The Wi-Fi access point (AP) has had a distinguished career performing a singular task: shuttling data between our devices and networks – quickly, accurately and reliably. That deceptively simple function transformed the Wi-Fi AP into a data network connecting our smartphones, televisions, appliances, sensors and other smart electronics to the cloud and each other.

Like much of today’s computing infrastructure, the advent of artificial intelligence is pushing Wi-Fi AP technology to expand its scope as residential gateways and enterprise appliances are called upon to do more than simply relay data.

Designers are pondering a new class of Wi-Fi APs equipped with neural processing units (NPUs) that execute AI inference locally. In this model, APs operate as small processing centers that allow connected devices to move compute-intensive workloads to the network edge instead of performing them on the client or sending them to a remote cloud server. Here, the AP begins to blur the line between networking and computing by functioning both as the pathway to AI and the AI coprocessor itself.

In this blog, we examine why AI compute is moving into the Wi-Fi network and what an AI-capable AP could mean for connected devices and edge computing.

The Benefits of Shared Access

Conceptually, the expanding role of the Wi-Fi AP reflects a broad trend toward greater local autonomy and less cloud dependency by shifting AI closer to the edge of the network.

Edge AI Architecture

By owning both the wireless connection and some of the AI execution, APs would evolve from “access points” to data “distribution hubs” imbued with contextual reasoning for deciding which workloads receive airtime, bandwidth and compute priority through traffic profiling and dynamic resource scheduling. The upside is less internet congestion and fewer routing interruptions by making signal latency more predictable and manageable.

The implications don’t end there. Manufacturers could design lighter weight endpoints by tapping AI-enhanced APs as a shared resource for executing inference, rather than packing dedicated AI processors into every watch, sensor and phone.

Bringing AI Closer to the User

The impetus for expanding the role and purpose of Wi-Fi access points is driven, in part, by the limitations of today’s AI architecture.

When a smartphone sends a request to an AI application, it can perform some processing locally or send the request through an AP to a cloud server. Cloud AI provides enormous computing resources, but the farther data travels, the greater the potential for latency and variability that are outside the user’s control.

The latency issue can be addressed by running more AI directly on client devices, but that creates other challenges. Smartphones, smart glasses and other small battery-powered devices have limited power, thermal and compute budgets. AI inference consumes battery power, generates heat and must be designed into already tight form factors.

AI-capable Wi-Fi APs present an attractive middle ground. APs are generally plugged into an external power source and have fewer space restrictions. This would allow inference to remain at the network edge without consuming precious client compute resources while also easing dependence on the backhaul connection between the AP and cloud.

Emerging applications such as augmented and virtual reality, smart glasses and home robotics could benefit from this model. The timing is important, too. Wi-Fi 7, emerging Wi-Fi 8 capabilities and the broad availability of NPUs are creating the technological foundation needed to make distributed AI architectures practical.

But putting AI compute into the access point is more than a design integration decision. It also changes what engineers need to know about Wi-Fi performance.

As connectivity and compute begin to converge at the network edge, familiar RF measurements take on a new context. Engineers must increasingly understand not only whether the wireless link performed as intended, but whether its performance enabled the AI application to deliver the expected outcome. Stay tuned for my next blog, where I will examine how that shift could change Wi-Fi testing – from evaluating graceful degradation and increasingly demanding uplink traffic to recreating multi-client AI workloads and connecting application behavior back to the physical layer.