain division of labor with "AMD design + GlobalFoundries manufacturing + ASE packaging," expected to launch in 2027. ● NVIDIA's moves are more rapid, investing a cumulative $6 billion in 30 days, all directed at optical interconnect and AI computing infrastructure. Its Rubin Ultra architecture will prioritize CPO solutions, while the longer-term Feynman generation plans a full transition to optical interconnect technology. ● TSMC's silicon photonics engine COUPE platform, based on SoIC-X chip stacking technology, directly 3D-stacks electrical chips with photonic chips. It completed validation of small form-factor pluggable optical modules in 2025 and will begin mass production in 2026. Simultaneously, the platform integrates the COUPE optical engine into CoWoS packaging in CPO form, marking a critical step for CPO implementation. ● South Korea's Samsung Electro-Mechanics and LG Innotek have officially moved from the conceptual research phase of CPO technology to early development stages, expecting to integrate various components on semiconductor substrates to enable final packaged semiconductor products with CPO functionality.
Domestic Practice: Multi-Dimensional "Industry-Academia-Research-Investment" Deployment The accelerated implementation of technology is creating a vast market space. According to a Goldman Sachs research report, the global addressable market size for optical interconnect will rise from the current $15 billion to $154 billion by 2027-2028, representing a ninefold overall increase. Facing this industry development opportunity, the LEGENDHOLDING system, leveraging its long-term accumulation in cutting-edge technology, has made forward-looking deployments in the optical interconnect field.
● The "Peking University - LEGENDHOLDING Advanced Photonic Integration Technology Joint Laboratory," co-established by LEGENDHOLDING and Peking University, conducts collaborative research focusing on photonic integration process exploration, low-power optical engine prototype validation, and optical interconnect application research for intelligent computing centers. In March 2026, the joint laboratory published its first academic paper in the internationally renowned journal *Photonics*. The research team broke through the technical limitations of short-wavelength grating couplers in coupling efficiency and process tolerance, with related technologies already under patent application protection. ● TuringQ, a company jointly invested in by Junlian Capital, Legend Star, and Legend Capital, is a pioneer in optical quantum computing with full-stack technical capabilities from chips and algorithms to integrated systems. Recently, TuringQ entered into a deep strategic partnership with Kiwimoore, a leader in AI network interconnect. They will jointly research, develop, and advance next-generation optical interconnect OIO technology projects, building a photoelectric fusion hardware system spanning from classical to quantum computing. This collaboration represents the downward compatibility of its core optical quantum computing technological capabilities to empower the optical connectivity field, drive breakthroughs in computing power bottlenecks, and establish a new paradigm of photoelectric fusion computing. ● OptiLinkCore, a company invested in by Junlian Capital, is an innovative leader in China's optical interconnect OIO field, focusing on inter-chip optical interconnect for high-performance computing. It has accelerated the OIO technology from R&D to practical deployment. Its multi-channel OIO silicon photonic chips, optical engines, and other products enable computing chips to emit light directly, significantly increasing inter-chip transmission bandwidth while reducing transmission energy consumption. Compared to data center optical modules, its bandwidth-energy efficiency product achieves a ten-thousand-fold improvement, greatly enhancing AI cluster training and inference efficiency, facilitating the widespread adoption of AI large models, and propelling intelligent computing centers toward all-optical interconnect. ● Xiri Photonics, a company invested in by Legend Capital, possesses deep international research background and industrialization capabilities, covering key areas such as material understanding, wafer-level processes, chip design, device engineering, and industry collaboration. The company's founder/Chief Scientist, Professor Wang Cheng, is one of the principal pioneers of global thin-film lithium niobate photonic technology and a core inventor of the thin-film lithium niobate electro-optic modulator. Leveraging source innovation accumulation and the engineering transformation capabilities of its industrialization team, Xiri Photonics is accelerating the transition of thin-film lithium niobate technology from laboratory achievements to practical applications in AI optical interconnect and next-generation communication networks.
The technological prospects of optical interconnect are highly clear. Its value lies not only in alleviating the current thirst for computing power but also in defining future computing architectures. When data transmission bottlenecks are eliminated, the fusion of distributed computing, in-memory computing, and even quantum computing can truly unleash its potential, injecting powerful momentum into AI infrastructure.
Comments