Innodisk Unveils Core Ultra Series 3 Edge AI Platforms for Local Multi-Model Inference

Innodisk announced a new generation of industrial edge AI hardware platforms powered by Intel Core Ultra Series 3 processors, including the APEX-E300, APEX-P300, and ASBC-3160 systems. Designed to deliver up to 180 total platform TOPS across heterogeneous CPU, GPU, and NPU architectures, the systems are built specifically to handle concurrent multi-modal edge workloads. During demonstrations at the AI Infra Summit, the company showcased live agentic AI orchestration running directly on the APEX-E300, running Vision Language Models (VLMs) and Large Language Models (LLMs) locally without upstream cloud dependencies. This release highlights an operational evolution in edge architecture: moving past narrow computer vision pipelines toward generalized, multi-agent systems. Traditionally, edge computing deployments have been constrained to lightweight heuristic processing or basic object detection, offloading complex inference, natural language understanding, or contextual reasoning back to centralized cloud providers. However, industrial environments, healthcare facilities, and automated retail networks operate under strict latency, privacy, and connectivity constraints. By packing 180 TOPS of heterogeneous compute into ruggedized form factors, platforms like these allow teams to consolidate visual anomaly detection, local telemetry parsing, and localized agent actions on a single node without cloud ingress/egress penalties. This development reflects the broader trend of edge computing and AI convergence. With industry analysts forecasting that over two-thirds of enterprise-managed data will be generated and processed outside centralized data centers, compute capacity is decentralizing rapidly. The emergence of highly optimized small language models and efficient neural processing units (NPUs) has created an inflection point, making local physical AI and autonomous agent execution viable within strict power envelopes. For DevOps and platform engineers, operationalizing multi-model workloads at the edge requires significant changes in deployment strategies. Teams must manage containerized VLM and LLM models on heterogeneous silicon using specialized runtime runtimes, implement robust over-the-air (OTA) model synchronization, and set up local telemetry buffers. Edge architects should evaluate how local agentic inferencing alters network topology, specifically assessing whether shifting to edge-native inferencing can eliminate egress cloud costs while meeting microsecond response requirements.