Six months ago, Amazon founder Jeff Bezos publicly stated that "orbital data centers will be the next step in the industrial transition from Earth to space." Google was soon reported to be evaluating space data center solutions, rapidly heating up the concept of space computing power. At the time, most attention focused on whether this was feasible. After all, ground-based data centers were still grappling with power supply issues, and sending chips into space seemed more like a distant prospect.

But six months later, the pace of change in this sector has far exceeded expectations. NVIDIA unveiled a dedicated space AI module at GTC 2026, SpaceX announced a $20 billion investment to build its own chip factory, Blue Origin submitted an application to the FCC for 50,000 data center satellites, and Chinese companies' computing constellations are already in orbit. In the US alone, the total number of space data center satellite applications submitted to the FCC has exceeded 1.2 million. From chip hardware and energy supply to communication networks and manufacturing, a complete industrial chain for space computing power is taking shape. What was still being debated for feasibility six months ago has now entered the stage of concrete industrialization with real financial commitments.

01

Musk's Vertical Integration and NVIDIA's Ecosystem Alliance

The fundamental contradiction in space computing power has always been chips. Traditional aerospace chips have development cycles spanning years, with fewer than five global manufacturers capable of radiation-hardened designs. These chips are expensive yet offer extremely low computing power, utterly unable to meet the demands of modern AI large models. This landscape is now changing.

By the end of 2025, Starcloud, a startup invested in by NVIDIA, sent a 60-kilogram satellite, Starcloud-1, into orbit via SpaceX's rideshare mission. This satellite was equipped with an NVIDIA H100 GPU. In December 2025, Starcloud announced the successful on-orbit operation of AI models like Google's Gemma using this satellite. This marked the first publicly reported instance of commercial AI chips completing model training in a space environment. The event validated a critical assumption: consumer-grade or data center-grade high-power GPUs can operate in orbital environments.

Building on this validation, chip giants began formally entering the field. In March, NVIDIA unveiled the Space-1 Vera Rubin module at its annual GTC conference. Based on the latest Vera Rubin architecture, the module is designed specifically for space environments with strict size, weight, and power constraints. According to official data, its inference performance is about 25 times higher than that of the H100 GPU. NVIDIA's strategy is to extend its dominant CUDA software ecosystem from ground-based data centers to orbit, enabling developers to deploy AI applications in space using the same toolchain. On the same day, NVIDIA announced its first batch of partners, including Aetherflux, Axiom Space, Kepler Communications, Planet Labs, Sophia Space, and Starcloud, covering key nodes in the current space data center industrial chain.

Unlike NVIDIA's approach of providing general-purpose modules, SpaceX opted for vertical integration. In March, Elon Musk announced that Tesla, SpaceX, and xAI would jointly invest $20-25 billion to build a chip factory named TeraFab in Austin, Texas. The factory plans to produce the D3 high-power radiation-hardened processors specifically designed for space environments to support SpaceX's orbital data center plans. Behind this decision lies SpaceX's acute anxiety about future computing power demands. Relying on external foundries would not only face production capacity bottlenecks but also entail high procurement costs. Vertical integration aligns with Musk's consistent business logic. However, TeraFab's 2nm process target is extremely aggressive—currently, only a handful of foundries worldwide can mass-produce 2nm chips, making this timeline highly uncertain from a semiconductor industry perspective.

From Starcloud's on-orbit validation to NVIDIA's standardized modules and SpaceX's self-built wafer fab, the application path for semiconductor hardware in space environments has become clear: instead of relying on traditional aerospace-specific chips, the most advanced commercial AI computing power on the ground is being directly brought into space through system-level redundancy designs and novel heat dissipation technologies.

02

Industrial Chain Takes Shape

Beyond chips, space computing power requires support from multiple sectors, including energy, heat dissipation, communications, and manufacturing. Over the past six months, a supporting industrial chain around these areas has begun to take shape.

In terms of energy supply, solar panels in space operate at 5-7 times the efficiency of those on the ground, unaffected by clouds, day-night cycles, or atmospheric attenuation. Startup Aetherflux is focusing on developing space solar power systems to address power supply issues for orbital data centers. The company has secured $50 million in funding and received grants from the US Department of Defense's Operational Energy Capability Improvement Fund. In March, Aetherflux established a new satellite development center in Seattle.

In computing platform design and heat dissipation, Sophia Space completed a $10 million seed funding round in February. The company specializes in orbital edge computing, developing modular, passively cooled managed computing platforms. The core of its technical approach is solving heat dissipation in space—in a vacuum environment without air convection, heat can only be dissipated through radiation, posing stringent requirements for the platform's thermal design.

In data transmission, Kepler Communications announced the deployment of space cloud infrastructure powered by NVIDIA. Kepler currently operates a commercial optical data relay constellation of 33 satellites, serving as a key infrastructure provider for space data center communication networks. Through Kepler's optical relay network, data can be directly transmitted to computing nodes in orbit for processing, with only the final analysis results sent back to the ground, significantly reducing communication pressure.

On the manufacturing side, traditional ICT companies are also making moves. In March, Hon Hai Precision Industry's subsidiary, Hon Brite Technology, announced an expanded partnership with Israeli space computing company Ramon.Space to jointly develop space data center infrastructure. Hon Brite, which has obtained AS9100 aerospace quality certification, will establish a production line for space computing products. This collaboration signifies a shift in space computing power hardware manufacturing from aerospace-specific models to industrial contract manufacturing modes. Bringing in a consumer electronics contract manufacturing giant like Foxconn is expected to replicate assembly line production experience in space hardware, achieving economies of scale and significantly reducing manufacturing costs.

03

Million-Scale Constellation Applications

Another battleground in the space computing power race lies in the file cabinets of the US Federal Communications Commission (FCC). The planned scale of satellite constellations has ballooned from hundreds to millions.

In January 2026, SpaceX led the way by submitting an application to the FCC to launch up to 1 million data center satellites. SpaceX stated in its application: "Orbital data centers are the most efficient way to meet the accelerating demand for AI computing power." This scale far exceeds its current in-orbit Starlink satellite count. SpaceX's move is widely interpreted by the industry as an extreme "land grab" strategy aimed at preemptively securing valuable low-Earth orbit resources.

Subsequently, Amazon submitted a petition to the FCC in early March, requesting the rejection of SpaceX's application on technical grounds such as spectrum interference and orbital safety. The FCC chairman later publicly criticized Amazon's opposition, arguing that its own satellite deployment progress lagged behind yet it sought to hinder competitors.

Following closely, Bezos-owned Blue Origin also submitted an application to the FCC, planning to launch 51,600 data center satellites named Project Sunrise. These satellites would operate in 500-1,800 km sun-synchronous orbits, transmitting data via laser links and coordinating with Blue Origin's previously planned 5,408 TeraWave communication satellites. Blue Origin's entry has turned the space data center competition into another direct clash between Musk and Bezos in the space sector.

Startups have also proposed aggressive plans. In March, Starcloud applied to the FCC for operational permits for 88,000 satellites. These satellites are planned for deployment in 600-850 km dawn-dusk sun-synchronous orbits to achieve near-continuous solar power generation.

In just two months, the total number of space data center satellite applications submitted to the FCC by US companies alone exceeded 1.2 million. These figures currently represent more strategic positioning—given limited orbital and spectrum resources, early applications mean early occupation. This "spectrum first, constellation later" strategy has precedents in the communication satellite field, and the space data center sector is now repeating this pattern on a larger scale.

04

China's On-Orbit Practices

Unlike the aggressive planning of US companies, Chinese enterprises focus more on technical validation and practical deployment of small-to-medium-scale constellations.

In 2024, the "Oriental Smart Eye" constellation achieved the first on-orbit application of a domestically produced general-purpose CPU+NPU architecture, sending the first AI large model into space. In May 2025, Chengdu Guoxing Aerospace and the Zhejiang Lab jointly launched the first 12 satellites of the "Three-Body Computing Constellation." Each satellite carries an 8-billion-parameter space-based AI model, with single-satellite computing power reaching 744 TOPS and the 12-satellite collaborative on-orbit computing power at 5 POPS. The plan aims to complete a 1,000-satellite computing network by 2032, with total computing power reaching 100 exaflops per second. Compared to US companies' million-satellite plans, Chinese constellations are relatively restrained in scale but extremely fast in implementation.

Guoxing Aerospace also has a longer-term "Star Computing" plan: deploying 2,800 computing satellites, including 2,400 for inference and 400 for training, with full deployment targeted by 2035.

In terms of hardware cost control, Chinese companies are exploring different technical routes. Oriental Space adopts a "triple redundancy architecture," using ordinary industrial-grade chips instead of expensive traditional aerospace-grade radiation-hardened chips. With three systems running simultaneously, if one chip is damaged by cosmic rays, the other two correct errors through a voting mechanism and automatically switch. Once industrial-grade chips can replace aerospace-grade ones, the supplier landscape shifts from a monopoly of fewer than five global manufacturers to a competitive market of dozens, reducing single-chip costs from hundreds of thousands of yuan to ten thousand yuan levels.

In March, Dreame Technology's Chip Crossing announced a plan for 2 million "Yao Tai" computing power satellites at the AWE 2026 forum and plans to launch its first self-developed space computing box for on-orbit validation soon. This figure surpasses SpaceX's 1 million-satellite plan and is currently the largest publicly announced computing constellation plan globally. However, some commentators have questioned the feasibility of this plan, noting that orbital capacity and spectrum resource limitations are unavoidable realities.

Additionally, Shanghai is systematically developing its space-based computing industry. Reports indicate that Shanghai has listed space-based computing as a key focus for future industries. Chinese Academy of Sciences academician Wang Jianyu pointed out that space-based computing, with its on-orbit data processing breaking ground dependence, will transform the role of satellites. China's advancement in space computing power is shifting from spontaneous corporate actions to an organized strategic layout.

05

Market Realities

According to Space IQ data, investment in low-Earth orbit sectors exceeded $45 billion in 2025, nearly doubling from nearly $25 billion in 2024. Space Capital statistics show that cumulative global space economy investment since 2009 has surpassed $400 billion, with the US contributing over half. Space Capital CEO Chad Anderson believes the space industry is still in the "early stages of a multi-decade infrastructure cycle."

BIS Research predicts that the global on-orbit data center market will reach $1.777 billion by 2029, growing at a 67.4% CAGR to $39.09 billion by 2035.

However, from the perspectives of semiconductor and aerospace engineering realities, these projections warrant caution. The engineering challenges facing space data centers have not disappeared due to capital enthusiasm. Heat dissipation in vacuum environments remains the core bottleneck—without air convection, heat can only be dissipated through radiation, requiring far larger heat dissipation areas than ground solutions. Damage to advanced-process chips from cosmic rays also requires sustained engineering investment to resolve.

Moreover, economic viability highly depends on further reductions in launch costs. According to Ars Technica analysis, for orbital data centers to be economically viable, the cost of placing one kilogram into orbit needs to fall below $1,000. Currently, SpaceX's Falcon 9 launch costs around $5,000 per kilogram, while the fully reusable Starship aims to reduce this to several hundred dollars, though it has not yet entered regular commercial operations.

Another noteworthy signal is SpaceX's IPO. Reports suggest SpaceX may file for listing as soon as this week. Space Capital CEO Anderson compares this to the "Netscape moment" for the space industry—if SpaceX successfully goes public, it could attract broader capital into space infrastructure, accelerating the commercialization of space data centers.

Space computing power has transformed from a concept requiring explanation into a crowded race. But a crowded track does not equal industrial maturity. The key to determining the outcome of this competition lies in who can first achieve substantive breakthroughs in hard indicators like launch costs, chip reliability, and heat dissipation technologies.