Akamai Technologies has added two services to its AI Cloud platform, Akamai Inference and Akamai Agent Sandbox. Together, they let enterprises serve AI models and run autonomous agents on any infrastructure, from Akamai’s cloud data centers to the network edge.
The Sandbox handles the security side, and the timing explains why. Akamai said autonomous agents now take on more software coding work, so organizations need harder boundaries around them. The service gives each agent an isolated execution space. Agents run on Akamai Functions, and their code stays inside WebAssembly environments that the open-source Spin framework manages. Enterprises can then set limits on an agent’s compute, memory, system APIs and outbound network access, which Akamai says keeps agents inside their boundaries.
Akamai Inference covers the serving side. The managed service offers dozens of open-weight and custom models through OpenAI-compatible application programming interfaces, and Akamai said teams can deploy new AI applications and agents at the edge without rewriting code. At launch, enterprises can choose from 12 open-weight models that range from 8 billion to 2.8 trillion parameters, including Nvidia’s Nemotron 3 Ultra, OpenAI’s gpt-oss-120b and Google’s Gemma 4. Embedding models for retrieval-augmented generation link those models to Akamai’s managed database services, so answers draw on enterprise data.
Both services sit on a cloud platform that Akamai built around running inference closer to users and their data, which cuts latency. The AI Cloud covers compute, cloud storage, networking and security, and customers can reach GPUs and other AI accelerators from more than 4,350 edge locations.
Adam Karon, chief operating officer and general manager of Akamai Cloud, said the company expects inference to become the main focus for most enterprises. “They want to build, run and protect AI on the same platform,” he said.
Akamai also widened its GPU fleet with Nvidia Blackwell Ultra clusters, including RTX PRO 6000 Blackwell Server Edition GPUs. Each eight-GPU node offers up to 2.3 terabytes of total HBM3e memory, which Akamai said supports long-context reasoning. The company plans to add Nvidia’s next-generation Vera Rubin platform in the new year.
For enterprises that hand more coding work to agents, the practical question is how tightly to restrict them. Akamai’s answer is to enforce limits on compute, memory and network access inside the sandbox.
