NVIDIA and AWS lead shift toward hybrid AI architectures for inference

The transition to hybrid AI architectures marks a critical shift for streaming and surveillance providers managing massive data volumes. By moving inference to the edge, organizations can bypass the latency and egress costs associated with continuous high-resolution video uploads. This strategy forces a new competitive landscape where hardware providers like Intel and Dell must integrate deeply with cloud ecosystems from Microsoft and Google to maintain relevance. As streaming professionals navigate market fragmentation, the ability to maintain local operations during network outages becomes a key differentiator for enterprise reliability. Watch for standardized benchmarks comparing model accuracy after quantization on edge devices versus full-scale cloud versions.

NVIDIA has positioned its Jetson platform and TensorRT inference engine as the default edge AI stack for video-intensive workloads. In March 2026, NVIDIA announced that Jetson Thor had entered volume production with over 40 design wins across robotics and smart-city deployments, targeting 800 TOPS of INT8 performance in a 100-watt envelope suitable for multi-camera video analytics. The company's broader strategy ties edge inference back to its cloud GPU clusters through NVIDIA AI Enterprise licensing, creating a unified development path from training to deployment. AWS has taken a complementary approach with IoT Greengrass and SageMaker Edge Manager, which allow models trained in the cloud to be compiled and deployed to on-premises hardware. In April 2026, AWS expanded SageMaker Edge Manager support to include NVIDIA Jetson Orin and Intel OpenVINO-compatible devices, reducing the friction for enterprises running mixed-vendor edge fleets.

Microsoft and Google are pursuing similar hybrid strategies but with different go-to-market emphasis. Azure IoT Edge integrates with Azure Machine Learning to provide a managed pipeline for model deployment to edge nodes, and in May 2026, Microsoft announced that Azure IoT Edge would support ONNX Runtime 2.0 with native INT4 quantization for large language models on edge devices, directly addressing the memory constraints of running generative AI inference locally. Google Distributed Cloud Edge, meanwhile, targets telecom operators who need inference at cell sites and central offices. In February 2026, Google Cloud confirmed that Distributed Cloud Edge had been deployed by three tier-one carriers for real-time video quality-of-experience monitoring, processing 4K streams at under 10 milliseconds of round-trip latency. Cloudflare Workers AI represents a different edge model, running inference on Cloudflare's global network rather than customer-owned hardware, and the company reported in Q2 2026 that Workers AI inference requests had grown 340% year over year, driven largely by image classification and video moderation workloads.

Independent benchmarking of edge inference performance remains sparse but is beginning to emerge. In June 2026, MLPerf published its first Edge Inference v5.0 results, showing that NVIDIA Jetson Orin Nano achieved 92% of cloud A100 accuracy on ResNet-50 and YOLOv8 at 15 watts, validating the viability of edge deployment for production video analytics. Dell NativeEdge and Red Hat Device Edge address the operational layer, providing fleet management and OTA model updates for distributed inference nodes. Dell announced in July 2026 that NativeEdge had been certified for NVIDIA Jetson Thor and Intel OpenVINO workloads, signaling that hardware vendors are converging on standardized edge AI reference architectures. Intel OpenVINO continues to serve as the inference runtime for x86-based edge deployments, and , up from nine a year earlier. For those seeking the latest hardware, has recently launched to further boost local processing capabilities.