---
type: "post"
id: "akamai-launches-akamai-inference-with-12-open-weight-be7d9b"
slug: "akamai-launches-akamai-inference-with-12-open-weight-be7d9b"
title: "Akamai launches Akamai Inference with 12 open-weight models and a WebAssembly Agent Sandbox"
description: "Akamai has expanded its AI Cloud with Akamai Inference, an OpenAI-compatible managed service for open-weight and custom models, and Akamai Agent Sandbox, an isolated WebAssembly environment for agent execution."
retrieval_nugget: "Akamai has expanded its AI Cloud with Akamai Inference, an OpenAI-compatible managed service for open-weight and custom models, and Akamai Agent Sandbox, an isolated WebAssembly environment for agent execution. For AI builders, the launch joins model serving, data grounding, edge placement, and runtime controls in one infrastructure proposition."
published_at: "2026-10-27"
updated_at: "2026-10-27"
record_date: "2026-10-27"
date_kind: "published_at"
topics: ["google"]
entities: ["Akamai Technologies","Akamai AI Cloud","Akamai Inference","Akamai Agent Sandbox","Spin","Nvidia Nemotron 3 Ultra","gpt-oss-120b","Google Gemma 4"]
related_patterns: ["agent-security-moves-to-runtime-boundaries","harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Akamai is launching its own inference service and agent sandbox, adding 12 open-weight models and WebAssembly-based agent execution; the priors cover other inference services or comparisons, not this Akamai launch."
prior_refs: ["memory:25a48078-8e49-83a7-a339-56f8a5ce56da","memory:2f75184f-b490-887b-a48f-c973e8e6bc33","memory:367450fa-24a1-883c-adcf-31bad5950ad9"]
source_url: "https://siliconangle.com/2026/10/27/akamai-expands-its-ai-cloud-platform-with-dedicated-inference-and-agent-execution-environments"
source_urls: ["https://siliconangle.com/2026/10/27/akamai-expands-its-ai-cloud-platform-with-dedicated-inference-and-agent-execution-environments"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:akamai-launches-akamai-inference-with-12-open-weight-be7d9b"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/akamai-launches-akamai-inference-with-12-open-weight-be7d9b/","markdown":"https://newruntime.com/posts/akamai-launches-akamai-inference-with-12-open-weight-be7d9b.md","json":"https://newruntime.com/posts/akamai-launches-akamai-inference-with-12-open-weight-be7d9b.json"}
---

# Akamai launches Akamai Inference with 12 open-weight models and a WebAssembly Agent Sandbox

## Retrieval answer

Akamai has expanded its AI Cloud with Akamai Inference, an OpenAI-compatible managed service for open-weight and custom models, and Akamai Agent Sandbox, an isolated WebAssembly environment for agent execution. For AI builders, the launch joins model serving, data grounding, edge placement, and runtime controls in one infrastructure proposition.

## Summary

Akamai has expanded its AI Cloud with Akamai Inference, an OpenAI-compatible managed service for open-weight and custom models, and Akamai Agent Sandbox, an isolated WebAssembly environment for agent execution. For AI builders, the launch joins model serving, data grounding, edge placement, and runtime controls in one infrastructure proposition.

## What happened

Akamai Technologies announced two new foundational components for its AI Cloud platform on October 27, 2026: Akamai Inference and Akamai Agent Sandbox. The company positions them as managed environments through which enterprise customers can serve AI workloads and execute agents across Akamai cloud data centers and its network edge. Its stated aim is to give customers a way to run AI close to users and data, while combining compute, storage, networking, and security on the same platform.

Akamai Inference is the model-serving component. It exposes dozens of open-weight and custom AI models through OpenAI-compatible application programming interfaces. Akamai says this compatibility lets teams deploy AI applications and agents at the edge without rewriting their code for a new API surface. At launch, the service offers 12 open-weight models ranging from 8 billion to 2.8 trillion parameters. Named examples include Nvidia Nemotron 3 Ultra, OpenAI gpt-oss-120b, and Google Gemma 4.

The launch also includes specialized embedding models intended for retrieval-augmented generation pipelines. Akamai says these models can connect the open-weight models to its managed database services so responses can be grounded in enterprise data. The source does not identify the complete model list, describe the databases in detail, or specify how customers select model versions and capacity.

Akamai Agent Sandbox addresses a different layer: agent execution. It is an isolated space for agents acting autonomously on behalf of human users. Agents run on Akamai Functions, while their code is contained in isolated WebAssembly environments managed by Spin, an open-source framework. Akamai says customers will be able to set granular controls over compute, memory, system APIs, and outbound network access. Those controls are intended to prevent an agent from exceeding its assigned operational boundary.

The company also expanded GPU capacity in Akamai AI Cloud with Nvidia RTX PRO 6000 Blackwell Server Edition GPUs in Blackwell Ultra clusters. Akamai says each eight-GPU node provides up to 2.3 terabytes of total HBM3e memory for long-context reasoning. It further says Nvidia's Vera Rubin GPU platform is planned for availability in the new year. The announcement provides neither pricing nor availability details for the new services, model endpoints, sandbox, or planned GPU capacity.

## What the archive adds

This is a new infrastructure event, but it intersects with our recent note on [Vals AI's Inference Provider Index](https://newruntime.com/posts/vals-ai-launched-the-inference-provider-index-comparing-60fbd1/), which found material cost differences among providers serving the same open-weight models. Akamai adds another OpenAI-compatible serving option rather than a new model family.

Our earlier Telegram coverage also described Cloudflare's beta Dynamic Worker Loader, which used V8 isolates for AI-generated code. Akamai's new sandbox is related in problem framing—isolated agent execution—but uses WebAssembly environments managed by Spin.

## Trend signal

The announcement supports [Agent security moves to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/). Akamai's stated controls over memory, system APIs, and outbound network access make the execution environment, rather than an instruction to the model, the proposed place to constrain an autonomous agent. It also modestly supports [Harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/): an OpenAI-compatible interface is designed to keep application integration stable while the served open-weight model can vary.

## Connections to other events

The Vals AI index is relevant because it framed inference-provider selection as a measurable operational choice: the same open-weight model may differ in cost and speed by host. Akamai's launch introduces a host whose differentiation is not only model access, but also edge placement, managed data services, and an execution sandbox. The two accounts do not establish how Akamai will compare on price, latency, or reliability.

## Why it matters

Agent deployments need two decisions that are often treated separately: where the model is served and where the agent is permitted to act. Akamai is packaging both. A team that already speaks an OpenAI-compatible API could, in principle, evaluate its inference endpoints without an application-level rewrite, then decide whether the same platform's function and WebAssembly environment provides useful controls for its agent tools and generated code.

The important architectural question is not whether WebAssembly is inherently safer than every alternative. It is whether the available controls map to the agent's actual permissions: which APIs it may call, what memory and compute it may consume, and whether it can reach the public network. Those are observable deployment choices that can be tested and reviewed. The launch therefore makes runtime isolation a first-class item in infrastructure evaluation, alongside model quality and hardware capacity.

## Limits and uncertainty

The supplied source is an announcement report and does not establish production performance, geographic availability, pricing, quota policy, service-level commitments, or customer adoption. It does not provide benchmarks for the 12 models, a full inventory of those models, or evidence that OpenAI-compatible integration covers every application behavior teams use. It also does not describe Spin configuration, isolation guarantees, identity management, secret handling, logging, or audit features. Akamai says controls will be available over egress and system APIs, but the exact policy language and enforcement behavior are not supplied.

## Skill to build now

Create a small deployment checklist for every agent that can execute code or call external services. List its required system APIs, maximum compute and memory use, allowed destinations for outbound network traffic, and the data stores it may access. Then test the agent against a denied destination and a denied API call. That exercise gives a team a concrete way to compare a sandbox such as Akamai's with its current runtime, independent of the model selected behind the application API.

## Reader takeaway

Akamai is turning its AI Cloud into a combined model-serving and agent-execution platform: open-weight inference can be reached through an OpenAI-compatible API, while agents can run in WebAssembly sandboxes with stated resource and network controls. Treat it as a new option to evaluate, not as proof yet of lower cost, better latency, or stronger security.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)