Author: Peng Cheng
The competition among AI (Artificial Intelligence) large models is pushing data center interconnects to their physical limits. As cluster scales evolve from 10,000 GPUs to 100,000 or even 1,000,000 GPUs, the bandwidth demand for optical interconnects is expanding exponentially: the current industry status quo is that 800G is already in commercial use at scale, 1.6T is beginning to ramp up, and the 3.2T roadmap is clear. Meanwhile, 3.2T after 1.6T equals 8×400G, while traditional silicon photonics faces the dilemma of "having no ready-made solutions" at 400G/lane, making TFLN (Thin Film Lithium Niobate) widely recognized as the most promising candidate. On the electrical interconnect side, the processing speed of DSP (Digital Signal Processor) can no longer keep up with the bandwidth expansion, and the power reduction dividend of each SerDes generation is narrowing day by day. "Copper retreating and optics advancing" is transforming from a slogan into an architectural reality: NPO (Near-Package Optics) and CPO (Co-Packaged Optics) are moving the optical engine from the front panel to the vicinity of the switching chips and GPUs, trading shorter electrical traces for higher bandwidth density and lower power consumption.
With these two threads intertwining, the CIOE (China Optoelectronic Expo) in Shenzhen this September is becoming the optimal window to observe the optical chip industry chain.
01 Light Source: Three Technical Routes
The laser is the heart of the optical module. At this expo, light source companies mainly focus on three technical routes.
The first route is VCSEL (Vertical-Cavity Surface-Emitting Laser). Based on GaAs (Gallium Arsenide) materials, it supports direct on-off modulation and emits multimode light. Its rate ceiling is relatively low, but it offers the lowest cost, primarily targeting short-distance applications. Huaxin Semiconductor showcased 850nm 10G/25G VCSEL chips and 1×4 56G/112G/224G VCSEL arrays, announcing that the cumulative delivery of chips running on the network has exceeded 100 million. The real highlight lies in the 224G, which is key to the 200G/lane multimode solution. Huaxin's mass production timeline is set for the second half of 2027, basically in sync with international giants.
The second route is EML (Electro-absorption Modulated Laser). According to introductions, EML integrates a DFB (Distributed Feedback) laser and an EA (Electro-absorption) modulator on the same chip. It requires no external modulator and can achieve optical signal modulation by directly loading electrical signals. It is suitable for pluggable optical modules and short-distance transmission scenarios in AI cabinets. However, a single unit only supports a one-to-one configuration and cannot achieve multi-channel one-to-many deployment. Dugen Core launched 1271/1291/1311/1331nm 100G PAM4 CWDM EML chips and supporting COC (Chip on Carrier) devices. The modulation bandwidth exceeds 40GHz, targeting AI data center 400G-FR4 and 800G-FR8 scenarios. Relying on its independent IDM (Integrated Device Manufacturer) full-chain platform, it simultaneously advances in bandwidth, extinction ratio, linear modulation, and spectral stability. Its external modulation characteristics avoid the chirp effect of direct laser modulation, ensuring signal quality for long-distance transmission. It is reported that they are also developing 400mW VCSEL/DFB light sources. Suzhou Dingxin, on the other hand, displayed the 100G/200G EML series, adopting an active butt-joint scheme to independently optimize the DFB and EA parts, achieving both high power output from the DFB and high modulation rate from the EA. It uses aluminum-containing material quantum wells, ensuring a high extinction ratio on the basis of high bandwidth through reasonable design. In addition, it also exhibited two ingeniously designed derivative products: the 50G Tunable EML can output light of different wavelengths by adjusting the grating refractive index. It can serve as a universal backup to reduce spare parts inventory costs in multi-wavelength scenarios, and directly replace a faulty product at a specific wavelength by adjusting the wavelength; the 56G EML+SOA integrates an SOA (Semiconductor Optical Amplifier) into the laser chip link to provide gain compensation for the weak optical signal after modulation.
The third route is high-power CW DFB. Silicon materials do not emit light themselves. In the silicon photonics era, the light source has evolved from "one laser per channel" to "one laser feeding multiple channels," making power the new bottleneck. HiSilicon showcased high-power CW (Continuous Wave) light source chips: the DRx series for DR4/DR8 parallel single-mode provides a complete power gradient of 400/200/130/100/80mW, and the one for 800G 2×FR4 provides 100/80mW, covering two mainstream high-speed module architectures with power tiers. Suzhou Dingxin's high-power CW lasers (70/100/200mW) adopt a low-cost ridge waveguide integrated passive butt-joint scheme, effectively reducing manufacturing costs and increasing capacity while ensuring high product reliability. Meanwhile, the divergence angle can be tuned through butt-joint material design to improve coupling efficiency. Huaxin Semiconductor's CW laser chips have been sampled in 2026 and will be mass-produced in 2027. CW lasers do not directly undertake modulation but only provide continuous light, and the high-speed electro-optical conversion is then completed by the silicon photonics modulator.
02 Modulator: Structural Contention, Yield Dilemma, and Material Transition
The light source solves "where the light comes from," while the modulator solves "how information is loaded onto the light." It is the most technologically intensive segment in optical chips. In principle, modulators are mainly divided into two structures: MZM (Mach-Zehnder Modulator) and microring. MZM offers robust performance and good linearity, but it is relatively large in size, difficult to integrate, and has higher power consumption. Microrings are compact, power-efficient, and naturally adaptable to multi-wavelength DWDM (Dense Wavelength Division Multiplexing), making them an ideal choice for NPO/CPO multi-channel optical engines. However, they are highly sensitive to processes, have large process deviations, and are extremely susceptible to temperature. As the industry moves towards NPO/CPO, the architectural balance is tilting towards microrings.
The primary constraint on microrings is yield. According to introductions, the current domestic yield for microring processes is only about 10%, and even ordinary overseas fabs fall short of 50%. The maskless programmable lithography system for silicon photonics exhibited by Guanglian Xinke could be a key variable. Unlike general-purpose laser trimming equipment on the market, this device is designed exclusively for microring trimming scenarios. It can trim the microring resonance peak position to within 10 pm, far superior to the inherent 1nm-level offset error of TSMC (Taiwan Semiconductor Manufacturing Company)'s own process. Distinct from the industry's common practice of screening good dies and discarding defective ones, this device can adjust all repairable microrings to a state that meets usage standards. In actual use, the average trimming time for a single microring is about ten-plus seconds, and the total time for trimming microrings across a full 12-inch wafer is approximately 5 hours. Moreover, microring devices trimmed by this equipment can eliminate the subsequent heater correction step in the TSMC process. This equipment breakthrough provides a systematic solution to the biggest mass production obstacle for the microring route for the first time. In addition to microring trimming technology, Guanglian Xinke has also developed supporting wavelength locking technology. This technology dynamically adjusts the temperature through heating resistors placed immediately adjacent to the microrings, with supporting feedback loops to counteract external temperature fluctuations and thermal diffusion interference between microrings. The detection loops, feedback algorithms, and driving circuits required for temperature measurement and control are all integrated inside the electrical chip, requiring no additional external independent modules. This can effectively solve the wavelength drift problem caused by the temperature sensitivity of microring devices.
Beyond structure, material is a larger variable for modulators. Silicon photonics modulation can still handle 200G and below, but when it comes to single-wave 400G and above, the silicon photonics solution is no longer feasible. TFLN is regarded as the core candidate to replace silicon photonics modulators in the 400G/lane (3.2T era). Domestic Niobate Optoelectronics is a leading IDM in this direction: its coherent modulator products adopt a DP-QPSK (Dual-Polarization Quadrature Phase Shift Keying) scheme and have been applied in long-distance transmission and backbone network scenarios. Facing data center scenarios, Niobate Optoelectronics exhibited products such as 400G IMDD, 1.6T DR8 IMDD, and 800G DR4 IMDD, which are currently in the small-batch validation stage and will gradually advance to mass production in the future.
However, the bottleneck determining the pace of deployment is capacity. Niobate Optoelectronics has currently built a 6-inch wafer mass production line and possesses 8-inch mass production capabilities. However, the expansion pace is constrained by equipment procurement, debugging cycles, and upstream supply chain limitations, making it impossible to ramp up rapidly in the short term—the factory construction cycle is rigid and cannot accommodate sudden large-scale orders in the short term. There is generally a two-to-three-year time lag between hype and capacity realization, which is a common realistic constraint across the entire TFLN sector: the direction is confirmed, but realization takes time.
03 From Chips to Systems: NPO and All-Optical Switching
Light sources and modulators constitute the device foundation of optical modules, but advancements at the device level must ultimately be realized at the system architecture level. According to introductions, 200G/lane technology matured last year. Currently, 800G optical modules generally still use 8 parallel 100G lasers; the 1.6T module adopting an 8-channel 200G scheme has high maturity; the 400G/lane scheme is expected to mature technologically in the first half of next year. Relevant discussions have been fermenting since last year's OFC (Optical Fiber Communication Conference), and public demonstrations are already available. However, the physical cost of 400G/lane is that transmission loss increases with the square of the rate. Existing transmission distances cannot adapt to current equipment architectures. It is necessary to shorten the distance between switching chips and optical modules, upgrade DSPs capable of processing long-distance high-bandwidth signals, or simply change the architecture.
HiSilicon exhibited a 7.2T/6.4T NPO optical engine solution with a built-in light source. According to introductions, this product adopts a SiPh (Silicon Photonics) single-mode plus built-in light source scheme, with power consumption of less than 30W. Among them, the 6.4T version targets switching scenarios with a DSP-free design, reducing insertion loss by shortening device spacing; the 7.2T version targets GPU scenarios, integrating the light source and modulator to eliminate additional insertion loss components such as external polarization-maintaining fibers and connectors; all optical chips have built-in digital diagnostic reporting interfaces, supporting the return of operational status data to facilitate rapid fault localization during the operation and maintenance phase. Currently, this solution is still in the sample stage and has not yet achieved large-scale mass production.
In addition, HiSilicon also showcased its layout roadmap for silicon photonics chips in computing centers. Its 400G/800G→800G/1.6T→1.6T/3.2T optical chips correspond to a three-step leap in single-channel rates from 100G→200G→400G, advancing along both DR parallel and FR wavelength division lines. The supporting laser solutions evolve from High Power CW DFB to TFLN modulators.
Haiguang Xinzheng is also completing this puzzle from the chip side. It not only brought an NPO optical switching system solution but also displayed supporting NPO silicon photonics transceiver chips. Its newly launched "Lingti" chip is a 32-channel NPO silicon photonics transceiver chip with a single-channel rate of no less than 200Gb/s, targeting 3.2T/6.4T NPO optical modules. The chip adopts a high-channel-density integrated transceiver design, supporting advanced packaging with high copper pillars or RDL (Redistribution Layer) + TMV (Through Mold Via), which can significantly shorten high-speed electrical interconnect distances. Meanwhile, it achieves excellent low-crosstalk characteristics through packaging architecture optimization, possessing advantages such as high bandwidth density, low power consumption, small size, high integration, and good signal integrity.
At the network level, another thread is OCS (Optical Circuit Switching). Traditional data center switching relies on electrical switching chips, where the optical signal at each port must undergo electro-optical conversion. In contrast, OCS completes routing switching directly at the optical layer without ever dropping to the electrical domain. Bopu Semiconductor, in addition to its silicon photonics IP licensing business, has launched its self-developed OCS all-optical switch. It adopts a silicon photonics waveguide plus Mach-Zehnder interferometer thermo-optic switch scheme. The whole machine power consumption is about 10W, and the switching command response time is 100-300 microseconds. With no mechanical parts, its response speed is much faster than traditional MEMS (Micro-Electro-Mechanical Systems) solutions, representing a leading level domestically for this route. Internationally, Google has already applied similar products, and the Spanish manufacturer iPronics is its international supplier. What Bopu Semiconductor exhibited this time is the already launched 32*32 port version. A 64*64 port version will be released in March next year, and a 128*128 port version will be launched by the end of next year. It is understood that Bopu Semiconductor's OCS all-optical switch started R&D last year, and industry demand rose synchronously last year. Currently, domestic customers are in the product validation stage, and large-scale commercial deployment has not yet occurred. The OCS all-optical switch is a brand-new business line currently being heavily promoted, with plans to formally advance full-scale commercial deployment starting next year.
All-optical switching and NPO are complementary rather than substitutive. NPO optimizes how light enters and exits the chip, while OCS optimizes how light flows within the network. When electricity can no longer keep up with light, the transformation will not stop at the device level but will propagate upwards along the chain of chips, modules, and switching systems.
However, between "samples" and "mass production" at the expo, there are hurdles such as yield, reliability, and customer validation. Whether it is HiSilicon's 7.2T NPO, Huaxin's 224G VCSEL, or Niobate's data center IMDD, none have reached the stage of scaled delivery. But the narrative shift is already undeniable. When electricity can no longer keep up with light, optical chips are no longer supporting actors in computing infrastructure but the protagonists determining how large clusters can scale and how deep models can be trained. This massive industrial migration from electricity to light has only just begun to unfold its curtain.