Akamai Technologies (Nasdaq: AKAM) has introduced Akamai Inference and Akamai Agent Sandbox, two services intended to give enterprises managed model serving and isolated execution for AI agents on Akamai Cloud. The company said the combination can reduce the infrastructure work required to deploy AI applications while placing a boundary around code that agents write or run.
Akamai Inference provides managed access to open-weight and custom models through OpenAI-compatible APIs. Customers can use a serverless, pay-per-use option or dedicated deployments, while Akamai handles model execution, orchestration and authentication, according to the company. At launch, the catalog includes 12 open-weight models, along with embedding models for retrieval-augmented generation.
That API compatibility is a practical feature for development teams with existing applications built around OpenAI-style interfaces. Akamai said it can allow those applications to move without rewrites, although a production migration would still require teams to test model behavior, throughput, observability and the handling of prompts and enterprise data.
The service sits on Akamai’s network and is designed to connect with cloud platforms, managed databases and object storage. The company said customers can use its core GPU and CPU clusters as well as edge locations, depending on the workload. It also said the service handles high availability and predictable performance, claims that prospective users would need to evaluate against their own latency, region and resiliency requirements.
Akamai Agent Sandbox addresses a separate operational risk: an AI agent that writes or executes code. The service runs agents on Akamai Functions and places the code they produce in a separate isolated environment. Akamai said the open-source Spin framework component uses WebAssembly isolation to control compute, memory, outbound connections and system capabilities.
The distinction between an agent’s runtime and the environment for its generated code is significant for security and platform teams. Agentic applications often need tools, data or network access to complete work, but broad access can turn an incorrect instruction or compromised agent into a path to production systems. An isolation layer can constrain what code reaches; it does not replace identity policy, code review, logging or carefully scoped credentials.
Akamai said developers can choose among several architectures, including managed Kubernetes, Akamai Functions and dedicated infrastructure, based on their workload. It plans to add controls that govern what agents can reach and what can reach them. The company did not specify a general-availability date for those future controls.
The company also expanded its dedicated GPU fleet with Blackwell Ultra clusters and said it is preparing its data centers for the next NVIDIA (Nasdaq: NVDA) platform. Those resources are aimed at customers that want to build, tune or self-host models on infrastructure they control. Akamai said each eight-GPU Blackwell Ultra node carries 2.3 TB of aggregate HBM3e memory, but it did not release comparative benchmark data for the new services.
Akamai Inference and Agent Sandbox are available in beta, the company said. For infrastructure teams, the new services reflect a move toward a more integrated AI operating stack: model access, compute, storage, networking and an execution boundary for agents from one provider. The operational benefit will depend on whether the beta services provide the required security controls, model options and regional coverage for a team’s particular production workload.