Frontier AI Is Coming to the Edge: Telcos Need to Own the Orchestration Layer

Anthropic has committed US$11.6 billion over 7 years to Akamai to support its Central Processing Unit (CPUs) workloads through Akamai's distributed AI infrastructure and software. The contract builds on Akamai’s distributed AI inference platform across its global edge, regional, and core infrastructure, using NVIDIA's Artificial Intelligence (AI) Grid design architecture to route workloads according to latency, cost, performance, and other requirements. Its infrastructure spans more than 4,400 edge locations alongside regional and core data centers, with CPUs and thousands of NVIDIA RTX PRO 6000 Blackwell Graphic Processing Units (GPUs) being deployed across the platform.

Akamai's approach highlights the role of orchestration in managing distributed AI infrastructure. Its platform can determine where an inference workload should run based on factors such as workload requirements, available compute, latency, and infrastructure location, allowing workloads to be placed across a geographically distributed footprint. The exact role that Anthropic's workloads will play across Akamai's network is not yet clear, but the agreement is an important signal that frontier AI companies are engaging with infrastructure providers beyond traditional centralized cloud environments. Akamai's model recognizes that centralized AI factories will remain best suited for training and frontier workloads, and distributed inference creates a continuous workload-placement problem that can be addressed through intelligent orchestration. In parallel, telco Application Programming Interfaces (APIs) are evolving. CAMARA and Open Gateway APIs such as Simple Edge Discovery already let applications find a suitable edge zone by network proximity, and Ericsson is now discussing AI-native, agent-friendly APIs as a basis for AI platform monetization. For telcos, Akamai provides an early commercial example of how orchestration can turn geographically distributed AI infrastructure into a monetized service beyond connectivity.

As inference is increasingly distributed across devices, Radio Access Network (RAN), telco edge, regional data centers, hyperscalers, and centralized AI factories, there will rarely be one optimal location for every workload. The appropriate location can change based on latency, accelerator availability, model availability, network congestion, power constraints, data sovereignty, cost, and Service-Level Agreements (SLAs). This makes workload placement an increasingly important problem because while compute capacity determines what AI inference can be supported, the orchestration determines where, when, and how it is run. Additionally, telcos have network information that can influence these decisions, including device location, mobility, radio conditions, congestion, backhaul performance, and network topology. For latency-sensitive applications, this information could help determine whether an inference workload should run at a local edge site, a regional data center, or a more centralized facility. This creates an opportunity to evolve Edge Discovery from identifying the nearest suitable site toward determining the best location for a specific AI workload.

This creates an opportunity for Tier One telcos and telco partnerships, such as the European Edge Continuum, to monetize their distributed sites, power, connectivity, and network intelligence as infrastructure for AI workloads. Edge APIs could provide the interface through which AI platforms discover, access, and dynamically orchestrate these distributed resources as part of an AI Grid. Existing APIs can help applications discover suitable edge locations, but an Edge API layer could expose compute availability, model availability, latency, Quality of Service (QoS), data-residency constraints, and other factors needed to select a location. The role of the API would therefore shift from "access the edge" to "place and optimize the AI workload." This would allow an application to request an outcome, such as keeping an AI agent below a defined latency threshold. Operators could also combine their own sites with infrastructure from partners, hyperscalers, and third-party GPU providers, giving applications access to a broader pool of compute. The opportunity is to make this infrastructure appear as a single resource to the customer, with the operator managing the placement decision behind the API.

The strategic objective for telcos should be to own the intelligence that coordinates the AI Grid, even when the operator does not own every asset within it. ABI Research sees an opportunity for Tier One telcos and telco partnerships to become infrastructure partners to frontier labs and AI-native companies, as those firms distribute workloads across more locations, jurisdictions, and providers. Telcos may face a disadvantage in global reach and open access; most operators have geographically limited footprints and their infrastructure is often tied to their subscriber bases. Global Internet backbone providers and Content Delivery Networks (CDNs) such as Akamai can offer broader, more openly accessible infrastructure across markets. Operators should build or partner for an orchestration platform that combines the network telemetry with compute availability, model availability, workload requirements, latency, QoS, cost, and data-residency policies to determine where inference should run. The first step is to give the platform visibility across available compute and network conditions so it can identify suitable locations for specific workloads. Operators should also extend existing Edge APIs beyond discovery toward workload placement, enabling applications and AI agents to request an execution environment based on specific performance and policy requirements. This gives operators a foundation for selling defined AI outcomes, such as latency, availability, or data residency, across their infrastructure.

The Akamai-Anthropic agreement signals that frontier labs are seriously looking into distributing their workloads, creating an opening for telcos to pursue similar partnerships. Telcos should approach labs with their power-ready sites, in-country capacity, and network-level performance. Because the Akamai contract is for CPU capacity, operators should not assume a GPU-heavy edge is the only entry point, and existing telco cloud and central office capacity may be relevant to agentic workloads. This could give operators an entry point into the AI infrastructure market without deploying large amounts of dedicated GPU capacity.

Over the longer term, operators should use partnerships to expand the infrastructure available through the AI Grid. Partnerships with hyperscalers, GPU providers, neoclouds, and other edge infrastructure providers could allow operators to expand available compute capacity without making the capital investment required to deploy GPUs at every location. Commercial offers should therefore focus on outcomes such as latency, availability, data sovereignty, and cost optimization rather than charging customers simply for API calls or access to individual sites. Agentic AI, Physical AI, and other latency-sensitive workloads should be early targets because their dynamic requirements make intelligent placement particularly valuable. Operators will need to prove that their network intelligence provides a meaningful advantage over orchestration platforms from cloud providers, CDNs, and AI infrastructure vendors. The operators that capture the most AI Grid value will be those that make distributed compute easy to access and program, and optimize them through an orchestration layer.