Prince Mario-Max Schaumburg-Lippe: Akamai Puts AI Agents in a Safe Sandbox

Akamai picked a good morning to make its point. On October 27, 2026, from Cambridge, Massachusetts, the company launched two products at once – Akamai Inference and Akamai Agent Sandbox – and the pairing tells you exactly where enterprise AI stands right now. Teams do not just need a model. They need a safe place for that model to work.

Inference is the managed serving layer. Agent Sandbox is the locked room where agents can run code without wandering off. Put them together on Akamai Cloud and you get a full stack: serve a model, ground it in your own data, let an agent act, and keep a hard boundary around every step.

What Akamai Inference actually does

Start with the basics. Akamai Inference gives companies managed access to open-weight and custom models through OpenAI-compatible APIs. That last detail matters more than it sounds. If your app already talks to an API in that format, you can move it without a rewrite. No rip-and-replace project. No six-month migration plan gathering dust in a shared drive.

At launch the catalog holds 12 open-weight models, ranging from 8 billion to 2.8 trillion parameters. The names will ring a bell: NVIDIA Nemotron 3 Ultra, OpenAI gpt-oss-120b and Google Gemma 4 are all in there, along with embedding models built for retrieval-augmented generation. Customers can pick serverless, pay-per-use inference or a dedicated deployment. Small team testing an idea this quarter? Go serverless. Bank with strict controls? Go dedicated.

The models do not sit in isolation, either. Akamai Cloud connects to managed databases and object storage, and it links to private and public clouds, so a model can draw on real company data – product catalogs, support histories, internal documents – while Akamai handles orchestration, execution and authentication. Requests run across the Akamai global network, which keeps latency low and performance steady.

It is a practical pitch, and it rhymes with what readers saw in Microsoft scoring choices in 85 milliseconds – speed only counts if a business can actually deploy it on a Tuesday morning.

A sandbox with real walls

Now the second half, and honestly the more interesting half. Agents are starting to write and run their own code. That is useful. It is also the moment a sensible chief technology officer sits up a little straighter.

Akamai Agent Sandbox gives that code a separate, isolated home. Agents run on Akamai Functions, and the code they write runs in its own environment, sequestered from the agent runtime itself. The sandbox is delivered as an open source component of the Spin framework and uses WebAssembly isolation to set clear limits on compute, memory, outbound connections and system capabilities. In plain English: the agent can do its job, and it cannot reach things it was never given.

Akamai is not dogmatic about the method, either. The release notes that isolation is converging on a few models – microVMs with snapshot and fork capability, policy-enforced containers, WebAssembly-isolated functions – and says developers can choose the architecture that fits the workload. Options include NVIDIA-accelerated and dedicated CPU infrastructure, managed Kubernetes, and Akamai Functions for serverless execution across edge and core. More governance tools, covering what agents can reach and what can reach them, are planned next.

Why the edge count matters

Here is the number to remember: more than 4,350 edge locations. Akamai describes a continuum of compute that runs from dense GPU and CPU clusters in core regions out to that edge footprint. For a shop assistant answering a customer in Ohio or a factory system reacting on a line in Bavaria, distance is delay. Running inference close to the user and the data trims that delay in a way a single distant data center simply cannot match.

That same instinct showed up in Uber giving AI tools to all 30,000 of its staff – the winners are not running a lab demo, they are putting AI where the work already happens.

New GPUs under the floor

Akamai also widened the hardware bench. The dedicated GPU fleet now includes NVIDIA Blackwell Ultra clusters, and the company is preparing its data centers to support the NVIDIA Vera Rubin platform in 2027. The fleet already spans NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs for low-latency inference.

Each eight-GPU Blackwell Ultra node carries 288 GB of HBM3e memory per GPU – 2.3 TB in total. That is enough room to hold frontier-class models in memory, which is exactly what test-time scaling, chain-of-thought reasoning and long-context agent work demand. Clusters run without rate limits or shared-tenancy queueing, and customers keep control over model versions and deprecation timelines. Anyone who has had a provider retire a model mid-project will read that line twice, happily.

The $14 billion signal

Companies vote with contracts, and the vote here is large. Akamai says the infrastructure underneath these launches has already secured multiyear Akamai Cloud contracts collectively worth more than $14 billion. That figure does not prove Inference and Agent Sandbox will win – both are in beta, with Inference offering serverless options in beta and Agent Sandbox available via the Spin SDK – but it does prove customers are buying the foundation.

Adam Karon, Akamai chief operating officer and general manager of cloud, put the customer pull plainly in the release: customers want to build, run and protect AI on the same platform. Jon Alexander, senior vice president of cloud technology, added the other half of the thought – teams should not have to choose between speed and control.

What a smart team does with this

Three takeaways if you run a product or IT team. First, start serverless and measure. Pay-per-use inference lets you learn what your users actually ask before you commit to dedicated hardware. Second, sandbox first, agent second. Decide what an agent may touch – which APIs, which data, which outbound connections – before you celebrate what it can do. Third, ground the model in your own data early. A mid-size model fed your real catalog and support notes will beat a giant model guessing from general knowledge, and it will do it cheaper.

Akamai spent decades moving and protecting other people’s content at global scale. Inference and Agent Sandbox apply that same habit – distribution plus boundaries – to AI. It is not a flashy promise. It is plumbing, done at 4,350 locations, with the doors locked where they should be. For enterprises trying to get agents out of the demo and into daily work, plumbing is exactly the point.

Also published on: