High-Speed Printing Breakthrough Accelerates the Production of Energy-Efficient AI Chips
Beijing, Friday, 17 July 2026.
Researchers successfully printed four million optical neurons on a one-square-millimeter chip in fifteen minutes, enabling the cost-effective mass production of energy-efficient AI vision processors.
A Paradigm Shift in AI Hardware Manufacturing
On July 16, 2026, researchers led by Prof. Shih-Chi Chen and Prof. Chaoran Huang from The Chinese University of Hong Kong published a landmark study detailing a high-throughput manufacturing platform for optical neural networks (ONNs) [1]. Published in the journal Light: Advanced Manufacturing, the study introduces a multi-focus two-photon lithography (TPL) platform designed to overcome the physical and economic barriers of traditional silicon-based electronic chips [1]. As artificial intelligence applications expand globally, traditional silicon chips face growing limitations in handling massive AI demands [4]. Computing directly with light via integrated photonic neural networks offers a promising path to bypass these hardware limitations by significantly relieving bandwidth, latency, and energy constraints [3].
Overcoming Energy and Architectural Bottlenecks
At the core of this breakthrough is a highly efficient, task-agnostic optical encoder that utilizes a 3D-printed diffractive layer for untrained random projection [1]. This design offloads the heavy computational lifting to the optical domain, meaning that task adaptation is achieved exclusively by retraining a lightweight digital readout layer [1]. This digital readout layer requires only 1,000 weights, which drastically reduces the energy consumption and computational overhead compared to conventional electronic processors [1]. Such architectural innovations are critical as researchers push to transition neuromorphic photonics from isolated device classifications to full-system, real-world utility [3].
Commercial Scalability and the Competitive Landscape
The ability to rapidly print optical neural networks could disrupt the competitive landscape of AI hardware, where companies are actively seeking alternatives to traditional silicon [GPT]. For instance, startups like NeoOptics, Inc., led by founder Charles Dove, are already working to build ultrathin metasurface optics to replace bulky lens stacks with application-tuned optical alternatives [2]. While NeoOptics focuses on transforming consumer electronics, healthcare, and autonomous sensing through wider fields of view and flatter camera modules [2], the high-speed printing technique developed in Hong Kong could drastically lower the production costs of similar optical systems [1]. This convergence of design and manufacturing advances signals a broader shift toward commercially viable photonic computing [1][3].