---
type: "post"
id: "akamai-launches-akamai-inference-and-wasm-based-agent-e2a166"
slug: "akamai-launches-akamai-inference-and-wasm-based-agent-e2a166"
title: "Akamai launches Akamai Inference and Wasm-based Agent Sandbox in beta"
description: "Akamai has launched Akamai Inference and Akamai Agent Sandbox in beta, adding managed model serving and isolated agent-code execution to Akamai Cloud."
retrieval_nugget: "Akamai has launched Akamai Inference and Akamai Agent Sandbox in beta, adding managed model serving and isolated agent-code execution to Akamai Cloud. For AI builders, the announcement matters because it combines an OpenAI-compatible inference endpoint with a WebAssembly boundary intended to constrain what agent-generated code can consume, connect to, and do."
published_at: "2026-10-29"
updated_at: "2026-10-29"
record_date: "2026-10-29"
date_kind: "published_at"
topics: ["google","agent-runtime"]
entities: ["Akamai","Akamai Inference","Akamai Agent Sandbox","Spin","NVIDIA","NVIDIA Blackwell Ultra","OpenAI gpt-oss-120b","Google Gemma 4"]
related_patterns: ["agent-security-moves-to-runtime-boundaries","harness-architecture-outlives-model-choice","agent-ready-software-exposes-capabilities"]
novelty_label: "same_event"
what_changed: "Same event as Akamai launches Akamai Inference and WebAssembly-isolated Agent Sandbox in beta"
prior_refs: ["memory:c5cadb0b-2ff0-83d4-a8a5-779e3af81ac6","memory:02407428-d5a6-89a3-a6bf-5832c139f881","memory:e3708595-88be-862b-a49b-6d7ed54c974d","memory:0334cbc5-e5bd-801f-ac3b-b75572425bcd"]
source_url: "https://www.ir.akamai.com/news-releases/news-release-details/akamai-launches-ai-inference-service-and-agent-execution"
source_urls: ["https://www.ir.akamai.com/news-releases/news-release-details/akamai-launches-ai-inference-service-and-agent-execution"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:akamai-launches-akamai-inference-and-wasm-based-agent-e2a166"
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-and-wasm-based-agent-e2a166/","markdown":"https://newruntime.com/posts/akamai-launches-akamai-inference-and-wasm-based-agent-e2a166.md","json":"https://newruntime.com/posts/akamai-launches-akamai-inference-and-wasm-based-agent-e2a166.json"}
---

# Akamai launches Akamai Inference and Wasm-based Agent Sandbox in beta

## Retrieval answer

Akamai has launched Akamai Inference and Akamai Agent Sandbox in beta, adding managed model serving and isolated agent-code execution to Akamai Cloud. For AI builders, the announcement matters because it combines an OpenAI-compatible inference endpoint with a WebAssembly boundary intended to constrain what agent-generated code can consume, connect to, and do.

## Summary

Akamai has launched Akamai Inference and Akamai Agent Sandbox in beta, adding managed model serving and isolated agent-code execution to Akamai Cloud. For AI builders, the announcement matters because it combines an OpenAI-compatible inference endpoint with a WebAssembly boundary intended to constrain what agent-generated code can consume, connect to, and do.

## What happened

On October 27, 2026, Akamai announced two beta products: Akamai Inference and Akamai Agent Sandbox. The company presents them as additions to its cloud platform for enterprises that want to serve models, connect them to company data, and operate agents on infrastructure that extends from core regions to edge locations. Akamai says its network has more than 4,350 edge locations and frames the new services as managed layers on top of its existing compute, storage, networking, edge, and security offerings.

Akamai Inference is a managed service for open-weight and custom models exposed through OpenAI-compatible APIs. That compatibility is intended to let applications move without rewrites, according to the announcement. Customers can choose serverless, pay-per-use inference or dedicated deployments. The launch catalog contains 12 open-weight models ranging from 8 billion to 2.8 trillion parameters; Akamai names NVIDIA Nemotron 3 Ultra, OpenAI gpt-oss-120b, and Google Gemma 4 among them, alongside embedding models for retrieval-augmented generation. Akamai says the service handles orchestration, model execution, and authentication, while its managed databases, object storage, and connections to private and public clouds can provide access to enterprise data.

The second product, Akamai Agent Sandbox, addresses a different layer: execution of code created or used by an agent. Agents themselves run on Akamai Functions, while code they write runs in a separate isolated environment rather than in the agent runtime. The sandbox is delivered as an open-source component of the Spin framework and uses WebAssembly isolation. Akamai says developers can set limits around compute, memory, outbound connections, and system capabilities. It also says it plans to add controls governing what agents can reach and what can reach them, which means the described boundary is not yet the full access-control roadmap.

The release also expands Akamai's dedicated GPU fleet with NVIDIA Blackwell Ultra clusters and says the company is preparing for NVIDIA Vera Rubin support in 2027. Akamai describes each eight-GPU Blackwell Ultra node as having 288 GB of HBM3e per GPU, or 2.3 TB total. It positions dedicated clusters as an option for teams that want control over model versions and deprecation schedules and do not want shared-tenancy queues or rate limits. Both Akamai Inference and the Agent Sandbox via the Spin SDK are beta products; the announcement gives no public price, usage limits, region list, or self-service signup path. Akamai directs prospective users to sales for NVIDIA B300 cluster availability.

## What the archive adds

This is a new product event, but it overlaps directly with the earlier Field Note, [Akamai launches Akamai Inference with 12 open-weight models and a WebAssembly Agent Sandbox](https://newruntime.com/posts/akamai-launches-akamai-inference-with-12-open-weight-be7d9b/). That prior note identifies the same two product surfaces: a 12-model inference offering and WebAssembly-based agent execution. The present release supplies the primary-source detail on beta status, deployment modes, model range, and the stated sandbox controls.

## Trend signal

The release supports the pattern that [agent security moves to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/). Akamai does not describe safety as a prompt-level property: it separates the agent from code the agent writes and names compute, memory, network egress, and system capabilities as control points. The claim should remain qualified. The announcement describes an architecture and future plans, not independent evidence that the boundary withstands hostile workloads or is straightforward to configure correctly.

## Why it matters

Teams building agents commonly assemble model endpoints, execution sandboxes, storage, identity, networking, and security controls from separate providers. Akamai is attempting to make two of those choices more integrated: an API-compatible endpoint for hosted open-weight models and an execution boundary for code agents generate. That can reduce integration work for organizations already using Akamai Cloud, particularly where latency, network placement, or a single operational vendor matter.

The more consequential element is the separation of agent logic from agent-written code. An agent that can produce and run code creates a different risk profile from a chatbot that merely returns text. A runtime boundary can make permissions and egress explicit engineering decisions. Builders should evaluate that boundary as an operational control surface, however, not assume that WebAssembly alone settles authorization, secret handling, data access, or review of high-impact actions.

## What happens next

- **Akamai will add more explicit reachability and access-governance controls to Agent Sandbox.** (6 months, 65%). Watch for: documentation of identities, permission policies, or inbound and outbound access rules.
- **The beta will gain clearer commercial and operational terms for inference.** (6 months, 70%). Watch for: public pricing, region availability, quotas, or a self-service onboarding path.

## Limits and uncertainty

This is a company announcement, so its performance, availability, and security framing are Akamai's claims. It does not publish benchmark results, inference pricing, supported regions, API quotas, model-by-model availability, or service-level terms. It also does not specify the default permissions of the sandbox, how developers configure outbound connections and system capabilities, how secrets are handled, or what security testing has validated the isolation model. “Open source component” describes delivery through Spin but does not establish the maturity, support commitments, or portability of the complete managed service. Beta status further means product behavior and access can change.

## Skill to build now

Map agent execution into two planes this week: the agent runtime and every environment where it can execute generated code. For each execution environment, write down the allowed network destinations, available secrets, filesystem scope, compute and memory ceiling, and the actions requiring an external approval step. That inventory makes products such as Agent Sandbox comparable on real controls rather than on the generic claim that an agent is “sandboxed.”

## Reader takeaway

Akamai's beta turns a cloud and edge platform into a more complete agent stack: OpenAI-compatible serving on one side, WebAssembly-isolated execution on the other. The useful question for builders is not whether the stack is integrated, but whether its concrete runtime permissions, egress controls, pricing, and operational limits fit the agent they intend to deploy.

## Open question

Will Akamai expose sufficiently granular identity, secret, and network policies for the sandbox to serve regulated or high-impact agent workloads, rather than only low-risk code execution?

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.
- topic hub: [Agent Runtime](https://newruntime.com/topics/agent-runtime.md): Explore the agent-runtime 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)