Massive AI Model Orchestration for 6G Networks: Architecture, Optimization, and Energy-Efficient Deployment

Sixth-generation (6G) networks will serve large pre-trained artificial intelligence (AI) foundation models from cloud, edge and device tiers, but placing, compressing and scheduling them under latency, energy and accuracy constraints remains open. This paper proposes Massive AI Model Orchestration (MAIMO), a three-tier framework that manages model residency and request routing inside the radio access and core networks. The orchestration problem is formulated as a mixed-integer multi-objective program for which a…

By the Truvace deskOctober 21, 20262 min read1 sourcePeer-reviewedTRV-2026-2122· cite & verify

In brief

The paper proposes MAIMO, a three-tier framework that decides where large pre-trained AI models live and how requests are routed across cloud, edge and device tiers inside 6G radio access and core networks. An online controller pairs a BiLSTM traffic predictor with proximal policy optimization to make placement and routing decisions.

Serving ever larger AI models over future mobile networks will be constrained by latency, energy and accuracy, so results showing lower delay and energy use matter for network design. However, every result comes from a purpose-built simulator rather than a live network, so real-world performance remains unproven.

Main points

Mean end-to-end latency was 11.3 ms versus 22.7 ms for cloud-only service across 20 paired replications.
Ultra-reliable low-latency vehicle-to-everything requests were served in 2.1 ms on average with warm caches.
An analytic energy model gives 5.46 J per inference versus 16.73 J for cloud-only service.
Constrained carbon-aware workload shifting reduced operational carbon by 27.2%.

The gain

In simulation, the MAIMO orchestration framework roughly halves end-to-end latency and cuts energy per inference to about a third of cloud-only serving, with carbon-aware shifting lowering operational carbon.

The rundown

The authors formulate orchestration as a mixed-integer multi-objective program. They prove a Pareto-optimal solution exists and show the offline placement-routing version is NP-hard, which motivates an online learned controller instead of exact optimization.

In the comparison, MAIMO had the lowest service-level-agreement violation rate among schemes that met the accuracy floor. It only tied deep Q-network and contextual-bandit controllers on mean latency, so its latency edge over those learned baselines was not shown.

The simulator is calibrated to standardized urban-macrocell channel models, and the energy figures come from a transparent analytic model rather than hardware measurements.