A systems vision for AI infrastructure in space is developed as the systems layer that manages AI capabilities across spacecraft, orbital networks, ground stations, and cloud backends, while treating orbital and physical state as part of the resource model.
Abstract
Satellites are becoming programmable computing platforms capable of running increasingly demanding AI workloads. This shift raises a systems problem: how can AI services remain deployable, manageable, and recoverable after launch when compute capacity, connectivity, energy, and thermal headroom vary over orbital time? This paper develops a systems vision for AI infrastructure in space. We define it as the systems layer that manages AI capabilities across spacecraft, orbital networks, ground stations, and cloud backends, while treating orbital and physical state as part of the resource model. We synthesize relevant foundations from terrestrial AI infrastructure, satellite networking, and satellite edge computing, and examine the physical constraints that directly shape system design. We further ground this vision in three in-orbit case studies spanning the node, platform, and service levels. Telemetry from BUPT-1 satellite shows that usable compute capacity is bounded by thermal and energy envelopes. SateLight on BUPT-2 satellite reduces application-update transmission latency by 56.54% on average and up to 91.18%, with 100% update correctness. A stateful VLM serving case further shows that thermal interruptions make execution-state recovery a first-class systems problem. These observations motivate a research agenda for space-native resource management, lifecycle support, and sustained AI service across space and ground.
Mega-constellations of Low Earth Orbit (LEO) satellites are enabling a new Space Cloud paradigm in which edge servers hosted on-board process tasks autonomously, reducing ground-segment latency for globally dispersed users. Realizing this vision requires Collaborative Edge Intelligence (CEI): distributed algorithms that coordinate Multi-access Edge Computing (MEC) server activation, task offloading, and routing across a time-varying orbital topology. Existing evaluation tools are inadequate—simulation platforms abstract away protocollevel behavior, while emulation testbeds lack orbital dynamics. This paper presents MeteorNet, an open-source, continuous-time emulation platform that integrates SGP4 orbital propagation, Mininet/ONOS network emulation, Docker-containerized MEC services, and MongoDB telemetry. Supporting eight orchestration strategies, including Fuzzy Logic and distributed Reinforcement Learning controllers, MeteorNet enables apples-to-apples CEI benchmarking under realistic orbital conditions. Experiments show that intelligent controllers halve the MEC activation cost at low load, while revealing an orbital-visibility bottleneck that limits task success at high load, regardless of the orchestration policy.
J. Fraire, Camilo Rojas, Fabio Patrone et al.· International Mediterranean...· 0 citations
Frontier autonomous-driving models -- especially vision-language-action (VLA) models, whose forward pass approaches $\sim$60~TFLOPs -- are outgrowing economical onboard deployment, since peak hardware sits idle most of the day. Cloud inference can instead share GPUs across active vehicles, but the vehicle must upload through a capacity-limited uplink, reach a GPU without queueing, and return a decision within the closed-loop budget. This paper asks: can the cloud drive? We answer with an analytical framework coupling communication limits, a roofline GPU service model, stochastic latency, and utilization-aware cost across three model classes, three offloading strategies, and three communication generations, applied to New York City. Separating a reactive 100~ms budget from a 300~ms deliberative tier (presuming an onboard reactive fallback), we find three \emph{nested} binding regimes. Communication binds first in dense cells: 5G fails early, 5G-Advanced is the practical threshold for feature-level offloading, and 6G adds headroom. Compute binds next under the reactive budget: near-term VLA is latency-infeasible regardless of bandwidth, because autoregressive FP16 decode is memory-bandwidth-bound (~114 ms on 2025 hardware). Its floor clears 100 ms around 2027; 6G then admits feature-level VLA by ~2028, 5G-Advanced only at light loading and not the dense corridor, and the deliberative tier from 2026. Cost binds last: once admissible, utilization-pooled cloud GPUs undercut onboard hardware for VLA, whose baseline (up to \$8,500 per vehicle-year) is expensive and idle; feature-level offloading (S2) is where the VLA cost crossover concentrates. Latency decides which model is admissible in which year; cost decides whether it is economical.
While compute and application state for networked services are typically hosted on terrestrial servers, a specific subset of workloads such as multi-party interactive applications can benefit from in-orbit compute, particularly in regions with sparse terrestrial network infrastructure. Hosting stateful services on LEO satellites, however, introduces a fundamental challenge: the satellite serving a user changes every few minutes due to orbital motion, threatening to disrupt users' active connection sessions. This paper presents TETHER, a network abstraction that preserves application-visible connectivity despite continual satellite movement. TETHER combines persistent virtual IPs, anticipatory state migration, and transport continuity mechanisms to transparently relocate application state between satellites before performance degradation occurs. The system leverages Linux Checkpoint/Restore In Userspace (CRIU) for process migration and exploits predictable orbital trajectories to proactively schedule migrations. We explore two migration policies with different trade-offs between latency and overhead. directly-above migrates application state to the satellite closest to the user, minimizing latency at the cost of more frequent migrations. In contrast, edge-of-reachability migrates state to a satellite that can stay under an acceptable latency budget the longest, reducing migration frequency while tolerating higher latency. We evaluate both policies using the xeoverse LEO network emulator. Our results show that anticipatory migration preserves long-lived TCP sessions across satellite transitions and enables applications to trade latency tolerance against migration overhead.
Emi Digby, Nishanth R. Sastry· Proceedings of the ACM SIGCO...· 0 citations
Space Computing Power Networks (SCPNs), also termed as Satellite Comptuting Power Networks, as an integration of satellite networks, orbital computing, and terrestrial infrastructure, have been becoming an emerging architecture and attracting growing research attention during the past few years. Beyond meeting the differentiated intelligent communication, computing, and caching service requirements from users and terminals across space, air, ground, and sea, SCPNs hold significant importance for space exploration, earth observation, environment monitoring, remote user activities, and so on. There is no doubt that SCPNs will be the critical part of 6G to realize the ubiquitous and seamless intelligence. However, compared to traditional Terrestrial Computing Power Networks (TCPNs) and Satellite Computing Networks (SCNs), SCPNs holds the uniqueness, such as the cycled node movements, hierarchical network topology, extremely large network scalability, ubiquitous resource heterogeneity and constraints, and particular space computing environment. The system integration, protocol optimization, service orchestration, and sustainable operation of SCPN have inspired many meaningful research and projects. Considering existing survey papers mainly focus on scenarios of TCPNs carrying large-scale and complex computing tasks, this paper presents a comprehensive survey of state-of-the-art research on SCPN, covering various aspects ranging from system architecture, applications and challenges, and diversified Quality of Service (QoS) metric analysis and optimization. Finally, a number of potential future research directions have also been discussed to enlighten more innovative works.
Shi Chen, Y. Wu, Bomin Mao et al.· IEEE Communications Surveys...· 0 citations
Evaluating end-to-end network performance in Low Earth Orbit (LEO) satellite mega-constellations requires systemlevel testbeds capable of capturing extreme satellite mobility, rapidly evolving topologies, and realistic protocol behavior at scale. Existing approaches rely on analytical models or eventdriven simulations that abstract away protocol-level interactions and time-varying orbital dynamics. This paper demonstrates the capability of our developing hybrid emulation-simulation testbed for end-to-end system-level evaluation of LEO megaconstellations, using ground station placement as a representative case study. The testbed instantiates each satellite as an isolated network namespace with a real protocol stack, computes all propagation delays from instantaneous satellite distances and the speed of light, and evaluates performance across a full orbital period rather than a single static snapshot. Using the testbed, we evaluate four placement strategies across Starlink Gen1, OneWeb, and Amazon Kuiper architectures. The results reveal previously unreported system-level insights, including latency saturation beyond approximately 100 ground stations, a reduction in orbital sensitivity from 10-15 ms to less than 1 ms, and fundamental performance differences between ISL-enabled and non-ISL architectures.
Murat Parlakisik, Ertan Ozturk· International Mediterranean...· 0 citations
The AI-RAN Alliance has proposed the coexistence of Artificial Intelligence (AI) workloads and Radio Access Network (RAN) functions on shared edge infrastructure, referred to as AI-and-RAN. However, practical understanding of resource contention between compute-intensive AI tasks and timing-constrained RAN operations remains limited, largely due to the lack of end-to-end platforms that enable controlled co-location and measurement across system and RAN layers. To fill this gap, this paper presents an end-to-end AI-and-RAN testbed for controlled co-location experiments on an unified CPU–GPU platform. We deploy a containerized 5G RAN stack with an Open Radio Access Network (O-RAN) compliant E2 interface for monitoring alongside edge Large Language Model (LLM) inference workloads. To expose resource sensitivity, we intentionally constrain the CPU budget and vary AI workload intensity. Through measurements of Medium Access Control (MAC) layer key performance indicators (KPIs) and edge-level resource utilization, we demonstrate that LLM inference reduces downlink (DL) throughput and increases Block Error Rate (BLER), providing empirical insights to inform future AI-and-RAN orchestration strategies.
Taeil Jung, Hyun-Min Yoo, Kyung-Sook Kim et al.· International Conference on...· 0 citations
Related blog posts
Microsoft Research Blog· microsoft.comAug 31, 2026
What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.
MIT News · Artificial Intelligence· news.mit.eduAug 27, 2026
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.