Akamai moves AI inference to the edge for autonomous agents
Akamai launched its AI Inference and Agent Sandbox in beta, distributing model execution to the edge. It signals a shift away from centralized clouds.
On October 27, 2026, Akamai Technologies (opens in a new tab) announced the beta launch of two additions to its cloud platform: Akamai Inference and Akamai Agent Sandbox. The release shifts artificial intelligence model execution and isolated runtimes away from centralized data centers and pushes them across a global edge network. Software development increasingly relies on autonomous agents to execute complex tasks. Relying on a few massive server farms for these tasks creates latency and introduces security bottlenecks. Moving compute closer to end users and localized data sources allows technology leads to deliver faster workflows. This launch illustrates how edge infrastructure will reshape software architecture over the next five years.
Distributing model execution
The traditional approach to artificial intelligence requires sending user requests back to a central server. This round trip adds milliseconds of delay that compound during complex tasks. Akamai changes this geography by placing compute across more than 4,350 edge locations. The new inference service gives developers managed access to open models and custom models through application programming interfaces that are compatible with OpenAI. This compatibility means companies can move their existing applications to the edge without rewriting their code.
At launch, the beta catalog includes 12 open models. These range from 8 billion to 2.8 trillion parameters. The list features NVIDIA Nemotron 3 Ultra, OpenAI gpt-oss-120b, and Google Gemma 4. It also includes embedding models for retrieval tasks. Customers can choose between serverless inference or dedicated deployments. By running requests on a global network instead of a central hub, the architecture offloads compute overhead from centralized cloud data centers. The result is reduced latency and highly available performance for end users.
Securing autonomous code
Artificial intelligence agents are moving beyond simple text generation. They now take actions on behalf of users and write their own code. This autonomy introduces significant security risks if the generated code runs unchecked. The new sandbox provides a secure environment for developers to run these agents.
The agents operate on the serverless functions platform of the company. When an agent writes code, that code executes in a separate, isolated environment. This keeps the generated code completely sequestered from the runtime of the agent itself. The sandbox is built as an open source component of the Spin framework and relies on WebAssembly for isolation. This technology enforces strict boundaries on compute usage, memory allocation, outbound network connections, and system capabilities. Developers can deploy agents knowing exactly where they run and what data they can reach. Akamai also plans to add capabilities to govern what agents can access and what can access them.
Hardware for the next generation
Supporting this distributed architecture requires substantial hardware investments. Akamai expanded its dedicated graphics processing unit fleet to include NVIDIA Blackwell Ultra clusters. A single node contains eight GPUs and provides 2.3 terabytes of total memory. This provides enough capacity to hold frontier models.
The architecture specifically targets workloads that require massive memory. These include test-time scaling, reasoning tasks, and long-context agentic work. For low-latency inference, the fleet also spans NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. These dedicated clusters give enterprises control over model versions and deprecation timelines. They run without rate limits or shared queueing. "Enterprises should not have to choose between speed and control," said Jon Alexander, Senior Vice President of Cloud Technology at Akamai.
What it means for your business
Mid-sized companies building software products face a difficult balance between speed and control. Centralized cloud providers offer managed services but often at the cost of high latency for global users. Building a private infrastructure requires massive capital expenditure. The Akamai release provides a middle ground for technology teams.
By utilizing serverless edge inference, a technology lead can deploy artificial intelligence features globally without managing the underlying hardware stack. The API compatibility ensures that integration requires minimal engineering effort. For finance leads, the serverless model shifts the cost from a fixed capital expense to a variable operational cost tied directly to usage. Moving compute closer to localized data sources also helps companies process user data at a local edge node rather than a distant central server. This reduces the risk of data exposure across long network transits.
The sandbox solves a specific operational headache. When a business deploys an agent to automate customer support or internal data processing, that agent needs strict boundaries. WebAssembly isolation ensures that a malfunctioning or compromised agent cannot access sensitive corporate databases or consume unlimited compute resources. This allows smaller teams to adopt autonomous workflows safely.
What to watch
Both the inference service and the sandbox are currently in beta. The serverless deployment options are available in the beta phase, while the NVIDIA B300 GPU clusters require contacting sales for availability. Technology leads should monitor the transition from beta to general availability to assess pricing and service level agreements.
The hardware evolution will also dictate the pace of edge computing. Akamai is preparing its data centers to support the NVIDIA Vera Rubin platform in 2027. This signals a continuous upgrade cycle that businesses can use without buying the hardware themselves. Monitoring these hardware deployments will help technology leads plan their own software roadmaps.
Finally, observe how agent isolation standards converge. The industry is currently experimenting with several models. These include microVMs with snapshot capabilities, policy-enforced containers, and WebAssembly-isolated functions. Akamai chose WebAssembly for its sandbox, but the broader market consensus over the next few years will determine the standard for secure agent deployment.
The next five years of software development will be defined by autonomous agents operating close to the user. Distributing model inference and sandboxed runtimes to the network edge removes the latency and security bottlenecks of centralized clouds. As infrastructure adapts to these demands, businesses gain the tools to build faster and safer applications.