- Akamai pitched a new inference platform that offers a middle ground betwnee hyperscale and neocloud capabilities.
- It also debuted a new Agent Sandbox that will contain the code agents write.
- Gartner expects spending on AI agents and assistants to more than double, from $29.2 billion in 2026 to $65.5 billion next year
Enterprises are torn between hyperscalers and inference specialists. But Akamai is pitching its new Inference platform as a third, middle of the road option that gives companies access to models, agents and security without having to minutely manage compute options.
“Most enterprises we talk to feel stuck between two imperfect choices,” Danielle Cook, Akamai’s director of product, told Fierce.
Cook argued hyperscalers offer scale and trust but slow access to new open-weight models and not enough help getting inference workloads up and running. Specialist platforms (i.e. neoclouds like CoreWeave, Lambda and Together.AI) move quickly on models, but leave teams to work out where other applications run, how they're all is secured and whether the provider is reliable enough for production, she said.
Launched in beta this week, Akamai's Inference Service is designed to walk the line between the two. Customers can choose serverless, pay-per-use inference or dedicated deployments. The launch catalog has 12 open-weight models ranging from 8 billion to 2.8 trillion parameters, including Nvidia’s Nemotron 3 Ultra, OpenAI’s gpt-oss-120b and Google’s Gemma 4, plus embedding models for retrieval-augmented generation.
“It’s the completeness of a hyperscaler without the complexity, and the speed of a specialist without the risk,” Cook said. “The goal is to give developers a range of [model] options to cover the most common needs rather than lock them into one model or provider. We expect those needs to keep evolving, and we'll continue adding to the catalog as new models and customer requirements emerge.”
Setting the stage for inference and agents
Akamai has spent the past year laying the groundwork for inference. It launched Akamai Inference Cloud last fall, deploying Nvidia Blackwell 6000 GPUs in 17 cities on a network that already spanned about 4,300 points of presence in more than 700 cities.
In March, it followed up with an AI Grid orchestrator, its implementation of Nvidia's AI Grid reference design. The orchestrator routes inference requests by intent, latency, cost and data sovereignty requirements.
Cook said customer demand for more control over models, costs and how they build their applications drove Akamai to add its managed Inference Service to the list.
The launch comes as enterprise spending on AI and continues to ramp. According to Gartner’s September 2026 forecast, worldwide spending on AI is expected to hit $2.7 trillion in 2026 and rise to $3.6 trillion in 2027. While the bulk of that is going toward AI infrastructure, investments into AI agents and assistants, models and cybersecurity are all growing apace in their own right.
For instance, spending on AI agents and assistants is expected to jump from $29.2 billion in 2026 to $65.5 billion in 2027. AI cybersecurity spending was tipped to increase from $51.3 billion to nearly $86 billion.
Locking down agents
As it happens, Akamai is also bringing security capabilities to the table with Akamai Agent Sandbox.
Cook said Sandbox contains the code an agent writes so that it can’t reach other systems, exceed its resources or send data where it shouldn’t. Developers remain responsible for deciding what an agent is allowed to do: which capabilities and connections to grant, what data and credentials it can access, and whether its output is correct. “The sandbox enforces the boundaries. The developer still sets them,” she said.
Access and authentication are some of the biggest open questions in agentic AI right now. Cook noted that while agents need access to real systems to be useful, most enterprise security models were built for people and applications.
Cook said the safest starting point is to treat an agent like an untrusted contractor. It should only be able to reach the specific systems it needs for the task, with outbound connections and capabilities limited by default. That’s the idea behind sandboxing: containing what an agent and the code it writes can do, so a mistake or a manipulated prompt can’t spread beyond what it’s been granted.
She added there’s still work to be done by the industry, particularly around audit trails. “Teams need to know not just what an agent did, but why it did it and on whose behalf. We’re seeing a wave of governance and control plane work across the ecosystem to tackle exactly that, and it’s an area we’re paying close attention to as well,” Cook concluded.
Read more about inference on Fierce Network
Rebellions courts telcos with cheaper AI inference and an open-source pitch
SambaNova targets AI inference boom with chips built for existing data centers
Arrcus weaves AI inference fabric PoC for Telus
This startup is building a new CDN for AI inferencing
Akamai's AI orchestrator puts inference where latency matters