Strategic Assessment of On-Orbit High-Performance Computing: Starcloud and NVIDIA Initiatives

Executive Summary

The Starcloud and NVIDIA on-orbit HPC initiative is positioned as a strategic response to the fundamental crisis of power, cooling, and water resource constraints in terrestrial data centers, driven by the exponential demand for AI and Large Language Models (LLMs). With the global AI market projected to reach $3.49726 trillion by 2033 1, traditional infrastructure is failing to keep pace with the scale of this demand. Starcloud proposes a first-principles solution to energy and cooling issues by leveraging unlimited solar energy and the infinite heat sink of deep space, claiming a 10x reduction in carbon dioxide over the lifetime of the data center 2.
However, this disruptive technology faces significant engineering and regulatory challenges that must be overcome. Fundamental physics analysis indicates that heat dissipation in space is limited to a lower power density of 10-20 kW per rack compared to terrestrial liquid-cooled facilities 3. Furthermore, using commercial GPUs (NVIDIA H100) in orbit requires advanced software-defined resilience (checkpoint/restart, ECC) to mitigate radiation-induced malfunctions 4. The optimal market is latency-tolerant workloads or data-adjacent processing, such as in-orbit processing of Earth observation data (SAR), which eliminates downlink bottlenecks and significantly reduces response times 2. The long-term success of this venture depends on the continued reduction of launch costs, the implementation of DPU-leveraged cloud-native security (Zero-Trust architecture) 5, and regulatory approval regarding orbital debris mitigation, particularly concerning 5-gigawatt-class mega-structures 2.

Section 1: Terrestrial Infrastructure Bottlenecks and the Strategic Necessity of On-Orbit HPC Infrastructure

1.1. Exponential Demand Curve for AI and Large Language Models (LLMs)

Demand for High-Performance Computing (HPC) infrastructure is increasing at an unprecedented rate due to the rapid development of Artificial Intelligence (AI). The global AI market is projected to expand at a remarkable compound annual growth rate (CAGR) of 31.5% from 2025 to 2033, reaching $3.49726 trillion by 2033 1. Within this growth, the Large Language Model (LLM) segment is expanding particularly rapidly, with a projected CAGR of 36.9% from 2025 to 2030 8.
This explosive market expansion is placing significant pressure on infrastructure. Training foundation models is the most compute-intensive task, with resource requirements increasing by 10x to 100x per model generation 9. For example, it is estimated that training a large model like GPT-4 required running approximately 25,000 NVIDIA A100 GPUs for 90 to 100 days 10. This exponential growth rate in the AI/LLM sector shows a severe divergence compared to the projected growth rate of the traditional HPC as a Service (HPCaaS) market (6.39% CAGR from 2025 to 2034 11). This divergence suggests an infrastructure gap where existing supply chains and terrestrial infrastructure are not keeping up with the increase in AI-driven computational demand, forming a strategic window of opportunity for alternative solutions like Starcloud that can bypass terrestrial constraints 3.

1.2. Terrestrial Data Center Constraints: The Energy, Water, and Cooling Crisis

Current terrestrial data centers face compound constraints regarding both scaling and sustainability. U.S. data center capacity is required to double from 17 GW in 2022 to 35 GW by 2027, but AI workloads alone could demand 60-100 GW of power by 2030 3. In addition to energy supply constraints, cooling is also a severe bottleneck. Cooling systems in terrestrial data centers consume 1.2 trillion liters of water annually, which is sparking environmental opposition and concerns over water scarcity 3.
To handle the thermal load of high-density workloads, particularly modern GPU clusters, the industry is shifting from traditional air cooling to liquid cooling 13. Liquid cooling is emerging as an essential solution that enables power densities of 30-100+ kW per rack 3. Starcloud's proposal offers a direct, first-principles solution to this most difficult terrestrial scaling problem. They claim that space provides "unlimited low-cost renewable energy" to maintain power supply and utilizes the vacuum of deep space as an "infinite heat sink" to reduce water usage to zero 2. This model is projected to provide a 10x reduction in carbon dioxide over the lifetime of the data center compared to terrestrial power supply, excluding the initial environmental cost of the launch phase 2, providing a strong competitive advantage from an Environmental, Social, and Governance (ESG) perspective.

1.3. Introduction to Starcloud: A Case Study in Space-Based Computing

Starcloud was founded in 2024 14 and, as a member of the NVIDIA Inception program 2, is pioneering the frontier of space-based computing infrastructure 14. The founding team combines expertise in building large-scale GPU clusters (co-founder and Chief Engineer Adi Oltean was a Principal Software Engineer at Microsoft and led the SpaceX Starlink tracking beam program) with experience in operational space systems 14, which significantly de-risks the venture.
The company has raised $21 million in seed funding from strategic investors such as NFX, In-Q-Tel (the U.S. government's defense-related venture capital firm), and the NVIDIA Inception program 14. In-Q-Tel's involvement indicates that this venture has high strategic value beyond commercial LLM training, strongly supporting the vision of "sovereign cloud" computing 14. As its most recent mission, the launch of the Starcloud-1 satellite (60 kilograms), equipped with NVIDIA H100 GPUs, is scheduled for November 2024 2. This will be the first instance of a state-of-the-art data center-class GPU being deployed in space 2. Starcloud's ultimate goal is to build a 5-gigawatt-class on-orbit data center requiring ultra-large solar and cooling panels approximately 4 kilometers square 2.

1.4. NVIDIA's Strategic Role: From Chip Vendor to Infrastructure Enabler

The partnership between Starcloud and NVIDIA is deeply tied to NVIDIA's broader corporate strategy. NVIDIA is rapidly shifting from simply selling semiconductors to providing comprehensive AI solutions and a vertically integrated ecosystem that includes its own cloud services 16. The company is leveraging its hardware dominance (H100 and GH200) and its software lock-in mechanism, CUDA, to act as a "gatekeeper" against the hyperscaler "Big Three" (AWS, Azure, GCP) cloud providers 16.
By prioritizing the shipment of high-demand H100 chips to smaller partners like Starcloud (as well as Coreweave, Equinix, and Lambda) 16, NVIDIA is expanding its influence and helping to build alternative cloud capacity 16. By supporting Starcloud, NVIDIA is sponsoring the creation of a new, highly specialized vertical layer in the AI infrastructure stack: the orbital layer. Through this partnership, NVIDIA can validate the suitability of state-of-the-art commercial off-the-shelf (COTS) hardware (H100) in extreme environments 2 and secure early market share in a future mega-market. It also enables further competitive pressure on the "Big Three" by controlling access to the technology required for the most advanced LLM training, regardless of the physical location of the computing 16.

Section 2: First-Principles Analysis: Engineering Feasibility and Physical Constraints

2.1. Thermal Management in a Vacuum: Limits of Radiative Cooling

The engineering feasibility of an on-orbit data center is largely defined by the fundamental physical constraint of heat dissipation. While space offers nearly unlimited solar energy 2, heat removal—that is, cooling—is the true bottleneck 3. On-orbit facilities rely on passive radiative cooling, which utilizes the vacuum of deep space as an infinite heat sink 2.
Fundamental physics-based analysis (based on the Stefan-Boltzmann law) indicates that radiative cooling limits power density to 10-20 kW per rack 3. This is significantly lower than the 30-100+ kW achievable in terrestrial liquid-cooled facilities 3. Due to this low density limit, achieving Starcloud's ambitious goal of a 5-gigawatt-class facility 2 would require a massive physical structure (panels approximately 4 kilometers square) that scales in proportion to power 2. This "density scaling paradox" significantly increases launch costs and complexity, and unless the dramatic reduction in launch costs through reusable rocket technology (a key success factor mentioned in 3) is realized, it may offset the economic benefits of operational costs.
Comparing the main constraints of terrestrial HPC facilities and on-orbit HPC, while energy supply, cooling water, and land permits are the primary bottlenecks on Earth, surface area for heat dissipation and launch costs are the primary constraints in space 3. Space addresses the power supply issue but replaces water and land issues with mass and area issues. Terrestrial cooling systems use large amounts of water, whereas on-orbit systems reduce water usage to zero through passive radiative cooling using the vacuum of deep space, making it a key differentiator for ESG and in water-stressed regions 2. Terrestrial facilities enable high heat densities of 30–100+ kW/rack through liquid cooling, whereas on-orbit facilities are limited to 10–20 kW/rack due to radiative cooling, meaning on-orbit computing must focus on high FLOPs/Watt efficiency to maximize utilization within a limited thermal envelope 3. Long-term economic viability depends on whether the reduction in operational costs—a 10x reduction in carbon dioxide over the lifetime—outweighs the initial launch CapEx 2.

2.2. Hardware Resilience: Radiation Hardening and Reliability Engineering

The success of on-orbit computing depends on ensuring the reliability of commercial hardware in radiation-exposed environments. Commercial GPUs like the NVIDIA H100 are not designed with radiation tolerance for space use 4. Therefore, operational reliability in orbit must be ensured through a multi-layered approach, including physical shielding, error detection and correction (ECC) memory, watchdog timers, and software-level resilience such as checkpoint/restart (CR) mechanisms to mitigate bit flips caused by single-event upsets (SEUs) 4.
Starcloud adopts this strategy because it recognizes that full radiation hardening is expensive and time-consuming, and is not practical for rapidly evolving GPU technology lineups like the H100 and Blackwell 4. By relying on COTS hardware and sophisticated software mitigation measures, Starcloud can deploy cutting-edge GPU technology into space more rapidly. However, this approach increases software complexity and operational overhead. Therefore, the expertise of GPU cluster specialists like Mr. Oltean is essential to maintain uptime and data integrity in high-failure-rate environments 14. The success of the H100's "space debut" 2 depends more on the robustness of this software-defined resilience infrastructure than on the hardware itself.

2.3. Orbital Mechanics and HPC Workload Latency Management

The physical placement in orbit determines the types of workloads Starcloud targets. Latency is one of the primary constraints for the adoption of space-based computing 18. Geostationary orbit (GEO) is fundamentally unsuitable for real-time consumer applications due to round-trip latency of 500-650 milliseconds 18. Low Earth orbit (LEO) satellites have lower latency (50–150 milliseconds) but are still significantly slower compared to terrestrial fiber optic networks (less than 50 milliseconds) 18.
Due to this latency constraint, on-orbit facilities must be optimized for latency-tolerant workloads where real-time interactivity is not essential, such as AI training, offline rendering, and disaster recovery 19. These latency-tolerant workloads constitute terrestrial batch processingwhere power/cost efficiency is the primary input 19. Starcloud strategically solves this latency challenge by pursuing another market: space edge computing. In the latter case, because data is collected by the satellites themselves (Earth observation, SAR data, etc.), on-orbit placement dramatically reduces effective latency by eliminating downlink bottlenecks 2. This transforms Starcloud's value proposition from cost reduction to increased speed of insight, which terrestrial data centers cannot match regardless of power capacity 2.

Section 3: Starcloud and NVIDIA: In-depth Architectural Analysis and Cloud Strategy

3.1. Hardware and Software Co-design for Cloud-Native Space Architecture

Starcloud's on-orbit computing platform functions as an extension of NVIDIA's integrated cloud-native supercomputing architecture 21. The platform is built around NVIDIA H100 GPUs utilizing the Hopper architecture, featuring the Transformer Engine and 4th generation Tensor Cores, which accelerate trillion-parameter LLM training by up to 4x compared to the previous generation 9.
At the core of this infrastructure are NVIDIA BlueField data processing units (DPUs) and high-speed, low-latency Quantum InfiniBand networking 5. While security is a complex issue in terrestrial data centers, it becomes a survival issue in orbit, which is a remotely located, multi-tenant facility with no physical access. BlueField DPUs are critical because they offload infrastructure tasks (networking, security, storage virtualization) from the host processor 5. This capability enables multi-node tenant isolation and Zero-Trust architecture while maintaining bare-metal performance 5. After hosting an untrusted tenant, the DPU ensures a "clean boot image" involving a full cleanup and re-establishment of trust for a newly scheduled tenant 5. This feature facilitates the high level of security required for Starcloud's "sovereign cloud" ambitions 14. Infrastructure developers can use the NVIDIA DOCA SDK to rapidly build network, storage, security, AI, and HPC applications on this DPU 5.

3.2. Competitive Positioning: "Gatekeeper" Strategy and Market Disruption

The rise of generative AI has caused demand for computing hardware to skyrocket, resulting in a crisis for the "Big Three" public clouds (AWS, Azure, GCP) 16. NVIDIA is leveraging its dominant position regarding H100 chips 16 and strengthening its control over the AI computing supply chain by diversifying hardware distribution channels 16. By prioritizing the supply of high-demand H100 chips to smaller partners like Starcloud, NVIDIA is helping to build alternative cloud capacity and exerting competitive pressure on hyperscalers 16.
AWS and Google Cloud already offer access to H100s 23, with rates ranging from approximately $2.00 to $2.30 per hour depending on commitment 23. Therefore, Starcloud's competitive value proposition is not low rental costs that compete directly with hyperscaler rates. Rather, its value lies in providing access to specialized capacity that is constantly constrained on the ground by power and cooling limitations, as well as the unique advantages of orbital placement (physical data isolation, low-carbon operations, and suitability for extremely latency-tolerant workloads) 2. This is a positioning as a niche vertical market player providing specialized, high-demand services in the AI infrastructure market.

Section 4: Market Feasibility and Optimal Use Cases for Space-Based HPC

4.1. Global Market Demand Analysis and Starcloud's Niche

The market feasibility of on-orbit HPC is supported by the robust growth of the global AI infrastructure market. The explosive growth of the AI market (31.5% CAGR) is primarily driven by AI adoption by large enterprises, with the enterprise segment expected to grow at the highest CAGR (30.50%) during the forecast period 1. Companies are leveraging AI to improve decision-making, automate operations, and enhance customer service 25.
In particular, hybrid cloud solutions are considered essential as they provide the flexibility for organizations to use cloud resources to scale during peak times, and are projected to grow at a CAGR of 31.90% 25. Starcloud is positioned as a highly niche, resource-unconstrained orbital layer within this rapidly growing hybrid infrastructure model. This layer is specialized to absorb the most power-intensive, latency-tolerant, and geopolitically sensitive workloads that existing providers struggle to handle 19. Additionally, the healthcare and life sciences segment has become a major driver of HPC adoption as it deploys HPC to process massive biological datasets for drug discovery 26.

4.2. Workload Suitability: Latency Tolerance vs. Space Edge Applications

The value of on-orbit computing is maximized not by competing with the traditional HPC market, but by focusing on specific problems that can be solved by its unique placement. On-orbit facilities are ideal for latency-tolerant workloads such as AI training, offline rendering, and disaster recovery 19. These tasks consume large amounts of power but do not require millisecond-level response times.
The most powerful use case is space edge computing, particularly the real-time processing of Earth observation data 2. Earth observation methods such as Synthetic Aperture Radar (SAR) imaging 2 generate massive amounts of data, approximately 10 gigabytes per second 2. Starcloud-2 eliminates downlink bottlenecks by enabling this data to be analyzed in real-time on-orbit 14. This dramatically reduces response times for critical applications like wildfire detection and distress signal response from hours to minutes 2. This "data-adjacent processing" approach effectively solves the downlink bandwidth problem rather than general HPC problems, creating a unique and defensible market moat that terrestrial data centers cannot provide 14. Starcloud also intends to provide secure global data storage and sovereign cloud computing, which is attractive for highly sensitive national workloads 14.

4.3. Economic Modeling: Comparison of Cost Structures

Starcloud projects that energy costs in space will be 10 times cheaper than terrestrial options, even when accounting for launch costs 2. This claim is based on a fundamental restructuring of operating expenditures (OpEx). In orbit, the costs of cooling and power are different from those on the ground, becoming nearly zero 2.
However, this OpEx benefit is offset by a significant increase in initial capital expenditure (CapEx). CapEx is heavily front-loaded through launch costs and the need for specialized hardware redundancy (shielding, ECC, CR systems) for radiation mitigation 4. For this model to remain competitive, the operational lifespan of on-orbit hardware must be significantly extended to amortize the high launch CapEx. Therefore, Starcloud's economic viability depends on rigorous engineering to ensure a long Mean Time Between Failures (MTBF) and continued cost reductions in the launch sector (e.g., the maturation of reusable rocket technology) 3.

Section 5: Sustainability, Policy, and Regulatory Constraints

5.1. Environmental Impact and Alignment with Corporate Sustainability

On-orbit data centers offer a clear advantage regarding the major sustainability challenges faced by terrestrial data centers, particularly water usage and carbon emissions. By utilizing the vacuum of deep space for on-orbit cooling, Starcloud completely avoids the 1.2 trillion liters of water usage required by terrestrial facilities annually 2. Furthermore, the company projects a 10-fold reduction in carbon dioxide over the lifetime of the data center compared to powering it on the ground 2.
This approach aligns with the broader sustainability commitments of partner companies like NVIDIA. NVIDIA has established a strong sustainability track record, including achieving 100% of its global electricity consumption from renewable power in fiscal year 2025 27. While Starcloud's operational footprint is superior to terrestrial ones, the total carbon load is shifted to the launch vehicle 2. Therefore, the calculation of environmental impact is highly dependent on the efficiency of reusable heavy-lift launch systems like the SpaceX Starship and their ability to achieve target economics (2028-2030) 3.

5.2. Legal and Regulatory Frameworks: Data Sovereignty and Security

Space-based computing presents complex legal challenges regarding data sovereignty and export controls 19. On-orbit computing must seamlessly integrate with national regulations and cloud compliance frameworks that dictate where models are trained and where datasets reside 19. Starcloud's focus on sovereign cloud computing 14 indicates demand in this area.
Despite physical isolation, the data itself still retains its national origin and regulatory requirements 19. For Starcloud to meet the stringent requirements of national security and defense customers like In-Q-Tel 14, which invests on behalf of U.S. intelligence agencies, complex legal and technical challenges must be resolved. This includes providing robust technical evidence of geographic boundaries for physical locations, encryption, and verified tenant isolation via NVIDIA BlueField DPUs 5.

5.3. Orbital Debris Mitigation and Compliance

The deployment of large-scale infrastructure in space is strictly monitored by international regulators. Regulators (such as the Federal Communications Commission (FCC) and the Department of Commerce) strictly monitor compliance with orbital debris mitigation and end-of-life deorbit rules 4. The FCC has adopted guidance on orbital debris rules and manages licensing requirements 6.
Starcloud's 5-gigawatt-class vision implies a massive structure of approximately 4 square kilometers 2, which significantly increases the risk of collisions and subsequent debris generation in the shared space environment 3. For commercial operation of a facility of such massive mass and scale, compliance with debris mitigation is essential 4. Before construction permits are granted, reliable propulsion and disposal systems to guarantee end-of-life deorbit or facility management must convince regulators worldwide. This is a significant regulatory headwind for Starcloud's scaling plans.

Section 6: Strategic Outlook and Conclusion

6.1. Projected Trajectory of Space-Based HPC Adoption (2025-2035)

The adoption curve for on-orbit HPC is likely to follow a highly specialized trajectory, distinct from general cloud computing models. Initial adoption will be led by anchor customers in the government and defense sectors, prioritizing data isolation and physical security provided by orbital deployment 3. Additionally, the aerospace industry, particularly the real-time Earth observation and geospatial intelligence sectors, will lead the way by leveraging edge computing capabilities 2.
Broad commercial viability, particularly achieving the ultimate goal of 5 gigawatts, depends on whether multiple Critical Success Factors (CSFs) are realized simultaneously 3.

6.2. Critical Success Factors (CSFs) and Threshold Technologies

Based on this analysis, the key CSFs required for Starcloud's long-term success are as follows 3.

1. Reduction in Launch Costs: Starcloud's economic model relies on reusable heavy-lift launch systems, such as the SpaceX Starship, achieving target economics by 2028-2030 to amortize initial CapEx and dramatically lower launch costs 3.

2. Long-term Success in Radiation Mitigation: The long-term reliability of COTS hardware (H100/Blackwell) must be proven over multi-year missions through the effectiveness of multi-layered software-defined resilience (checkpoint/restart, ECC) 4. This is key to mitigating operational risk.

3. Breakthrough in Thermal Scaling: Overcoming the 10-20 kW/rack radiative cooling limit imposed by basic physics is essential for making orbital facilities cost-competitive in density with terrestrial liquid-cooled rivals 3. Innovative thermal management technologies are required to break this limit 3.

6.3. Conclusions and Recommendations

The Starcloud and NVIDIA initiative represents a fundamental rethinking of where computation resides in the AI era. Space-based HPC offers a unique strategic advantage in solving the power and water resource bottlenecks of terrestrial infrastructure 2.
Conclusion:

• Starcloud's primary competitive advantage lies not in price competition with terrestrial hyperscalers, but in providing unlimited, sustainable capacity for power-intensive, latency-tolerant AI training workloads 19, and unique capabilities in data-adjacent processing (space edge computing) 2.

• Through its support of Starcloud, NVIDIA is strengthening its dominance over the hardware supply chain 16, establishing itself as a key competitive pressure point 16, and sponsoring the creation of an orbital layer in the future AI infrastructure stack.

Strategic Recommendations:

1. Prioritizing Investment in Software and Security: Since hardware resilience in orbit is a primary operational risk, investment should be prioritized in the software layer (L4-L7), focusing on fault tolerance, scheduling, and secure multi-tenancy via BlueField DPUs (DOCA integration) 5.

2. Niche Market Entry Strategy: It is recommended to avoid direct competition with hyperscalers in the general compute market 19. Immediate commercialization efforts should focus on two high-margin, latency-intolerant niches: sovereign AI training for national security clients 14, and in-orbit data reduction services for the rapidly growing Earth observation and geospatial intelligence sector 2.

References

1. Artificial Intelligence Market Size | Industry Report, 2033 - Grand View Research, accessed October 24, 2025, https://www.grandviewresearch.com/industry-analysis/artificial-intelligence-ai-market

2. How Starcloud Is Bringing Data Centers to Outer Space - NVIDIA Blog, accessed October 24, 2025, https://blogs.nvidia.com/blog/starcloud/

3. Space-Based Data Centers: First Principles Deconstruction and Market Ecosystem Analysis Through 2035, accessed October 24, 2025, https://vixra.org/pdf/2510.0065v1.pdf

4. Nvidia H100 Goes to Orbit for Space Data Center Test - FindArticles, accessed October 24, 2025, https://www.findarticles.com/nvidia-h100-goes-to-orbit-for-space-data-center-test/

5. Cloud Native Supercomputing - NVIDIA, accessed October 24, 2025, https://www.nvidia.com/en-us/networking/products/cloud-native-supercomputing/

6. Space and satellite wrap up – Legal and regulatory developments in 2024 - Bird & Bird, accessed October 24, 2025, https://www.twobirds.com/en/insights/2025/global/space-and-satellite-wrap-up-legal-and-regulatory-developments-in-2024

7. Space Modernization for the 21st Century Notice of Proposed Rulemaking – SB Docket No. 25-306 - Federal Communications Commission, accessed October 24, 2025, https://docs.fcc.gov/public/attachments/DOC-415048A1.pdf

8. Large Language Models Market Size | Industry Report, 2030 - Grand View Research, accessed October 24, 2025, https://www.grandviewresearch.com/industry-analysis/large-language-model-llm-market-report

9. Efficient Fine-Grained GPU Performance Modeling for Distributed Deep Learning of LLM This work was supported in part by the NSF research grant #2320952, #2117439, #2112606, and #2117439. - arXiv, accessed October 24, 2025, https://arxiv.org/html/2509.22832v1

10. Scaling Intelligence: Designing Data Centers for Next-Gen Language Models - arXiv, accessed October 24, 2025, https://arxiv.org/html/2506.15006v3

11. High Performance Computing as a Service Market Size to Hit USD 72.12 Bn by 2034, accessed October 24, 2025, https://www.precedenceresearch.com/high-performance-computing-as-a-service-market

12. Smart Solutions to Overcome Data Center Cooling Challenges - Badger Meter, accessed October 24, 2025, https://www.badgermeter.com/blog/data-center-cooling-challenges/

13. Beyond Air: The Perks of Liquid Cooling in Data Centers - TierPoint, accessed October 24, 2025, https://www.tierpoint.com/blog/data-center-liquid-cooling/

14. Starcloud: Shaping the Future of Space-Based Data Centers, accessed October 24, 2025, https://www.future-of-computing.com/starcloud-shaping-the-future-of-space-based-data-centers/

15. Team - Starcloud, accessed October 24, 2025, https://www.starcloud.com/team

16. NVIDIA vs Cloud Providers - AWS, Azure, GCP - YouTube, accessed October 24, 2025, https://www.youtube.com/watch?v=yHBjG8umPcM

17. H100 Tensor Core GPU - NVIDIA, accessed October 24, 2025, https://www.nvidia.com/en-us/data-center/h100/

18. Low Earth Orbit (LEO) Satellites vs. Geostationary Satellites — A Technical Comparison | by RocketMe Up Networking | Medium, accessed October 24, 2025, https://medium.com/@RocketMeUpNetworking/low-earth-orbit-leo-satellites-vs-geostationary-satellites-a-technical-comparison-2a4f15aaedc9

19. Bezos Envisions Gigawatt Data Centers In Outer Space - FindArticles, accessed October 24, 2025, https://www.findarticles.com/bezos-envisions-gigawatt-data-centers-in-outer-space/

20. Starcloud - Hacker News - Y Combinator, accessed October 24, 2025, https://news.ycombinator.com/item?id=45667458

21. NVIDIA Cloud-Native Supercomputing, accessed October 24, 2025, https://www.nvidia.com/en-us/lp/networking/cloud-native-supercomputing-technology/

22. Azure vs. AWS vs. Google Cloud: Who Wins the AI Cloud War? - Stansberry Research, accessed October 24, 2025, https://stansberryresearch.com/stock-market-trends/azure-vs-aws-vs-google-cloud-whos-winning-the-cloud-ai-war-in-2025

23. NVIDIA GPU Pricing | Nebius AI Cloud, accessed October 24, 2025, https://nebius.com/prices

24. GPU pricing | Google Cloud, accessed October 24, 2025, https://cloud.google.com/compute/gpus-pricing

25. AI Infrastructure Market Size, Share | Growth Report [2032] - Fortune Business Insights, accessed October 24, 2025, https://www.fortunebusinessinsights.com/ai-infrastructure-market-110456

26. High Performance Computing Market Size, Growth Report [2032] - Fortune Business Insights, accessed October 24, 2025, https://www.fortunebusinessinsights.com/industry-reports/high-performance-computing-hpc-and-high-performance-data-analytics-hpda-market-100636

27. NVIDIA Sustainability Report Fiscal Year 2025, accessed October 24, 2025, https://images.nvidia.com/aem-dam/Solutions/documents/NVIDIA-Sustainability-Report-Fiscal-Year-2025.pdf