As the wave of digitalization reshapes global industry, the explosive growth of artificial intelligence applications is redrawing the boundaries of productivity at unprecedented speed. Computing power, the core infrastructure behind that shift, is now facing a widening supply-demand strain that is becoming a major bottleneck for continued industrial upgrading.

As Moore’s Law pushes traditional electronic chips closer to physical limits, performance gains have slowed sharply. AI computing power demand, by contrast, doubles every 3.4 months. That imbalance is driving a surge in data center energy consumption. The International Energy Agency’s Energy and AI report cited OpenAI’s GPT-4 as an example: during a 14-week model training run, it consumed 42.4 GWh of electricity, or 0.43 GWh per day, roughly equivalent to the average daily power use of 28,500 households in Europe and the United States. Globally, data centers consumed 415 TWh of electricity in 2024, accounting for 1.5% of global power use, and that figure is expected to more than double to 945 TWh by 2030.

128×128 matrix-scale optical computing chip

Against this backdrop, optical computing is gradually entering the industry’s field of view as a new computing paradigm. By using light as the carrier of information, it brings inherent advantages including light-speed transmission, high computing power and low power consumption, making it a promising path for breaking the computing power bottleneck. At this critical point in the computing power revolution, Guangbenwei Technology emerged with an innovative technical route based on the heterogeneous integration of silicon photonics and phase-change materials, positioning itself as an important force in bringing optical computing into commercial use. Founded by two post-95 returnees from overseas study, the company successfully taped out the world’s first 128×128 matrix-scale optical computing chip in June 2024. The chip not only broke through the matrix-size bottleneck that had long constrained the optical computing industry, but also used an in-memory computing architecture to push optical computing from the lab toward product-level applications, opening a new computing paradigm for the next generation of computing power competition.

An Eight-Year Pact Between Two Young Technologists

Xiong Yinjiang, then 18, and Cheng Tangsheng, then 17, met while volunteering as teachers in a Qiang village in Nanbaoshan, Sichuan. At night, squeezed together on joined bed boards, they talked. When Cheng said he wanted to start a technology company, he gave voice to an idea Xiong had long kept buried. The seed planted by that meeting quietly grew in the years that followed.

Three years later, the two again worked together on an ecological restoration project in Inner Mongolia, leading teenagers into the field to experience how technology could change the environment. Although the project was educational in nature, the experience further strengthened their shared belief in using innovative technology to confront real problems: a truly meaningful startup had to address core industrial needs and achieve a breakthrough from zero to one.

In the following years, they pursued separate academic paths. Cheng went to the University of Oxford, studying under Harish Bhaskaran, a Fellow of the Royal Academy of Engineering, and led or participated in research on phase-change-material optical computing chips and new ultra-low-power nanoscale phase-change materials. Xiong focused at the University of Chicago on AI algorithm development and commercialization, taking part in early large-model inference and training and directly experiencing the limits of existing computing paradigms in computing power and energy efficiency.

The turning point came in 2021. Cheng made a lab breakthrough, using phase-change materials to enable large-scale matrix photonic in-memory computing. Almost at the same time, Xiong found in practical work that AI training demand for computing power was surging, and the energy cost of existing computing paradigms had become a clear weakness. In frequent transoceanic calls, they realized that optical computing’s thousandfold energy-efficiency advantage over electronic chips in matrix operations aligned closely with AI’s massive need for computing power.

On the basis of that technical insight and market judgment, they chose the track of “optical computing + AI.” The decision rested on two considerations: first, global technology routes had not yet been settled, while China’s accumulated strengths in the optical communications supply chain gave it an early-mover advantage; second, optical computing had a high enough ceiling and was by no means limited to a niche, overcrowded market.

Cheng Tangsheng and Xiong Yinjiang returned to China one after the other, and Guangbenwei Technology was formally established in April 2022. In the company’s early days, the two reached a shared view: “What we want to do is not short-term monetization, but to push optical computing from the lab into industrialization and become a key participant in the next computing power revolution.”

The First Tape-Out Determined Whether the Project Would Survive

In its earliest stage, Guangbenwei Technology faced a decisive test: its first tape-out would determine whether the project could continue. When the company launched in 2022, its funding was enough to support only one tape-out. Starting in April, the team threw itself into chip design, spending three to four months just completing simulations and design plans. With limited conditions, the two founders and core R&D staff stationed themselves in the lab. They ultimately completed functional verification of a small-matrix chip, ensuring that both key devices and the overall result landed safely.

“Semiconductors are a long-cycle industry. In the early stage, you have to prove not only that the technology is feasible, but also that the market can believe in its commercial value,” Cheng Tangsheng said, capturing the dual challenge of a hard-tech startup. Once the direction was chosen, the real fight had only begun. Matrix scale and the performance of core devices are key factors shaping an optical chip’s computing power and computational efficiency. The team carried out comprehensive optimization across phase-change materials, core optical devices and optical chip architecture. It not only reduced transmission loss in the optical chip and shrank device size, but also significantly improved computational efficiency and accuracy. Cheng explained: “Expanding the matrix scale of an optical chip can directly increase computing power and computational efficiency, but it also brings higher optical loss and challenges around the stability and precision of on-chip optical devices. The overall system can operate efficiently only when even the shortest stave in the barrel meets the requirement.”

Demonstration of the Crossbar technology route

While expanding matrix scale, the team used an innovative Crossbar photonic matrix computing structure to improve chip-area utilization. The design allowed the 128×128 chip to integrate more than 16,000 nodes, with each node programmable in real time, breaking the limitation of traditional optical chips whose fixed weights can handle only a single task. In June 2024, the chip was successfully taped out, becoming the world’s first optical computing chip to meet commercial standards. The industry regards 128×128 as the “critical point” for optical computing commercialization because any matrix scale below 128×128 lacks sufficient computing power and density to support complex applications such as large-model inference and training.

Laying the Groundwork for Product Commercialization

Guangbenwei Technology is now taking a key step toward product commercialization. Cheng Tangsheng said the company’s first-generation optoelectronic hybrid computing card will soon be sampled to downstream users. At the same time, tape-out plans for 256×256 and larger matrix-scale optical computing chips are advancing rapidly. More importantly, the company has already built a complete optical computing product system.

In supply-chain collaboration, Guangbenwei Technology has chosen a “two-way rooting” strategy. Upstream, it works closely with multiple domestic silicon photonics production lines, covering everything from 8-inch to 12-inch lines, and has secured advanced manufacturing resources in advance to ensure process stability and supply-chain control. Downstream, it is jointly developing customized products with internet giants while also participating in standardized local-government AI data center projects.

Guangbenwei Technology’s optoelectronic hybrid computing card

This forward-looking commercialization strategy is also reflected in Guangbenwei Technology’s financing approach. Xiong Yinjiang believes that, in the long-cycle optical computing track, it is especially important to work with capital that understands the laws of technological evolution and shares a commitment to long-term value. In his view, investors with industrial insight and a long-term perspective can better accompany a company through the full process from technology R&D to scaled commercial deployment.

In June 2023, Guangbenwei Technology received angel-round investment from Yunqi Partners, FreeS Fund, Xiaomiao Langcheng, MiraclePlus and others. Eight months later, it completed an angel+ round, accelerating the 128×128 chip tape-out process. In December 2024, the company reached a strategic partnership with a domestic internet giant with extensive application-scenario resources, enabling deeper ecosystem coordination. In June 2025, a new round led by Dunhong Asset Management and joined by state-backed funds including the Pudong Technology Angel Fund of Funds further integrated industrial-chain resources in Shanghai, Suzhou and other places, providing solid support for the subsequent production ramp. This means Guangbenwei Technology has taken the lead in building full-chain capabilities in optical computing, from materials and design to manufacturing and application.

A Chance to Help Define a New AI Computing Paradigm

For now, optical computing chips still rely on electrical drive, with light and electricity working together through analog chips: optical chips handle AI linear operations, while electronic chips are responsible for scheduling and nonlinear processing, forming a tightly coordinated optoelectronic hybrid computing system.

Guangbenwei Technology’s optoelectronic hybrid computing system

“We hope to let light participate more and more in the entire computing architecture,” Xiong Yinjiang said in an interview. Although “all-optical computing” remains a long-term vision, at this stage the clear direction is to keep increasing the share of optical computing and use optoelectronic collaboration to continuously improve computational efficiency and reduce system energy consumption. That has become an inevitable path for building the next generation of green, efficient AI data center infrastructure.

A breakthrough in optical computing could turn low-carbon or even zero-carbon AI large-model inference and training from fantasy into reality. China may have a chance to overtake on a new track in the contest to define the next AI computing paradigm.

In the near future, AI data centers may no longer need massive amounts of electricity to stay online. Their cooling pressure could be sharply reduced, operating noise significantly lowered, and they could evolve into efficient “urban digital hearts” that coexist with cities in a more friendly way. In autonomous driving, the nanosecond-level processing speed of optical chips could help vehicles parse road-condition data faster. In complex scenarios such as slippery roads in the rain or night driving, the system’s environmental perception and decision response would become more precise and efficient, adding an important layer of safety. In medical imaging centers, systems powered by optical computing could accelerate model reconstruction and analysis, making more inclusive healthcare built on “early detection and early diagnosis” easier to achieve.

The industrialization of optical computing is not the endpoint, but the starting point of a new AI technology revolution. It will open an intelligent era of abundant computing power and controllable energy consumption. When we no longer have to pay a steep energy and environmental price for computing power, the human value of the intelligent era will be released as well.