SenseTime × Guoxing Yuhang Thousand-Stars Space Computing Constellation: Earth-Space Integrated AI Infrastructure Deep Dive

1. Introduction: When AI Compute Reaches for the Stars

On July 18, 2026, at WAIC 2026’s “Foundation Model Architecture Innovation and Ecosystem Cooperation Forum,” Guoxing Yuhang and SenseTime announced a partnership that left the industry in awe — the construction of the “SenseTime Space Computing Constellation,” with a scale of thousands of satellites and over 10,000 PetaFlops of total compute power.

This is not a “future concept” — within 2026, the first batch of “SenseTime” compute satellites will launch and begin network verification. By 2028, a satellite-ground hybrid cloud commercial compute backbone will be operational. By 2030, a global network of thousands of compute satellites will cover the entire planet.

Why does AI compute need to go to space?

The answer lies in three words: coverage, latency, and cost.

Over 70% of Earth’s surface is ocean. Three billion people live in areas with poor network coverage. Ocean-going vessels, polar research stations, disaster response, remote field operations — these scenarios either have no internet or latency too high for real-time AI inference. A single low-orbit satellite, at 300-500km altitude, can cover a 1000km+ radius with millisecond-level latency.

Yang Fan, SenseTime’s co-founder and president of its Large Model Infrastructure division, stated at the forum: “For weak-network environments like emergency rescue, ocean vessels, and field operations, space computing power can effectively fill the blind spots of ground network coverage. Its value extends beyond capability boundaries — it’s infrastructure supporting the global reach of Chinese AI capabilities, potentially significantly improving the coverage of domestic AI services going global.”

2. Technical Architecture of the Space Computing Constellation

2.1 Three-Phase Evolution Path

The constellation follows a “single-satellite verification → platform construction → constellation networking” path:

2026: First “SenseTime” satellite launch, on-orbit AI inference verification 2028: Satellite-ground hybrid cloud backbone, 10-50 satellite mesh network 2030: Thousand-satellite global coverage, 10,000+ PFlops total compute

2.2 Onboard Computing Platform

The core of each “SenseTime” compute satellite is a high-performance, autonomous, space-grade computing platform with high compute power, large storage, and real-time intelligent processing capability.

Key components:

• AI Inference Chip (domestic): Radiation-hardened, low-power
• General CPU (radiation-tolerant): System management, communication
• FPGA Accelerator: Flexible reconfiguration for diverse workloads
• Large Capacity Storage (>10TB): Model parameters + inference cache
• Communication: Ka-band satellite-ground (10Gbps), laser inter-satellite (100Gbps)

2.3 Unique Challenges of Space-Based AI Inference

Space-based AI inference faces three challenges unseen on Earth:

1. Radiation Effects: High-energy particles cause Single Event Upsets (SEU). Requires Triple Modular Redundancy (TMR) and ECC
2. Strict Power Limits: Satellite solar panels provide ~5kW total, with AI inference budget typically 100-500W
3. Model Update Difficulty: Only possible through bandwidth-limited satellite-ground links during short communication windows

type SpaceInferenceEngine struct {
    modelWeights    []byte
    powerBudget     float64
    radiationLevel  float64
    tmrEnabled      bool
}

// TMR (Triple Modular Redundancy) voting mechanism
type TMRUnit struct {
    outputs [3]float32
}

func (t *TMRUnit) VotedOutput() float32 {
    if t.outputs[0] == t.outputs[1] || t.outputs[0] == t.outputs[2] {
        return t.outputs[0]
    }
    if t.outputs[1] == t.outputs[2] {
        return t.outputs[1]
    }
    // All不一致, take median
    sorted := make([]float32, 3)
    copy(sorted, t.outputs[:])
    // sort
    if sorted[0] > sorted[1] { sorted[0], sorted[1] = sorted[1], sorted[0] }
    if sorted[1] > sorted[2] { sorted[1], sorted[2] = sorted[2], sorted[1] }
    if sorted[0] > sorted[1] { sorted[0], sorted[1] = sorted[1], sorted[0] }
    return sorted[1]
}

3. Satellite-Ground Collaborative Scheduling

3.1 Hybrid Cloud Architecture

The core innovation isn’t “on-orbit inference” itself, but satellite-ground collaborative scheduling — treating ground data centers and space computing satellites as a unified compute pool, dynamically allocating based on task requirements.

class SpaceGroundCoordinator:
    def select_optimal_node(self, task, satellites, ground_stations):
        candidates = []
        
        # Check satellite candidates
        for sat_id, sat in satellites.items():
            latency = self.calculate_link_latency(sat, task.user_location)
            if latency > task.max_latency_ms:
                continue
            if sat.power_available_w < 100:
                continue
            
            score = (0.4 * latency/task.max_latency_ms + 
                    0.3 * (1 - min(sat.compute_capacity/10, 1)) +
                    0.3 * (1 - sat.power_available_w/500))
            candidates.append((score, sat_id, "satellite"))
        
        # Check ground station candidates
        for gs_id, gs in ground_stations.items():
            distance = self._haversine_distance(gs.location, task.user_location)
            fiber_latency = distance * 5 / 1000
            total_latency = fiber_latency + 5
            
            if total_latency > task.max_latency_ms:
                continue
            score = 0.5 * total_latency/task.max_latency_ms + 0.5 * (1 - min(gs.compute/100, 1))
            candidates.append((score, gs_id, "ground"))
        
        candidates.sort()
        return candidates[0] if candidates else None

3.2 Application Scenarios

Ocean Shipping AI: Real-time route optimization, weather prediction, equipment fault预警 for vessels on global routes

Disaster Emergency Communication: AI-powered emergency communication and decision support within 10 minutes of disasters

Polar Research: On-orbit processing of research data, only transmitting key results back to ground stations

AI Going Global: Providing AI infrastructure coverage for Southeast Asia, Africa, and South America where ground networks are insufficient

4. SenseTime’s Three New Computing Pathways

The space computing constellation is part of SenseTime’s broader “three new computing pathways” strategy:

SenseTime Infrastructure Data:

• Daily token service: 2.42 trillion
• 2026 projected growth: 25x, Q4 reaching 10 trillion daily tokens
• Heterogeneous inference cost-performance: 1.25x NVIDIA H-series
• Domestic chip MFU improvement: 85%-152%

5. The Economics of Space Computing

5.1 Cost Structure

A single compute satellite: satellite platform ($3-8M), AI compute payload ($1.5-4.5M), launch ($1.5-3M), ground station ($0.75M). Total for 1000 satellites: ~$7-15B.

Unique cost advantages vs. ground infrastructure:

1. Coverage cost: One satellite covers 1000km radius = hundreds of ground stations
2. Marginal cost: Each additional satellite costs less (shared launch, ground stations)
3. Energy cost: Space solar power, no electricity bills

5.2 Business Model

• Satellite-ground hybrid cloud: Token-based billing, unified pricing
• Industry-specific services: Ocean shipping, polar research, emergency communication
• AI going global infrastructure: Last-mile compute coverage for Chinese AI services

6. Conclusion

The SenseTime-Guoxing Yuhang “Thousand Stars, Ten Thousand PetaFlops” space computing constellation is far more than a PR event. It marks a milestone in Chinese AI infrastructure’s transition from “ground-based” to “earth-space integrated.”

From a technical perspective, on-orbit AI inference faces challenges of radiation tolerance, power limits, and model updates, but the partnership has delivered engineering solutions. From a commercial perspective, space computing fills the blue ocean market that ground networks cannot reach — ocean, polar, emergency, and global expansion — where AI demand is rapidly growing.

When the first “SenseTime” satellite launches in 2026, AI compute will truly enter the era of “stars and the sea.”

Sources: People’s Daily, Xinhua News, 36Kr, Science and Technology Board Daily

• ←Previous
• Next→