The telecoms industry has spent years searching for a use case that genuinely requires edge computing, ultra-low latency and a more intelligent RAN. Multi-access Edge Computing (MEC) has long been presented as a foundational element of future network architecture, yet many early deployments struggled to move beyond limited trials and niche applications. AI-RAN has followed a similar path. While the concept has generated significant industry interest, many operators have continued to question where the real commercial and technical value lies.
Recent demonstrations by SoftBank and Ericsson around network-enabled Physical AI suggest the industry may finally have found a use case that naturally brings these technologies together.
Unlike earlier MEC demonstrations focused on content delivery, video optimisation or traffic offload, Physical AI introduces a closed-loop interaction between robots, AI inference and the network itself. Rather than treating AI as an application that operates independently on top of the network, SoftBank is positioning AI as something that must be tightly integrated with connectivity, compute and radio infrastructure. This has important implications for MEC, distributed compute and the future evolution of telecoms infrastructure.
The concept behind Physical AI is relatively straightforward. Autonomous robots, drones and industrial systems need to perceive their environment, process information and make decisions in real time. Some AI inference can run locally on the device, but onboard compute is often constrained by size, thermal limits, cost and power consumption. More advanced AI models may require significantly more processing capability than can realistically be embedded into every endpoint.
SoftBank’s approach is to treat MEC as an extension of the robot’s compute environment. The network effectively becomes part of the control loop rather than simply acting as a transport layer between devices and the cloud.
This model requires a much tighter relationship between the RAN and application workloads than traditional mobile networks were originally designed to support. In conventional architectures, the RAN largely operates independently of application behaviour. SoftBank’s AI-RAN architecture, developed in collaboration with Ericsson, instead monitors radio conditions, compute availability and application requirements simultaneously. AI workloads can then move dynamically between the device and MEC infrastructure depending on latency, congestion and processing demands.
Ericsson’s role is particularly important because it demonstrates how differentiated connectivity becomes part of the AI execution framework itself. Network slicing, priority handling and RAN automation are integrated into the orchestration process so that the network can continuously adapt to changing AI workload requirements rather than relying on static QoS policies.
A robot performing real-time perception and navigation may initially process workloads locally. If additional compute power is required, inference tasks can be offloaded to nearby MEC resources. If network conditions deteriorate or latency increases, workloads may shift back onto the device. The network continuously orchestrates these decisions while attempting to maintain predictable, deterministic performance levels.
This is significantly different from the fixed slicing models often discussed during the early 5G era. Instead of static policies, the network dynamically adjusts connectivity and workload placement according to application behaviour and infrastructure conditions.
SoftBank’s broader strategy reinforces this direction through its Telco AI Cloud architecture, which combines GPU infrastructure, MEC platforms and RAN intelligence under a unified orchestration framework. The objective is to support AI workloads that span cloud, edge and endpoint devices across sectors such as industrial automation, logistics and construction.
A critical milestone in this development is the interworking between SoftBank’s AITRAS orchestrator and the Ericsson Intelligent Automation Platform (EIAP). This collaboration reflects a future where AI orchestration platforms and RAN automation systems operate through tightly integrated control and policy frameworks. By integrating these layers, the orchestrator can make informed decisions about where to place an AI workload based on a real-time view of the RAN’s capacity and health.
This shift highlights why many previous MEC deployments struggled to achieve large-scale commercial success. Early edge computing use cases rarely demanded strict latency guarantees or continuous workload mobility between device and edge. In many scenarios, centralised cloud platforms remained entirely sufficient.
Physical AI changes that. Real-time robotics and autonomous systems require predictable latency and continuous responsiveness. They also require the ability to distribute AI workloads dynamically across multiple compute domains without interrupting operations. Achieving this level of coordination demands a far deeper integration between MEC, AI orchestration and the RAN than most operators currently deploy.
If the network is expected to participate directly in AI execution and workload placement decisions, the RAN itself becomes part of the distributed AI infrastructure platform.
The implications for telecoms infrastructure are substantial. Operators aiming to support Physical AI applications may need to rethink network architecture around distributed compute rather than centralised capacity. MEC deployments would require tighter integration with radio infrastructure, orchestration systems and AI frameworks.
The industry has discussed this convergence for years through initiatives such as the AI-RAN Alliance and broader work around AI-native networking. What makes the SoftBank and Ericsson work particularly interesting is that it connects these ideas into a practical end-to-end implementation focused on real-world autonomous systems.
Physical AI may finally provide the commercial and technical justification the industry has been searching for to move MEC and AI RAN from research projects and isolated trials into mainstream network architecture. It hints at a future where operators are no longer simply connectivity providers, but providers of distributed AI infrastructure platforms capable of supporting real-time autonomous systems.
Related Posts:
- Telecoms Infrastructure Blog: Understanding the Internet’s Hidden Infrastructure
- Telecoms Infrastructure Blog: SK Telecom's Vision for Future Telco Infrastructure in the AI Era
- Telecoms Infrastructure Blog: 'Story of the Edge' by Akamai
- The 3G4G Blog: Edge Computing Tutorial from Transforma Insights
- Telecoms Infrastructure Blog: AWS for Public and Private 5G Networks
- Telecoms Infrastructure Blog: SK Telecom’s 5G MEC Status and Plan
- Operator Watch Blog: How is KT Reducing the Latency of their 5G Network
