2026年及未来：边缘人工智能（AI）转型之路

Edge AI is the fastest growing part of the AI wave, with a 37%CAGR through 20301compared to 28% CAGR2of the overall AI market. It has gained significant traction in the last 5+ years - even asthe world has begun a transformation based on our collective “ChatGPT By collaborating with leading technology developers, academia andresearchers and industry leaders, we can explore what lies ahead in 2026 Introduction For the past decade, “cloud-first”has been the default mantra intechnology. The cloud gave usscale, flexibility, and global reach, When meeting with customers, weoften hear “We love the power of AI (meaning the cloud), but we can’t afford the latency.” When you need tostop a robotaxi for a wandering pedestrian, or a conveyor belt to avoid aproduct defect, or detect a safety risk in real time, waiting for round trips to Edge AI has now entered thediscussion at scale,challenging the previous real-time decisions and low-power implementations matter most. Thefuture is not cloud versus edge, but a new balance – where the edge can While there are many contributing factors to the increasing popularity ofedge AI, three drivers stand out to be the main enablers.The first driverbeing an explosion of available data:The sheer amount of data that is being generated by connected devices (75% of data is created at theedge3) and edge deployed devices and sensors continue to grow. Edge AI The second driver is the surge in compute performance:Neural Processing Units (NPUs) bring a tremendous increase in Machine Learningperformance to the Edge - all the way from MPUs to extremely constrained devices, like Microcontrollers (MCUs). This paradigm shift enables a The third driver is an unsustainable energy and resource path for cloud AI. Much has been written about the dramatic strain that datacenters areputting on the energy grids and systems of communities around the world. Data centers can consume 40% of a community electricity budget. Inaddition, a single large-scale data center can consume up to 5,000,000 What is edge AI and why should we care? Edge AI means the deployment of AI algorithms and models directly onlocal edge devices to enable real-time data processing and responses tinyML and more. This can range from sensor to server, although wetypically define “edge” as computing outside the multi-tenant data center. When is edge AI the best choice in a use case? There is a fantastic, easy toremember acronym that Daniel Situnayake & Jenny Plunkett (Speelman) B = Bandwidth: Reduced need for constant cloud connectivity, minimizing L = Latency:Real-time processing and analysis enables fasterdecision-making in applications where time is criticalE = Economics:Optimized resource utilization, reduced energyconsumption, and connecting only when it matters If any of the above are relevant to the application at hand, edge AI is likelythe better solution to the problem. An example for where “E”, “R”, and “P”matter are smart wearables - typically in a wearable we care aboutbattery power (the economics), so doing processing locally on device available without connectivity. Last but not least, the privacy aspect: forwearables today, most of the data still gets shared with the cloud via an How does edge AI work? Edge AI begins with on‑device data collection, where sensors capture rawsignals that are immediately preprocessed to de-noise, resize, and Models are developed and trained (usually in the cloud) on curateddatasets that reflect real edge conditions such as variable lighting, motion, Before deployment, models are typically optimized through techniques likequantization or pruning, and then compiled for the target runtime and In the field, inference is scheduled to respect resource and power limits,with strategies like streaming or batching and fallbacks to simpler models Why is 2026 a key year foredge AI? In 2026 we will see the firstbroad wave of IoT devicesembedded with edge AIacceleration, as well as edge AI becoming a key predictions. As AI-enabled hardware becomes more accessible andcorresponding AI software stacks more easy-to-use, the developer Indeed, there is a lot of movement in the market, visible especially incontinuing acquisitions (e.g. Qualcomm acquiring Edge Impulse & Arduino,NXP acquiring Kinara, Infineon acquiring Imagimob), key investment fromestablished companies like, Ceva, Microchip, EmbedUR, STMicroelectronics, growth in organizations like RISC-V International, the AI RAN Alliance, andour own EDGE AI FOUNDATION, who have seen a dramatic increase in What can we expect for edge AI in 2026? ●SLMs and Generative Edge AI: Generative and Agentic AI is no longerbound to the cloud - several companies have shown languagemodels running at the edge, even on devices as constrained as ●Agentic Edge AI:It was only November 2022, a little over three yearsago when OpenAI publicly launched ChatGPT and brought AIchatbots to the mainstream. As the models evolved to be able to ●Sensor to