--- title: "[Optical Chip Decoding - The Foundation of Computing Power] In the AI era, why do we need optical chips?" type: "Topics" locale: "en" url: "https://longbridge.com/en/topics/43145393.md" description: "$ZJ INNOLIGHT(03308.HK) has become a hot topic in the capital market this year. As the world's largest optical module supplier, a large portion of the core optical modules for AI giants like $NVIDIA(NVDA.US) and $Amazon(AMZN.US) come from it. However, the true "heart" and "soul" of an optical module are actually the optical chips hidden inside. Zhongji Innolight's rise to become the global leader in optical modules relies on its precise packaging technology to integrate various optical chips into high-performance modules; yet, the core active optical chips that determine speed and mass production yield remain firmly in the hands of upstream overseas enterprises..." datetime: "2026-08-03T02:12:34.000Z" locales: - [en](https://longbridge.com/en/topics/43145393.md) - [zh-CN](https://longbridge.com/zh-CN/topics/43145393.md) - [zh-HK](https://longbridge.com/zh-HK/topics/43145393.md) author: "[财华社](https://longbridge.com/en/profiles/11651030.md)" --- # [Optical Chip Decoding - The Foundation of Computing Power] In the AI era, why do we need optical chips? This year, $ZJ INNOLIGHT(03308.HK) has become a hot topic in the capital markets. As the world's largest optical module supplier, a significant portion of the core optical modules for AI giants such as $NVIDIA(NVDA.US) and $Amazon(AMZN.US) comes from it. However, the true "heart" and "soul" of optical modules are actually hidden within—the optical chips. InnoLight's rise to become the global leader in optical modules relies on its precise packaging technology to integrate various optical chips into high-performance optical modules. However, the core active optical chips that determine speed and mass production yield remain under the control of upstream overseas enterprises. This is precisely why leading companies like InnoLight have been aggressively investing in self-developed optical chips in recent years: only by mastering the "heart" of optical modules independently can they truly hold the initiative in competition. **What exactly do optical chips do?** The GPUs and CPUs we are familiar with daily are electronic chips responsible for computing data; optical chips, on the other hand, do not perform calculations. Their core function is to convert optical-electrical signals and manage the transmission and routing control of optical signals. When tens of thousands of GPUs are assembled into large-scale AI computing clusters, massive amounts of data need to be transmitted every moment between graphics cards and between graphics cards and switches. Traditional copper cables transmitting electrical signals suffer from severe heating, signal attenuation, and limited transmission distance when speeds are increased, making them completely incapable of handling the huge traffic demands of AI. Fiber optics transmit via light waves, offering high speed, low power consumption, and the ability to achieve medium-to-long-distance transmission, making them the core solution for computing interconnectivity. However, GPUs only recognize electrical signals, while fiber optics can only propagate optical signals; the two cannot connect directly. Optical chips play the role of translators: converting current into laser light sent into the fiber on one side, and converting the returning laser light back into electrical signals for the processor to read on the other. Beyond traditional communication interconnects, optical chips are extending into the field of "optical computing." Next-generation optical computing chips can use photons as carriers to complete matrix multiplication and addition inference operations for AI cores, possessing natural advantages of ultra-low latency and extremely low power consumption. In the future, optoelectronic hybrid computing—where GPUs handle complex logic and general computations while optical chips take on high-concurrency matrix calculations—is expected to become the ultimate solution to breaking through the bottlenecks of traditional electronic AI computing power. Optical chips are divided into two major categories, with vast differences in technical difficulty and value: **Active Optical Chips (Industry Core, Highest Barrier):** These require electricity to emit or receive optical signals, mainly including laser diode chips and photodetector chips. In the total cost of high-speed AI optical modules such as 800G/1.6T, active optical chips may account for more than half. They determine the speed and stability of computing transmission and are the core link where overseas enterprises exert chokehold pressure. **Passive Optical Chips (Lower Threshold, Basically Self-Sufficient Domestically):** These do not require electricity and are only responsible for splitting light, guiding light flow, and filtering different wavelength beams, similar to water pipe diversion valves. They are mostly used in the optical path integration of CPO (Co-Packaged Optics). Mass production has already been achieved domestically, with no supply bottlenecks. **Technical Difficulty Levels of Optical Chips: From Entry-Level to Future Top-Tier Routes** Classified progressively from low to high based on process structure, application scenarios, and speed generations, this corresponds to the industry's technological iteration and product upgrade path: **1) VCSEL Chips (Lowest Threshold: Short-Distance Internal Interconnect)** Like countless miniature spotlights, they are low-cost and low-power, but suffer from severe beam divergence. Transmission distances are usually within 100 meters, primarily used for short-distance high-speed interconnection between data center racks, servers, and switches, serving as the main force of the "capillaries" within AI computing clusters. **2) DFB Laser Diode Chips (Mid-to-Low End: Broadband, Universal for 5G Base Stations)** Spotlights equipped with focusing covers, producing focused and regular beams. Data is transmitted by adjusting current to switch light transmission. They cover all networks at 25G and below, widely used in home broadband and 5G base stations, and have been fully localized in China. The drawback is that after increasing speed, severe laser wavelength jitter caused by direct modulation leads to signal confusion, making it difficult to support the ultra-high bandwidth of 800G and 1.6T AI applications. **3) EML Electro-Absorption Modulated Lasers (Current Mass Production Technology Ceiling: Standard for AI Computing)** Currently the most core optical chip in global AI data centers, featuring the highest barriers and strongest overseas monopoly. Light emission and signal modulation work separately: the DFB light source continuously emits stable laser light, paired with an electrically controlled shutter to manage light on/off states. The laser has almost no jitter, achieving single-channel speeds of up to 100G or 200G, with transmission distances reaching 2 to 80 kilometers. An 800G optical module typically requires eight 100G EML chips, while a 1.6T module requires eight 200G EML chips. NVIDIA's (NVDA.US) entire line of AI servers adopts this solution, and its indium phosphide integrated carving process has long been controlled by the US and Japan. **4) Two Major Frontier Top-Tier Technologies for the Next Generation (Targeting 3.2T+ Speeds, CPO Architecture)** **Silicon Photonic Integrated Chips:** Leveraging mature silicon semiconductor processes, hundreds of optical components such as modulators and waveguides are integrated onto a single silicon wafer, resulting in small size and lower mass production costs. The drawback is that silicon cannot emit light, requiring an external independent indium phosphide laser light source (CW light source). It is one of the mainstream integration solutions for CPO (Co-Packaged Optics). In the future, it may allow the optical engine to be directly attached to the switching chip, solving power consumption bottlenecks. **Lithium Niobate on Insulator (LNOI) Modulation Chips (Bandwidth Performance Ceiling):** The modulation rate is twice that of Indium Phosphide EML, with lower power consumption and minimal signal loss, suitable for future 3.2T and 6.4T ultra-high-speed optical modules. The disadvantage is that thin-film precision processing is extremely difficult; the current stage is a critical period transitioning from small-scale commercialization to large-scale mass production. **Conclusion** From basic optoelectronic conversion and computing cluster interconnects to frontier optoelectronic hybrid computing, optical chips have become the underlying rigid foundation for AI computing iteration. Meanwhile, the clear technical stratification reveals that while China has achieved autonomous controllability in basic optical chip fields such as low-speed and passive chips, the core capacity and 话语权 (discourse power) for the EML high-speed active chips supporting current 800G/1.6T high-end AI computing, as well as the 3.2T+ frontier optical chip technologies facing the future, remain firmly in the hands of overseas manufacturers. This gap in technological generations and the misalignment of upstream and downstream industrial capabilities are not merely issues of single processes, but comprehensive barriers involving the entire industrial chain division of labor, material processes, and wafer fabrication systems. Accompanied by the extreme supply-demand shortage brought about by the explosion of AI computing power, this presents a rare window for domestic substitution for Chinese manufacturers. However, it also raises core questions: How is the complete optical chip industrial chain divided? Which key links are we stuck at? What is the core difference between optical chips and traditional silicon wafers? We will dismantle the barriers of the entire optical chip industrial chain layer by layer in the second part of this special report, clarifying the real shortcomings and breakthrough dilemmas of the domestic industry. Author: Wu Yan ### Related Stocks - [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md) - [03308.HK](https://longbridge.com/en/quote/03308.HK.md) - [300308.CN](https://longbridge.com/en/quote/300308.CN.md) - [AMZN.US](https://longbridge.com/en/quote/AMZN.US.md) - [NVDL.US](https://longbridge.com/en/quote/NVDL.US.md) - [07788.HK](https://longbridge.com/en/quote/07788.HK.md) - [07388.HK](https://longbridge.com/en/quote/07388.HK.md) - [NVDY.US](https://longbridge.com/en/quote/NVDY.US.md) - [NVDD.US](https://longbridge.com/en/quote/NVDD.US.md) - [NVDX.US](https://longbridge.com/en/quote/NVDX.US.md) - [NVDQ.US](https://longbridge.com/en/quote/NVDQ.US.md) - [AMZU.US](https://longbridge.com/en/quote/AMZU.US.md) - [NVD.DE](https://longbridge.com/en/quote/NVD.DE.md)