China’s Orbital AI Supercomputer: A Bold Leap Toward Space-Based Intelligence

China is making an audacious move to revolutionize computing by shifting the battleground for AI supremacy into space. Through its “Three-Body Computing Constellation”, the country is building the world’s first large-scale orbital AI supercomputing network — a futuristic step that could redefine how data is processed, shared, and leveraged globally.

🌌 A Supercomputer Above Earth: What’s Being Built

In May 2025, China launched the first 12 satellites of a planned 2,800-satellite AI constellation from the Jiuquan Satellite Launch Center. This project, led by commercial space firm ADA Space in collaboration with Zhijiang Laboratory and Neijiang High-Tech Zone, is not just about infrastructure — it’s about creating a fully distributed, autonomous AI network in low Earth orbit.

Each satellite functions as a computing node, carrying an AI model with 8 billion parameters and capable of 744 TOPS (tera operations per second). When fully deployed, the network is expected to achieve 1,000 POPS (peta operations per second) — rivaling the world’s most powerful terrestrial supercomputers.

🧠 Intelligence in Orbit: Why It Matters

This isn’t just a space tech milestone — it’s a shift in how humanity might approach global-scale AI tasks:

  • Low-latency edge processing: Instead of routing data back to Earth, satellites can process it on the spot, dramatically reducing time and transmission costs.

  • Space-to-space AI inference: Inter-satellite communication via high-speed 100 Gbps laser links allows the constellation to function as a cohesive neural network.

  • Data sovereignty in orbit: With storage capacities of 30 TB per satellite, China can manage massive datasets outside terrestrial jurisdictions — a move with significant geopolitical implications.

🛰 Real-World Applications and Use Cases

The satellites will not just compute; they will observe, analyze, and model. Equipped with advanced sensors such as X-ray polarization detectors, the network will monitor cosmic events like gamma-ray bursts. On Earth, it will generate real-time 3D digital twin models of terrains, cities, and environments — useful for:

  • Disaster response and relief

  • Military reconnaissance

  • Immersive tourism and gaming experiences

  • Smart urban planning and development

By processing these models in space, China is sidestepping the latency, bandwidth constraints, and environmental impact of traditional Earth-based data centers.

☀️ Clean, Scalable, and Sustainable

Unlike power-hungry terrestrial server farms, this orbital network:

  • Runs entirely on solar energy

  • Utilizes the cold vacuum of space for passive cooling

  • Eliminates the need for massive water consumption and cooling infrastructure

This makes it a greener alternative at a time when data center emissions are projected to become a major global concern.

🌍 Strategic and Global Implications

This is not just a scientific endeavor — it’s a strategic maneuver in the evolving space and AI race. China’s investment in space-based AI:

  • Challenges US and EU dominance in cloud and supercomputing

  • Expands its presence in space infrastructure, setting new precedents for sovereignty and control over orbital AI

  • Potentially creates a military advantage, with autonomous, AI-powered sensing and computing nodes functioning globally and independent of Earth-bound assets

If successful, it could redefine the architecture of cloud computing — from centralized terrestrial data centers to decentralized orbital AI nodes, making today’s infrastructure look outdated.

🔭 What’s Next?

As more satellites are launched, watch for:

  • How other global powers respond — will we see a “cloud wars in orbit” era?

  • The technical feasibility of scaling, updating, and maintaining complex models in space

  • The ethical and regulatory challenges of AI operating autonomously outside Earth’s jurisdiction

🧩 Final Thought

China’s orbital supercomputer project is more than technological ambition — it’s a statement. It reflects a paradigm shift toward off-world computation, combining AI, aerospace, and geostrategy into a bold vision of the future. Whether it succeeds or not, the ripple effects of this initiative are already influencing how nations think about data, intelligence, and the future of computing.

Alibaba’s ZeroSearch allows AI search without engines

Alibaba’s ZeroSearch represents a paradigm shift in AI training methodologies, enabling large language models (LLMs) to develop sophisticated search capabilities through self-simulation rather than reliance on external search engines. Here’s a restructured analysis with key insights:

Core Innovation: Self-Sufficient Search Training

ZeroSearch eliminates dependency on commercial search APIs by transforming LLMs into autonomous retrieval systems. This approach leverages:

  • Internal Knowledge Utilization: Pre-trained LLMs generate simulated search results using their existing knowledge base:
  • Controlled Environment: Developers precisely manage document quality during training, avoiding unpredictable real-world search results

Curriculum-Based Rollout Strategy

Progressive Complexity Scaling:

  • Starts with high-quality document generation, gradually introducing noise and irrelevant data.
  • Enhances reasoning skills by exposing models to increasingly challenging retrieval scenarios.
  • Achieves Google Search-level performance with a 7B-parameter model (33.06 vs. Google’s 32.47)

Key Outcomes:

  • 14B-parameter model outperforms Google Search in benchmarks (33.97 score)
  • Models learn to distinguish useful information from noise through structured prompt engineering.

Economic Impact: 88% Cost Reduction

Resource Optimization:

  • Shared simulation servers maximize GPU utilization during low-activity periods
  • Scalable model sizes (3B to 14B parameters) let users balance performance and computational needs

Technical Architecture

Simulated Retrieval Pipeline:

  1. Lightweight Fine-Tuning: Converts base LLMs into retrieval modules using annotated interaction data.
  2. Dual-Sample Training:
    • Positive samples: Trajectories leading to correct answers.
    • Negative samples: Introduces controlled noise through prompt adjustments.
  3. Multi-Turn Interaction Template: Guides query processing through structured reasoning-search-answer cycles.

Algorithm Flexibility: Supports PPO, GRPO, and Reinforce++ frameworks

Strategic Implications

  • Democratized AI Development: Makes advanced search training accessible to startups by removing API cost barriers
  • Reduced Platform Dependency: Reduces reliance on major tech companies’ search infrastructure
  • Enhanced Control: Enables precise calibration of training data quality for specialized applications

This breakthrough demonstrates how self-simulated training environments could redefine AI development economics, particularly for resource-constrained organizations. By combining cost efficiency with performance parity to commercial search engines, ZeroSearch sets a new standard for building autonomous, knowledge-rich AI systems.