Akamai Technologies introduced Akamai Inference and Akamai Agent Sandbox on 27 October 2026, adding model-serving and isolated code execution to its AI Cloud platform, SiliconANGLE reported. The cloud services company is offering enterprises a way to run AI models and the code written by autonomous agents on infrastructure extending from its data centres to its network edge. Both products are available in beta.
Key points
- Akamai Inference launches with 12 open-weight models accessible through OpenAI-compatible APIs.
- Agent Sandbox runs agent-generated code in isolated WebAssembly environments, with controls over resources and outbound connections.
- Both services are in beta, and Akamai has added Nvidia Blackwell Ultra clusters to its GPU fleet.
Akamai Inference offers 12 open-weight models
Akamai Inference is a managed service for serving open-weight and custom models through APIs compatible with OpenAI’s. Its launch catalogue contains 12 open-weight models, ranging from 8 billion to 2.8 trillion parameters, including Nvidia Nemotron 3 Ultra, OpenAI gpt-oss-120b and Google Gemma 4. Customers can choose serverless, pay-per-use inference or dedicated deployments, StreetInsider reported. The beta includes serverless deployment, while access to dedicated Nvidia B300 GPU clusters requires contacting Akamai sales.
Compatibility concerns the way an application sends requests to a model. Akamai says developers could use its inference service without rewriting code that already calls an OpenAI-compatible API. The choice of model remains separate from that interface: an application can send a request in a familiar format while the service runs one of the models Akamai offers. That gives builders a route to its managed infrastructure without making the models interchangeable in what they produce.
Akamai also offers specialised embedding models, which turn material into representations used to retrieve relevant information. The company intends these for retrieval-augmented generation, in which a model receives selected enterprise data alongside a request rather than relying only on what it learned during training. Akamai says they can connect its open-weight models with its managed databases. Its wider cloud platform combines compute, storage, networking and security across a network of more than 4,350 edge locations, according to SiliconANGLE.
The expansion follows another reported cloud commitment: AI Affairs reported Anthropic’s $11.6 billion Akamai cloud deal and stake warrant. Adam Karon, Akamai Cloud’s chief operating officer and general manager, said customers want to build, run and protect AI on one platform. The new service supplies the model-serving part of that offering; the accompanying sandbox addresses what happens when an agent writes code to carry out a task.
Spin isolates code in Akamai Agent Sandbox
Agent Sandbox runs AI agents on Akamai Functions and places their generated code in separate WebAssembly environments managed by Spin, an open-source framework. WebAssembly provides a contained place for a program to execute — closer to a fenced workbench than unrestricted access to the surrounding system. Customers can set limits on computing power, memory, system capabilities and outbound network connections. Those controls matter when an agent can write a program and then run it, rather than merely produce text for someone else to inspect.
Writing and running code for a work task could therefore take place inside a bounded environment. The code could use the resources and connections permitted for that run, while the task could still draw on a model reached through Akamai’s compatible API. How much access that arrangement would permit depends on the limits set for the sandbox.
The product is an open-source component of the Spin framework, StreetInsider reported. Its controls concern the execution of code and the resources that code can reach. In a separate example of the risks facing agents, Zscaler found web pages carrying hidden instructions directed at AI agents in July 2026; four of the 26 models it tested complied with a purported licence-payment instruction, Decrypt reported. That attack sought to influence an agent’s instructions, whereas Akamai’s sandbox sets boundaries around code the agent executes.
Jon Alexander, Akamai’s senior vice president of cloud technology, said teams running agents need a boundary around code those agents write. The sandbox gives customers controls over outbound connections, alongside limits on memory and computing power. Those settings make the execution environment a distinct part of an agent deployment: a model may decide to produce code, while the sandbox governs the conditions in which that code runs.
Blackwell Ultra expands Akamai’s GPU fleet
Akamai has also added Nvidia Blackwell Ultra clusters to its AI Cloud GPU fleet. Each eight-GPU node provides up to 2.3 terabytes of HBM3e memory, which Akamai says can support long-context reasoning workloads. The hardware addition sits alongside the managed inference service and the sandbox rather than replacing either: model execution needs compute capacity, while agent-generated code needs its own execution boundary.
Akamai is preparing its data centres to support Nvidia’s Vera Rubin platform in 2027, StreetInsider reported.
