AI is increasingly being designed to run on cameras, sensors, wearables and other devices close to where data is created. This shift is changing more than where models execute. It is changing how products respond, how much data must leave a device, and how teams manage a distributed computing footprint.
The attraction is clear: local inference can improve responsiveness, reinforce data control and avoid sending every raw input across a network. But edge AI also exchanges centralized flexibility for tighter hardware constraints and a more difficult operational model.
Gartner expects more than two-thirds of enterprise-managed data to be created and processed outside the data center or cloud by 2028. It also expects more than two-thirds of enterprises globally to deploy edge AI by 2029. IDC expects half of enterprise AI inference workloads to run on endpoints or edge nodes by 2030.
Why devices are becoming AI targets
The immediate driver is the volume of data produced by always-on devices. Gartner counted about 11.7 billion IoT devices in 2025, a total it expects to grow by 9% annually. Security cameras, traffic cameras and infrastructure sensors can generate continuous streams that are costly or impractical to transmit elsewhere.
Edge processing offers a way to filter, interpret or retain information before it reaches centralized infrastructure. Local inference can also reduce dependency on network connectivity for time-sensitive decisions. That matters for applications involving safety, security or operational continuity.
Privacy, data residency and sovereignty provide an additional pull. Sensor information may include audio, video, fingerprints or facial imaging, making control over where it is processed commercially and politically relevant. Local execution does not eliminate privacy obligations, but it can limit how much raw information needs to cross the network.
Hardware and software are becoming one decision
The hardware challenge is substantial. Endpoint devices typically operate within tight limits on power, memory, compute and physical space. A model that runs economically in a data center may not fit the battery, thermal or memory requirements of a wearable or industrial sensor.
Synaptics’ new xMCU class illustrates how chipmakers are responding. Its Astra SR200 Series combines AI-focused compute, integrated acceleration and custom DRAM that can scale to hundreds of megabytes. The architecture is intended to support larger AI models and richer multimodal workloads entirely on-device.
The series launches with three pin-compatible product lines designed for different workloads. The SR210 targets power-optimized, always-on voice and sensing at 200 GOPS. The SR230 adds vision, display and optional audio at 300 GOPS, while the SR250 supports 400 GOPS and dual vision and display paths.
That product structure reflects an important delivery principle: teams should not choose between a single endpoint configuration and a completely different architecture. A common hardware foundation and shared software development kit can allow performance to scale while preserving software investment and system architecture.
Our analysis: location is an architectural decision
Moving inference toward the edge is not a free shortcut. Teams can gain faster responses, tighter data control and reduced network transfer, but they inherit limitations in memory, power and compute. They also take responsibility for updates, monitoring and failures across hardware that may be physically dispersed, intermittently connected or difficult to access.
A centralized model can often be changed in one place. A fleet of edge devices may require coordinated rollout, version control, rollback planning and a policy for devices that are temporarily offline. Deployment therefore becomes part of the product rather than a final handoff to operations.
The deeper trade-off is flexibility. Central infrastructure offers broader compute capacity and simpler coordination. Edge inference offers proximity and local control. The strongest designs use both: devices handle immediate, context-sensitive decisions, while centralized systems retain the data and processing needed for broader analysis, coordination and model improvement.
This also means hardware selection should follow workload requirements. Voice-only sensing, camera-based vision and premium multimodal systems have different compute, memory and power needs. Choosing silicon before defining those boundaries risks either paying for unnecessary capability or discovering later that the device cannot support the intended model.
Decision checklist
Before committing to an edge AI design, project teams should answer:
- Which decisions require immediate local response?** Time-sensitive events should not depend on a distant cloud round trip.
- Which data must remain local?** Privacy, sovereignty and network-cost requirements should determine what leaves the device.
- Which workloads fit the device envelope?** Model size, memory use, power draw and thermal limits need to be evaluated together.
- Where is the boundary between tiers?** Teams should specify what runs on the device, what belongs on a nearby edge node and what remains centralized.
- How will software change after deployment?** Update delivery, versioning, rollback and offline behavior must be planned before rollout.
- How will the system behave when connectivity fails?** A device should have a defined operating mode for unreliable or unavailable network connections.
- Will hardware support the workload’s next stage?** Teams should consider whether performance can scale without discarding the software and architecture already built.
What to watch next
The important question is not whether edge AI replaces the cloud. It is how processing is divided among devices, nearby edge infrastructure and centralized systems.
Model capacity, power efficiency and software support will determine how far local inference can expand. Operational measures will matter just as much: update speed, privacy requirements, device lifecycle management and the cost of supporting distributed hardware.
For connected devices, AI wearables, industrial vision and voice systems, the architectural pattern is becoming clearer. The decision point is moving closer to the data, but success depends on treating that location as part of architecture, governance and delivery rather than as a technical afterthought.
Key takeaways
- Edge AI enables inference closer to sensors, users and connected devices.
- Potential benefits include lower latency, stronger data control, privacy and reduced network transfer.
- Endpoint projects must account for power, memory, model capacity, security and distributed operations.
- Teams should define device, edge-node and centralized workloads before selecting hardware.
- Edge AI is not a wholesale replacement for cloud computing; its strongest form is a coordinated division of work.
Frequently asked questions
What is edge AI?
Edge AI is AI processing performed close to where data is created, such as on an endpoint or nearby edge node, instead of sending every input to a centralized cloud environment.
Why are organizations moving AI toward the edge?
Key factors include growing data volumes, the need for faster responses, network-transfer costs, resilience requirements, and concerns about data residency, privacy and sovereignty.
What is the main hardware challenge for edge AI?
Devices have tight limits on power, memory and compute. Teams must ensure that the model, sensor workload and device architecture fit those constraints without creating unacceptable cost or operational complexity.
How should a project team choose what to process locally?
Identify decisions that require immediate responses or stronger data control, then determine what information the central system actually needs. Define the workload boundary before selecting hardware.