Akamai adds managed AI model access and isolated workspaces for agents

Akamai is adding managed model serving and isolated agent execution to its AI Cloud, putting the systems that generate AI responses alongside the environments where agents carry out tasks. Announced October 27, Akamai Inference and Akamai Agent Sandbox pair access to open-weight and custom models with controls over code execution, resources and network access.

The coordinated expansion adds software services to a platform that already combines computing, storage, networking and security. Akamai’s stated strategy is to place inference—the work of producing a model response—closer to users and the data an application needs, with the aim of reducing delays.

A model-serving layer with familiar interfaces

Akamai Inference provides managed access to open-weight and custom models through OpenAI-compatible application programming interfaces, or APIs—the connections software uses to send requests and receive results. Akamai says that compatibility lets teams deploy AI applications and agents at the edge without rewriting their code.

At launch, the service offers 12 open-weight models, ranging from 8 billion to 2.8 trillion parameters. The named options span three providers:

• Nvidia’s Nemotron 3 Ultra.
• OpenAI’s gpt-oss-120b.
• Google’s Gemma 4.

The catalog also includes specialized embedding models for retrieval-augmented generation, an approach that brings relevant information into a model’s response process. Akamai intends these models to connect the open-weight offerings with its managed databases, grounding answers in enterprise data rather than treating model access as a standalone service.

They want to build, run and protect AI on the same platform.

Adam Karon, Akamai chief operating officer and general manager of Akamai Cloud

A separate boundary for agents doing work

Agent Sandbox addresses a different part of the workload: executing code for agents acting on a user’s behalf. It supplies an isolated execution space rather than the model-serving interface. Agents run on Akamai Functions, with their code contained in WebAssembly environments managed by Spin, an open-source framework.

Within those environments, enterprises can control how much computing power and memory an agent receives. They can also restrict access to system APIs and outbound network connections. The controls cover both the resources an agent consumes and the systems it can reach, rather than relying only on instructions about how it should behave.

Akamai frames the sandbox as a response to agents taking on more software-coding responsibilities. Its pitch combines that execution boundary with governance over proprietary and third-party data. Those are related but distinct requirements: connecting an AI application to business information, and limiting what its agent can do while working.

New GPU access, with Rubin still ahead

The same AI Cloud announcement also broadens the hardware offering. Akamai said it now provides access to Nvidia Blackwell Ultra GPU clusters, adding another infrastructure option beneath the new model-serving and agent-execution services.

Nvidia’s Vera Rubin platform remains a future addition. Akamai said it plans to make that hardware available in the new year, placing it on a different timetable from the Blackwell access announced now. The roadmap extends the computing choices behind AI Cloud, but Rubin availability is still a plan, not a delivered part of this update.