To run reliably in the cloud, AI agents cannot be simple synchronous API calls. Their long-running, stateful nature requires an asynchronous architecture. This typically involves a message broker and task queue to farm out agentic loops to ephemeral workers, preventing process failures and enabling scalability.

The dominant AI use case will shift from real-time, human-in-the-loop chatbots to long-running background agents. For these agents, which work for hours or days, an extra few seconds of latency is meaningless, unlocking massive cost-saving opportunities by prioritizing throughput over speed.

The need for low-latency services for agents and real-time applications in finance and healthcare is driving a shift towards distributed data centers. Instead of remote gigawatt facilities, companies are deploying smaller, power-efficient, air-cooled racks like SambaNova's in existing metropolitan data centers, closer to users.

Simply "scaling up" (adding more GPUs to one model instance) hits a performance ceiling due to hardware and algorithmic limits. True large-scale inference requires "scaling out" (duplicating instances), creating a new systems problem of managing and optimizing across a distributed fleet.

While AI training requires massive, centralized data centers, the growth of inference workloads is creating a need for a new architecture. This involves smaller (e.g., 5 megawatt), decentralized clusters located closer to users to reduce latency. This shift impacts everything from data center design to the software required to manage these distributed fleets.

A 2-second delay is acceptable for a single user prompt. However, in an agentic system where 20 agents communicate sequentially, that delay compounds to 40 seconds, rendering the application unusable. This shift necessitates infrastructure with sub-second response times, driving hardware deployment to urban centers.

The initial assumption of a centralized AI model (large hub, large spoke) is wrong. The new model will involve large foundational hubs, enterprise-specific training hubs, and distributed "spokes" of on-premise hardware for inference. This shift is driven by the need for data control and cost efficiency.

While the idea of distributed compute pools is appealing, it's not feasible for AI training due to high latency demands; GPUs must be physically co-located. However, AI inference is less sensitive to this lag, making a distributed network of compute (like home GPUs) a much more viable and exciting model.

To manage costs, the optimal architecture isn't running everything on the most powerful model. Instead, a smart orchestrator agent should break down complex problems and dispatch simpler sub-tasks to smaller, cheaper models, optimizing for both cost and performance.