Hybrid Computing: How AI Is Reshaping Cloud, Edge, and On-Device Processing

This article synthesizes recent literature (2022–2025) to examine how advances in artificial intelligence (AI) are driving hybrid computing paradigms that distribute computation across cloud, edge, and on-device environments. It discusses architectural trends, technical trade-offs (latency, privacy, energy), enabling technologies, and empirical indicators of adoption. Policy and research implications for deploying hybrid AI systems are identified.

1. Introduction

The proliferation of AI applications and the concomitant growth in data volumes have catalyzed a migration from monolithic cloud architectures toward hybrid computing models that combine cloud, edge, and on-device resources. Hybrid computing denotes architectures in which workload placement is dynamic and determined by constraints such as latency, privacy, energy consumption, and cost (Andriulo et al., 2024; Ye et al., 2023). Recent market forecasts and technical surveys indicate rapid investment in both cloud and edge infrastructures as organizations pursue distributed AI solutions (Gartner, 2024; Singh & Gill, 2023).

2. Conceptual framework and defining characteristics

Hybrid AI computing can be characterized by three strata: (1) centralized cloud data centers offering high-capacity training and large-scale aggregation; (2) proximate edge nodes providing low-latency inference and pre-processing; and (3) on-device execution enabling immediate responses and improved privacy (Wang et al., 2025; Su et al., 2022). Task partitioning across these strata is guided by application SLAs, communication cost, model size, and the availability of hardware accelerators (FPGAs, GPUs, NPUs).

2.1 Workload orchestration

Orchestration mechanisms—scheduling policies, dynamic offloading, and adaptive compression—determine how AI workloads migrate between tiers (Ye et al., 2023). Intelligent schedulers use telemetry and learned policies to allocate time-sensitive inference workloads to the edge while reserving cloud resources for batch training and model updates (Ye et al., 2023).

3. Technical trade-offs: latency, privacy, and energy

The hybrid approach is an optimization across competing objectives. Edge and on-device inference reduce end-to-end latency and lower network bandwidth consumption, critical for real-time systems (Andriulo et al., 2024). Conversely, cloud resources remain essential for large-scale model training, dataset aggregation, and complex analytics where latency is tolerable (Gartner, 2024).

Privacy is improved by local processing and federated paradigms that keep raw data on devices while sharing model updates (Wang et al., 2025). However, federated methods introduce communication overhead and heterogeneity challenges requiring careful compression and adaptive learning strategies (Singh & Gill, 2023).

Energy constraints drive different optimization targets: on-device inference emphasizes model compression (pruning, quantization), runtime efficiency, and hardware acceleration; edge nodes provide intermediate compute capability that balances energy and latency; cloud platforms emphasize throughput and energy efficiency at datacenter scale. Recent empirical analyses highlight that system-level energy per inference may be lowest when an optimized hybrid split is used for many IoT workloads (Su et al., 2022).

4. Enabling technologies and methods

Several technological advances underpin hybrid AI adoption. Model efficiency techniques (knowledge distillation, pruning, quantization) make on-device and edge inference practical (Wang et al., 2025). Containerization, lightweight virtualization, and orchestration frameworks enable rapid deployment and migration between tiers (Andriulo et al., 2024). Hardware trends—specialized NPUs in mobile SoCs and micro-data centers at the edge—support inference workloads that previously required cloud resources (Ye et al., 2023).

4.1 Federated and split learning

Federated learning provides a privacy-preserving method for distributed training across devices and edge nodes, while split learning partitions model execution across device and server to reduce device load. Empirical studies demonstrate improved privacy but require new protocols to cope with heterogeneous capacity and unreliable connectivity (Singh & Gill, 2023; Wang et al., 2025).

5. Empirical signals: market and deployment trends

Industry and market reports indicate concurrent expansion of both cloud and edge investments. Gartner projected global end-user spending on public cloud services to reach approximately $723.4 billion in 2025, and forecast that the majority of organizations will adopt hybrid cloud strategies in the near term (Gartner, 2024). IDC and related analyses report accelerating edge spending as firms prioritize low-latency, AI-driven inference at the network periphery (IDC, cited in industry analyses; see Ye et al., 2023).

Academic deployments corroborate industry trends: hybrid cloud–edge systems have been successfully deployed for large scientific facilities and industrial Internet-of-Things (IIoT) applications, yielding measurable reductions in end-to-end processing time and network load (Ye et al., 2023; Andriulo et al., 2024).

6. Challenges and open research directions

Despite progress, several challenges constrain hybrid AI systems. First, heterogeneity of devices and network conditions complicate robust orchestration. Second, standardization of telemetry and cross-tier APIs is immature, hindering portability. Third, rigorous evaluation methodologies for hybrid deployments (measuring privacy, latency, energy, and cost jointly) remain underdeveloped (Singh & Gill, 2023).

Research priorities include adaptive partitioning algorithms that consider user privacy preferences; energy-aware schedulers integrating hardware power models; secure protocols for federated learning resistant to poisoning and inference attacks; and socio-technical work to ensure equitable access to hybrid AI benefits across geographies (Wang et al., 2025; Su et al., 2022).

7. Practical and policy implications

Practitioners should adopt hybrid designs when applications require low latency, privacy preservation, or reduced bandwidth. Policymakers must consider data-sovereignty implications, since edge and on-device processing can support regulatory compliance by restricting data flows. Standards efforts for hybrid orchestration and federated governance will accelerate safe deployment (Andriulo et al., 2024).

8. Conclusion

Hybrid computing—driven by the maturation of efficient AI methods, specialized hardware, and orchestration frameworks—represents a practical architecture for reconciling the divergent requirements of latency, privacy, and scale. Continued interdisciplinary research and standardized engineering practices will determine the extent to which hybrid AI transforms critical sectors such as healthcare, transportation, and industrial automation.

A concise academic visualisation of distributed AI across cloud, edge, and devices.

Vector illustration showing the cloud, edge, and on-device layers connected by data flow and a stylised neural network overlay.

---

References (APA style)

1. Andriulo, F. C., Fiore, M., Mongiello, M., Traversa, E., & Zizzo, V. (2024). Edge computing and cloud computing for Internet of Things: A review. Informatics, 11(4), 71. https://doi.org/10.3390/informatics11040071
2. Gartner. (2024, November 19). Gartner forecasts worldwide public cloud end-user spending to total $723 billion in 2025. Gartner Newsroom. https://www.gartner.com/en/newsroom/press-releases/2024-11-19-gartner-forecasts-worldwide-public-cloud-end-user-spending-to-total-723-billion-dollars-in-2025
3. Singh, R., & Gill, S. S. (2023). Edge AI: A survey. Internet of Things and Cyber-Physical Systems, 3, 71–92. https://doi.org/10.1016/j.iotcps.2023.02.004
4. Su, W., et al. (2022). AI on the edge: A comprehensive review. International Journal (Springer). https://doi.org/10.1007/s10462-022-10141-4
5. Wang, X., Tang, Z., Guo, J., Meng, T., Wang, T., & Jia, W. (2025). Empowering edge intelligence: A comprehensive survey on on-device AI models. arXiv:2503.06027. https://arxiv.org/abs/2503.06027
6. Ye, J., Wang, C., Chen, J., Wan, R., Li, X., Sepe, A., & Tai, R. (2023). Cloud–edge hybrid computing architecture for large-scale scientific facilities augmented with an intelligent scheduling system. Applied Sciences, 13(9), 5387. https://doi.org/10.3390/app13095387

Note: Selected market figures cited (Gartner, 2024) refer to public-cloud spending forecasts. Where possible, scholarly references from 2022–2025 were prioritized (MDPI, Elsevier, Springer, ACM/arXiv). This article is original and intended for academic dissemination; readers are encouraged to consult the cited sources for deeper technical detail.