Akamai has been distributing infrastructure since 1998, long before “edge” became a category of computing and certainly before anyone was thinking about where an AI agent should run.
The original problem was straightforward then. The internet could not scale if every request had to travel back to centralized infrastructure. Move content closer to users, route intelligently around constraints, and the network itself becomes part of the architecture.
And now, nearly three decades later, AI inference is presenting a familiar problem, only this time the thing being distributed is intelligence.
That is the thread running through the most recent episode of the Shared Everything podcast featuring a conversation with Akamai’s John Alexander, SVP of Cloud Product Management, and Ari Weil, VP of Product Marketing and Developer Relations.
Training gave us the AI factory, enormous concentrations of GPUs, power, networking and data assembled to build models. Inference changes the requirements. A robot making a physical decision, a fraud system evaluating a transaction or an agent responding to a user cannot always afford a round trip to a distant GPU cluster. The closer AI gets to operating in real time, the more placement matters.
Akamai calls the emerging architecture the AI grid, and the interesting part is not simply putting GPUs in more places. Alexander describes workload placement as a multidimensional optimization across compute availability, data locality, model availability, latency and cost. That sounds remarkably similar to the network optimization problem Akamai has been solving since the late 1990s, except now the payload and the infrastructure underneath it are considerably more complicated.
This is also where distributed inference becomes a data engineering problem. Models need context and applications need current data, but spreading compute across locations cannot mean creating isolated data islands everywhere it goes.
Alexander points to VAST’s global namespace as a way to abstract that regional complexity, while placement, caching, replication and data movement happen underneath. The application can address the data without having to know precisely where it lives.
That brings an old infrastructure question back around: do you move compute to the data or data to the compute? For distributed AI, the answer depends on the workload. Some inference belongs on dense GPU infrastructure, some can run efficiently on CPUs, and some needs to happen extremely close to the point of interaction.
Akamai, VAST and compute partners are approaching this as a continuum, distributing compute, models and data pipelines while trying to hide that underlying complexity from developers.
The AI factory is not going away, but it is no longer the whole AI infrastructure story. Around it, a distributed layer is taking shape in which CPUs, GPUs, networks and data infrastructure have to operate as one system. There is some symmetry in Akamai finding itself here. It began by pushing content outward to make the internet scale.
Now AI is creating another problem shaped by geography, latency and scale, and Akamai has spent nearly three decades building for exactly that kind of problem.
