Akamai is introducing two new services to the Akamai Cloud: Akamai Inference for managed model serving and Akamai Agent Sandbox for the isolated execution of code written by AI agents themselves.
With these two services, Akamai aims to offer businesses a single platform to build, run, and secure AI applications and agents. According to the company, most teams currently piece these components together from different providers and then have to secure the connections between them themselves. Akamai Cloud spans from GPU and CPU clusters in core regions to more than 4,350 edge locations.
“We have been implementing an AI infrastructure strategy based on the notion that AI inference will represent the majority of enterprise focus and investment. Our customers proved that thesis and pulled us further,” says Adam Karon, COO and GM of Akamai Cloud.
Open-weight models via OpenAI-compatible APIs
Akamai Inference provides managed access to open-weight and custom models via OpenAI-compatible APIs. As a result, existing applications do not need to be rewritten. Customers can choose between serverless inference on a pay-per-use basis or dedicated deployments. At launch, the catalog includes twelve open-weight models, ranging from 8 billion to 2.8 trillion parameters. These include Nvidia Nemotron 3 Ultra, OpenAI gpt-oss-120b, and Google Gemma 4, supplemented by embedding models for RAG. The service handles orchestration, model execution, and authentication.
WebAssembly as a boundary for agent code
Agents running on Akamai Agent Sandbox operate on Akamai Functions, while the code they write is executed in its own isolated environment, separate from the agent’s runtime. The sandbox is an open-source component of the Spin framework. That framework comes from Fermyon, which Akamai previously acquired to strengthen serverless edge computing. Through WebAssembly isolation, the sandbox sets limits on compute, memory, outbound connections, and system privileges.
Akamai also allows developers to opt for managed Kubernetes and dedicated CPU or GPU infrastructure. In addition, the company plans to add features that control what agents can access and what can access them.
Blackwell Ultra now, Vera Rubin in 2027
On the hardware front, this is a major step forward. Akamai started with GPU services based on Nvidia RTX 4000 Ada cards. Now, the fleet consists of RTX PRO 6000 Blackwell Server Edition GPUs for low-latency inference and Blackwell Ultra clusters. Each node with eight GPUs has 288 GB of HBM3e per GPU, for a total of 2.3 TB. These clusters run without rate limits or shared queues. Akamai is also preparing its data centers for the Nvidia Vera Rubin platform in 2027.
Akamai Inference and Agent Sandbox (via the Spin SDK) are now available in beta. Anyone interested in using the NVIDIA B300 clusters can contact Akamai.