Telco for AI: Six Use Cases Telecoms Can Deploy Now

Each year a telco sits on the sidelines, hyperscalers quietly capture billions in AI compute revenue that telcos were positioned to win. The window is narrow: global edge computing investment is forecast to reach nearly $380 billion by 2028, according to IDC, as enterprises push compute closer to where data is generated and decisions made. Miss this shift, and telcos risk seeing their fiber and edge footprints become dumb pipes, as enterprises pay a premium elsewhere for performance and low-latency AI infrastructure that telcos already own.

In Part 1 of this series, we laid out the architectural case: distributed inference creates opportunity, and a standards-based orchestration layer can turn existing fiber, MEC, and 5G assets into a monetizable AI fabric. The short version: telcos already own the real estate that hyperscalers are racing to build.

Part 2 is where theory meets billing cycle.

This piece gets concrete. We’ll walk through the specific scenarios where telcos can generate direct revenue from that infrastructure: selling GPU capacity with latency guarantees, running federated inference platforms, and offering dynamic network slices for enterprises that cannot ship petabytes to centralized clouds.

• Distributed AI training
• Edge AI inference
• GPU roaming
• GPU-as-a-Service (GaaS)
• AI express lanes
• AIaaS marketplace

Edge AI monetization in telecom networks

Three interlocking pieces make these scenarios bankable revenue:

Operators willing to extend infrastructure into AI workloads. This means offering low-latency, high-bandwidth network capacity (e.g., 100Gbps+) to support the data synchronization demands of distributed training, as purpose-built AI services: edge inference for industrial robotics, real-time decision support for autonomous systems, and similar use cases with strict performance requirements.

Open standards for federation and interoperability. Standards are what make multi-operator AI services viable. Mplify has begun defining NaaS standards for AI workloads, creating the foundation for exposing GPUs, edge compute, and network slices as monetizable services. Mplify LSO’s Sonata interface enables inter-provider federation, Cantata handles B2B marketplace exposure, and TM Forum’s Open Digital Architecture covers the full service lifecycle, from catalog to billing. These aren’t box-checking exercises. They’re what separate a scalable, multi-operator AI platform from a one-off integration project that dies after one customer. You can build a private single-operator edge AI deployment without them. You cannot build a menace to hyperscalers.

Intelligent orchestration. Not just basic rule-based automation, but policy-driven, network-aware orchestration designed to optimize workload placement, synchronization, security isolation, and end-to-end SLA enforcement. This turns promised 99.999% uptime and sub-1ms paths into billable outcomes: GPU-hours with deterministic latency, sovereign GDPR-compliant processing, and federated learning without raw data leaving premises.

The scenarios below fuse cloud-native compute orchestration (Kubernetes + GPU scheduling) with telco-grade network orchestration (slicing, Mplify LSO Cantata/Sonata for service chaining). And the payoff isn’t hypothetical: McKinsey pegs the addressable GPU-as-a-Service market for telcos alone at $35–70 billion annually by 2030, much of it from North America and Asia, while broader AI infrastructure demand could hit $200 billion globally by 2028. Telcos that move here aren’t just defending their connectivity margins, but positioning to own the high-margin intelligence layer on top.

Distributed AI training (federated learning)

Distributed AI training coordinates model development across edge nodes, on-premises clusters, and regional data centers—rather than centralizing everything in a single cloud. The impact is material: workloads that would take months on a single machine can be reduced to weeks, with the flexibility to place them according to cost, performance, or regulatory constraints across on-prem, public cloud, and edge environments.

Three structural limits make this architecture necessary, not just useful.

Hardware memory: the largest AI models can’t fit on a single GPU—distributed training partitions model state across multiple GPUs, turning a physical constraint into a scalable infrastructure problem. Data volume: at a petabyte scale, multiple model replicas must train in parallel, with optimized pipelines preventing storage and I/O from becoming the bottleneck. Data sovereignty: in regulated sectors, federated learning allows training at the edge, sharing only model parameters while raw data stays local. Compliance is baked in from the start.

Gradient synchronization is bandwidth-intensive and highly sensitive to latency, jitter, and packet loss. Training efficiency depends directly on network quality. Operators can provision deterministic, high-capacity paths (100Gbps+) between sites, ensuring the conditions required for distributed training.

Where GPU capacity is insufficient locally, inter-operator federation frameworks such as Mplify LSO Sonata enable real-time sourcing of partner infrastructure. Kubernetes-based orchestration with GPU scheduling then scales workloads across regions as a single logical cluster, even if the underlying infrastructure spans multiple operators.

Key components: deterministic high-bandwidth connectivity, intelligent workload orchestration, inter-operator federation interfaces, secure aggregation and governance controls

Areas of application: Healthcare diagnostics, cross-bank fraud detection, multi-site industrial optimization, connected vehicle model refinement, sovereign AI infrastructure and government clouds, LLM developers, and academic research institutions

↑ Back to top

Edge AI inference

For operational applications, inference latency isn’t a performance metric—it’s a functional requirement. A computer vision system on a factory floor can’t wait for a cloud round-trip, and transmitting raw data at scale (4K video streams from thousands of cameras, for instance) is both prohibitively expensive and a reliable way to saturate your network.

Moving inference to the edge enables closed-loop systems in which decisions and actions occur almost simultaneously at the source. Data volumes across the network shrink; sensitive operational data stays within controlled boundaries.

Operators make this work at scale by combining edge compute, deterministic connectivity, and automated orchestration.

Kubernetes platforms integrated with Nephio manage thousands of distributed edge nodes as a unified environment—models deployed, updated, or rolled back across regions with zero-touch, inference pods scaled and network slices adjusted dynamically as demand shifts.

TM Forum APIs turn this into commercial products. TMF620/633 lets operators bundle AI models, compute, and 5G slices into catalog SKUs; TMF641 lets enterprises instantiate inference services across multiple sites simultaneously, with slices auto-configured for sub-10ms latency, bandwidth guarantees, and tenant isolation. The result is SLA-backed AI execution that holds even during upstream cloud or WAN disruptions.

Key components: MEC compute co-located with RAN, zero-touch model deployment, Kubernetes and Nephio for edge orchestration, TMF620/633 for service cataloging, TMF641 for multi-site service ordering, dynamic slice management for latency and isolation

Areas of application: Industrial computer vision, smart grid monitoring, retail analytics, connected vehicle perception systems, real-time fraud detection, remote asset management for Industry 4.0 and smart city operators

↑ Back to top

GPU roaming

GPU roaming works like mobile roaming: when an operator lacks local GPU capacity, they “lease” power from a partner provider—without latency spikes or session interruption, and in a way that’s unnoticed by end users.

What makes it distinct is session continuity.

When an asset crosses network boundaries, Multi-Domain Orchestration relocates workloads to the nearest edge GPU cluster while preserving performance in real time. Transport paths, QoS parameters, and compute placement are adjusted simultaneously—so autonomous drones, connected vehicles, or robotic fleets keep operating across coverage zones without reinitialization or decision gaps. Sub-10ms targets hold throughout the transition.

Capacity constraints are handled through federation. If Operator A’s GPU pool is saturated, Mplify LSO Sonata automates quote-to-order with Operator B in real time. Interlude coordinates cross-domain path setup and quality of service enforcement; TMF628 gives enterprises a single operational view across operator domains; TMF670/671 handle GPU-hour settlement.

From the enterprise perspective, it behaves as a single distributed GPU fabric. Underneath, it may span multiple operators and jurisdictions. That’s only possible because the network owner controls both mobility and compute placement—a combination hyperscalers, operating within fixed infrastructure footprints, simply cannot replicate.

Key components: multi-domain orchestration for workload mobility, federated GPU capacity pools, MEF LSO Sonata for automated inter-operator sourcing, Interlude for cross-domain control and QoS coordination, TMF628 for unified resource monitoring, intent-based workload placement

Areas of application: Autonomous vehicle fleets, drone and robotic fleet management, cross-border logistics tracking, distributed real-time video analytics, global enterprises and Tier 2/3 operators requiring burst GPU capacity

↑ Back to top

GPU-as-a-Service (GaaS)

Enterprises need elastic access to high-performance accelerators without overcommitting capital or defaulting entirely to hyperscalers. GaaS is where telcos can compete directly—and win on terms hyperscalers can’t easily match.

Users can access fractional GPUs or bare-metal instances on demand, scaling from individual slices to full clusters.

Orchestration layers partition VRAM and compute cycles with strict isolation, binding them directly to Kubernetes pods or virtual machines. Unlike cloud-only offerings, telcos can pair GPU allocation with guaranteed network characteristics—latency ceilings, lossless fabrics, or dedicated east-west bandwidth—turning raw compute into SLA-backed infrastructure.

Commercial automation makes this operationally viable at scale. TMF622 enables instant ordering based on defined hardware profiles (e.g., H100-class GPUs with RoCE-enabled transport). TMF641 activates containers or VMs programmatically across distributed sites, while TMF639 maintains real-time inventory visibility across federated GPU pools. The result is not just rentable silicon, but geographically distributed, network-aware GPU capacity exposed as a telecom-grade service.

Key components: fractional GPU partitioning with isolation, Kubernetes-integrated provisioning, deterministic network attachment, TMF622 ordering automation, TMF641 service activation, TMF639 real-time inventory control

Areas of application: Enterprise LLM fine-tuning, AI-driven R&D workloads, burst compute for media rendering, regional sovereign AI initiatives, AI startups and research institutions requiring high-performance computing without upfront investment 

↑ Back to top

AI express lanes

Large-scale AI training generates intense east-west traffic during gradient synchronization, often involving billions of parameters exchanged repeatedly between nodes. In conventional environments, this traffic competes with other workloads, introducing jitter and packet loss that directly degrade training efficiency.

Telecom operators can engineer dedicated “AI express lanes” by provisioning dynamic network slices optimized specifically for synchronization bursts.

When a distributed training job initiates, orchestration automatically triggers topology optimization—selecting shortest paths, reserving bandwidth and configuring low-loss transport profiles to maintain deterministic performance.

Mplify LSO Legato configures the underlay network elements in real time, while TMF628 provides continuous monitoring for congestion or performance deviation. If thresholds are breached, slices are expanded or rerouted automatically. The network does not merely carry AI traffic; it adapts to it.

This capability is structurally telecom-driven. Only operators control both the physical transport fabric and the service-layer orchestration required to guarantee synchronization performance across metro, regional and cross-border domains.

Key components: dynamic AI-dedicated slices, topology-aware path optimization, LSO Legato underlay automation, TMF628 congestion monitoring, closed-loop slice scaling.

Areas of application: Distributed LLM training, multi-site model parallelism, cross–data center AI research collaboration, AI data center operations, high-frequency financial model updates for trading firms

↑ Back to top

AIaaS marketplace

Enterprises increasingly prefer packaged AI capabilities over raw infrastructure. Telecom operators can aggregate compute, models and storage into standardized, discoverable offerings, creating an AI-as-a-Service marketplace that extends beyond single-provider limits.

Self-service portals expose bundled services such as “LLM fine-tune package” or “edge computer vision suite.”

They combines GPU capacity, pre-trained models, storage, and network slices into unified SKUs. TMF620 standardizes these product definitions, enabling catalog-driven commercialization. LSO Cantata exposes eastbound APIs for BSS integration, while TMF671 automates usage-based billing, including GPU-hour consumption across federated domains.

Federation through Mplify LSO Sonata enables operators to dynamically source partner resources, turning individual footprints into a liquid, multi-operator AI market. Enterprises see a unified service catalog; underneath, capacity and capabilities may span multiple jurisdictions and providers.

Telecom’s structural advantage here is one hyperscalers can’t replicate: the industry already operates global roaming, settlement, and interconnect frameworks. Extending that model to AI services enables scale without bespoke integrations or bilateral complexity.

Key components: standardized AI product catalog (TMF620), federated resource sourcing via Sonata, BSS integration through Cantata, usage-based settlement with TMF671, multi-tenant lifecycle orchestration

Areas of application: Cross-border enterprise AI deployment, sovereign AI marketplaces, public-sector AI programs, SME access to enterprise-grade AI infrastructure, turnkey AI solutions for small and medium enterprises (SMEs)

↑ Back to top

Industrializing AI: telecom’s moment

This is a once-in-a-generation moment for telecom operators. AI is becoming infrastructure, and infrastructure is the telecom domain. But without orchestration, it does not scale, federate, or monetize.

GaaS, AI express lanes, and AIaaS marketplaces only work if a unifying control plane translates business intent into automated execution: allocating GPUs, configuring deterministic network fabrics, enforcing latency SLAs, and assuring real-time performance. Without orchestration, every deployment becomes a bespoke integration exercise with manual provisioning, static capacity planning, and fragmented monitoring. That model collapses under AI’s dynamic demand patterns.

Standards are equally non-negotiable.

Mplify LSO and TM Forum Open APIs eliminate bilateral integrations and vendor lock-in, allowing operators to interconnect compute, network, and commercial systems globally. They create portability across domains and predictability across vendors. Without standards, federation stalls. With standards, markets form.

These aren’t theoretical frameworks waiting for validation. They are built collaboratively by leading operators and solution providers, validated through proofs of concept and industry Catalyst programs, and hardened in multi-party environments before commercial launch. That cooperative validation cycle turns architecture intoa  deployable business.

Telecom operators uniquely control mobility, transport determinism, regional presence, and interconnect frameworks. When unified through standards-based orchestration, those assets become a distributed AI platform.

The infrastructure is already yours. The standards exist. The demand is real and growing fast.

If you’re mapping your AI infrastructure strategy—or looking for the right partners to accelerate it—we’d like to be part of that conversation.