Computer Networks
Youtube search... ...Google search
- Telecommunications ... Computer Networks ... 5G ... Satellite Communications ... Quantum Communications ... Communication Agents ... Smart Cities ... Digital Twin ... Internet of Things (IoT)
- Risk, Compliance and Regulation ... Ethics ... Privacy ... Law ... AI Governance ... AI Verification and Validation
- Cybersecurity ... OSINT ... Frameworks ... References ... Offense ... NIST ... DHS ... Screening ... Law Enforcement ... Government ... Defense ... Lifecycle Integration ... Products ... Evaluating
- Zero Trust
- Analytics ... Visualization ... Graphical Tools ... Diagrams & Business Analysis ... Requirements ... Loop ... Bayes ... Network Pattern
- Development ... Notebooks ... AI Pair Programming ... Codeless ... Hugging Face ... AIOps/MLOps ... AIaaS/MLaaS
- Virtualization - Dynamic Spectrum Sharing (DSS)
- Self-Organizing
- Containers; Docker, Kubernetes & Microservices
- Excel ... Documents ... Database; Vector & Relational ... Graph ... LlamaIndex
- Policy ... Policy vs Plan ... Constitutional AI ... Trust Region Policy Optimization (TRPO) ... Policy Gradient (PG) ... Proximal Policy Optimization (PPO)
- Architectures for AI ... Generative AI Stack ... Enterprise Architecture (EA) ... Enterprise Portfolio Management (EPM) ... Architecture and Interior Design
Transport Agnostic - capability is achieved by splitting the control/management ‘planes’ from the data ‘plane’, using software defined networking (SDN) mechanisms to dynamically redirect traffic to the best transport, as opposed to having control/management/data on the same ‘plane’. The users/edge devices delegate (are agnostic) the tasks of assigning transport addresses/routes/protocols/mechanisms are used. From a security point of view, users/edge never access the control/management ‘planes’!
Contents
- 1 Cognitive Network (CN)
- 2 Intent-Based Networking (IBN)
- 3 Multiprotocol Label Switching (MPLS)
- 4 Software-Defined Enterprise (SDE) / Software-Defined Networking (SDN) / Software-defined Wide Area Network (SD-WAN)
- 5 Space-based Data Center
Cognitive Network (CN)
In communication networks, cognitive network is a new type of data network that makes use of cutting edge technology from several research areas to solve some problems current networks are faced with. Cognitive network is different from cognitive radio as it covers all the layers of the OSI model. Cognitive Network | Wikipedia
Intent-Based Networking (IBN)
- Intelligent Automation
- Intelligent Assurance
- Understanding what's on the Network
- Detecting Threats in encrypted traffic
Multiprotocol Label Switching (MPLS)
Multiprotocol Label Switching (MPLS) is a networking technology that enhances the efficiency and speed of data transmission across networks by using labels to route packets instead of traditional IP addresses. This technology is particularly useful in large networks where the traditional routing methods can become complex and inefficient. MPLS operates independently of the underlying IP addressing and routing protocols, allowing for more flexible and efficient routing of traffic.
MPLS works by assigning labels to packets, which are then used to determine the path the packet should take through the network. This label-switched path (LSP) is determined by the first device (usually a router) that processes the packet, which then forwards the packet along the LSP to its destination. This process is much faster and more efficient than traditional routing methods, which require each device in the path to perform a routing lookup for each packet.
MPLS supports a variety of protocols and technologies, including IP, ATM, and Frame Relay, and it interfaces with existing routing protocols such as RSVP and OSPF. It also provides mechanisms for traffic engineering, quality of service (QoS), and the creation of virtual private networks (VPNs) both at Layer 2 and Layer 3.
In addition to improving network performance, MPLS also offers features like traffic engineering, which allows for the optimization of network paths based on various constraints such as bandwidth availability, and the creation of VPNs that can transport different types of traffic over the same network infrastructure.
MPLS is widely used in enterprise and service provider networks to deliver advanced, value-added services over a single infrastructure. It can be integrated seamlessly with existing infrastructure and supports a wide range of platforms, making it a versatile solution for both service providers and enterprises.
For example, in a service provider network, MPLS can be used to aggregate subscribers with differing access links on an MPLS edge without changing their current environments. This allows for the delivery of a wide variety of services over a single infrastructure, including Layer 3 VPNs, Layer 2 VPNs, Traffic Engineering, QoS, GMPLS, and IPv6.
In summary, MPLS is a powerful technology that enhances network efficiency and performance by using labels to route packets, supporting a wide range of protocols and technologies, and offering advanced features like traffic engineering and VPN creation.
Software-Defined Enterprise (SDE) / Software-Defined Networking (SDN) / Software-defined Wide Area Network (SD-WAN)
Youtube search... ...Google search
- Artificial Intelligence Enabled Software Defined Networking: A Comprehensive Overview | Majd Latah and Levent Toker
- SDN, AI, and DevOps | Russ White - Rule11 Reader
- Defense Information Systems Agency (DISA)
- 3 Use Cases for Machine Learning Within SD-WAN | Lanner
- SDN, AI and DevOps | Russ White - Juniper
Virtual network architecture that allows enterprises to leverage any combination of transport services to securely connect users to applications. SD-WAN simplifies the management and operation of a WAN by decoupling the networking hardware from its control mechanism. This concept is similar to how software-defined networking implements virtualization technology to improve data center management and operation. Wikipedia
- WAN Optimization
- Fault Prediction
- Network Management
- Security
|
|
|
|
|
|
|
|
Explained
|
|
Network Functions Virtualization (NFV)
Youtube search... ...Google search
NFV or Virtual Network Function (VNF) allows network operators to manage and expand their network capabilities on demand using virtual, software based applications where physical boxes once stood in the network architecture. This makes it easier to load-balance, scale up and down, and move functions across distributed hardware resources. With continual updates, operators can keep things running on the latest software without interruption to their customers. On the road to NFV deployment | Ericsson
For example, a virtual session border controller could be deployed to protect a network without the typical cost and complexity of obtaining and installing physical network protection units. Other examples of NFV include virtualized load balancers, firewalls, intrusion detection devices and WAN accelerators. ...The NFV framework consists of three main components:
- Virtualized network functions (VNFs) are software implementations of network functions that can be deployed on a network functions virtualization infrastructure (NFVI).
- Network functions virtualization infrastructure (NFVI) is the totality of all hardware and software components that build the environment where NFVs are deployed. The NFV infrastructure can span several locations. The network providing connectivity between these locations is considered as part of the NFV infrastructure.
- Network functions virtualization management and orchestration architectural framework (NFV-MANO Architectural Framework) is the collection of all functional blocks, data repositories used by these blocks, and reference points and interfaces through which these functional blocks exchange information for the purpose of managing and orchestrating NFVI and VNFs.
The building block for both the NFVI and the NFV-MANO is the NFV platform. In the NFVI role, it consists of both virtual and physical processing and storage resources, and virtualization software. Network function virtualization | Wikipedia
|
|
|
|
Space-based Data Center
Space-based data centers (also orbital data centers, ODCs, or orbital compute) are computing facilities deployed in low Earth orbit (LEO) or beyond, in which spacecraft act as processing nodes rather than as passive sensors or communications relays. Interest in the concept accelerated sharply between 2024 and 2026 as terrestrial data center construction ran into grid interconnection queues, water constraints and permitting resistance, while the volume of data produced by Earth-observation and communications constellations began to exceed available downlink bandwidth.
Huawei was among the earliest large infrastructure vendors to name space data centers as a structural pattern for the coming decade, doing so in its Data Center 2030 technology forecast. As of September 2026 the field spans state research programmes, hyperscalers, launch providers and a cohort of venture-funded startups, with regulatory filings on file for well over one million satellites in aggregate.
Rationale
Proponents advance four principal arguments.
- Power
- Solar arrays in sun-synchronous dawn–dusk orbits receive near-continuous, unattenuated sunlight, avoiding cloud cover, atmospheric losses and night. Estimates of the per-panel advantage over terrestrial installations range from roughly five-fold to eight-fold depending on the comparison baseline. Because such orbits avoid eclipse for most of the year, battery storage requirements are also reduced.
- Cooling
- Waste heat is rejected radiatively rather than by chilled air or water, eliminating the 10–30% of terrestrial facility power typically consumed by cooling plant, and eliminating water consumption entirely. Critics note that this converts an operating-cost problem into a mass-and-area problem (see below).
- Downlink bottleneck
- This is the driver most often cited by operators rather than by AI-infrastructure investors. Satellites now generate more data than they can transmit; researchers at Zhejiang Lab have stated that as much as 90% of data generated on orbit is never effectively processed. Processing in orbit and downlinking only derived results — detections, classifications, alerts — collapses the bandwidth requirement. The commonly used analogy is editing video on the capture device rather than uploading raw footage.
- Latency and sovereignty
- Placing compute adjacent to the point of data generation reduces round-trip time for time-critical applications (disaster response, maritime monitoring, defence tipping-and-cueing). Several programmes, notably in Europe and China, additionally frame orbital compute as a matter of digital sovereignty.
Huawei
Data Center 2030
Huawei released the Data Center 2030 report at HUAWEI CONNECT 2023 in Shanghai on 20 September 2023. Presented by Michael Ma, then Vice President of Huawei and President of the ICT Product Portfolio Management & Solutions Department, the report was the product of roughly three years of consultation with more than one hundred academics, customers, partners and research institutes across more than fifty workshops.
The report frames the central problem of the decade as a widening gap between compute demand — growing faster than Moore's law — and the resource constraints on supply. It defines five future scenarios and six key technical characteristics of future data centers. Among these, the "new patterns" characteristic explicitly names underwater data centers and space data centers as construction patterns to be developed for varied application scenarios, alongside "big clusters" (shifting the unit of construction and operations from the server to the rack and then the whole facility) and "lightweight edges" (pushing capacity from core to edge for low-latency and data-residency reasons).
Huawei's framing is notable for treating orbital compute as an extension of an edge-computing continuum rather than as a replacement for terrestrial hyperscale capacity — a materially more conservative position than that later taken by SpaceX or Blue Origin.
Cloud-native satellites and the Tiansuan Constellation
Huawei's most concrete flight heritage in this area comes through KubeEdge, the Kubernetes-based edge orchestration project that Huawei Cloud initiated and open-sourced in November 2018 and later donated to the Cloud Native Computing Foundation, together with its edge-AI subproject Sedna and the MindSpore deep-learning framework.
Huawei Cloud joined as one of the first co-construction partners of the Tiansuan Constellation (天算星座), an open in-orbit research platform initiated in October 2021 by Beijing University of Posts and Telecommunications (BUPT) with the commercial satellite manufacturer Spacety. The first satellite carrying the reconstructed KubeEdge stack launched from Jiuquan Satellite Launch Center on 7 December 2021 and was described as the world's first cloud-native satellite.
Reported results from the platform include:
- Ground-target identification accuracy improved by more than 50% through collaborative inference between satellite and ground station, using a lightweight model on orbit and a high-precision model on the ground with confidence-based escalation;
- A reduction of approximately 90% in the volume of data returned to Earth;
- Support for continuous over-the-air update of on-orbit AI models, incremental deep learning and federated learning.
The work was presented as a keynote at KubeCon + CloudNativeCon Europe 2022. The architectural claim is that containerisation breaks the traditional "one satellite, one mission, one payload" model in which software is fixed for the life of the spacecraft.
On-orbit hardware and stated obstacles
Research associated with Huawei's technology stack has explored radiation-tolerant commercial off-the-shelf CPUs and GPUs hosting virtual machines, with neighbouring satellites linked by optical laser terminals to form cooperative computing clusters. Huawei's own AI silicon line (Ascend) and the MindSpore framework provide the software-hardware pairing for Chinese in-orbit inference work, though Huawei has not announced a dedicated space-qualified accelerator comparable to Nvidia's Space-1 module.
Huawei's published material is candid about the obstacles: the launch cost of heavy server racks; securing continuous solar power; and rejecting heat in vacuum, where convective and evaporative cooling are unavailable. These are the same three constraints identified independently by the European ASCEND study and by sceptical analysts (see Criticism).
United States
Starcloud
Starcloud (Redmond, Washington; formerly Lumen Orbit) is the most advanced dedicated orbital data center startup by flight heritage. Founded in 2024 and a graduate of Y Combinator's Summer 2024 cohort, it published an early white paper proposing multiple gigawatts of orbital AI compute.
- Starcloud-1 — a 60 kg satellite launched by SpaceX in November 2025, carrying the first Nvidia H100 data-center GPU to operate in low Earth orbit. Starcloud reported training an AI model on orbit, a first for the sector.
- FCC filing — February 2026, for a constellation of up to 88,000 satellites in sun-synchronous dusk–dawn orbits between 600 and 850 km.
- Funding — $170 million Series A at a $1.1 billion valuation in March 2026 (Benchmark, EQT); a $250 million extension announced 21 August 2026 at a $2.3 billion valuation, led by Manhattan West, with Nvidia and Cisco Investments joining as new investors. Total capital raised: approximately $450 million.
- Starcloud-2 — a 450 kg spacecraft carrying Nvidia Blackwell-generation GPUs, planned for launch on a Falcon 9 in January 2027, intended to run commercial workloads for early customers including the AI infrastructure provider Crusoe.
- Starcloud-3 — approximately 3 tonnes and 200 kW, designed for deployment by Starship, with production lines being established at a 100,000 sq ft facility in Woodinville, Washington.
- Long-term concept — spacecraft with solar arrays on the order of 4 km across supporting gigawatt-class compute, and roughly 20 GW of orbital capacity at full constellation scale.
Starcloud is working with Nvidia on the Space-1 Vera Rubin Module, and hopes to fly it in late 2028. Chief executive Philip Johnston has been explicit that the company's scaling is gated on Starship availability, describing pre-Starship operations on Falcon 9 as treading water.
Nvidia
Nvidia is a supplier and, since August 2026, an investor rather than an operator. At GTC in March 2026 it announced a space computing line:
- Space-1 Vera Rubin Module — its first purpose-built accelerator for orbital data centers, claimed to deliver up to 25× the AI compute of the H100 in orbit, with a tightly integrated CPU-GPU architecture intended to run large language and foundation models on orbit. Cooling is by large radiators rather than air or liquid loops.
- IGX Thor and Jetson Orin — edge-AI inferencing platforms for geospatial intelligence and autonomous space operations.
Named launch-slate partners include Aetherflux (now Cowboy Space), Axiom Space, Kepler Communications, Planet Labs, Sophia Space and Starcloud. This positions Nvidia as the common silicon layer beneath several nominally competing constellations, and is the reason the phrase "Rubin servers" has become shorthand for the category.
Nvidia chief executive Jensen Huang gave a notably measured assessment on the company's Q4 FY2026 earnings call in February 2026, saying that the economics of orbital data centers are poor at present but should improve, that space offers abundant energy and room, and that the absence of airflow makes heat rejection the binding constraint, requiring fairly large radiators.
SpaceX and xAI
SpaceX has made orbital compute the strategic centrepiece of its corporate reorganisation.
- FCC filing — 30 January 2026, seeking authority for up to one million orbital data center satellites at altitudes between 500 and 2,000 km. The filing projects that launching one million tonnes of satellites annually would yield on the order of 100 GW of AI compute capacity.
- xAI acquisition — closed 2 February 2026 in an all-stock transaction (one xAI share converting to 0.1433 SpaceX shares), creating a combined entity valued at approximately $1.25 trillion. Musk framed the rationale explicitly around space-based AI and orbital data centers, with xAI functioning as the implicit anchor customer.
- IPO — SpaceX filed for a listing targeted for mid-2026, reportedly seeking to raise upwards of $75 billion at a valuation of $1.5 trillion or more, with proceeds supporting orbital data center development.
- Terafab — announced at an Austin event on 21–22 March 2026 as a joint venture of SpaceX, Tesla and xAI. The proposed facility would produce one terawatt of processors annually, which Musk characterised as roughly fifty times current combined advanced-chip production. Initial cost estimates of $20–25 billion have been revised substantially upward in later filings. The site would span approximately 100 million square feet and draw more than 10 GW. Around 80% of output is earmarked for orbital deployment and 20% for terrestrial use in Tesla's robotics and autonomy programmes.
- Spacecraft — publicly disclosed concepts include the AI Sat Mini and a larger AI1-class node with a wingspan on the order of 70 m at roughly 100–150 kW per satellite. Musk closed the Terafab presentation with a concept video of data center satellites manufactured on the Moon and launched by electromagnetic mass driver.
Blue Origin
Blue Origin entered with two coupled programmes:
- Project Sunrise — FCC application SAT-LOA-20260310-00118, filed 19 March 2026, for up to 51,600 data center satellites in sun-synchronous orbits between 500 and 1,800 km, relying primarily on optical inter-satellite links with limited Ka-band use for control and reliability functions.
- TeraWave — a 5,408-satellite connectivity constellation announced in January 2026, providing the backhaul layer for Project Sunrise as well as serving terrestrial data center, enterprise and government customers, with deployment scheduled to begin in Q4 2027.
Blue Origin's ownership of New Glenn gives it launch independence that most competitors lack. The filing produced an unusual regulatory episode: Amazon — like Blue Origin, controlled by Jeff Bezos — had petitioned the FCC to deny SpaceX's million-satellite application, whereupon SpaceX asked the Commission to apply Amazon's own arguments to Blue Origin's substantially similar filing. NASA filed an objection to Project Sunrise in May 2026, as did DarkSky International on light-pollution grounds.
Google — Project Suncatcher
Announced 4 November 2025, Project Suncatcher is a research "moonshot" investigating constellations of solar-powered satellites carrying Google Tensor Processing Units linked by free-space optical communications.
Key elements of the accompanying preprint:
- A reference architecture of roughly 81 satellites within a cluster radius of about 1 km, in low Earth orbit at approximately 640 km.
- Radiation testing of the Trillium (TPU v6e) generation in a 67 MeV proton beam. High-bandwidth memory was the most sensitive subsystem but showed irregularities only after a cumulative dose of about 2 krad(Si), roughly three times the expected shielded five-year mission dose of 750 rad(Si). No hard failures attributable to total ionising dose were observed up to the maximum tested dose of 15 krad(Si).
- A conclusion that the concept is not precluded by fundamental physics or insurmountable economics, contingent on solving thermal management, high-bandwidth ground communication and on-orbit reliability.
- Economic parity with terrestrial systems estimated to become plausible around 2035 if launch costs fall below roughly $200/kg.
Google is partnering with Planet Labs to launch two prototype satellites by early 2027 to test TPU hardware in orbit and validate optical inter-satellite links for distributed machine-learning tasks.
Axiom Space
Axiom Space has pursued the incremental, kilowatt-scale path and holds a claim to the first commercial computing hardware operated aboard a crewed facility.
- AxDCU-1 — a data-processing prototype powered by Red Hat Device Edge, launched to the International Space Station in August 2025 and deployed that autumn.
- AxODC Node (ISS) — developed with Spacebilt, with an optical communication terminal from Skyloom and hardware from Phison Electronics and Microchip Technology, as a step toward 100 Gbps connectivity.
- ODC Nodes 1 and 2 — the first two dedicated orbital data center nodes launched to LEO on 11 January 2026, hosted on Kepler Communications spacecraft at approximately 400 km, giving round-trip latencies in the 5–20 ms range.
- Optical links — integration with Kepler Communications US and Skyloom Global relay constellations, targeting up to 10 Gbit/s and compliance with Space Development Agency interoperability standards.
- ODC T1 — a roughly half-cubic-metre server rack planned for launch by 2027, with subsequent modules on Axiom Station. Axiom has stated an intention to scale from kilowatts to megawatts and has floated lunar and Martian deployments as follow-ons.
Axiom received a Space Exploration & Aeronautics Research Fund grant of up to $5.5 million from the Texas Space Commission for ODC development, and has worked with Amazon Web Services on edge hardware.
Cowboy Space (formerly Aetherflux)
Founded in 2024 by Robinhood co-founder Baiju Bhatt as a space-based solar power company, Aetherflux rebranded as Cowboy Space Corporation in May 2026 alongside a pivot to orbital compute, following a $275 million Series B at a roughly $2 billion valuation.
Its distinguishing architectural choice is vertical integration of launch and payload: the upper stage of a company-built rocket becomes the data center once in orbit. The Stampede constellation (FCC application SAT-LOA-20260323-00135, up to 20,000 satellites) specifies:
- Dawn–dusk sun-synchronous orbits between 700 and 1,000 km;
- Satellites of 20,000–25,000 kg each, generating approximately 1 MW of usable power;
- Just under 800 GPUs per satellite, built around the Nvidia Space-1 Vera Rubin module;
- Optical inter-satellite links as the primary data path, with first proprietary launch of a one-megawatt node targeted before the end of 2028.
Cowboy has signed a Space Act Agreement with NASA's Stennis Space Center for propulsion testing.
Other US ventures
| Organization | Location | Approach | Status (Sept 2026) |
|---|---|---|---|
| Orbital Compute | US | 100,000 modular 100 kW satellites supporting ~10 GW; incremental deployment using third-party launch | FCC filing; first launch targeted 2027 |
| Sophia Space | US | TILE (Thermal-Integrated LEO Edge) — tabletop-sized tiles combining solar generation with passive radiative cooling, assembled into racks; Caltech-linked patent | ~$13.9M raised incl. $10M seed (Feb 2026); orbital demo targeted late 2027/early 2028 on an Apex Space bus |
| Lonestar Data Holdings | Florida | Off-planet data storage for resilience rather than AI compute; lunar and cislunar | Operated storage hardware aboard an Intuitive Machines lunar lander, Feb 2025; $120M agreement with Sidus Space for six satellites, first with 15 PB at an Earth–Moon libration point, launching 2027–2030 |
| OrbitsEdge | Cocoa Beach, FL | Radiation-hardened, datacenter-grade compute enclosures sold to satellite operators; partnership with HPE | First orbital demonstration planned 2026 |
| Kepler Communications | Canada/US | Optical relay and compute backbone rather than a standalone ODC | 10 satellites launched 11 Jan 2026 (Falcon 9, Vandenberg); 300 kg each, ≥4 optical terminals, multi-GPU modules; $233M+ raised |
| Skyloom Global / Spacebilt | US | Optical communication terminals and platform integration for Axiom nodes | Flying |
| Ramon.Space | Israel/US | Radiation-tolerant space-resilient computing systems; enabler layer | Established supplier |
| HPE | US | Spaceborne Computer-2 aboard the ISS; consortium member in ASCEND; OrbitsEdge partner | Long-running ISS heritage |
China
Three-Body Computing Constellation
The Three-Body Computing Constellation (三体计算星座), led by ADA Space (国星宇航, Chengdu) with Zhejiang Lab, is the first dedicated orbital computing constellation to reach orbit.
- First launch — 12 satellites on a Long March 2D from Jiuquan on 14 May 2025, providing a combined 5 POPS (peta-operations per second) and 30 TB of onboard storage, with individual satellites rated around 744 TOPS. Inter-satellite laser links operate at 100 Gbit/s.
- Programme scale — part of a wider Star-Compute Program targeting 2,800 satellites, with an ITU filing in place; 100 satellites planned by 2027 and a stated eventual capability of 1,000 POPS at 1,000-plus satellites.
- On-orbit AI — by February 2026, after roughly nine months of in-orbit testing, Zhejiang Lab reported ten AI models deployed on orbit, including an 8-billion-parameter remote sensing model and an 8-billion-parameter astronomical time-domain model, with inter-satellite networking validated and laser links reported at 99.99% availability over an eight-day test.
- Alibaba Qwen3 — deployed to the constellation in January 2026; a complex task uploaded from the ground was reportedly processed jointly across satellites and returned in under two minutes.
- Commercial partners include Alibaba Cloud, Kepu Cloud and iSoftStone. ADA Space has pursued a Hong Kong listing and is separately advancing its own constellation split between inference and training spacecraft.
Beijing Astro-Future Institute
The Beijing Astro-Future Institute of Space Technology (北京星辰未来空间技术研究院) is building what Beijing has publicly designated a space data center programme, announced at a municipal working conference in November 2025. It raised at least ¥140 million (approximately US$20 million) in June 2025, with reported backing from Lenovo and the Beijing municipal government, and leads a "space data center innovation consortium".
Its stated three-phase roadmap:
- 2025–2027 — solve on-orbit power supply and heat rejection; launch a demonstration satellite (Chenguang-1); build a first-phase constellation of 200 kW total power and 1,000 POPS, supporting "space data computed in space".
- 2028–2030 — master on-orbit assembly and construction to reduce build and operating cost; second-phase constellation supporting "ground data computed in space".
- 2031–2035 — mass production and on-orbit docking to build very large space data centers, with core compute supplied from orbit.
Longer term the institute has described dedicated data centers in 700–800 km sun-synchronous orbit in the first half of the 2030s: few in number but large, with concepts for a sixteen-spacecraft constellation of interlinked gigawatt-scale facilities.
State programmes
- CASC / 15th Five-Year Plan — ahead of the 2026–2030 plan, the China Aerospace Science and Technology Corporation identified space infrastructure as one of four focus areas, describing gigawatt-scale space-based digital and intelligent infrastructure and a new architecture integrating cloud, edge and endpoint compute, with the goal of enabling space-data/space-compute, ground-data/space-compute and simultaneous space-ground computing.
- Xingshu Plan (星枢计划) — Shanghai's flagship initiative, unveiled with Fudan University at the 2026 World AI Conference in July 2026. Three phases scale from a verification constellation of 2 computing plus 12 edge satellites, to 50 computing and 100 edge satellites, to a 1,000-satellite on-demand orbital computing service — effectively a rental model for governments and enterprises in weather forecasting, disaster response, maritime monitoring and grid inspection. A space computing hub opened in Songjiang on 31 August 2026.
- Tiansuan Constellation — BUPT/Spacety open research platform, with Huawei Cloud as a co-construction partner (see above).
- CAICT — the China Academy of Information and Communications Technology has established a space computing power professional committee to coordinate standards.
Chinese programmes benefit from a decade and a half of research on space-based solar power, begun in 2008 with technology testing from 2013, since a space data center reuses much of the same power-generation and thermal-management hardware.
Europe
ASCEND
ASCEND (Advanced Space Cloud for European Net zero emission and Data sovereignty) is a European Commission feasibility study funded under Horizon Europe, contracted to a consortium led by Thales Alenia Space in 2022 and launched in 2023. Partners span environmental analysis (Carbone 4, VITO), cloud and IT (Orange Business, CloudFerro, Hewlett Packard Enterprise), launch (ArianeGroup) and orbital systems (DLR, Airbus Defence and Space, Thales Alenia Space).
Findings published in June 2024:
- Space-based data centers are technically feasible and can deliver a net environmental benefit provided a launcher roughly ten times less emissive than current vehicles is developed — validated as achievable with ArianeGroup input and analysis from ESA's PROTEIN study.
- Orbital facilities consume no water for cooling, a material advantage under increasing drought pressure.
- A target of 1 GW of orbital capacity by 2050, with the study concluding the project economically viable. (Reported return-on-investment figures vary considerably between sources, from several billion to several hundred billion euros by 2050; the primary study documentation should be consulted.)
- Modular infrastructure would be robotically assembled on orbit using technologies from the EROSS IOD (European Robotic Orbital Support Services In-Orbit Demonstrator), led by Thales Alenia Space, with a first flight in 2026.
An ASCEND in-orbit demonstration deploying a small-scale orbital data center module to validate European technologies is targeted for 2028. The programme is framed principally around the European Green Deal and European digital sovereignty rather than around AI compute scaling.
Japan
Space Compass
Space Compass Corporation is a 50/50 joint venture of NTT and SKY Perfect JSAT, agreed in April 2022 following a 2021 collaboration and formally established in July 2022 in Chiyoda-ku, Tokyo, with an initial investment of ¥6 billion.
Its Space Integrated Computing Network is a multi-orbit, optical-communication-based infrastructure combining:
- An optical data relay service carrying observation-satellite data to the ground via GEO, overcoming the capacity and contact-window limits of radio downlink;
- A space data center layer, progressively adding satellites with advanced computing functions so that data is analysed on orbit and only useful results returned;
- NTT's IOWN all-photonics technology as the transport substrate;
- HAPS (high-altitude platform station) services for low-latency coverage, disaster response and remote areas.
Space Compass has been selected under JAXA's Space Strategy Fund for next-generation optical data relay and holds a Japanese Ministry of Defense contract for a geostationary optical communication technology demonstration. The venture predates most Western ODC announcements by several years, though its schedule has slipped from the originally stated 2025 service start.
Other regions
- India — SkyServe develops orbital edge computing software, working with NASA's Jet Propulsion Laboratory to test AI models on D-Orbit spacecraft.
- ESA — beyond PROTEIN and ASCEND, ESA's earlier PhiSat-1 cubesat (with Intel and Ubotica) demonstrated onboard AI inference for Earth observation.
Regulatory landscape
Orbital data centers occupy an awkward regulatory category: spacecraft whose primary commercial function is compute capacity rather than communications, Earth observation or navigation. Existing spectrum and licensing frameworks were not written for this case, and the reliance on optical inter-satellite links means comparatively little radio spectrum is requested relative to constellation size.
Filings on record
| Applicant | System | Satellites | Filed | Altitude / orbit |
|---|---|---|---|---|
| SpaceX | Orbital data center constellation | up to 1,000,000 | 30 Jan 2026 | 500–2,000 km |
| Starcloud | (ODC constellation) | up to 88,000 | Feb 2026 | 600–850 km SSO dusk–dawn |
| Blue Origin | Project Sunrise | up to 51,600 | 19 Mar 2026 | 500–1,800 km SSO |
| Blue Origin | TeraWave (connectivity backbone) | 5,408 | Jan 2026 | — |
| Cowboy Space | Stampede Data Center System | up to 20,000 | Mar–May 2026 | 700–1,000 km dawn–dusk SSO |
| Orbital Compute | (ODC constellation) | ~100,000 | 2026 | LEO |
Filings are proposals under review, not approvals. Objections on file include Amazon's petition to deny SpaceX's application; SpaceX's reciprocal request that the same standards be applied to Blue Origin; a NASA objection to Project Sunrise in May 2026; and DarkSky International's objection on the grounds that such constellations would permanently alter the night sky.
Criticism and open questions
Thermal
Heat rejection is the most frequently cited technical objection. In vacuum, radiation is the only mechanism available, and radiator area scales with dissipated power.
- The ISS active thermal control system rejects up to about 70 kW across roughly 422 m² of ammonia-loop radiators — approximately 166 W/m² in practice, well below theoretical maxima once solar exposure, Earth infrared and system losses are accounted for.
- Modelling published in mid-2026 puts an H100-class (~700 W) GPU at roughly 1.4 m² of radiator area and a 40 kW rack at roughly 80 m², with five-year surface degradation adding around 40% to required area. Optimised high-temperature radiator designs reach an effective PUE near 1.3 at approximately 2.5 kg/kW of thermal mass.
- Some analyses of gigawatt-class single facilities have produced radiator area figures in the millions of square feet.
- A design spiral is often described: packing accelerators densely concentrates heat faster than it can be moved to radiators, while spreading them out inflates interconnect and shielding mass. Thermal cycling roughly every 90 minutes adds fatigue loading.
The counter-argument from operators is that thermal management at 100 kW per spacecraft is a well-understood engineering trade with decades of LEO heritage, and that 2025–2026 flight data suggests power delivery, not cooling, is the current limiter on hardware actually in orbit. The dispute is therefore less about physics than about the scale at which the trade turns unfavourable.
Economics
The case rests almost entirely on launch cost. Falcon 9 costs are on the order of $2,700/kg; Starship targets approximately $200/kg; the single publicly contracted Starship price to date implies roughly $600/kg at full payload. Most published models place the inflection point for orbital compute somewhere between $200 and $500/kg.
Additional economic objections:
- Hardware obsolescence — AI accelerators have a useful competitive life of roughly three to five years, against satellite platforms designed for 15–25 years. This has driven proposals for modular plug-in compute cards swappable under a long-lived bus, though on-orbit servicing at constellation scale is itself unproven.
- Radiation premium — the cost and performance penalty of space-qualifying hardware is arguably the single variable determining whether the case closes.
- Maintenance — terrestrial facilities benefit from component replacement; distributed orbital assets do not.
- Operator projections should be treated as advocacy: Starcloud, for instance, has projected that operating a 40 MW orbital cluster over ten years would cost roughly $8.2 million against approximately $167 million for a terrestrial equivalent.
Sceptical commentary through 2026 has been substantial, including a widely circulated February 2026 video by science communicator Kyle Hill and remarks by Voyager Technologies chief executive Dylan Taylor identifying cooling as the fundamental unsolved problem. A dedicated economics session at SmallSat Europe in May 2026 marked a shift from vision-led to cost-led discussion of the category.
Space environment
- Debris and collision risk — roughly 14,000 active satellites currently operate in LEO. A one-million-satellite constellation represents an increase of nearly two orders of magnitude, materially raising collision probability. A full Kessler syndrome cascade would be effectively irreversible on human timescales. Approximately 44,000 tracked objects larger than 10 cm are already capable of destroying a satellite.
- Astronomy — leading astronomers have warned that data center constellations, combined with proposed orbiting mirror projects, would severely degrade ground-based observation and the visible night sky.
- Reentry chemistry — the atmospheric effects of large-scale satellite reentry, particularly alumina and other metal oxides in the mesosphere, remain poorly characterised.
- Material degradation — unattenuated ultraviolet exposure degrades radiator surfaces, reducing performance over mission life.
Governance and sovereignty
Brookings and others have argued that the gap between projected capability and demonstrated capability is itself a governance risk, encouraging regulatory decisions on the basis of claims not yet validated. A distinct concern is jurisdictional: if citizen-generated data is processed in orbit, it is unclear whether sovereignty rests with the country of origin, the launching state, or the constellation operator — a question with particular force for nations that lack independent launch capability.
Market context
Space startup funding reached a record $20.3 billion in 2026, separate from SpaceX's IPO, with orbital compute emerging as a distinct investment category. Goldman Sachs published "The Second Space Age" on 13 August 2026, projecting a $1.8 trillion space economy by 2035. Earlier market estimates for the in-orbit data center segment specifically were considerably more modest, in the low billions by the end of the decade — a divergence that reflects genuine uncertainty about whether the category is a niche edge-processing market or a replacement layer for terrestrial hyperscale.
A useful analytical framing divides the field into three tiers:
- Near-term operators with hardware in orbit — Axiom Space, Starcloud, ADA Space/Zhejiang Lab, Kepler.
- Platform builders and enablers — Nvidia, Ramon.Space, Sophia Space, Skyloom, OrbitsEdge, HPE.
- Vision-led mega-architectures — SpaceX/xAI, Blue Origin Project Sunrise, Cowboy Space, CASC.
The economics may therefore resemble terrestrial cloud buildouts, in which value accrues not only to eventual orbital hyperscalers but to whoever controls bottleneck technologies such as radiation-tolerant AI compute, low-SWaP optical terminals and launch capacity.
Timeline
| Date | Event |
|---|---|
| 2008 | China begins research into space-based solar power, later relevant to orbital data center power and thermal systems |
| Nov 2018 | Huawei Cloud initiates and open-sources KubeEdge |
| Dec 2021 | First Tiansuan Constellation satellite launched with KubeEdge; described as the world's first cloud-native satellite |
| Apr–Jul 2022 | NTT and SKY Perfect JSAT establish Space Compass |
| Nov 2022 | European Commission contracts Thales Alenia Space to lead the ASCEND study |
| Sept 2023 | Huawei releases Data Center 2030 at HUAWEI CONNECT, naming space data centers a key structural pattern |
| Jun 2024 | ASCEND feasibility results published |
| Sept 2024 | Starcloud (as Lumen Orbit) publishes white paper on gigawatt-scale orbital AI compute |
| Feb 2025 | Lonestar operates data storage hardware en route to the Moon aboard an Intuitive Machines lander |
| 14 May 2025 | First 12 satellites of the Three-Body Computing Constellation launched |
| Aug–Sept 2025 | Axiom AxDCU-1 deployed aboard the ISS |
| 4 Nov 2025 | Google announces Project Suncatcher |
| Nov 2025 | Starcloud-1 launched; first Nvidia H100 in low Earth orbit |
| Nov 2025 | Beijing announces its space data center construction programme |
| 11 Jan 2026 | Axiom's first two dedicated ODC nodes and Kepler's 10 optical relay satellites launched |
| Jan 2026 | Alibaba Qwen3 deployed to the Three-Body constellation; Blue Origin announces TeraWave |
| 30 Jan 2026 | SpaceX files with the FCC for up to one million ODC satellites |
| 2 Feb 2026 | SpaceX completes acquisition of xAI at a combined ~$1.25 trillion valuation |
| Feb 2026 | Starcloud files for 88,000 satellites; Zhejiang Lab reports ten AI models running on orbit; Sophia Space raises $10M |
| Feb 2026 | Jensen Huang publicly characterises orbital data center economics as poor today but improving |
| Mar 2026 | Nvidia unveils Space-1 Vera Rubin Module at GTC; Starcloud raises $170M at $1.1B |
| 19 Mar 2026 | Blue Origin files Project Sunrise (51,600 satellites) |
| 21–22 Mar 2026 | Musk unveils Terafab; SpaceX details orbital data center spacecraft |
| May 2026 | Aetherflux becomes Cowboy Space, raises $275M, files for 20,000-satellite Stampede; NASA objects to Project Sunrise |
| Jul 2026 | Shanghai unveils the Xingshu Plan at the World AI Conference |
| 13 Aug 2026 | Goldman Sachs publishes "The Second Space Age" |
| 21 Aug 2026 | Starcloud raises $250M at $2.3B with Nvidia and Cisco participating |
| 31 Aug 2026 | Shanghai opens its Songjiang space computing hub |
| Early 2027 | Planned: Google/Planet Suncatcher prototype satellites; Starcloud-2 |
| 2028 | Planned: ASCEND in-orbit demonstration; Cowboy Space first 1 MW node; Starcloud Space-1 Vera Rubin flight |
External links
- Huawei — Data Center 2030
- Google Research — Project Suncatcher
- Axiom Space — Orbital Data Centers
- ASCEND (Horizon Europe)
- KubeEdge — Cloud Native Edge Computing Satellite case study
- Nvidia — Space computing platforms
- Space Compass Corporation
- SpaceNews — Orbital Data Centers coverage
- Thales Alenia Space — ASCEND feasibility results
