Foreword:

When the competition for AI computing power extends beyond data centers and power grids, it begins to approach new physical boundaries.

The launch of optical computing into space by Photonics-Based Technology and Dongfang Tiansuan marks a crucial leap for China's space-based computing from conceptual validation to engineering closure.

Author | Fang Wensan

Image Source | Internet

Optical Computing in Space: Changing the Deployment of Computing Power

Recently, Dongfang Tiansuan and Photonics-Based Technology jointly established the Space-Based Optical Computing Innovation Center and initiated the development of the world's first optical computing satellite and the world's first space-based optical computing payload. This move signifies the transition of optical computing from large-scale model applications to space-based engineering applications.

The value of space computing power lies in enabling data to undergo initial judgment at the point of generation. This collaboration integrates two technological lines onto a single satellite payload, meaning optical computing must undergo validation in the space environment. It also signifies that space-based computing is seeking a more suitable underlying chip route for space constraints than traditional electronic computing.

Low-Earth orbit satellites, remote sensing constellations, inter-satellite laser communications, and on-orbit intelligent services are pushing space from a "data collection layer" to a "data processing layer."

As the number of satellites increases, the model of transmitting all massive raw data to the ground for computation will be constrained by link bandwidth, latency, ground station resources, and task response speed.

The industrial logic of on-board AI inference is thus established. Tasks such as image recognition, abnormal target screening, disaster monitoring, inspection of marine and energy facilities, and emergency communication scheduling can reduce downlink pressure and improve response efficiency when filtered at the orbital end as early as possible.

AI computing power itself is also facing energy constraints. The International Energy Agency predicts in "Energy and AI" that global data center electricity consumption will reach approximately 945 TWh by 2030, more than doubling from around 415 TWh in 2024, with AI being a significant driver of this growth.

This does not mean that space computing will quickly replace ground-based intelligent computing centers, but it does indicate that computing infrastructure is evolving from a single ground-based data center model to a multi-layered collaboration among cloud, edge, end, and satellite.

Dongfang Tiansuan Handles Space Engineering, Photonics-Based Technology Provides the Computing Foundation

The collaboration between Photonics-Based Technology and Dongfang Tiansuan has progressed from joint announcements to engineering implementation, with clear division of labor along the industrial chain.

Dongfang Tiansuan oversees payload development, space radiation hardening, efficient thermal control, energy adaptation, and on-orbit environmental verification, covering the entire process from payload development, system integration, satellite assembly, to on-orbit verification. Photonics-Based Technology provides the optical computing chip architecture, computing power engines, and software ecosystem, laying a solid technological foundation for space-based optical computing payloads.

The former focuses on space engineering entry points, while the latter concentrates on novel AI computing cores. Only through collaboration can they form verifiable, iterable, and scalable space computing power product forms.

Dongfang Tiansuan is an industrialization platform for the Space Computing Joint Laboratory established under the guidance of the Shanghai Science and Technology Commission. It aims to explore industry-academia-research integration and accelerate the industrialization of space-based computing networks, playing a crucial role in Shanghai's space computing industry layout.

Behind this project lies a systematic experiment jointly promoted by local future industries, research institutions, commercial aerospace, and AI computing power enterprises.

A broader industrial ecosystem is being simultaneously established. The Yangtze River Delta Space-Based Computing Innovation Consortium is in the preparatory stage, with Dongfang Tiansuan collaborating with over twenty entities, including Shanghai Jiao Tong University, Photonics-Based Technology, Alibaba Cloud, and Moore Threads, to tackle seven core directions such as radiation-hardened computing chips and novel space-based energy sources, covering all key links of space-based computing power from standalone payloads to network services.

The long-term competition in domestic space-based computing ultimately hinges on the system closure capability from chips, payloads, satellites, links, platforms, to applications.

Photonics-Based Technology's optical computing route has undergone multiple rounds of validation in ground scenarios. Since its establishment in 2022, the company completed tape-outs of two optical computing chips in 2023, launched optoelectronic co-packaged chips and board prototypes; successfully tape-outed a 128×128 optical computing chip in 2024; and plans to advance the tape-out of a 512×512 optical computing chip in 2026, iterating optoelectronic fusion computing cards.

Its first-generation optoelectronic fusion computing card adopts a phase-change material route, enabling direct storage of AI model parameters within the computing card, eliminating the need for frequent read/write operations to external storage in traditional architectures, and reducing computing latency to one-tenth of traditional solutions. It has already secured orders from domestic large-scale model enterprises and has been deployed in financial vertical large-scale models.

These low-power characteristics precisely match the stringent constraints of space-based scenarios. In the space environment, power consumption, weight, data transfer, and temperature rise all contribute to system engineering costs.

Optical computing chips, with their low power consumption and minimal heat dissipation requirements, can compress the volume and weight of on-board equipment, reducing the scale of solar panel configurations and launch payload burdens.

Currently, both parties have initiated the development of optical computing payloads, with single-card computing power reaching 300 TOPS, supporting multi-precision inference. They plan to integrate inter-satellite laser communications to enable efficient data interaction within and between satellites, supporting the construction of a distributed space-based computing power network.

The core metric for space-based computing is the effective delivery of computing power under constrained environments, encompassing multiple key indicators such as computing power per unit of power consumption, thermal design margin, and effective performance after radiation exposure.

Only when satellites transition from mere data collection nodes to spatial facilities with intelligent understanding capabilities can the value of orbital resources be truly unlocked. This collaboration represents a bidirectional convergence of computing power forms and spatial platforms.

Space Environment Forces Route Reassessment: Optical Computing Meets Constraints

Ground-based data centers can rely on stable power grids, liquid cooling systems, routine maintenance, and hardware iterations to support high-power chip operations, but this logic does not hold in space.

Space energy comes solely from solar power, heat dissipation relies only on thermal radiation, payload weight directly determines launch costs and platform design, and chips must withstand long-term space radiation, electromagnetic interference, extreme temperature variations, and unattended environments.

The unique constraints of the space environment have propelled optical computing to the core track of space-based computing power. Space-based optical computing naturally possesses radiation resistance, electromagnetic interference resistance, low power consumption, and low latency, making it suitable for lightweight satellite platforms and enabling technological synergy with inter-satellite laser communications.

The lack of air convection in space makes heat dissipation extremely challenging for traditional electronic chips. Optical computing completes operations through optical waveguide transmission, significantly reducing heat generation pressure. Paired with a phase-change in-memory computing architecture, it can also minimize additional energy consumption and heat generated by data transfer.

The core computational loads of AI inference concentrate on matrix multiplication, convolution, and vector computing, which are precisely the advantage scenarios for optical computing. The parallel propagation, ultra-high bandwidth, and extremely low latency of light endow it with architectural-level advantages in specific AI inference tasks. Long-term experiments in academia on optical neural networks and photonic convolution accelerators have also validated the technical potential of photonic computing in deep learning acceleration.

Industrial implementation cannot rely solely on technical potential. Optical computing has long faced challenges in precision, programmability, scalable manufacturing, system packaging, software ecosystems, and synergy with electronic systems.

Optoelectronic fusion is currently the most feasible implementation path: using light to handle highly parallel computing loads and electricity to complete control, logic, storage interfaces, and system scheduling, integrating optical chips into mass-producible computing cards and systems.

The launch of optical computing payloads into space represents a profound transformation in space-based data processing modes. Traditional satellites rely on a "collection + ground processing" core logic, where massive amounts of raw data obtained from remote sensing observations must be transmitted back to the ground for analysis. Bandwidth bottlenecks result in invalid data occupying transmission resources, and response times in emergency scenarios can extend to several hours.

On-board optical computing systems can directly complete intelligent data analysis and screening in orbit. AI models perform target recognition, anomaly detection, and feature extraction on-board, transmitting only high-value results back to the ground, significantly compressing invalid transmission volumes and improving bandwidth utilization.

Combined with inter-satellite laser communication links, multiple optical computing satellites can construct a distributed space-based computing power network, enabling dynamic computing power scheduling and collaborative computing.

The shift from space-based data transmission to on-board in-situ computing represents an upfront release of space data value. In the future, with the implementation of on-board large-scale models and multi-agent collaboration, space-based computing power will transition from an auxiliary tool to a core production system, supporting more possibilities for deep space exploration and commercial aerospace.

Conclusion:

According to industry agency predictions, the global market size for space-based edge computing is expected to reach nearly 200 billion USD by 2026. With the rapid development of commercial aerospace, the demand for space computing power will continue to expand.

China's early technological advantages in optical computing, coupled with a complete aerospace engineering industrial chain, are expected to position it proactively in the new round of space computing power competition.

Partial References: Shanghai Observer: "Development Initiated for the World's First Optical Computing Satellite," Xinmin Evening News: "Shanghai Enterprise Initiates Development of the World's First Space-Based Optical Computing Satellite," Shanghai Municipal Science and Technology Commission: "Development Initiated for the World's First Optical Computing Satellite, Enabling Optoelectronic Fusion AI Inference in Space," EE World: "Photonics-Based Technology: The World's Only Commercial Company to Achieve In-Memory Computing with Optical Chips"