Standing at the forefront of the AI computing power boom, Moore’s Law is approaching its physical limits, and traditional electrical interconnects and electronic computing are running headlong into an insurmountable “power wall” and “bandwidth wall.” Against this industry backdrop, Lightelligence, a global unicorn in the field of silicon photonic chips, has launched an IPO on the Hong Kong Stock Exchange — an event that is not only a focal point for capital markets but also an important signal of a foundational paradigm shift in semiconductors.
Financial Profile: The Logic of “Losses” and “Growth” Driven by Heavy R&D Investment
As a “Specialist Technology Company,” Lightelligence’s financial statements exhibit the typical “high input, high barriers, high elasticity” characteristics of the hard technology track.
Revenue growth and structural optimization: From 2023 to 2025, Lightelligence’s revenue was RMB 38 million (≈ USD 5.28 million), RMB 60 million (≈ USD 8.33 million), and RMB 106 million (≈ USD 14.72 million) respectively, representing a compound annual growth rate of 66.9%. Although the absolute figures remain at the hundred-million-yuan scale, optical computing business revenue in 2025 reached RMB 20.2 million (≈ USD 2.81 million), accounting for approximately 21% of total revenue, a year-on-year surge of 579%, and has become the company’s second growth curve.
“Paper loss” optimization: Although the annual losses for 2023 to 2025 were RMB 413 million (≈ USD 57.36 million), RMB 735 million (≈ USD 102.08 million), and RMB 1.342 billion (≈ USD 186.39 million) respectively, these are primarily composed of two non-operating, non-cash items: changes in the fair value of financial instruments issued to investors (recorded as an increase in liabilities as the company’s valuation rises) and share-based compensation expenses. After adjusting for these items, the Non-IFRS adjusted net losses were RMB 264 million (≈ USD 36.67 million), RMB 296 million (≈ USD 41.11 million), and RMB 271 million (≈ USD 37.64 million) respectively. This means that as revenue has grown, operating leverage has begun to take effect, and the true operating losses have not spiraled out of control with the expansion of scale.
R&D investment continues to flow toward foundational technology high ground: The company’s R&D expenditures for 2023 to 2025 were RMB 280 million (≈ USD 38.89 million), RMB 352 million (≈ USD 48.89 million), and RMB 479 million (≈ USD 66.53 million) respectively. The two primary application areas are: 1) TSV (Through-Silicon Via) packaging technology, which is the core enabler of silicon photonic chip commercialization. By drilling holes through silicon wafers and filling them with metal, vertical electrical connections between chip layers are achieved. Through this technology, Lightelligence has achieved the integration of more than 40,000 photonic devices on a single chip, greatly improving signal integrity and reducing module volume by 90%. 2) PACE 3 (Photonic Arithmetic Computing Engine), as Lightelligence’s third-generation core architecture, its R&D objective is to achieve matrix operations for trillion-parameter models (1,000B+), with significantly optimized stability in optical path design compared to its predecessor generation. The successful tape-out of PACE 3 marks the point at which optical computing has formally acquired the capability for large-scale commercial inference.
In-Depth Business Breakdown: The Synergistic Effect of Opto-Electronic Integration
Lightelligence has built a complete computing power ecosystem through “optical interconnects + optical computing.”
Optical Interconnects: “Expanding at the Speed of Light” on the Computing Highway
The technical principle involves using silicon photonic technology to achieve high-speed data transmission between chips (Chip-to-Chip) and between racks, integrating the optical engine and electrical chips in the same package through CPO (Co-Packaged Optics) technology to reduce the losses associated with electrical signal transmission over copper wiring. Its core product, the Photonic Leap 128 (光跃128) system, has successfully resolved the communication barrier between GPUs, reducing the transmission latency of clusters of tens of thousands of cards by more than 90%.
Optical Computing: Doing Mathematics With “Light Waves,” Reshaping the Computing Foundation
The technical principle is based on the Mach-Zehnder Interferometer (MZI) principle, using the interference and diffraction of photons in waveguides to directly complete matrix operations — this is fundamentally “replacing electricity with light.” In scenarios such as deep learning inference, optical computing demonstrates an energy efficiency ratio far exceeding that of conventional GPUs due to its low latency and near-zero computational power consumption. For example, the OptiHummingbird acceleration card is specifically optimized for the Transformer architecture (such as the underlying architecture of ChatGPT), delivering approximately 15 times the performance improvement over conventional GPU plus HBM solutions.
Investment Highlights: Lightelligence’s Breakthroughs
As large AI models scale toward trillion parameters, conventional electrical signal transmission is facing severe bandwidth reduction due to resistive heat generation and electromagnetic interference — the so-called “power wall” and “bandwidth wall.” Silicon photonic technology, leveraging the zero rest mass properties of photons, is currently the industry-acknowledged definitive pathway for breaking through the computing power bottleneck. Lightelligence’s capability to integrate 40,000-plus photonic devices constructs an extremely high barrier to entry.
On the other hand, Lightelligence is standing at the intersection of domestic substitution and national strategy. According to the BCC Research, the state is promoting the construction of AI training centers in five cities including Shanghai and Beijing. For example, in the Shanghai training center being established by the Yidian Group, Lightelligence has become a key partner in the ten-thousand-card cluster testing, leveraging its technological advantages in optical modules and PCIe switches. Its optical modules are used not only for intra-server connections but also for inter-server interconnects, significantly improving the efficiency and stability of large-scale domestically produced AI chips (such as those from Tianshu Zhixin, Biren, and Muxin) in supporting large model training.
As model scale advances from tens of billions of parameters (32B/72B) toward hundreds of billions (1,000B) and even trillions, electrical interconnects can no longer support ultra-node configurations above 400 GB/s. With an 88.3% market share in China’s independent Scale-up optical interconnect market, Lightelligence possesses extremely strong flexibility and commercial influence in addressing the AI computing power predicament.
Future Outlook: The Silicon Photonic High Ground in the Era of Large Models
Looking ahead, Lightelligence and the silicon photonics industry in which it operates are on the eve of an explosion:
Ultra-node trend: The BCC Research predicts that physical ultra-node configurations of 288 cards or even larger scales will emerge in the future. As issues of heat dissipation and stability are progressively resolved, optical interconnects will become the key to overcoming physical limitations.
Domestic substitution and industrial co-opetition: Against the backdrop of constraints on domestically produced computing power, using silicon photonic technology to achieve a “curve-overtaking” strategy and increase the proportion of domestically produced AI chips applied in training centers has become a clear policy direction.
Blurring of hardware boundaries: As cloud vendors such as Google begin to develop their own chips and optical switching technologies in-house, the “boundary delineation” between Lightelligence and AI chip manufacturers will become clearer — that is, Lightelligence provides the ultimate transmission and auxiliary computing foundation, defining clear boundaries of cooperation.
In the domestic computing power track, Huawei is an unavoidable competitor. Huawei’s strength lies in its full-stack ecosystem (Ascend chips plus proprietary switches), primarily targeting Scale-out (large-scale horizontal expansion). Lightelligence, as an independent supplier, has an optical interconnect solution that can broadly accommodate various domestically produced and international GPU brands. In the sub-track of Scale-up (upward expansion within nodes), it currently holds an advantage in response speed and adaptability through its vertically integrated silicon photonic technology.
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.