Akamai launches AI inference and agent sandbox services across its cloud network

Akamai is adding a managed AI inference service and an isolated environment for running agents to its AI Cloud platform. The pitch is practical: deploy models through OpenAI-compatible APIs, then give agents tighter limits on what they can access and do.

Watch Desk analysis

What happened

Akamai Inference will offer 12 open-weight models at launch, from 8 billion to 2.8 trillion parameters, along with specialised embedding models. Akamai says teams can deploy AI applications and agents at the edge without rewriting code. The service draws on compute and accelerators available across more than 4,350 edge locations, according to SiliconANGLE’s report.

Akamai Agent Sandbox runs agent code in isolated WebAssembly environments managed by the open-source Spin framework. Organisations can set controls over compute, memory, system APIs and outbound network access. The report also says Akamai is expanding its GPU fleet with NVIDIA Blackwell Ultra clusters; eight-GPU nodes provide up to 2.3 terabytes of HBM3e memory. The company plans to offer NVIDIA’s Vera Rubin platform in the new year.

Our top picks

• An OpenAI-compatible inference API
  Teams can access a choice of models without rebuilding existing integrations around a new interface.
• A dozen models at launch
  The selection ranges from 8-billion-parameter models to a 2.8-trillion-parameter option, plus embedding models for retrieval systems.
• A sandbox with adjustable boundaries
  Operators can limit agent resources, API access and outbound connections rather than simply hoping an agent stays in its lane.
• Inference near users and data
  Akamai says its edge locations let customers run workloads closer to where they are needed.
• More GPU capacity
  Blackwell Ultra clusters are available, with Vera Rubin planned for the new year.

Why it matters

The announcement brings model serving and agent execution into the same infrastructure offer. That could appeal to organisations trying to keep inference close to users or proprietary data while applying tighter controls to autonomous software. The sandbox’s specific limits are more useful than a general promise of “secure agents”, though the announcement alone does not show how those controls perform in customer deployments.

Our read

Akamai is making a credible infrastructure play: not a new frontier model, but a place to serve models and run agents with explicit operational boundaries. The edge footprint and API compatibility are the hooks; the real test is whether customers find the controls straightforward enough to use and the service compelling enough to move workloads.

What to watch

• Which models and embedding options are available beyond the initial 12.
• How customers configure and monitor sandbox restrictions in practice.
• When Vera Rubin capacity arrives, and what hardware and pricing options Akamai offers.

Discussion spark: For enterprise agents, is a tightly controlled sandbox a meaningful safety boundary, or should organisations keep autonomous agents away from sensitive systems altogether?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.