NVIDIA CEO Jensen Huang dominated the headlines with a surprise appearance during the opening of COMPUTEX in Taipei this week, but the message his presence conveyed was no surprise: AI infrastructure is the defining technology story of 2026.
NVIDIA CEO Jensen Huang dominated the headlines with a surprise appearance during the opening of COMPUTEX in Taipei this week, but the message his presence conveyed was no surprise: AI infrastructure is the defining technology story of 2026.
The list of companies that make AI-powered security and building automation possible, however, goes far beyond hyperscalers such as NVIDIA, which provides SoCs of its Jetson series to edge AI cameras for brands like Hanwha Vision and Avigilon, and data center GPUs on which VMS and cloud brands such as Genetec and Milestone rely.
Other video security brands, or different series of the same brands, rely on a technology stack dominated by specialized companies that wowed COMPUTEX visitors who looked beyond the main stage.
Compute power
DEEPX makes purpose-built NPUs optimized for low-power, cost-efficient AI inference, posing a direct alternative to GPU-based solutions for edge devices like cameras and access readers. At their booth, the South Korean company showcases use cases of its DX-M1 M.2 module—from real-live people counting on the crowded show floor, to robotic vision for smart manufacturing, security incident detection for smart cities and even the detection of malignant tumors in CT scans.
The demos demonstrate the strong specs of the M1: 25 TOPS (Tera operations per second) AI performance, 4GB LPDDR5 memory and power consumption as low as 1W-5W. The next generation of DEEPX M-series NPU—the DX-M2—will push the company’s mission further: “As AI models grow, on-device AI increasingly replaces data center reliance.”
Made on 2nm nodes, the M2 features 80 TOPS, enabling AI models with parameter counts of 20 billion to 100 billion, at a power consumption of about 5 Watts.
While DEEPX NPUs power AI cameras by boutique brands such as Hi-Sharp and smart city infrastructures in Singapore, they are also to be found in AI boxes by Aetina.
Aetina is a subsidiary of Innodisk, whose booth revolves around the theme “building blocks of AI,” from industrial-grade storage and memory modules] to [ruggedized camera solutions and edge AI inference systems. As part of the Innodisk booth, Aetina showcases DEEPX-powered compact AI inference units for high-performance, low-temperature real-time vision AI edge computing for multi-model and multi-stream applications, giving manufacturers and integrators and alternative to NVIDIA or memryX-powered Aetina hardware.
The Taiwan-headquartered company also showcases edge AI vision solutions based on a new lightweight Vision Language Model (VLM). The solution, named Aetina Mini, addresses high hardware and operating costs related to traditional full-scale VLMs, which require large amounts of memory and electricity, limiting their usage on mobile and wireless devices, for example.
Aetina Mini is powered by NVIDIA’s Jetson Orin Nano, which consume 7W-15W at peak performance, and supports up to two PoE cameras. Use cases focus mainly on industrial settings, for example detection whether workers wear the required safety equipment or the AI-augmented inspections of equipment for potential leakage.
US-Taiwan company Kneron also makes NPUs that enable AI-powered ISP for security cameras and other edge AI workloads.
“We want to enable manufacturers to run AI models directly on the edge who previously relied on on-prem AI solutions such as enabled recorders or AI analytics in the cloud,” a company representative said at the Kneron booth. “Edge AI reduces hardware footprints for on-prem solutions and bandwidth and latency for cloud and hybrid solutions.”
Among camera manufacturers Kneron partners with are OMNIEye and Sunmore Smart Technology. Kneron also showcases NPUs for integration in near-edge AI solutions such as lightweight AI boxes that help users run analytics on legacy glassware that cannot be replaced easily, such as cameras integrated in ATMs.
“Our NPUs for lightweight AI boxes focus on very specific use cases. Equipped with 2GB memory, they help users who don’t need the inference power of all-purpose AI models save on costly resources, especially energy,” the representative said. “Think of ATM operators who need to capture the images of user and ensure the ATM rejects withdrawal attempts by persons who cover their faces, for example by wearing motorcycle helmets. It’s a specific, limited use case that needs the inference power of a purpose-build lightweight AI model, which our partners can build on OpenClaw, for example, and run on our NPUs.”
Storage trends: From NAS to the cloud
Aside from AI inference on the edge, storage is another big talking point in video security these days—and the latter is closely connected to the former, because AI metadata is often written in bursts that are challenging to handle for traditional storage solutions.
Several companies that have long been known for NAS solutions showcase innovations at COMPUTEX that go far beyond traditional on-prem storage. Chiefly among them is QNAP, which is integrating increasingly intricate video management software (VMS) into its NAS-based solution, featuring AI-powered event search, person and vehicle detection, and cross-camera tracking. Meanwhile, the company is also seeking to offer more flexibility, through architecture-level integration of edge storage, for example using microSD cards, and cloud integration. Building open systems, QNAP users can use third-party hardware just as easily as they can manage their system in the clouds of SaaS providers such as Genetec or Milestone.
“Hybrid solutions do not just mean edge-plus-cloud,” a company representative said at the QNAP booth. “Think, for example, of industrial automation systems, where forensic investigations into faulty products might involve more data than any edge storage solution can handle. Operators cannot continuously stream and store all data to the cloud. For them on-prem storage remains the solution of choice.”
“NAS offers local redundancy, which is important to users who prefer on-prem storage for reasons of failover-safety,” the representative said. “Especially when you add cloud management and edge storage in latency-sensitive systems, NAS offers great flexibility.”
Similarly, Latticework showcased solutions that combine cloud and edge analytics and storage with its well-known NAS solutions. Its VMS, too, offers AI analytics from motion detection and object classification to license plate and facial recognition. In contrast to QNAP, however, the US company has also pivoted toward building proprietary cloud infrastructure, with data centers in the US, Europe and APAC. It’s own VMS therefore also runs in the cloud.
“For video security applications, our vision is to reduce the reliance of users on on-prem hardware, including as NAS,” a company representative said at the Latticework booth. “In the future, our users will be able to bridge their edge devices directly to the cloud based on open ONVIF standards. NAS, or other on-prem hardware such as AI boxes will be optional for them—offering near-edge compute power and redundancy storage for those who need them.”
Product Adopted:Image Processors