Akamai launches AI inference and agent sandbox tools for enterprises

A unified AI stack for business

Akamai has introduced two new foundational products to its cloud platform: Akamai Inference and Akamai Agent Sandbox. These tools are designed to help enterprises move beyond just using models to actually building and operating AI applications at scale. The company aims to provide a trusted global infrastructure that handles the serving, running, and securing of AI workloads.

The move comes as companies look for ways to integrate AI into their existing workflows without managing every piece of the infrastructure separately. Currently, many teams have to assemble different components from various providers to handle data, routing, and governance. Akamai is attempting to solve this by offering a connected environment on one platform.

By combining managed services with its existing network fabric, Akamai wants to give customers a predictable experience. The company leverages both CPUs and GPUs alongside cloud storage and networking to support these new AI capabilities. This strategy is backed by multiyear contracts for Akamai Cloud worth more than $14 billion.

Managed access via Akamai Inference

Akamai Inference provides a way for enterprises to access open-weight and custom models through OpenAI-compatible APIs. This design allows existing applications to be integrated with new models without requiring significant code rewrites. Customers can choose between serverless, pay-per-use options or dedicated deployments depending on their specific needs.

The initial catalog features 12 different open-weight models, ranging in size from 8 billion to 2.8 trillion parameters. These include notable models like Google Gemma 4, NVIDIA Nemotron 3 Ultra, and OpenAI gpt-oss-120b. The service also includes embedding models specifically designed for retrieval-augmented generation, which helps models ground their answers in specific data.

To ensure high availability and low latency, Akamai runs these requests across its global network. The platform also connects to private and public clouds, allowing models to draw on internal enterprise data. Akamai Inference handles the heavy lifting of orchestration, model execution, and authentication for these requests.

Securing AI agents with a sandbox

As AI agents begin to perform real-world tasks, they are increasingly writing and running their own code. This creates a significant security risk if that code is allowed to reach parts of a network it should not access. Akamai Agent Sandbox provides a secure environment where developers can run these agents and the code they generate in an isolated space.

The sandbox is delivered as an open-source component of the Spin framework and utilizes WebAssembly isolation. This technology sets strict limits on the memory, compute power, and outbound connections available to the agent. By sequestering the code from the agent's runtime, developers can deploy agents with more confidence about their reach.

Akamai is offering several architecture choices for this isolation, including microVMs and policy-enforced containers. The platform supports both NVIDIA-accelerated and dedicated CPU infrastructure for these workloads. The company also plans to add specific capabilities to govern exactly what agents can reach and what can reach them.

Expanding the GPU hardware fleet

To support these new services, Akamai is significantly expanding its dedicated GPU hardware. The company recently added NVIDIA Blackwell Ultra GPU clusters to its fleet for enterprises that need to build and tune models. These clusters are designed for high-demand tasks like test-time scaling and long-context agentic work.

Each eight-GPU Blackwell Ultra node provides 2.3 TB of total memory, which is enough to hold frontier-class models. For lower-latency needs, Akamai also offers NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. These options allow customers to choose between managed platforms for quick features or dedicated hardware for more control.

Looking ahead, Akamai is already preparing its data centers to support the NVIDIA Vera Rubin platform. This next-generation hardware is expected to be supported by the company in 2027. This long-term hardware roadmap ensures that the infrastructure can keep up with the increasing memory demands of advanced AI models.